2.1.826

.gitignore

@@ -8,6 +8,8 @@ __pycache__
 *.backup
 *.corrupted
 sorted_styles.json
+/input
+/cache
 /language/default.json
 lena.png
 lena_result.png
@@ -16,8 +18,6 @@ config.txt
 config_modification_tutorial.txt
 user_path_config.txt
 user_path_config-deprecated.txt
-build_chb.py
-experiment.py
 /modules/*.png
 /repositories
 /venv

args_manager.py

@@ -1,33 +1,31 @@
-from fcbh.options import enable_args_parsing
+import ldm_patched.modules.args_parser as args_parser
-enable_args_parsing(False)
-import fcbh.cli_args as fcbh_cli
 
 
-fcbh_cli.parser.add_argument("--share", action='store_true', help="Set whether to share on Gradio.")
+args_parser.parser.add_argument("--share", action='store_true', help="Set whether to share on Gradio.")
-fcbh_cli.parser.add_argument("--preset", type=str, default=None, help="Apply specified UI preset.")
+args_parser.parser.add_argument("--preset", type=str, default=None, help="Apply specified UI preset.")
 
-fcbh_cli.parser.add_argument("--language", type=str, default='default',
+args_parser.parser.add_argument("--language", type=str, default='default',
-                             help="Translate UI using json files in [language] folder. "
+                                help="Translate UI using json files in [language] folder. "
                                   "For example, [--language example] will use [language/example.json] for translation.")
 
 # For example, https://github.com/lllyasviel/Fooocus/issues/849
-fcbh_cli.parser.add_argument("--enable-smart-memory", action="store_true",
+args_parser.parser.add_argument("--disable-offload-from-vram", action="store_true",
-                             help="Force loading models to vram when the unload can be avoided. "
+                                help="Force loading models to vram when the unload can be avoided. "
                                   "Some Mac users may need this.")
 
-fcbh_cli.parser.add_argument("--theme", type=str, help="launches the UI with light or dark theme", default=None)
+args_parser.parser.add_argument("--theme", type=str, help="launches the UI with light or dark theme", default=None)
-fcbh_cli.parser.add_argument("--disable-image-log", action='store_true',
+args_parser.parser.add_argument("--disable-image-log", action='store_true',
-                             help="Prevent writing images and logs to hard drive.")
+                                help="Prevent writing images and logs to hard drive.")
 
-fcbh_cli.parser.set_defaults(
+args_parser.parser.set_defaults(
     disable_cuda_malloc=True,
     auto_launch=True,
     port=None
 )
 
-fcbh_cli.args = fcbh_cli.parser.parse_args()
+args_parser.args = args_parser.parser.parse_args()
 
 # (Disable by default because of issues like https://github.com/lllyasviel/Fooocus/issues/724)
-fcbh_cli.args.disable_smart_memory = not fcbh_cli.args.enable_smart_memory
+args_parser.args.always_offload_from_vram = not args_parser.args.disable_offload_from_vram
 
-args = fcbh_cli.args
+args = args_parser.args

extras/face_crop.py

@@ -25,11 +25,11 @@ def crop_image(img_rgb):
     global faceRestoreHelper
     
     if faceRestoreHelper is None:
-        from fooocus_extras.facexlib.utils.face_restoration_helper import FaceRestoreHelper
+        from extras.facexlib.utils.face_restoration_helper import FaceRestoreHelper
         faceRestoreHelper = FaceRestoreHelper(
             upscale_factor=1,
             model_rootpath=modules.config.path_controlnet,
-            device='cpu'  # use cpu is safer since we are out of fcbh management
+            device='cpu'  # use cpu is safer since we are out of memory management
         )
 
     faceRestoreHelper.clean_all()

extras/facexlib/detection/__init__.py

@@ -1,7 +1,7 @@
 import torch
 from copy import deepcopy
 
-from fooocus_extras.facexlib.utils import load_file_from_url
+from extras.facexlib.utils import load_file_from_url
 from .retinaface import RetinaFace
 
 

extras/facexlib/detection/retinaface.py

@@ -6,9 +6,9 @@ import torch.nn.functional as F
 from PIL import Image
 from torchvision.models._utils import IntermediateLayerGetter as IntermediateLayerGetter
 
-from fooocus_extras.facexlib.detection.align_trans import get_reference_facial_points, warp_and_crop_face
+from extras.facexlib.detection.align_trans import get_reference_facial_points, warp_and_crop_face
-from fooocus_extras.facexlib.detection.retinaface_net import FPN, SSH, MobileNetV1, make_bbox_head, make_class_head, make_landmark_head
+from extras.facexlib.detection.retinaface_net import FPN, SSH, MobileNetV1, make_bbox_head, make_class_head, make_landmark_head
-from fooocus_extras.facexlib.detection.retinaface_utils import (PriorBox, batched_decode, batched_decode_landm, decode, decode_landm,
+from extras.facexlib.detection.retinaface_utils import (PriorBox, batched_decode, batched_decode_landm, decode, decode_landm,
                                                  py_cpu_nms)
 
 

extras/facexlib/parsing/__init__.py

@@ -1,6 +1,6 @@
 import torch
 
-from fooocus_extras.facexlib.utils import load_file_from_url
+from extras.facexlib.utils import load_file_from_url
 from .bisenet import BiSeNet
 from .parsenet import ParseNet
 

extras/facexlib/utils/face_restoration_helper.py

@@ -4,9 +4,9 @@ import os
 import torch
 from torchvision.transforms.functional import normalize
 
-from fooocus_extras.facexlib.detection import init_detection_model
+from extras.facexlib.detection import init_detection_model
-from fooocus_extras.facexlib.parsing import init_parsing_model
+from extras.facexlib.parsing import init_parsing_model
-from fooocus_extras.facexlib.utils.misc import img2tensor, imwrite
+from extras.facexlib.utils.misc import img2tensor, imwrite
 
 
 def get_largest_face(det_faces, h, w):

extras/facexlib/utils/face_utils.py

@@ -211,9 +211,9 @@ def paste_face_back(img, face, inverse_affine):
 if __name__ == '__main__':
     import os
 
-    from fooocus_extras.facexlib.detection import init_detection_model
+    from extras.facexlib.detection import init_detection_model
-    from fooocus_extras.facexlib.utils.face_restoration_helper import get_largest_face
+    from extras.facexlib.utils.face_restoration_helper import get_largest_face
-    from fooocus_extras.facexlib.visualization import visualize_detection
+    from extras.facexlib.visualization import visualize_detection
 
     img_path = '/home/wxt/datasets/ffhq/ffhq_wild/00009.png'
     img_name = os.splitext(os.path.basename(img_path))[0]

extras/ip_adapter.py

@@ -1,12 +1,12 @@
 import torch
-import fcbh.clip_vision
+import ldm_patched.modules.clip_vision
 import safetensors.torch as sf
-import fcbh.model_management as model_management
+import ldm_patched.modules.model_management as model_management
 import contextlib
-import fcbh.ldm.modules.attention as attention
+import ldm_patched.ldm.modules.attention as attention
 
-from fooocus_extras.resampler import Resampler
+from extras.resampler import Resampler
-from fcbh.model_patcher import ModelPatcher
+from ldm_patched.modules.model_patcher import ModelPatcher
 from modules.core import numpy_to_pytorch
 
 
@@ -82,7 +82,7 @@ class IPAdapterModel(torch.nn.Module):
         self.ip_layers.load_state_dict_ordered(state_dict["ip_adapter"])
 
 
-clip_vision: fcbh.clip_vision.ClipVisionModel = None
+clip_vision: ldm_patched.modules.clip_vision.ClipVisionModel = None
 ip_negative: torch.Tensor = None
 ip_adapters: dict = {}
 
@@ -91,7 +91,7 @@ def load_ip_adapter(clip_vision_path, ip_negative_path, ip_adapter_path):
     global clip_vision, ip_negative, ip_adapters
 
     if clip_vision is None and isinstance(clip_vision_path, str):
-        clip_vision = fcbh.clip_vision.load(clip_vision_path)
+        clip_vision = ldm_patched.modules.clip_vision.load(clip_vision_path)
 
     if ip_negative is None and isinstance(ip_negative_path, str):
         ip_negative = sf.load_file(ip_negative_path)['data']
@@ -165,7 +165,7 @@ def preprocess(img, ip_adapter_path):
     global ip_adapters
     entry = ip_adapters[ip_adapter_path]
 
-    fcbh.model_management.load_model_gpu(clip_vision.patcher)
+    ldm_patched.modules.model_management.load_model_gpu(clip_vision.patcher)
     pixel_values = clip_preprocess(numpy_to_pytorch(img).to(clip_vision.load_device))
 
     if clip_vision.dtype != torch.float32:
@@ -173,25 +173,20 @@ def preprocess(img, ip_adapter_path):
     else:
         precision_scope = lambda a, b: contextlib.nullcontext(a)
 
-    with precision_scope(fcbh.model_management.get_autocast_device(clip_vision.load_device), torch.float32):
+    with precision_scope(ldm_patched.modules.model_management.get_autocast_device(clip_vision.load_device), torch.float32):
-        outputs = clip_vision.model(pixel_values=pixel_values, output_hidden_states=True)
+        outputs = clip_vision.model(pixel_values=pixel_values, intermediate_output=-2)
 
     ip_adapter = entry['ip_adapter']
     ip_layers = entry['ip_layers']
     image_proj_model = entry['image_proj_model']
     ip_unconds = entry['ip_unconds']
 
-    if ip_adapter.plus:
+    cond = outputs[1].to(device=ip_adapter.load_device, dtype=ip_adapter.dtype)
-        cond = outputs.hidden_states[-2]
-    else:
-        cond = outputs.image_embeds
 
-    cond = cond.to(device=ip_adapter.load_device, dtype=ip_adapter.dtype)
+    ldm_patched.modules.model_management.load_model_gpu(image_proj_model)
-
-    fcbh.model_management.load_model_gpu(image_proj_model)
     cond = image_proj_model.model(cond).to(device=ip_adapter.load_device, dtype=ip_adapter.dtype)
 
-    fcbh.model_management.load_model_gpu(ip_layers)
+    ldm_patched.modules.model_management.load_model_gpu(ip_layers)
 
     if ip_unconds is None:
         uncond = ip_negative.to(device=ip_adapter.load_device, dtype=ip_adapter.dtype)

extras/vae_interpose.py

@@ -4,9 +4,9 @@ import os
 import torch
 import safetensors.torch as sf
 import torch.nn as nn
-import fcbh.model_management
+import ldm_patched.modules.model_management
 
-from fcbh.model_patcher import ModelPatcher
+from ldm_patched.modules.model_patcher import ModelPatcher
 from modules.config import path_vae_approx
 
 
@@ -76,17 +76,17 @@ def parse(x):
         model.eval()
         sd = sf.load_file(vae_approx_filename)
         model.load_state_dict(sd)
-        fp16 = fcbh.model_management.should_use_fp16()
+        fp16 = ldm_patched.modules.model_management.should_use_fp16()
         if fp16:
             model = model.half()
         vae_approx_model = ModelPatcher(
             model=model,
-            load_device=fcbh.model_management.get_torch_device(),
+            load_device=ldm_patched.modules.model_management.get_torch_device(),
             offload_device=torch.device('cpu')
         )
         vae_approx_model.dtype = torch.float16 if fp16 else torch.float32
 
-    fcbh.model_management.load_model_gpu(vae_approx_model)
+    ldm_patched.modules.model_management.load_model_gpu(vae_approx_model)
 
     x = x_origin.to(device=vae_approx_model.load_device, dtype=vae_approx_model.dtype)
     x = vae_approx_model.model(x).to(x_origin)

face_experiments.py

@@ -1,5 +1,5 @@
 import cv2
-import fooocus_extras.face_crop as cropper
+import extras.face_crop as cropper
 
 
 img = cv2.imread('lena.png')

fooocus_version.py

@@ -1 +1 @@
-version = '2.1.825'
+version = '2.1.826'

launch.py

@@ -5,10 +5,9 @@ import sys
 print('[System ARGV] ' + str(sys.argv))
 
 root = os.path.dirname(os.path.abspath(__file__))
-backend_path = os.path.join(root, 'backend', 'headless')
+sys.path.append(root)
-sys.path += [root, backend_path]
-
 os.chdir(root)
+
 os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
 os.environ["PYTORCH_MPS_HIGH_WATERMARK_RATIO"] = "0.0"
 os.environ["GRADIO_SERVER_PORT"] = "7865"
@@ -88,19 +87,19 @@ def download_models():
     return
 
 
-def ini_fcbh_args():
+def ini_args():
     from args_manager import args
     return args
 
 
 prepare_environment()
 build_launcher()
-args = ini_fcbh_args()
+args = ini_args()
 
 
-if args.cuda_device is not None:
+if args.gpu_device_id is not None:
-    os.environ['CUDA_VISIBLE_DEVICES'] = str(args.cuda_device)
+    os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu_device_id)
-    print("Set device to:", args.cuda_device)
+    print("Set device to:", args.gpu_device_id)
 
 
 download_models()

ldm_patched/contrib/external_canny.py

@@ -0,0 +1,301 @@
+# https://github.com/comfyanonymous/ComfyUI/blob/master/nodes.py 
+
+#From https://github.com/kornia/kornia
+import math
+
+import torch
+import torch.nn.functional as F
+import ldm_patched.modules.model_management
+
+def get_canny_nms_kernel(device=None, dtype=None):
+    """Utility function that returns 3x3 kernels for the Canny Non-maximal suppression."""
+    return torch.tensor(
+        [
+            [[[0.0, 0.0, 0.0], [0.0, 1.0, -1.0], [0.0, 0.0, 0.0]]],
+            [[[0.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, -1.0]]],
+            [[[0.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, -1.0, 0.0]]],
+            [[[0.0, 0.0, 0.0], [0.0, 1.0, 0.0], [-1.0, 0.0, 0.0]]],
+            [[[0.0, 0.0, 0.0], [-1.0, 1.0, 0.0], [0.0, 0.0, 0.0]]],
+            [[[-1.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, 0.0]]],
+            [[[0.0, -1.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, 0.0]]],
+            [[[0.0, 0.0, -1.0], [0.0, 1.0, 0.0], [0.0, 0.0, 0.0]]],
+        ],
+        device=device,
+        dtype=dtype,
+    )
+
+
+def get_hysteresis_kernel(device=None, dtype=None):
+    """Utility function that returns the 3x3 kernels for the Canny hysteresis."""
+    return torch.tensor(
+        [
+            [[[0.0, 0.0, 0.0], [0.0, 0.0, 1.0], [0.0, 0.0, 0.0]]],
+            [[[0.0, 0.0, 0.0], [0.0, 0.0, 0.0], [0.0, 0.0, 1.0]]],
+            [[[0.0, 0.0, 0.0], [0.0, 0.0, 0.0], [0.0, 1.0, 0.0]]],
+            [[[0.0, 0.0, 0.0], [0.0, 0.0, 0.0], [1.0, 0.0, 0.0]]],
+            [[[0.0, 0.0, 0.0], [1.0, 0.0, 0.0], [0.0, 0.0, 0.0]]],
+            [[[1.0, 0.0, 0.0], [0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]],
+            [[[0.0, 1.0, 0.0], [0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]],
+            [[[0.0, 0.0, 1.0], [0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]],
+        ],
+        device=device,
+        dtype=dtype,
+    )
+
+def gaussian_blur_2d(img, kernel_size, sigma):
+    ksize_half = (kernel_size - 1) * 0.5
+
+    x = torch.linspace(-ksize_half, ksize_half, steps=kernel_size)
+
+    pdf = torch.exp(-0.5 * (x / sigma).pow(2))
+
+    x_kernel = pdf / pdf.sum()
+    x_kernel = x_kernel.to(device=img.device, dtype=img.dtype)
+
+    kernel2d = torch.mm(x_kernel[:, None], x_kernel[None, :])
+    kernel2d = kernel2d.expand(img.shape[-3], 1, kernel2d.shape[0], kernel2d.shape[1])
+
+    padding = [kernel_size // 2, kernel_size // 2, kernel_size // 2, kernel_size // 2]
+
+    img = torch.nn.functional.pad(img, padding, mode="reflect")
+    img = torch.nn.functional.conv2d(img, kernel2d, groups=img.shape[-3])
+
+    return img
+
+def get_sobel_kernel2d(device=None, dtype=None):
+    kernel_x = torch.tensor([[-1.0, 0.0, 1.0], [-2.0, 0.0, 2.0], [-1.0, 0.0, 1.0]], device=device, dtype=dtype)
+    kernel_y = kernel_x.transpose(0, 1)
+    return torch.stack([kernel_x, kernel_y])
+
+def spatial_gradient(input, normalized: bool = True):
+    r"""Compute the first order image derivative in both x and y using a Sobel operator.
+    .. image:: _static/img/spatial_gradient.png
+    Args:
+        input: input image tensor with shape :math:`(B, C, H, W)`.
+        mode: derivatives modality, can be: `sobel` or `diff`.
+        order: the order of the derivatives.
+        normalized: whether the output is normalized.
+    Return:
+        the derivatives of the input feature map. with shape :math:`(B, C, 2, H, W)`.
+    .. note::
+       See a working example `here <https://kornia-tutorials.readthedocs.io/en/latest/
+       filtering_edges.html>`__.
+    Examples:
+        >>> input = torch.rand(1, 3, 4, 4)
+        >>> output = spatial_gradient(input)  # 1x3x2x4x4
+        >>> output.shape
+        torch.Size([1, 3, 2, 4, 4])
+    """
+    # KORNIA_CHECK_IS_TENSOR(input)
+    # KORNIA_CHECK_SHAPE(input, ['B', 'C', 'H', 'W'])
+
+    # allocate kernel
+    kernel = get_sobel_kernel2d(device=input.device, dtype=input.dtype)
+    if normalized:
+        kernel = normalize_kernel2d(kernel)
+
+    # prepare kernel
+    b, c, h, w = input.shape
+    tmp_kernel = kernel[:, None, ...]
+
+    # Pad with "replicate for spatial dims, but with zeros for channel
+    spatial_pad = [kernel.size(1) // 2, kernel.size(1) // 2, kernel.size(2) // 2, kernel.size(2) // 2]
+    out_channels: int = 2
+    padded_inp = torch.nn.functional.pad(input.reshape(b * c, 1, h, w), spatial_pad, 'replicate')
+    out = F.conv2d(padded_inp, tmp_kernel, groups=1, padding=0, stride=1)
+    return out.reshape(b, c, out_channels, h, w)
+
+def rgb_to_grayscale(image, rgb_weights = None):
+    r"""Convert a RGB image to grayscale version of image.
+
+    .. image:: _static/img/rgb_to_grayscale.png
+
+    The image data is assumed to be in the range of (0, 1).
+
+    Args:
+        image: RGB image to be converted to grayscale with shape :math:`(*,3,H,W)`.
+        rgb_weights: Weights that will be applied on each channel (RGB).
+            The sum of the weights should add up to one.
+    Returns:
+        grayscale version of the image with shape :math:`(*,1,H,W)`.
+
+    .. note::
+       See a working example `here <https://kornia-tutorials.readthedocs.io/en/latest/
+       color_conversions.html>`__.
+
+    Example:
+        >>> input = torch.rand(2, 3, 4, 5)
+        >>> gray = rgb_to_grayscale(input) # 2x1x4x5
+    """
+
+    if len(image.shape) < 3 or image.shape[-3] != 3:
+        raise ValueError(f"Input size must have a shape of (*, 3, H, W). Got {image.shape}")
+
+    if rgb_weights is None:
+        # 8 bit images
+        if image.dtype == torch.uint8:
+            rgb_weights = torch.tensor([76, 150, 29], device=image.device, dtype=torch.uint8)
+        # floating point images
+        elif image.dtype in (torch.float16, torch.float32, torch.float64):
+            rgb_weights = torch.tensor([0.299, 0.587, 0.114], device=image.device, dtype=image.dtype)
+        else:
+            raise TypeError(f"Unknown data type: {image.dtype}")
+    else:
+        # is tensor that we make sure is in the same device/dtype
+        rgb_weights = rgb_weights.to(image)
+
+    # unpack the color image channels with RGB order
+    r: Tensor = image[..., 0:1, :, :]
+    g: Tensor = image[..., 1:2, :, :]
+    b: Tensor = image[..., 2:3, :, :]
+
+    w_r, w_g, w_b = rgb_weights.unbind()
+    return w_r * r + w_g * g + w_b * b
+
+def canny(
+    input,
+    low_threshold = 0.1,
+    high_threshold = 0.2,
+    kernel_size  = 5,
+    sigma = 1,
+    hysteresis = True,
+    eps = 1e-6,
+):
+    r"""Find edges of the input image and filters them using the Canny algorithm.
+    .. image:: _static/img/canny.png
+    Args:
+        input: input image tensor with shape :math:`(B,C,H,W)`.
+        low_threshold: lower threshold for the hysteresis procedure.
+        high_threshold: upper threshold for the hysteresis procedure.
+        kernel_size: the size of the kernel for the gaussian blur.
+        sigma: the standard deviation of the kernel for the gaussian blur.
+        hysteresis: if True, applies the hysteresis edge tracking.
+            Otherwise, the edges are divided between weak (0.5) and strong (1) edges.
+        eps: regularization number to avoid NaN during backprop.
+    Returns:
+        - the canny edge magnitudes map, shape of :math:`(B,1,H,W)`.
+        - the canny edge detection filtered by thresholds and hysteresis, shape of :math:`(B,1,H,W)`.
+    .. note::
+       See a working example `here <https://kornia-tutorials.readthedocs.io/en/latest/
+       canny.html>`__.
+    Example:
+        >>> input = torch.rand(5, 3, 4, 4)
+        >>> magnitude, edges = canny(input)  # 5x3x4x4
+        >>> magnitude.shape
+        torch.Size([5, 1, 4, 4])
+        >>> edges.shape
+        torch.Size([5, 1, 4, 4])
+    """
+    # KORNIA_CHECK_IS_TENSOR(input)
+    # KORNIA_CHECK_SHAPE(input, ['B', 'C', 'H', 'W'])
+    # KORNIA_CHECK(
+    #     low_threshold <= high_threshold,
+    #     "Invalid input thresholds. low_threshold should be smaller than the high_threshold. Got: "
+    #     f"{low_threshold}>{high_threshold}",
+    # )
+    # KORNIA_CHECK(0 < low_threshold < 1, f'Invalid low threshold. Should be in range (0, 1). Got: {low_threshold}')
+    # KORNIA_CHECK(0 < high_threshold < 1, f'Invalid high threshold. Should be in range (0, 1). Got: {high_threshold}')
+
+    device = input.device
+    dtype = input.dtype
+
+    # To Grayscale
+    if input.shape[1] == 3:
+        input = rgb_to_grayscale(input)
+
+    # Gaussian filter
+    blurred: Tensor = gaussian_blur_2d(input, kernel_size, sigma)
+
+    # Compute the gradients
+    gradients: Tensor = spatial_gradient(blurred, normalized=False)
+
+    # Unpack the edges
+    gx: Tensor = gradients[:, :, 0]
+    gy: Tensor = gradients[:, :, 1]
+
+    # Compute gradient magnitude and angle
+    magnitude: Tensor = torch.sqrt(gx * gx + gy * gy + eps)
+    angle: Tensor = torch.atan2(gy, gx)
+
+    # Radians to Degrees
+    angle = 180.0 * angle / math.pi
+
+    # Round angle to the nearest 45 degree
+    angle = torch.round(angle / 45) * 45
+
+    # Non-maximal suppression
+    nms_kernels: Tensor = get_canny_nms_kernel(device, dtype)
+    nms_magnitude: Tensor = F.conv2d(magnitude, nms_kernels, padding=nms_kernels.shape[-1] // 2)
+
+    # Get the indices for both directions
+    positive_idx: Tensor = (angle / 45) % 8
+    positive_idx = positive_idx.long()
+
+    negative_idx: Tensor = ((angle / 45) + 4) % 8
+    negative_idx = negative_idx.long()
+
+    # Apply the non-maximum suppression to the different directions
+    channel_select_filtered_positive: Tensor = torch.gather(nms_magnitude, 1, positive_idx)
+    channel_select_filtered_negative: Tensor = torch.gather(nms_magnitude, 1, negative_idx)
+
+    channel_select_filtered: Tensor = torch.stack(
+        [channel_select_filtered_positive, channel_select_filtered_negative], 1
+    )
+
+    is_max: Tensor = channel_select_filtered.min(dim=1)[0] > 0.0
+
+    magnitude = magnitude * is_max
+
+    # Threshold
+    edges: Tensor = F.threshold(magnitude, low_threshold, 0.0)
+
+    low: Tensor = magnitude > low_threshold
+    high: Tensor = magnitude > high_threshold
+
+    edges = low * 0.5 + high * 0.5
+    edges = edges.to(dtype)
+
+    # Hysteresis
+    if hysteresis:
+        edges_old: Tensor = -torch.ones(edges.shape, device=edges.device, dtype=dtype)
+        hysteresis_kernels: Tensor = get_hysteresis_kernel(device, dtype)
+
+        while ((edges_old - edges).abs() != 0).any():
+            weak: Tensor = (edges == 0.5).float()
+            strong: Tensor = (edges == 1).float()
+
+            hysteresis_magnitude: Tensor = F.conv2d(
+                edges, hysteresis_kernels, padding=hysteresis_kernels.shape[-1] // 2
+            )
+            hysteresis_magnitude = (hysteresis_magnitude == 1).any(1, keepdim=True).to(dtype)
+            hysteresis_magnitude = hysteresis_magnitude * weak + strong
+
+            edges_old = edges.clone()
+            edges = hysteresis_magnitude + (hysteresis_magnitude == 0) * weak * 0.5
+
+        edges = hysteresis_magnitude
+
+    return magnitude, edges
+
+
+class Canny:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"image": ("IMAGE",),
+                                "low_threshold": ("FLOAT", {"default": 0.4, "min": 0.01, "max": 0.99, "step": 0.01}),
+                                "high_threshold": ("FLOAT", {"default": 0.8, "min": 0.01, "max": 0.99, "step": 0.01})
+                                }}
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "detect_edge"
+
+    CATEGORY = "image/preprocessors"
+
+    def detect_edge(self, image, low_threshold, high_threshold):
+        output = canny(image.to(ldm_patched.modules.model_management.get_torch_device()).movedim(-1, 1), low_threshold, high_threshold)
+        img_out = output[1].to(ldm_patched.modules.model_management.intermediate_device()).repeat(1, 3, 1, 1).movedim(1, -1)
+        return (img_out,)
+
+NODE_CLASS_MAPPINGS = {
+    "Canny": Canny,
+}

ldm_patched/contrib/external_clip_sdxl.py

@@ -0,0 +1,58 @@
+# https://github.com/comfyanonymous/ComfyUI/blob/master/nodes.py 
+
+import torch
+from ldm_patched.contrib.external import MAX_RESOLUTION
+
+class CLIPTextEncodeSDXLRefiner:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+            "ascore": ("FLOAT", {"default": 6.0, "min": 0.0, "max": 1000.0, "step": 0.01}),
+            "width": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}),
+            "height": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}),
+            "text": ("STRING", {"multiline": True}), "clip": ("CLIP", ),
+            }}
+    RETURN_TYPES = ("CONDITIONING",)
+    FUNCTION = "encode"
+
+    CATEGORY = "advanced/conditioning"
+
+    def encode(self, clip, ascore, width, height, text):
+        tokens = clip.tokenize(text)
+        cond, pooled = clip.encode_from_tokens(tokens, return_pooled=True)
+        return ([[cond, {"pooled_output": pooled, "aesthetic_score": ascore, "width": width,"height": height}]], )
+
+class CLIPTextEncodeSDXL:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+            "width": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}),
+            "height": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}),
+            "crop_w": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION}),
+            "crop_h": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION}),
+            "target_width": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}),
+            "target_height": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}),
+            "text_g": ("STRING", {"multiline": True, "default": "CLIP_G"}), "clip": ("CLIP", ),
+            "text_l": ("STRING", {"multiline": True, "default": "CLIP_L"}), "clip": ("CLIP", ),
+            }}
+    RETURN_TYPES = ("CONDITIONING",)
+    FUNCTION = "encode"
+
+    CATEGORY = "advanced/conditioning"
+
+    def encode(self, clip, width, height, crop_w, crop_h, target_width, target_height, text_g, text_l):
+        tokens = clip.tokenize(text_g)
+        tokens["l"] = clip.tokenize(text_l)["l"]
+        if len(tokens["l"]) != len(tokens["g"]):
+            empty = clip.tokenize("")
+            while len(tokens["l"]) < len(tokens["g"]):
+                tokens["l"] += empty["l"]
+            while len(tokens["l"]) > len(tokens["g"]):
+                tokens["g"] += empty["g"]
+        cond, pooled = clip.encode_from_tokens(tokens, return_pooled=True)
+        return ([[cond, {"pooled_output": pooled, "width": width, "height": height, "crop_w": crop_w, "crop_h": crop_h, "target_width": target_width, "target_height": target_height}]], )
+
+NODE_CLASS_MAPPINGS = {
+    "CLIPTextEncodeSDXLRefiner": CLIPTextEncodeSDXLRefiner,
+    "CLIPTextEncodeSDXL": CLIPTextEncodeSDXL,
+}

ldm_patched/contrib/external_compositing.py

@@ -0,0 +1,204 @@
+# https://github.com/comfyanonymous/ComfyUI/blob/master/nodes.py 
+
+import numpy as np
+import torch
+import ldm_patched.modules.utils
+from enum import Enum
+
+def resize_mask(mask, shape):
+    return torch.nn.functional.interpolate(mask.reshape((-1, 1, mask.shape[-2], mask.shape[-1])), size=(shape[0], shape[1]), mode="bilinear").squeeze(1)
+
+class PorterDuffMode(Enum):
+    ADD = 0
+    CLEAR = 1
+    DARKEN = 2
+    DST = 3
+    DST_ATOP = 4
+    DST_IN = 5
+    DST_OUT = 6
+    DST_OVER = 7
+    LIGHTEN = 8
+    MULTIPLY = 9
+    OVERLAY = 10
+    SCREEN = 11
+    SRC = 12
+    SRC_ATOP = 13
+    SRC_IN = 14
+    SRC_OUT = 15
+    SRC_OVER = 16
+    XOR = 17
+
+
+def porter_duff_composite(src_image: torch.Tensor, src_alpha: torch.Tensor, dst_image: torch.Tensor, dst_alpha: torch.Tensor, mode: PorterDuffMode):
+    if mode == PorterDuffMode.ADD:
+        out_alpha = torch.clamp(src_alpha + dst_alpha, 0, 1)
+        out_image = torch.clamp(src_image + dst_image, 0, 1)
+    elif mode == PorterDuffMode.CLEAR:
+        out_alpha = torch.zeros_like(dst_alpha)
+        out_image = torch.zeros_like(dst_image)
+    elif mode == PorterDuffMode.DARKEN:
+        out_alpha = src_alpha + dst_alpha  - src_alpha * dst_alpha
+        out_image = (1 - dst_alpha) * src_image + (1 - src_alpha) * dst_image + torch.min(src_image, dst_image)
+    elif mode == PorterDuffMode.DST:
+        out_alpha = dst_alpha
+        out_image = dst_image
+    elif mode == PorterDuffMode.DST_ATOP:
+        out_alpha = src_alpha
+        out_image = src_alpha * dst_image + (1 - dst_alpha) * src_image
+    elif mode == PorterDuffMode.DST_IN:
+        out_alpha = src_alpha * dst_alpha
+        out_image = dst_image * src_alpha
+    elif mode == PorterDuffMode.DST_OUT:
+        out_alpha = (1 - src_alpha) * dst_alpha
+        out_image = (1 - src_alpha) * dst_image
+    elif mode == PorterDuffMode.DST_OVER:
+        out_alpha = dst_alpha + (1 - dst_alpha) * src_alpha
+        out_image = dst_image + (1 - dst_alpha) * src_image
+    elif mode == PorterDuffMode.LIGHTEN:
+        out_alpha = src_alpha + dst_alpha - src_alpha * dst_alpha
+        out_image = (1 - dst_alpha) * src_image + (1 - src_alpha) * dst_image + torch.max(src_image, dst_image)
+    elif mode == PorterDuffMode.MULTIPLY:
+        out_alpha = src_alpha * dst_alpha
+        out_image = src_image * dst_image
+    elif mode == PorterDuffMode.OVERLAY:
+        out_alpha = src_alpha + dst_alpha - src_alpha * dst_alpha
+        out_image = torch.where(2 * dst_image < dst_alpha, 2 * src_image * dst_image,
+            src_alpha * dst_alpha - 2 * (dst_alpha - src_image) * (src_alpha - dst_image))
+    elif mode == PorterDuffMode.SCREEN:
+        out_alpha = src_alpha + dst_alpha - src_alpha * dst_alpha
+        out_image = src_image + dst_image - src_image * dst_image
+    elif mode == PorterDuffMode.SRC:
+        out_alpha = src_alpha
+        out_image = src_image
+    elif mode == PorterDuffMode.SRC_ATOP:
+        out_alpha = dst_alpha
+        out_image = dst_alpha * src_image + (1 - src_alpha) * dst_image
+    elif mode == PorterDuffMode.SRC_IN:
+        out_alpha = src_alpha * dst_alpha
+        out_image = src_image * dst_alpha
+    elif mode == PorterDuffMode.SRC_OUT:
+        out_alpha = (1 - dst_alpha) * src_alpha
+        out_image = (1 - dst_alpha) * src_image
+    elif mode == PorterDuffMode.SRC_OVER:
+        out_alpha = src_alpha + (1 - src_alpha) * dst_alpha
+        out_image = src_image + (1 - src_alpha) * dst_image
+    elif mode == PorterDuffMode.XOR:
+        out_alpha = (1 - dst_alpha) * src_alpha + (1 - src_alpha) * dst_alpha
+        out_image = (1 - dst_alpha) * src_image + (1 - src_alpha) * dst_image
+    else:
+        out_alpha = None
+        out_image = None
+    return out_image, out_alpha
+
+
+class PorterDuffImageComposite:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "source": ("IMAGE",),
+                "source_alpha": ("MASK",),
+                "destination": ("IMAGE",),
+                "destination_alpha": ("MASK",),
+                "mode": ([mode.name for mode in PorterDuffMode], {"default": PorterDuffMode.DST.name}),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE", "MASK")
+    FUNCTION = "composite"
+    CATEGORY = "mask/compositing"
+
+    def composite(self, source: torch.Tensor, source_alpha: torch.Tensor, destination: torch.Tensor, destination_alpha: torch.Tensor, mode):
+        batch_size = min(len(source), len(source_alpha), len(destination), len(destination_alpha))
+        out_images = []
+        out_alphas = []
+
+        for i in range(batch_size):
+            src_image = source[i]
+            dst_image = destination[i]
+
+            assert src_image.shape[2] == dst_image.shape[2] # inputs need to have same number of channels
+
+            src_alpha = source_alpha[i].unsqueeze(2)
+            dst_alpha = destination_alpha[i].unsqueeze(2)
+
+            if dst_alpha.shape[:2] != dst_image.shape[:2]:
+                upscale_input = dst_alpha.unsqueeze(0).permute(0, 3, 1, 2)
+                upscale_output = ldm_patched.modules.utils.common_upscale(upscale_input, dst_image.shape[1], dst_image.shape[0], upscale_method='bicubic', crop='center')
+                dst_alpha = upscale_output.permute(0, 2, 3, 1).squeeze(0)
+            if src_image.shape != dst_image.shape:
+                upscale_input = src_image.unsqueeze(0).permute(0, 3, 1, 2)
+                upscale_output = ldm_patched.modules.utils.common_upscale(upscale_input, dst_image.shape[1], dst_image.shape[0], upscale_method='bicubic', crop='center')
+                src_image = upscale_output.permute(0, 2, 3, 1).squeeze(0)
+            if src_alpha.shape != dst_alpha.shape:
+                upscale_input = src_alpha.unsqueeze(0).permute(0, 3, 1, 2)
+                upscale_output = ldm_patched.modules.utils.common_upscale(upscale_input, dst_alpha.shape[1], dst_alpha.shape[0], upscale_method='bicubic', crop='center')
+                src_alpha = upscale_output.permute(0, 2, 3, 1).squeeze(0)
+
+            out_image, out_alpha = porter_duff_composite(src_image, src_alpha, dst_image, dst_alpha, PorterDuffMode[mode])
+
+            out_images.append(out_image)
+            out_alphas.append(out_alpha.squeeze(2))
+
+        result = (torch.stack(out_images), torch.stack(out_alphas))
+        return result
+
+
+class SplitImageWithAlpha:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+                "required": {
+                    "image": ("IMAGE",),
+                }
+        }
+
+    CATEGORY = "mask/compositing"
+    RETURN_TYPES = ("IMAGE", "MASK")
+    FUNCTION = "split_image_with_alpha"
+
+    def split_image_with_alpha(self, image: torch.Tensor):
+        out_images = [i[:,:,:3] for i in image]
+        out_alphas = [i[:,:,3] if i.shape[2] > 3 else torch.ones_like(i[:,:,0]) for i in image]
+        result = (torch.stack(out_images), 1.0 - torch.stack(out_alphas))
+        return result
+
+
+class JoinImageWithAlpha:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+                "required": {
+                    "image": ("IMAGE",),
+                    "alpha": ("MASK",),
+                }
+        }
+
+    CATEGORY = "mask/compositing"
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "join_image_with_alpha"
+
+    def join_image_with_alpha(self, image: torch.Tensor, alpha: torch.Tensor):
+        batch_size = min(len(image), len(alpha))
+        out_images = []
+
+        alpha = 1.0 - resize_mask(alpha, image.shape[1:])
+        for i in range(batch_size):
+           out_images.append(torch.cat((image[i][:,:,:3], alpha[i].unsqueeze(2)), dim=2))
+
+        result = (torch.stack(out_images),)
+        return result
+
+
+NODE_CLASS_MAPPINGS = {
+    "PorterDuffImageComposite": PorterDuffImageComposite,
+    "SplitImageWithAlpha": SplitImageWithAlpha,
+    "JoinImageWithAlpha": JoinImageWithAlpha,
+}
+
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "PorterDuffImageComposite": "Porter-Duff Image Composite",
+    "SplitImageWithAlpha": "Split Image with Alpha",
+    "JoinImageWithAlpha": "Join Image with Alpha",
+}

ldm_patched/contrib/external_custom_sampler.py

@@ -0,0 +1,287 @@
+# https://github.com/comfyanonymous/ComfyUI/blob/master/nodes.py 
+
+import ldm_patched.modules.samplers
+import ldm_patched.modules.sample
+from ldm_patched.k_diffusion import sampling as k_diffusion_sampling
+import ldm_patched.utils.latent_visualization
+import torch
+import ldm_patched.modules.utils
+
+
+class BasicScheduler:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required":
+                    {"model": ("MODEL",),
+                     "scheduler": (ldm_patched.modules.samplers.SCHEDULER_NAMES, ),
+                     "steps": ("INT", {"default": 20, "min": 1, "max": 10000}),
+                      }
+               }
+    RETURN_TYPES = ("SIGMAS",)
+    CATEGORY = "sampling/custom_sampling/schedulers"
+
+    FUNCTION = "get_sigmas"
+
+    def get_sigmas(self, model, scheduler, steps):
+        sigmas = ldm_patched.modules.samplers.calculate_sigmas_scheduler(model.model, scheduler, steps).cpu()
+        return (sigmas, )
+
+
+class KarrasScheduler:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required":
+                    {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}),
+                     "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}),
+                     "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}),
+                     "rho": ("FLOAT", {"default": 7.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}),
+                    }
+               }
+    RETURN_TYPES = ("SIGMAS",)
+    CATEGORY = "sampling/custom_sampling/schedulers"
+
+    FUNCTION = "get_sigmas"
+
+    def get_sigmas(self, steps, sigma_max, sigma_min, rho):
+        sigmas = k_diffusion_sampling.get_sigmas_karras(n=steps, sigma_min=sigma_min, sigma_max=sigma_max, rho=rho)
+        return (sigmas, )
+
+class ExponentialScheduler:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required":
+                    {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}),
+                     "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}),
+                     "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}),
+                    }
+               }
+    RETURN_TYPES = ("SIGMAS",)
+    CATEGORY = "sampling/custom_sampling/schedulers"
+
+    FUNCTION = "get_sigmas"
+
+    def get_sigmas(self, steps, sigma_max, sigma_min):
+        sigmas = k_diffusion_sampling.get_sigmas_exponential(n=steps, sigma_min=sigma_min, sigma_max=sigma_max)
+        return (sigmas, )
+
+class PolyexponentialScheduler:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required":
+                    {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}),
+                     "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}),
+                     "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}),
+                     "rho": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}),
+                    }
+               }
+    RETURN_TYPES = ("SIGMAS",)
+    CATEGORY = "sampling/custom_sampling/schedulers"
+
+    FUNCTION = "get_sigmas"
+
+    def get_sigmas(self, steps, sigma_max, sigma_min, rho):
+        sigmas = k_diffusion_sampling.get_sigmas_polyexponential(n=steps, sigma_min=sigma_min, sigma_max=sigma_max, rho=rho)
+        return (sigmas, )
+
+class SDTurboScheduler:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required":
+                    {"model": ("MODEL",),
+                     "steps": ("INT", {"default": 1, "min": 1, "max": 10}),
+                      }
+               }
+    RETURN_TYPES = ("SIGMAS",)
+    CATEGORY = "sampling/custom_sampling/schedulers"
+
+    FUNCTION = "get_sigmas"
+
+    def get_sigmas(self, model, steps):
+        timesteps = torch.flip(torch.arange(1, 11) * 100 - 1, (0,))[:steps]
+        sigmas = model.model.model_sampling.sigma(timesteps)
+        sigmas = torch.cat([sigmas, sigmas.new_zeros([1])])
+        return (sigmas, )
+
+class VPScheduler:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required":
+                    {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}),
+                     "beta_d": ("FLOAT", {"default": 19.9, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}), #TODO: fix default values
+                     "beta_min": ("FLOAT", {"default": 0.1, "min": 0.0, "max": 1000.0, "step":0.01, "round": False}),
+                     "eps_s": ("FLOAT", {"default": 0.001, "min": 0.0, "max": 1.0, "step":0.0001, "round": False}),
+                    }
+               }
+    RETURN_TYPES = ("SIGMAS",)
+    CATEGORY = "sampling/custom_sampling/schedulers"
+
+    FUNCTION = "get_sigmas"
+
+    def get_sigmas(self, steps, beta_d, beta_min, eps_s):
+        sigmas = k_diffusion_sampling.get_sigmas_vp(n=steps, beta_d=beta_d, beta_min=beta_min, eps_s=eps_s)
+        return (sigmas, )
+
+class SplitSigmas:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required":
+                    {"sigmas": ("SIGMAS", ),
+                    "step": ("INT", {"default": 0, "min": 0, "max": 10000}),
+                     }
+                }
+    RETURN_TYPES = ("SIGMAS","SIGMAS")
+    CATEGORY = "sampling/custom_sampling/sigmas"
+
+    FUNCTION = "get_sigmas"
+
+    def get_sigmas(self, sigmas, step):
+        sigmas1 = sigmas[:step + 1]
+        sigmas2 = sigmas[step:]
+        return (sigmas1, sigmas2)
+
+class FlipSigmas:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required":
+                    {"sigmas": ("SIGMAS", ),
+                     }
+                }
+    RETURN_TYPES = ("SIGMAS",)
+    CATEGORY = "sampling/custom_sampling/sigmas"
+
+    FUNCTION = "get_sigmas"
+
+    def get_sigmas(self, sigmas):
+        sigmas = sigmas.flip(0)
+        if sigmas[0] == 0:
+            sigmas[0] = 0.0001
+        return (sigmas,)
+
+class KSamplerSelect:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required":
+                    {"sampler_name": (ldm_patched.modules.samplers.SAMPLER_NAMES, ),
+                      }
+               }
+    RETURN_TYPES = ("SAMPLER",)
+    CATEGORY = "sampling/custom_sampling/samplers"
+
+    FUNCTION = "get_sampler"
+
+    def get_sampler(self, sampler_name):
+        sampler = ldm_patched.modules.samplers.sampler_object(sampler_name)
+        return (sampler, )
+
+class SamplerDPMPP_2M_SDE:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required":
+                    {"solver_type": (['midpoint', 'heun'], ),
+                     "eta": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}),
+                     "s_noise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}),
+                     "noise_device": (['gpu', 'cpu'], ),
+                      }
+               }
+    RETURN_TYPES = ("SAMPLER",)
+    CATEGORY = "sampling/custom_sampling/samplers"
+
+    FUNCTION = "get_sampler"
+
+    def get_sampler(self, solver_type, eta, s_noise, noise_device):
+        if noise_device == 'cpu':
+            sampler_name = "dpmpp_2m_sde"
+        else:
+            sampler_name = "dpmpp_2m_sde_gpu"
+        sampler = ldm_patched.modules.samplers.ksampler(sampler_name, {"eta": eta, "s_noise": s_noise, "solver_type": solver_type})
+        return (sampler, )
+
+
+class SamplerDPMPP_SDE:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required":
+                    {"eta": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}),
+                     "s_noise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}),
+                     "r": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 100.0, "step":0.01, "round": False}),
+                     "noise_device": (['gpu', 'cpu'], ),
+                      }
+               }
+    RETURN_TYPES = ("SAMPLER",)
+    CATEGORY = "sampling/custom_sampling/samplers"
+
+    FUNCTION = "get_sampler"
+
+    def get_sampler(self, eta, s_noise, r, noise_device):
+        if noise_device == 'cpu':
+            sampler_name = "dpmpp_sde"
+        else:
+            sampler_name = "dpmpp_sde_gpu"
+        sampler = ldm_patched.modules.samplers.ksampler(sampler_name, {"eta": eta, "s_noise": s_noise, "r": r})
+        return (sampler, )
+
+class SamplerCustom:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required":
+                    {"model": ("MODEL",),
+                    "add_noise": ("BOOLEAN", {"default": True}),
+                    "noise_seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff}),
+                    "cfg": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.1, "round": 0.01}),
+                    "positive": ("CONDITIONING", ),
+                    "negative": ("CONDITIONING", ),
+                    "sampler": ("SAMPLER", ),
+                    "sigmas": ("SIGMAS", ),
+                    "latent_image": ("LATENT", ),
+                     }
+                }
+
+    RETURN_TYPES = ("LATENT","LATENT")
+    RETURN_NAMES = ("output", "denoised_output")
+
+    FUNCTION = "sample"
+
+    CATEGORY = "sampling/custom_sampling"
+
+    def sample(self, model, add_noise, noise_seed, cfg, positive, negative, sampler, sigmas, latent_image):
+        latent = latent_image
+        latent_image = latent["samples"]
+        if not add_noise:
+            noise = torch.zeros(latent_image.size(), dtype=latent_image.dtype, layout=latent_image.layout, device="cpu")
+        else:
+            batch_inds = latent["batch_index"] if "batch_index" in latent else None
+            noise = ldm_patched.modules.sample.prepare_noise(latent_image, noise_seed, batch_inds)
+
+        noise_mask = None
+        if "noise_mask" in latent:
+            noise_mask = latent["noise_mask"]
+
+        x0_output = {}
+        callback = ldm_patched.utils.latent_visualization.prepare_callback(model, sigmas.shape[-1] - 1, x0_output)
+
+        disable_pbar = not ldm_patched.modules.utils.PROGRESS_BAR_ENABLED
+        samples = ldm_patched.modules.sample.sample_custom(model, noise, cfg, sampler, sigmas, positive, negative, latent_image, noise_mask=noise_mask, callback=callback, disable_pbar=disable_pbar, seed=noise_seed)
+
+        out = latent.copy()
+        out["samples"] = samples
+        if "x0" in x0_output:
+            out_denoised = latent.copy()
+            out_denoised["samples"] = model.model.process_latent_out(x0_output["x0"].cpu())
+        else:
+            out_denoised = out
+        return (out, out_denoised)
+
+NODE_CLASS_MAPPINGS = {
+    "SamplerCustom": SamplerCustom,
+    "BasicScheduler": BasicScheduler,
+    "KarrasScheduler": KarrasScheduler,
+    "ExponentialScheduler": ExponentialScheduler,
+    "PolyexponentialScheduler": PolyexponentialScheduler,
+    "VPScheduler": VPScheduler,
+    "SDTurboScheduler": SDTurboScheduler,
+    "KSamplerSelect": KSamplerSelect,
+    "SamplerDPMPP_2M_SDE": SamplerDPMPP_2M_SDE,
+    "SamplerDPMPP_SDE": SamplerDPMPP_SDE,
+    "SplitSigmas": SplitSigmas,
+    "FlipSigmas": FlipSigmas,
+}

ldm_patched/contrib/external_freelunch.py

@@ -0,0 +1,115 @@
+# https://github.com/comfyanonymous/ComfyUI/blob/master/nodes.py 
+
+#code originally taken from: https://github.com/ChenyangSi/FreeU (under MIT License)
+
+import torch
+
+
+def Fourier_filter(x, threshold, scale):
+    # FFT
+    x_freq = torch.fft.fftn(x.float(), dim=(-2, -1))
+    x_freq = torch.fft.fftshift(x_freq, dim=(-2, -1))
+
+    B, C, H, W = x_freq.shape
+    mask = torch.ones((B, C, H, W), device=x.device)
+
+    crow, ccol = H // 2, W //2
+    mask[..., crow - threshold:crow + threshold, ccol - threshold:ccol + threshold] = scale
+    x_freq = x_freq * mask
+
+    # IFFT
+    x_freq = torch.fft.ifftshift(x_freq, dim=(-2, -1))
+    x_filtered = torch.fft.ifftn(x_freq, dim=(-2, -1)).real
+
+    return x_filtered.to(x.dtype)
+
+
+class FreeU:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model": ("MODEL",),
+                             "b1": ("FLOAT", {"default": 1.1, "min": 0.0, "max": 10.0, "step": 0.01}),
+                             "b2": ("FLOAT", {"default": 1.2, "min": 0.0, "max": 10.0, "step": 0.01}),
+                             "s1": ("FLOAT", {"default": 0.9, "min": 0.0, "max": 10.0, "step": 0.01}),
+                             "s2": ("FLOAT", {"default": 0.2, "min": 0.0, "max": 10.0, "step": 0.01}),
+                              }}
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "patch"
+
+    CATEGORY = "_for_testing"
+
+    def patch(self, model, b1, b2, s1, s2):
+        model_channels = model.model.model_config.unet_config["model_channels"]
+        scale_dict = {model_channels * 4: (b1, s1), model_channels * 2: (b2, s2)}
+        on_cpu_devices = {}
+
+        def output_block_patch(h, hsp, transformer_options):
+            scale = scale_dict.get(h.shape[1], None)
+            if scale is not None:
+                h[:,:h.shape[1] // 2] = h[:,:h.shape[1] // 2] * scale[0]
+                if hsp.device not in on_cpu_devices:
+                    try:
+                        hsp = Fourier_filter(hsp, threshold=1, scale=scale[1])
+                    except:
+                        print("Device", hsp.device, "does not support the torch.fft functions used in the FreeU node, switching to CPU.")
+                        on_cpu_devices[hsp.device] = True
+                        hsp = Fourier_filter(hsp.cpu(), threshold=1, scale=scale[1]).to(hsp.device)
+                else:
+                    hsp = Fourier_filter(hsp.cpu(), threshold=1, scale=scale[1]).to(hsp.device)
+
+            return h, hsp
+
+        m = model.clone()
+        m.set_model_output_block_patch(output_block_patch)
+        return (m, )
+
+class FreeU_V2:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model": ("MODEL",),
+                             "b1": ("FLOAT", {"default": 1.3, "min": 0.0, "max": 10.0, "step": 0.01}),
+                             "b2": ("FLOAT", {"default": 1.4, "min": 0.0, "max": 10.0, "step": 0.01}),
+                             "s1": ("FLOAT", {"default": 0.9, "min": 0.0, "max": 10.0, "step": 0.01}),
+                             "s2": ("FLOAT", {"default": 0.2, "min": 0.0, "max": 10.0, "step": 0.01}),
+                              }}
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "patch"
+
+    CATEGORY = "_for_testing"
+
+    def patch(self, model, b1, b2, s1, s2):
+        model_channels = model.model.model_config.unet_config["model_channels"]
+        scale_dict = {model_channels * 4: (b1, s1), model_channels * 2: (b2, s2)}
+        on_cpu_devices = {}
+
+        def output_block_patch(h, hsp, transformer_options):
+            scale = scale_dict.get(h.shape[1], None)
+            if scale is not None:
+                hidden_mean = h.mean(1).unsqueeze(1)
+                B = hidden_mean.shape[0]
+                hidden_max, _ = torch.max(hidden_mean.view(B, -1), dim=-1, keepdim=True)
+                hidden_min, _ = torch.min(hidden_mean.view(B, -1), dim=-1, keepdim=True)
+                hidden_mean = (hidden_mean - hidden_min.unsqueeze(2).unsqueeze(3)) / (hidden_max - hidden_min).unsqueeze(2).unsqueeze(3)
+
+                h[:,:h.shape[1] // 2] = h[:,:h.shape[1] // 2] * ((scale[0] - 1 ) * hidden_mean + 1)
+
+                if hsp.device not in on_cpu_devices:
+                    try:
+                        hsp = Fourier_filter(hsp, threshold=1, scale=scale[1])
+                    except:
+                        print("Device", hsp.device, "does not support the torch.fft functions used in the FreeU node, switching to CPU.")
+                        on_cpu_devices[hsp.device] = True
+                        hsp = Fourier_filter(hsp.cpu(), threshold=1, scale=scale[1]).to(hsp.device)
+                else:
+                    hsp = Fourier_filter(hsp.cpu(), threshold=1, scale=scale[1]).to(hsp.device)
+
+            return h, hsp
+
+        m = model.clone()
+        m.set_model_output_block_patch(output_block_patch)
+        return (m, )
+
+NODE_CLASS_MAPPINGS = {
+    "FreeU": FreeU,
+    "FreeU_V2": FreeU_V2,
+}

ldm_patched/contrib/external_hypernetwork.py

@@ -0,0 +1,121 @@
+# https://github.com/comfyanonymous/ComfyUI/blob/master/nodes.py 
+
+import ldm_patched.modules.utils
+import ldm_patched.utils.path_utils
+import torch
+
+def load_hypernetwork_patch(path, strength):
+    sd = ldm_patched.modules.utils.load_torch_file(path, safe_load=True)
+    activation_func = sd.get('activation_func', 'linear')
+    is_layer_norm = sd.get('is_layer_norm', False)
+    use_dropout = sd.get('use_dropout', False)
+    activate_output = sd.get('activate_output', False)
+    last_layer_dropout = sd.get('last_layer_dropout', False)
+
+    valid_activation = {
+        "linear": torch.nn.Identity,
+        "relu": torch.nn.ReLU,
+        "leakyrelu": torch.nn.LeakyReLU,
+        "elu": torch.nn.ELU,
+        "swish": torch.nn.Hardswish,
+        "tanh": torch.nn.Tanh,
+        "sigmoid": torch.nn.Sigmoid,
+        "softsign": torch.nn.Softsign,
+        "mish": torch.nn.Mish,
+    }
+
+    if activation_func not in valid_activation:
+        print("Unsupported Hypernetwork format, if you report it I might implement it.", path, " ", activation_func, is_layer_norm, use_dropout, activate_output, last_layer_dropout)
+        return None
+
+    out = {}
+
+    for d in sd:
+        try:
+            dim = int(d)
+        except:
+            continue
+
+        output = []
+        for index in [0, 1]:
+            attn_weights = sd[dim][index]
+            keys = attn_weights.keys()
+
+            linears = filter(lambda a: a.endswith(".weight"), keys)
+            linears = list(map(lambda a: a[:-len(".weight")], linears))
+            layers = []
+
+            i = 0
+            while i < len(linears):
+                lin_name = linears[i]
+                last_layer = (i == (len(linears) - 1))
+                penultimate_layer = (i == (len(linears) - 2))
+
+                lin_weight = attn_weights['{}.weight'.format(lin_name)]
+                lin_bias = attn_weights['{}.bias'.format(lin_name)]
+                layer = torch.nn.Linear(lin_weight.shape[1], lin_weight.shape[0])
+                layer.load_state_dict({"weight": lin_weight, "bias": lin_bias})
+                layers.append(layer)
+                if activation_func != "linear":
+                    if (not last_layer) or (activate_output):
+                        layers.append(valid_activation[activation_func]())
+                if is_layer_norm:
+                    i += 1
+                    ln_name = linears[i]
+                    ln_weight = attn_weights['{}.weight'.format(ln_name)]
+                    ln_bias = attn_weights['{}.bias'.format(ln_name)]
+                    ln = torch.nn.LayerNorm(ln_weight.shape[0])
+                    ln.load_state_dict({"weight": ln_weight, "bias": ln_bias})
+                    layers.append(ln)
+                if use_dropout:
+                    if (not last_layer) and (not penultimate_layer or last_layer_dropout):
+                        layers.append(torch.nn.Dropout(p=0.3))
+                i += 1
+
+            output.append(torch.nn.Sequential(*layers))
+        out[dim] = torch.nn.ModuleList(output)
+
+    class hypernetwork_patch:
+        def __init__(self, hypernet, strength):
+            self.hypernet = hypernet
+            self.strength = strength
+        def __call__(self, q, k, v, extra_options):
+            dim = k.shape[-1]
+            if dim in self.hypernet:
+                hn = self.hypernet[dim]
+                k = k + hn[0](k) * self.strength
+                v = v + hn[1](v) * self.strength
+
+            return q, k, v
+
+        def to(self, device):
+            for d in self.hypernet.keys():
+                self.hypernet[d] = self.hypernet[d].to(device)
+            return self
+
+    return hypernetwork_patch(out, strength)
+
+class HypernetworkLoader:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model": ("MODEL",),
+                              "hypernetwork_name": (ldm_patched.utils.path_utils.get_filename_list("hypernetworks"), ),
+                              "strength": ("FLOAT", {"default": 1.0, "min": -10.0, "max": 10.0, "step": 0.01}),
+                              }}
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "load_hypernetwork"
+
+    CATEGORY = "loaders"
+
+    def load_hypernetwork(self, model, hypernetwork_name, strength):
+        hypernetwork_path = ldm_patched.utils.path_utils.get_full_path("hypernetworks", hypernetwork_name)
+        model_hypernetwork = model.clone()
+        patch = load_hypernetwork_patch(hypernetwork_path, strength)
+        if patch is not None:
+            model_hypernetwork.set_model_attn1_patch(patch)
+            model_hypernetwork.set_model_attn2_patch(patch)
+        return (model_hypernetwork,)
+
+NODE_CLASS_MAPPINGS = {
+    "HypernetworkLoader": HypernetworkLoader
+}

ldm_patched/contrib/external_hypertile.py

@@ -0,0 +1,85 @@
+# https://github.com/comfyanonymous/ComfyUI/blob/master/nodes.py 
+
+#Taken from: https://github.com/tfernd/HyperTile/
+
+import math
+from einops import rearrange
+# Use torch rng for consistency across generations
+from torch import randint
+
+def random_divisor(value: int, min_value: int, /, max_options: int = 1) -> int:
+    min_value = min(min_value, value)
+
+    # All big divisors of value (inclusive)
+    divisors = [i for i in range(min_value, value + 1) if value % i == 0]
+
+    ns = [value // i for i in divisors[:max_options]]  # has at least 1 element
+
+    if len(ns) - 1 > 0:
+        idx = randint(low=0, high=len(ns) - 1, size=(1,)).item()
+    else:
+        idx = 0
+
+    return ns[idx]
+
+class HyperTile:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model": ("MODEL",),
+                             "tile_size": ("INT", {"default": 256, "min": 1, "max": 2048}),
+                             "swap_size": ("INT", {"default": 2, "min": 1, "max": 128}),
+                             "max_depth": ("INT", {"default": 0, "min": 0, "max": 10}),
+                             "scale_depth": ("BOOLEAN", {"default": False}),
+                              }}
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "patch"
+
+    CATEGORY = "_for_testing"
+
+    def patch(self, model, tile_size, swap_size, max_depth, scale_depth):
+        model_channels = model.model.model_config.unet_config["model_channels"]
+
+        apply_to = set()
+        temp = model_channels
+        for x in range(max_depth + 1):
+            apply_to.add(temp)
+            temp *= 2
+
+        latent_tile_size = max(32, tile_size) // 8
+        self.temp = None
+
+        def hypertile_in(q, k, v, extra_options):
+            if q.shape[-1] in apply_to:
+                shape = extra_options["original_shape"]
+                aspect_ratio = shape[-1] / shape[-2]
+
+                hw = q.size(1)
+                h, w = round(math.sqrt(hw * aspect_ratio)), round(math.sqrt(hw / aspect_ratio))
+
+                factor = 2**((q.shape[-1] // model_channels) - 1) if scale_depth else 1
+                nh = random_divisor(h, latent_tile_size * factor, swap_size)
+                nw = random_divisor(w, latent_tile_size * factor, swap_size)
+
+                if nh * nw > 1:
+                    q = rearrange(q, "b (nh h nw w) c -> (b nh nw) (h w) c", h=h // nh, w=w // nw, nh=nh, nw=nw)
+                    self.temp = (nh, nw, h, w)
+                return q, k, v
+
+            return q, k, v
+        def hypertile_out(out, extra_options):
+            if self.temp is not None:
+                nh, nw, h, w = self.temp
+                self.temp = None
+                out = rearrange(out, "(b nh nw) hw c -> b nh nw hw c", nh=nh, nw=nw)
+                out = rearrange(out, "b nh nw (h w) c -> b (nh h nw w) c", h=h // nh, w=w // nw)
+            return out
+
+
+        m = model.clone()
+        m.set_model_attn1_patch(hypertile_in)
+        m.set_model_attn1_output_patch(hypertile_out)
+        return (m, )
+
+NODE_CLASS_MAPPINGS = {
+    "HyperTile": HyperTile,
+}

ldm_patched/contrib/external_images.py

@@ -0,0 +1,177 @@
+# https://github.com/comfyanonymous/ComfyUI/blob/master/nodes.py 
+
+import ldm_patched.contrib.external
+import ldm_patched.utils.path_utils
+from ldm_patched.modules.args_parser import args
+
+from PIL import Image
+from PIL.PngImagePlugin import PngInfo
+
+import numpy as np
+import json
+import os
+
+MAX_RESOLUTION = ldm_patched.contrib.external.MAX_RESOLUTION
+
+class ImageCrop:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "image": ("IMAGE",),
+                              "width": ("INT", {"default": 512, "min": 1, "max": MAX_RESOLUTION, "step": 1}),
+                              "height": ("INT", {"default": 512, "min": 1, "max": MAX_RESOLUTION, "step": 1}),
+                              "x": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}),
+                              "y": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}),
+                              }}
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "crop"
+
+    CATEGORY = "image/transform"
+
+    def crop(self, image, width, height, x, y):
+        x = min(x, image.shape[2] - 1)
+        y = min(y, image.shape[1] - 1)
+        to_x = width + x
+        to_y = height + y
+        img = image[:,y:to_y, x:to_x, :]
+        return (img,)
+
+class RepeatImageBatch:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "image": ("IMAGE",),
+                              "amount": ("INT", {"default": 1, "min": 1, "max": 64}),
+                              }}
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "repeat"
+
+    CATEGORY = "image/batch"
+
+    def repeat(self, image, amount):
+        s = image.repeat((amount, 1,1,1))
+        return (s,)
+
+class SaveAnimatedWEBP:
+    def __init__(self):
+        self.output_dir = ldm_patched.utils.path_utils.get_output_directory()
+        self.type = "output"
+        self.prefix_append = ""
+
+    methods = {"default": 4, "fastest": 0, "slowest": 6}
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required":
+                    {"images": ("IMAGE", ),
+                     "filename_prefix": ("STRING", {"default": "ldm_patched"}),
+                     "fps": ("FLOAT", {"default": 6.0, "min": 0.01, "max": 1000.0, "step": 0.01}),
+                     "lossless": ("BOOLEAN", {"default": True}),
+                     "quality": ("INT", {"default": 80, "min": 0, "max": 100}),
+                     "method": (list(s.methods.keys()),),
+                     # "num_frames": ("INT", {"default": 0, "min": 0, "max": 8192}),
+                     },
+                "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"},
+                }
+
+    RETURN_TYPES = ()
+    FUNCTION = "save_images"
+
+    OUTPUT_NODE = True
+
+    CATEGORY = "_for_testing"
+
+    def save_images(self, images, fps, filename_prefix, lossless, quality, method, num_frames=0, prompt=None, extra_pnginfo=None):
+        method = self.methods.get(method)
+        filename_prefix += self.prefix_append
+        full_output_folder, filename, counter, subfolder, filename_prefix = ldm_patched.utils.path_utils.get_save_image_path(filename_prefix, self.output_dir, images[0].shape[1], images[0].shape[0])
+        results = list()
+        pil_images = []
+        for image in images:
+            i = 255. * image.cpu().numpy()
+            img = Image.fromarray(np.clip(i, 0, 255).astype(np.uint8))
+            pil_images.append(img)
+
+        metadata = pil_images[0].getexif()
+        if not args.disable_server_info:
+            if prompt is not None:
+                metadata[0x0110] = "prompt:{}".format(json.dumps(prompt))
+            if extra_pnginfo is not None:
+                inital_exif = 0x010f
+                for x in extra_pnginfo:
+                    metadata[inital_exif] = "{}:{}".format(x, json.dumps(extra_pnginfo[x]))
+                    inital_exif -= 1
+
+        if num_frames == 0:
+            num_frames = len(pil_images)
+
+        c = len(pil_images)
+        for i in range(0, c, num_frames):
+            file = f"{filename}_{counter:05}_.webp"
+            pil_images[i].save(os.path.join(full_output_folder, file), save_all=True, duration=int(1000.0/fps), append_images=pil_images[i + 1:i + num_frames], exif=metadata, lossless=lossless, quality=quality, method=method)
+            results.append({
+                "filename": file,
+                "subfolder": subfolder,
+                "type": self.type
+            })
+            counter += 1
+
+        animated = num_frames != 1
+        return { "ui": { "images": results, "animated": (animated,) } }
+
+class SaveAnimatedPNG:
+    def __init__(self):
+        self.output_dir = ldm_patched.utils.path_utils.get_output_directory()
+        self.type = "output"
+        self.prefix_append = ""
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required":
+                    {"images": ("IMAGE", ),
+                     "filename_prefix": ("STRING", {"default": "ldm_patched"}),
+                     "fps": ("FLOAT", {"default": 6.0, "min": 0.01, "max": 1000.0, "step": 0.01}),
+                     "compress_level": ("INT", {"default": 4, "min": 0, "max": 9})
+                     },
+                "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"},
+                }
+
+    RETURN_TYPES = ()
+    FUNCTION = "save_images"
+
+    OUTPUT_NODE = True
+
+    CATEGORY = "_for_testing"
+
+    def save_images(self, images, fps, compress_level, filename_prefix="ldm_patched", prompt=None, extra_pnginfo=None):
+        filename_prefix += self.prefix_append
+        full_output_folder, filename, counter, subfolder, filename_prefix = ldm_patched.utils.path_utils.get_save_image_path(filename_prefix, self.output_dir, images[0].shape[1], images[0].shape[0])
+        results = list()
+        pil_images = []
+        for image in images:
+            i = 255. * image.cpu().numpy()
+            img = Image.fromarray(np.clip(i, 0, 255).astype(np.uint8))
+            pil_images.append(img)
+
+        metadata = None
+        if not args.disable_server_info:
+            metadata = PngInfo()
+            if prompt is not None:
+                metadata.add(b"ldm_patched", "prompt".encode("latin-1", "strict") + b"\0" + json.dumps(prompt).encode("latin-1", "strict"), after_idat=True)
+            if extra_pnginfo is not None:
+                for x in extra_pnginfo:
+                    metadata.add(b"ldm_patched", x.encode("latin-1", "strict") + b"\0" + json.dumps(extra_pnginfo[x]).encode("latin-1", "strict"), after_idat=True)
+
+        file = f"{filename}_{counter:05}_.png"
+        pil_images[0].save(os.path.join(full_output_folder, file), pnginfo=metadata, compress_level=compress_level, save_all=True, duration=int(1000.0/fps), append_images=pil_images[1:])
+        results.append({
+            "filename": file,
+            "subfolder": subfolder,
+            "type": self.type
+        })
+
+        return { "ui": { "images": results, "animated": (True,)} }
+
+NODE_CLASS_MAPPINGS = {
+    "ImageCrop": ImageCrop,
+    "RepeatImageBatch": RepeatImageBatch,
+    "SaveAnimatedWEBP": SaveAnimatedWEBP,
+    "SaveAnimatedPNG": SaveAnimatedPNG,
+}

ldm_patched/contrib/external_latent.py

@@ -0,0 +1,112 @@
+# https://github.com/comfyanonymous/ComfyUI/blob/master/nodes.py 
+
+import ldm_patched.modules.utils
+import torch
+
+def reshape_latent_to(target_shape, latent):
+    if latent.shape[1:] != target_shape[1:]:
+        latent.movedim(1, -1)
+        latent = ldm_patched.modules.utils.common_upscale(latent, target_shape[3], target_shape[2], "bilinear", "center")
+        latent.movedim(-1, 1)
+    return ldm_patched.modules.utils.repeat_to_batch_size(latent, target_shape[0])
+
+
+class LatentAdd:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "samples1": ("LATENT",), "samples2": ("LATENT",)}}
+
+    RETURN_TYPES = ("LATENT",)
+    FUNCTION = "op"
+
+    CATEGORY = "latent/advanced"
+
+    def op(self, samples1, samples2):
+        samples_out = samples1.copy()
+
+        s1 = samples1["samples"]
+        s2 = samples2["samples"]
+
+        s2 = reshape_latent_to(s1.shape, s2)
+        samples_out["samples"] = s1 + s2
+        return (samples_out,)
+
+class LatentSubtract:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "samples1": ("LATENT",), "samples2": ("LATENT",)}}
+
+    RETURN_TYPES = ("LATENT",)
+    FUNCTION = "op"
+
+    CATEGORY = "latent/advanced"
+
+    def op(self, samples1, samples2):
+        samples_out = samples1.copy()
+
+        s1 = samples1["samples"]
+        s2 = samples2["samples"]
+
+        s2 = reshape_latent_to(s1.shape, s2)
+        samples_out["samples"] = s1 - s2
+        return (samples_out,)
+
+class LatentMultiply:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "samples": ("LATENT",),
+                              "multiplier": ("FLOAT", {"default": 1.0, "min": -10.0, "max": 10.0, "step": 0.01}),
+                             }}
+
+    RETURN_TYPES = ("LATENT",)
+    FUNCTION = "op"
+
+    CATEGORY = "latent/advanced"
+
+    def op(self, samples, multiplier):
+        samples_out = samples.copy()
+
+        s1 = samples["samples"]
+        samples_out["samples"] = s1 * multiplier
+        return (samples_out,)
+
+class LatentInterpolate:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "samples1": ("LATENT",),
+                              "samples2": ("LATENT",),
+                              "ratio": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}),
+                              }}
+
+    RETURN_TYPES = ("LATENT",)
+    FUNCTION = "op"
+
+    CATEGORY = "latent/advanced"
+
+    def op(self, samples1, samples2, ratio):
+        samples_out = samples1.copy()
+
+        s1 = samples1["samples"]
+        s2 = samples2["samples"]
+
+        s2 = reshape_latent_to(s1.shape, s2)
+
+        m1 = torch.linalg.vector_norm(s1, dim=(1))
+        m2 = torch.linalg.vector_norm(s2, dim=(1))
+
+        s1 = torch.nan_to_num(s1 / m1)
+        s2 = torch.nan_to_num(s2 / m2)
+
+        t = (s1 * ratio + s2 * (1.0 - ratio))
+        mt = torch.linalg.vector_norm(t, dim=(1))
+        st = torch.nan_to_num(t / mt)
+
+        samples_out["samples"] = st * (m1 * ratio + m2 * (1.0 - ratio))
+        return (samples_out,)
+
+NODE_CLASS_MAPPINGS = {
+    "LatentAdd": LatentAdd,
+    "LatentSubtract": LatentSubtract,
+    "LatentMultiply": LatentMultiply,
+    "LatentInterpolate": LatentInterpolate,
+}

ldm_patched/contrib/external_mask.py

@@ -0,0 +1,364 @@
+# https://github.com/comfyanonymous/ComfyUI/blob/master/nodes.py 
+
+import numpy as np
+import scipy.ndimage
+import torch
+import ldm_patched.modules.utils
+
+from ldm_patched.contrib.external import MAX_RESOLUTION
+
+def composite(destination, source, x, y, mask = None, multiplier = 8, resize_source = False):
+    if resize_source:
+        source = torch.nn.functional.interpolate(source, size=(destination.shape[2], destination.shape[3]), mode="bilinear")
+
+    source = ldm_patched.modules.utils.repeat_to_batch_size(source, destination.shape[0])
+
+    x = max(-source.shape[3] * multiplier, min(x, destination.shape[3] * multiplier))
+    y = max(-source.shape[2] * multiplier, min(y, destination.shape[2] * multiplier))
+
+    left, top = (x // multiplier, y // multiplier)
+    right, bottom = (left + source.shape[3], top + source.shape[2],)
+
+    if mask is None:
+        mask = torch.ones_like(source)
+    else:
+        mask = mask.clone()
+        mask = torch.nn.functional.interpolate(mask.reshape((-1, 1, mask.shape[-2], mask.shape[-1])), size=(source.shape[2], source.shape[3]), mode="bilinear")
+        mask = ldm_patched.modules.utils.repeat_to_batch_size(mask, source.shape[0])
+
+    # calculate the bounds of the source that will be overlapping the destination
+    # this prevents the source trying to overwrite latent pixels that are out of bounds
+    # of the destination
+    visible_width, visible_height = (destination.shape[3] - left + min(0, x), destination.shape[2] - top + min(0, y),)
+
+    mask = mask[:, :, :visible_height, :visible_width]
+    inverse_mask = torch.ones_like(mask) - mask
+
+    source_portion = mask * source[:, :, :visible_height, :visible_width]
+    destination_portion = inverse_mask  * destination[:, :, top:bottom, left:right]
+
+    destination[:, :, top:bottom, left:right] = source_portion + destination_portion
+    return destination
+
+class LatentCompositeMasked:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "destination": ("LATENT",),
+                "source": ("LATENT",),
+                "x": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 8}),
+                "y": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 8}),
+                "resize_source": ("BOOLEAN", {"default": False}),
+            },
+            "optional": {
+                "mask": ("MASK",),
+            }
+        }
+    RETURN_TYPES = ("LATENT",)
+    FUNCTION = "composite"
+
+    CATEGORY = "latent"
+
+    def composite(self, destination, source, x, y, resize_source, mask = None):
+        output = destination.copy()
+        destination = destination["samples"].clone()
+        source = source["samples"]
+        output["samples"] = composite(destination, source, x, y, mask, 8, resize_source)
+        return (output,)
+
+class ImageCompositeMasked:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "destination": ("IMAGE",),
+                "source": ("IMAGE",),
+                "x": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}),
+                "y": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}),
+                "resize_source": ("BOOLEAN", {"default": False}),
+            },
+            "optional": {
+                "mask": ("MASK",),
+            }
+        }
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "composite"
+
+    CATEGORY = "image"
+
+    def composite(self, destination, source, x, y, resize_source, mask = None):
+        destination = destination.clone().movedim(-1, 1)
+        output = composite(destination, source.movedim(-1, 1), x, y, mask, 1, resize_source).movedim(1, -1)
+        return (output,)
+
+class MaskToImage:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+                "required": {
+                    "mask": ("MASK",),
+                }
+        }
+
+    CATEGORY = "mask"
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "mask_to_image"
+
+    def mask_to_image(self, mask):
+        result = mask.reshape((-1, 1, mask.shape[-2], mask.shape[-1])).movedim(1, -1).expand(-1, -1, -1, 3)
+        return (result,)
+
+class ImageToMask:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+                "required": {
+                    "image": ("IMAGE",),
+                    "channel": (["red", "green", "blue", "alpha"],),
+                }
+        }
+
+    CATEGORY = "mask"
+
+    RETURN_TYPES = ("MASK",)
+    FUNCTION = "image_to_mask"
+
+    def image_to_mask(self, image, channel):
+        channels = ["red", "green", "blue", "alpha"]
+        mask = image[:, :, :, channels.index(channel)]
+        return (mask,)
+
+class ImageColorToMask:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+                "required": {
+                    "image": ("IMAGE",),
+                    "color": ("INT", {"default": 0, "min": 0, "max": 0xFFFFFF, "step": 1, "display": "color"}),
+                }
+        }
+
+    CATEGORY = "mask"
+
+    RETURN_TYPES = ("MASK",)
+    FUNCTION = "image_to_mask"
+
+    def image_to_mask(self, image, color):
+        temp = (torch.clamp(image, 0, 1.0) * 255.0).round().to(torch.int)
+        temp = torch.bitwise_left_shift(temp[:,:,:,0], 16) + torch.bitwise_left_shift(temp[:,:,:,1], 8) + temp[:,:,:,2]
+        mask = torch.where(temp == color, 255, 0).float()
+        return (mask,)
+
+class SolidMask:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}),
+                "width": ("INT", {"default": 512, "min": 1, "max": MAX_RESOLUTION, "step": 1}),
+                "height": ("INT", {"default": 512, "min": 1, "max": MAX_RESOLUTION, "step": 1}),
+            }
+        }
+
+    CATEGORY = "mask"
+
+    RETURN_TYPES = ("MASK",)
+
+    FUNCTION = "solid"
+
+    def solid(self, value, width, height):
+        out = torch.full((1, height, width), value, dtype=torch.float32, device="cpu")
+        return (out,)
+
+class InvertMask:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "mask": ("MASK",),
+            }
+        }
+
+    CATEGORY = "mask"
+
+    RETURN_TYPES = ("MASK",)
+
+    FUNCTION = "invert"
+
+    def invert(self, mask):
+        out = 1.0 - mask
+        return (out,)
+
+class CropMask:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "mask": ("MASK",),
+                "x": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}),
+                "y": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}),
+                "width": ("INT", {"default": 512, "min": 1, "max": MAX_RESOLUTION, "step": 1}),
+                "height": ("INT", {"default": 512, "min": 1, "max": MAX_RESOLUTION, "step": 1}),
+            }
+        }
+
+    CATEGORY = "mask"
+
+    RETURN_TYPES = ("MASK",)
+
+    FUNCTION = "crop"
+
+    def crop(self, mask, x, y, width, height):
+        mask = mask.reshape((-1, mask.shape[-2], mask.shape[-1]))
+        out = mask[:, y:y + height, x:x + width]
+        return (out,)
+
+class MaskComposite:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "destination": ("MASK",),
+                "source": ("MASK",),
+                "x": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}),
+                "y": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}),
+                "operation": (["multiply", "add", "subtract", "and", "or", "xor"],),
+            }
+        }
+
+    CATEGORY = "mask"
+
+    RETURN_TYPES = ("MASK",)
+
+    FUNCTION = "combine"
+
+    def combine(self, destination, source, x, y, operation):
+        output = destination.reshape((-1, destination.shape[-2], destination.shape[-1])).clone()
+        source = source.reshape((-1, source.shape[-2], source.shape[-1]))
+
+        left, top = (x, y,)
+        right, bottom = (min(left + source.shape[-1], destination.shape[-1]), min(top + source.shape[-2], destination.shape[-2]))
+        visible_width, visible_height = (right - left, bottom - top,)
+
+        source_portion = source[:, :visible_height, :visible_width]
+        destination_portion = destination[:, top:bottom, left:right]
+
+        if operation == "multiply":
+            output[:, top:bottom, left:right] = destination_portion * source_portion
+        elif operation == "add":
+            output[:, top:bottom, left:right] = destination_portion + source_portion
+        elif operation == "subtract":
+            output[:, top:bottom, left:right] = destination_portion - source_portion
+        elif operation == "and":
+            output[:, top:bottom, left:right] = torch.bitwise_and(destination_portion.round().bool(), source_portion.round().bool()).float()
+        elif operation == "or":
+            output[:, top:bottom, left:right] = torch.bitwise_or(destination_portion.round().bool(), source_portion.round().bool()).float()
+        elif operation == "xor":
+            output[:, top:bottom, left:right] = torch.bitwise_xor(destination_portion.round().bool(), source_portion.round().bool()).float()
+
+        output = torch.clamp(output, 0.0, 1.0)
+
+        return (output,)
+
+class FeatherMask:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "mask": ("MASK",),
+                "left": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}),
+                "top": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}),
+                "right": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}),
+                "bottom": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}),
+            }
+        }
+
+    CATEGORY = "mask"
+
+    RETURN_TYPES = ("MASK",)
+
+    FUNCTION = "feather"
+
+    def feather(self, mask, left, top, right, bottom):
+        output = mask.reshape((-1, mask.shape[-2], mask.shape[-1])).clone()
+
+        left = min(left, output.shape[-1])
+        right = min(right, output.shape[-1])
+        top = min(top, output.shape[-2])
+        bottom = min(bottom, output.shape[-2])
+
+        for x in range(left):
+            feather_rate = (x + 1.0) / left
+            output[:, :, x] *= feather_rate
+
+        for x in range(right):
+            feather_rate = (x + 1) / right
+            output[:, :, -x] *= feather_rate
+
+        for y in range(top):
+            feather_rate = (y + 1) / top
+            output[:, y, :] *= feather_rate
+
+        for y in range(bottom):
+            feather_rate = (y + 1) / bottom
+            output[:, -y, :] *= feather_rate
+
+        return (output,)
+    
+class GrowMask:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "mask": ("MASK",),
+                "expand": ("INT", {"default": 0, "min": -MAX_RESOLUTION, "max": MAX_RESOLUTION, "step": 1}),
+                "tapered_corners": ("BOOLEAN", {"default": True}),
+            },
+        }
+    
+    CATEGORY = "mask"
+
+    RETURN_TYPES = ("MASK",)
+
+    FUNCTION = "expand_mask"
+
+    def expand_mask(self, mask, expand, tapered_corners):
+        c = 0 if tapered_corners else 1
+        kernel = np.array([[c, 1, c],
+                           [1, 1, 1],
+                           [c, 1, c]])
+        mask = mask.reshape((-1, mask.shape[-2], mask.shape[-1]))
+        out = []
+        for m in mask:
+            output = m.numpy()
+            for _ in range(abs(expand)):
+                if expand < 0:
+                    output = scipy.ndimage.grey_erosion(output, footprint=kernel)
+                else:
+                    output = scipy.ndimage.grey_dilation(output, footprint=kernel)
+            output = torch.from_numpy(output)
+            out.append(output)
+        return (torch.stack(out, dim=0),)
+
+
+
+NODE_CLASS_MAPPINGS = {
+    "LatentCompositeMasked": LatentCompositeMasked,
+    "ImageCompositeMasked": ImageCompositeMasked,
+    "MaskToImage": MaskToImage,
+    "ImageToMask": ImageToMask,
+    "ImageColorToMask": ImageColorToMask,
+    "SolidMask": SolidMask,
+    "InvertMask": InvertMask,
+    "CropMask": CropMask,
+    "MaskComposite": MaskComposite,
+    "FeatherMask": FeatherMask,
+    "GrowMask": GrowMask,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "ImageToMask": "Convert Image to Mask",
+    "MaskToImage": "Convert Mask to Image",
+}

ldm_patched/contrib/external_model_advanced.py

@@ -0,0 +1,207 @@
+# https://github.com/comfyanonymous/ComfyUI/blob/master/nodes.py 
+
+import ldm_patched.utils.path_utils
+import ldm_patched.modules.sd
+import ldm_patched.modules.model_sampling
+import torch
+
+class LCM(ldm_patched.modules.model_sampling.EPS):
+    def calculate_denoised(self, sigma, model_output, model_input):
+        timestep = self.timestep(sigma).view(sigma.shape[:1] + (1,) * (model_output.ndim - 1))
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (model_output.ndim - 1))
+        x0 = model_input - model_output * sigma
+
+        sigma_data = 0.5
+        scaled_timestep = timestep * 10.0 #timestep_scaling
+
+        c_skip = sigma_data**2 / (scaled_timestep**2 + sigma_data**2)
+        c_out = scaled_timestep / (scaled_timestep**2 + sigma_data**2) ** 0.5
+
+        return c_out * x0 + c_skip * model_input
+
+class ModelSamplingDiscreteDistilled(torch.nn.Module):
+    original_timesteps = 50
+
+    def __init__(self):
+        super().__init__()
+        self.sigma_data = 1.0
+        timesteps = 1000
+        beta_start = 0.00085
+        beta_end = 0.012
+
+        betas = torch.linspace(beta_start**0.5, beta_end**0.5, timesteps, dtype=torch.float32) ** 2
+        alphas = 1.0 - betas
+        alphas_cumprod = torch.cumprod(alphas, dim=0)
+
+        self.skip_steps = timesteps // self.original_timesteps
+
+
+        alphas_cumprod_valid = torch.zeros((self.original_timesteps), dtype=torch.float32)
+        for x in range(self.original_timesteps):
+            alphas_cumprod_valid[self.original_timesteps - 1 - x] = alphas_cumprod[timesteps - 1 - x * self.skip_steps]
+
+        sigmas = ((1 - alphas_cumprod_valid) / alphas_cumprod_valid) ** 0.5
+        self.set_sigmas(sigmas)
+
+    def set_sigmas(self, sigmas):
+        self.register_buffer('sigmas', sigmas)
+        self.register_buffer('log_sigmas', sigmas.log())
+
+    @property
+    def sigma_min(self):
+        return self.sigmas[0]
+
+    @property
+    def sigma_max(self):
+        return self.sigmas[-1]
+
+    def timestep(self, sigma):
+        log_sigma = sigma.log()
+        dists = log_sigma.to(self.log_sigmas.device) - self.log_sigmas[:, None]
+        return (dists.abs().argmin(dim=0).view(sigma.shape) * self.skip_steps + (self.skip_steps - 1)).to(sigma.device)
+
+    def sigma(self, timestep):
+        t = torch.clamp(((timestep.float().to(self.log_sigmas.device) - (self.skip_steps - 1)) / self.skip_steps).float(), min=0, max=(len(self.sigmas) - 1))
+        low_idx = t.floor().long()
+        high_idx = t.ceil().long()
+        w = t.frac()
+        log_sigma = (1 - w) * self.log_sigmas[low_idx] + w * self.log_sigmas[high_idx]
+        return log_sigma.exp().to(timestep.device)
+
+    def percent_to_sigma(self, percent):
+        if percent <= 0.0:
+            return 999999999.9
+        if percent >= 1.0:
+            return 0.0
+        percent = 1.0 - percent
+        return self.sigma(torch.tensor(percent * 999.0)).item()
+
+
+def rescale_zero_terminal_snr_sigmas(sigmas):
+    alphas_cumprod = 1 / ((sigmas * sigmas) + 1)
+    alphas_bar_sqrt = alphas_cumprod.sqrt()
+
+    # Store old values.
+    alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
+    alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()
+
+    # Shift so the last timestep is zero.
+    alphas_bar_sqrt -= (alphas_bar_sqrt_T)
+
+    # Scale so the first timestep is back to the old value.
+    alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)
+
+    # Convert alphas_bar_sqrt to betas
+    alphas_bar = alphas_bar_sqrt**2  # Revert sqrt
+    alphas_bar[-1] = 4.8973451890853435e-08
+    return ((1 - alphas_bar) / alphas_bar) ** 0.5
+
+class ModelSamplingDiscrete:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model": ("MODEL",),
+                              "sampling": (["eps", "v_prediction", "lcm"],),
+                              "zsnr": ("BOOLEAN", {"default": False}),
+                              }}
+
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "patch"
+
+    CATEGORY = "advanced/model"
+
+    def patch(self, model, sampling, zsnr):
+        m = model.clone()
+
+        sampling_base = ldm_patched.modules.model_sampling.ModelSamplingDiscrete
+        if sampling == "eps":
+            sampling_type = ldm_patched.modules.model_sampling.EPS
+        elif sampling == "v_prediction":
+            sampling_type = ldm_patched.modules.model_sampling.V_PREDICTION
+        elif sampling == "lcm":
+            sampling_type = LCM
+            sampling_base = ModelSamplingDiscreteDistilled
+
+        class ModelSamplingAdvanced(sampling_base, sampling_type):
+            pass
+
+        model_sampling = ModelSamplingAdvanced()
+        if zsnr:
+            model_sampling.set_sigmas(rescale_zero_terminal_snr_sigmas(model_sampling.sigmas))
+
+        m.add_object_patch("model_sampling", model_sampling)
+        return (m, )
+
+class ModelSamplingContinuousEDM:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model": ("MODEL",),
+                              "sampling": (["v_prediction", "eps"],),
+                              "sigma_max": ("FLOAT", {"default": 120.0, "min": 0.0, "max": 1000.0, "step":0.001, "round": False}),
+                              "sigma_min": ("FLOAT", {"default": 0.002, "min": 0.0, "max": 1000.0, "step":0.001, "round": False}),
+                              }}
+
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "patch"
+
+    CATEGORY = "advanced/model"
+
+    def patch(self, model, sampling, sigma_max, sigma_min):
+        m = model.clone()
+
+        if sampling == "eps":
+            sampling_type = ldm_patched.modules.model_sampling.EPS
+        elif sampling == "v_prediction":
+            sampling_type = ldm_patched.modules.model_sampling.V_PREDICTION
+
+        class ModelSamplingAdvanced(ldm_patched.modules.model_sampling.ModelSamplingContinuousEDM, sampling_type):
+            pass
+
+        model_sampling = ModelSamplingAdvanced()
+        model_sampling.set_sigma_range(sigma_min, sigma_max)
+        m.add_object_patch("model_sampling", model_sampling)
+        return (m, )
+
+class RescaleCFG:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model": ("MODEL",),
+                              "multiplier": ("FLOAT", {"default": 0.7, "min": 0.0, "max": 1.0, "step": 0.01}),
+                              }}
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "patch"
+
+    CATEGORY = "advanced/model"
+
+    def patch(self, model, multiplier):
+        def rescale_cfg(args):
+            cond = args["cond"]
+            uncond = args["uncond"]
+            cond_scale = args["cond_scale"]
+            sigma = args["sigma"]
+            sigma = sigma.view(sigma.shape[:1] + (1,) * (cond.ndim - 1))
+            x_orig = args["input"]
+
+            #rescale cfg has to be done on v-pred model output
+            x = x_orig / (sigma * sigma + 1.0)
+            cond = ((x - (x_orig - cond)) * (sigma ** 2 + 1.0) ** 0.5) / (sigma)
+            uncond = ((x - (x_orig - uncond)) * (sigma ** 2 + 1.0) ** 0.5) / (sigma)
+
+            #rescalecfg
+            x_cfg = uncond + cond_scale * (cond - uncond)
+            ro_pos = torch.std(cond, dim=(1,2,3), keepdim=True)
+            ro_cfg = torch.std(x_cfg, dim=(1,2,3), keepdim=True)
+
+            x_rescaled = x_cfg * (ro_pos / ro_cfg)
+            x_final = multiplier * x_rescaled + (1.0 - multiplier) * x_cfg
+
+            return x_orig - (x - x_final * sigma / (sigma * sigma + 1.0) ** 0.5)
+
+        m = model.clone()
+        m.set_model_sampler_cfg_function(rescale_cfg)
+        return (m, )
+
+NODE_CLASS_MAPPINGS = {
+    "ModelSamplingDiscrete": ModelSamplingDiscrete,
+    "ModelSamplingContinuousEDM": ModelSamplingContinuousEDM,
+    "RescaleCFG": RescaleCFG,
+}

ldm_patched/contrib/external_model_downscale.py

@@ -0,0 +1,55 @@
+# https://github.com/comfyanonymous/ComfyUI/blob/master/nodes.py 
+
+import torch
+import ldm_patched.modules.utils
+
+class PatchModelAddDownscale:
+    upscale_methods = ["bicubic", "nearest-exact", "bilinear", "area", "bislerp"]
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model": ("MODEL",),
+                              "block_number": ("INT", {"default": 3, "min": 1, "max": 32, "step": 1}),
+                              "downscale_factor": ("FLOAT", {"default": 2.0, "min": 0.1, "max": 9.0, "step": 0.001}),
+                              "start_percent": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 1.0, "step": 0.001}),
+                              "end_percent": ("FLOAT", {"default": 0.35, "min": 0.0, "max": 1.0, "step": 0.001}),
+                              "downscale_after_skip": ("BOOLEAN", {"default": True}),
+                              "downscale_method": (s.upscale_methods,),
+                              "upscale_method": (s.upscale_methods,),
+                              }}
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "patch"
+
+    CATEGORY = "_for_testing"
+
+    def patch(self, model, block_number, downscale_factor, start_percent, end_percent, downscale_after_skip, downscale_method, upscale_method):
+        sigma_start = model.model.model_sampling.percent_to_sigma(start_percent)
+        sigma_end = model.model.model_sampling.percent_to_sigma(end_percent)
+
+        def input_block_patch(h, transformer_options):
+            if transformer_options["block"][1] == block_number:
+                sigma = transformer_options["sigmas"][0].item()
+                if sigma <= sigma_start and sigma >= sigma_end:
+                    h = ldm_patched.modules.utils.common_upscale(h, round(h.shape[-1] * (1.0 / downscale_factor)), round(h.shape[-2] * (1.0 / downscale_factor)), downscale_method, "disabled")
+            return h
+
+        def output_block_patch(h, hsp, transformer_options):
+            if h.shape[2] != hsp.shape[2]:
+                h = ldm_patched.modules.utils.common_upscale(h, hsp.shape[-1], hsp.shape[-2], upscale_method, "disabled")
+            return h, hsp
+
+        m = model.clone()
+        if downscale_after_skip:
+            m.set_model_input_block_patch_after_skip(input_block_patch)
+        else:
+            m.set_model_input_block_patch(input_block_patch)
+        m.set_model_output_block_patch(output_block_patch)
+        return (m, )
+
+NODE_CLASS_MAPPINGS = {
+    "PatchModelAddDownscale": PatchModelAddDownscale,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    # Sampling
+    "PatchModelAddDownscale": "PatchModelAddDownscale (Kohya Deep Shrink)",
+}

ldm_patched/contrib/external_model_merging.py

@@ -0,0 +1,283 @@
+# https://github.com/comfyanonymous/ComfyUI/blob/master/nodes.py 
+
+import ldm_patched.modules.sd
+import ldm_patched.modules.utils
+import ldm_patched.modules.model_base
+import ldm_patched.modules.model_management
+
+import ldm_patched.utils.path_utils
+import json
+import os
+
+from ldm_patched.modules.args_parser import args
+
+class ModelMergeSimple:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model1": ("MODEL",),
+                              "model2": ("MODEL",),
+                              "ratio": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}),
+                              }}
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "merge"
+
+    CATEGORY = "advanced/model_merging"
+
+    def merge(self, model1, model2, ratio):
+        m = model1.clone()
+        kp = model2.get_key_patches("diffusion_model.")
+        for k in kp:
+            m.add_patches({k: kp[k]}, 1.0 - ratio, ratio)
+        return (m, )
+
+class ModelSubtract:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model1": ("MODEL",),
+                              "model2": ("MODEL",),
+                              "multiplier": ("FLOAT", {"default": 1.0, "min": -10.0, "max": 10.0, "step": 0.01}),
+                              }}
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "merge"
+
+    CATEGORY = "advanced/model_merging"
+
+    def merge(self, model1, model2, multiplier):
+        m = model1.clone()
+        kp = model2.get_key_patches("diffusion_model.")
+        for k in kp:
+            m.add_patches({k: kp[k]}, - multiplier, multiplier)
+        return (m, )
+
+class ModelAdd:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model1": ("MODEL",),
+                              "model2": ("MODEL",),
+                              }}
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "merge"
+
+    CATEGORY = "advanced/model_merging"
+
+    def merge(self, model1, model2):
+        m = model1.clone()
+        kp = model2.get_key_patches("diffusion_model.")
+        for k in kp:
+            m.add_patches({k: kp[k]}, 1.0, 1.0)
+        return (m, )
+
+
+class CLIPMergeSimple:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "clip1": ("CLIP",),
+                              "clip2": ("CLIP",),
+                              "ratio": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}),
+                              }}
+    RETURN_TYPES = ("CLIP",)
+    FUNCTION = "merge"
+
+    CATEGORY = "advanced/model_merging"
+
+    def merge(self, clip1, clip2, ratio):
+        m = clip1.clone()
+        kp = clip2.get_key_patches()
+        for k in kp:
+            if k.endswith(".position_ids") or k.endswith(".logit_scale"):
+                continue
+            m.add_patches({k: kp[k]}, 1.0 - ratio, ratio)
+        return (m, )
+
+class ModelMergeBlocks:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model1": ("MODEL",),
+                              "model2": ("MODEL",),
+                              "input": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}),
+                              "middle": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}),
+                              "out": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01})
+                              }}
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "merge"
+
+    CATEGORY = "advanced/model_merging"
+
+    def merge(self, model1, model2, **kwargs):
+        m = model1.clone()
+        kp = model2.get_key_patches("diffusion_model.")
+        default_ratio = next(iter(kwargs.values()))
+
+        for k in kp:
+            ratio = default_ratio
+            k_unet = k[len("diffusion_model."):]
+
+            last_arg_size = 0
+            for arg in kwargs:
+                if k_unet.startswith(arg) and last_arg_size < len(arg):
+                    ratio = kwargs[arg]
+                    last_arg_size = len(arg)
+
+            m.add_patches({k: kp[k]}, 1.0 - ratio, ratio)
+        return (m, )
+
+class CheckpointSave:
+    def __init__(self):
+        self.output_dir = ldm_patched.utils.path_utils.get_output_directory()
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model": ("MODEL",),
+                              "clip": ("CLIP",),
+                              "vae": ("VAE",),
+                              "filename_prefix": ("STRING", {"default": "checkpoints/ldm_patched"}),},
+                "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"},}
+    RETURN_TYPES = ()
+    FUNCTION = "save"
+    OUTPUT_NODE = True
+
+    CATEGORY = "advanced/model_merging"
+
+    def save(self, model, clip, vae, filename_prefix, prompt=None, extra_pnginfo=None):
+        full_output_folder, filename, counter, subfolder, filename_prefix = ldm_patched.utils.path_utils.get_save_image_path(filename_prefix, self.output_dir)
+        prompt_info = ""
+        if prompt is not None:
+            prompt_info = json.dumps(prompt)
+
+        metadata = {}
+
+        enable_modelspec = True
+        if isinstance(model.model, ldm_patched.modules.model_base.SDXL):
+            metadata["modelspec.architecture"] = "stable-diffusion-xl-v1-base"
+        elif isinstance(model.model, ldm_patched.modules.model_base.SDXLRefiner):
+            metadata["modelspec.architecture"] = "stable-diffusion-xl-v1-refiner"
+        else:
+            enable_modelspec = False
+
+        if enable_modelspec:
+            metadata["modelspec.sai_model_spec"] = "1.0.0"
+            metadata["modelspec.implementation"] = "sgm"
+            metadata["modelspec.title"] = "{} {}".format(filename, counter)
+
+        #TODO:
+        # "stable-diffusion-v1", "stable-diffusion-v1-inpainting", "stable-diffusion-v2-512",
+        # "stable-diffusion-v2-768-v", "stable-diffusion-v2-unclip-l", "stable-diffusion-v2-unclip-h",
+        # "v2-inpainting"
+
+        if model.model.model_type == ldm_patched.modules.model_base.ModelType.EPS:
+            metadata["modelspec.predict_key"] = "epsilon"
+        elif model.model.model_type == ldm_patched.modules.model_base.ModelType.V_PREDICTION:
+            metadata["modelspec.predict_key"] = "v"
+
+        if not args.disable_server_info:
+            metadata["prompt"] = prompt_info
+            if extra_pnginfo is not None:
+                for x in extra_pnginfo:
+                    metadata[x] = json.dumps(extra_pnginfo[x])
+
+        output_checkpoint = f"{filename}_{counter:05}_.safetensors"
+        output_checkpoint = os.path.join(full_output_folder, output_checkpoint)
+
+        ldm_patched.modules.sd.save_checkpoint(output_checkpoint, model, clip, vae, metadata=metadata)
+        return {}
+
+class CLIPSave:
+    def __init__(self):
+        self.output_dir = ldm_patched.utils.path_utils.get_output_directory()
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "clip": ("CLIP",),
+                              "filename_prefix": ("STRING", {"default": "clip/ldm_patched"}),},
+                "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"},}
+    RETURN_TYPES = ()
+    FUNCTION = "save"
+    OUTPUT_NODE = True
+
+    CATEGORY = "advanced/model_merging"
+
+    def save(self, clip, filename_prefix, prompt=None, extra_pnginfo=None):
+        prompt_info = ""
+        if prompt is not None:
+            prompt_info = json.dumps(prompt)
+
+        metadata = {}
+        if not args.disable_server_info:
+            metadata["prompt"] = prompt_info
+            if extra_pnginfo is not None:
+                for x in extra_pnginfo:
+                    metadata[x] = json.dumps(extra_pnginfo[x])
+
+        ldm_patched.modules.model_management.load_models_gpu([clip.load_model()])
+        clip_sd = clip.get_sd()
+
+        for prefix in ["clip_l.", "clip_g.", ""]:
+            k = list(filter(lambda a: a.startswith(prefix), clip_sd.keys()))
+            current_clip_sd = {}
+            for x in k:
+                current_clip_sd[x] = clip_sd.pop(x)
+            if len(current_clip_sd) == 0:
+                continue
+
+            p = prefix[:-1]
+            replace_prefix = {}
+            filename_prefix_ = filename_prefix
+            if len(p) > 0:
+                filename_prefix_ = "{}_{}".format(filename_prefix_, p)
+                replace_prefix[prefix] = ""
+            replace_prefix["transformer."] = ""
+
+            full_output_folder, filename, counter, subfolder, filename_prefix_ = ldm_patched.utils.path_utils.get_save_image_path(filename_prefix_, self.output_dir)
+
+            output_checkpoint = f"{filename}_{counter:05}_.safetensors"
+            output_checkpoint = os.path.join(full_output_folder, output_checkpoint)
+
+            current_clip_sd = ldm_patched.modules.utils.state_dict_prefix_replace(current_clip_sd, replace_prefix)
+
+            ldm_patched.modules.utils.save_torch_file(current_clip_sd, output_checkpoint, metadata=metadata)
+        return {}
+
+class VAESave:
+    def __init__(self):
+        self.output_dir = ldm_patched.utils.path_utils.get_output_directory()
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "vae": ("VAE",),
+                              "filename_prefix": ("STRING", {"default": "vae/ldm_patched_vae"}),},
+                "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"},}
+    RETURN_TYPES = ()
+    FUNCTION = "save"
+    OUTPUT_NODE = True
+
+    CATEGORY = "advanced/model_merging"
+
+    def save(self, vae, filename_prefix, prompt=None, extra_pnginfo=None):
+        full_output_folder, filename, counter, subfolder, filename_prefix = ldm_patched.utils.path_utils.get_save_image_path(filename_prefix, self.output_dir)
+        prompt_info = ""
+        if prompt is not None:
+            prompt_info = json.dumps(prompt)
+
+        metadata = {}
+        if not args.disable_server_info:
+            metadata["prompt"] = prompt_info
+            if extra_pnginfo is not None:
+                for x in extra_pnginfo:
+                    metadata[x] = json.dumps(extra_pnginfo[x])
+
+        output_checkpoint = f"{filename}_{counter:05}_.safetensors"
+        output_checkpoint = os.path.join(full_output_folder, output_checkpoint)
+
+        ldm_patched.modules.utils.save_torch_file(vae.get_sd(), output_checkpoint, metadata=metadata)
+        return {}
+
+NODE_CLASS_MAPPINGS = {
+    "ModelMergeSimple": ModelMergeSimple,
+    "ModelMergeBlocks": ModelMergeBlocks,
+    "ModelMergeSubtract": ModelSubtract,
+    "ModelMergeAdd": ModelAdd,
+    "CheckpointSave": CheckpointSave,
+    "CLIPMergeSimple": CLIPMergeSimple,
+    "CLIPSave": CLIPSave,
+    "VAESave": VAESave,
+}

ldm_patched/contrib/external_post_processing.py

@@ -0,0 +1,277 @@
+# https://github.com/comfyanonymous/ComfyUI/blob/master/nodes.py 
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from PIL import Image
+import math
+
+import ldm_patched.modules.utils
+
+
+class Blend:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "image1": ("IMAGE",),
+                "image2": ("IMAGE",),
+                "blend_factor": ("FLOAT", {
+                    "default": 0.5,
+                    "min": 0.0,
+                    "max": 1.0,
+                    "step": 0.01
+                }),
+                "blend_mode": (["normal", "multiply", "screen", "overlay", "soft_light", "difference"],),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "blend_images"
+
+    CATEGORY = "image/postprocessing"
+
+    def blend_images(self, image1: torch.Tensor, image2: torch.Tensor, blend_factor: float, blend_mode: str):
+        if image1.shape != image2.shape:
+            image2 = image2.permute(0, 3, 1, 2)
+            image2 = ldm_patched.modules.utils.common_upscale(image2, image1.shape[2], image1.shape[1], upscale_method='bicubic', crop='center')
+            image2 = image2.permute(0, 2, 3, 1)
+
+        blended_image = self.blend_mode(image1, image2, blend_mode)
+        blended_image = image1 * (1 - blend_factor) + blended_image * blend_factor
+        blended_image = torch.clamp(blended_image, 0, 1)
+        return (blended_image,)
+
+    def blend_mode(self, img1, img2, mode):
+        if mode == "normal":
+            return img2
+        elif mode == "multiply":
+            return img1 * img2
+        elif mode == "screen":
+            return 1 - (1 - img1) * (1 - img2)
+        elif mode == "overlay":
+            return torch.where(img1 <= 0.5, 2 * img1 * img2, 1 - 2 * (1 - img1) * (1 - img2))
+        elif mode == "soft_light":
+            return torch.where(img2 <= 0.5, img1 - (1 - 2 * img2) * img1 * (1 - img1), img1 + (2 * img2 - 1) * (self.g(img1) - img1))
+        elif mode == "difference":
+            return img1 - img2
+        else:
+            raise ValueError(f"Unsupported blend mode: {mode}")
+
+    def g(self, x):
+        return torch.where(x <= 0.25, ((16 * x - 12) * x + 4) * x, torch.sqrt(x))
+
+def gaussian_kernel(kernel_size: int, sigma: float, device=None):
+    x, y = torch.meshgrid(torch.linspace(-1, 1, kernel_size, device=device), torch.linspace(-1, 1, kernel_size, device=device), indexing="ij")
+    d = torch.sqrt(x * x + y * y)
+    g = torch.exp(-(d * d) / (2.0 * sigma * sigma))
+    return g / g.sum()
+
+class Blur:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "image": ("IMAGE",),
+                "blur_radius": ("INT", {
+                    "default": 1,
+                    "min": 1,
+                    "max": 31,
+                    "step": 1
+                }),
+                "sigma": ("FLOAT", {
+                    "default": 1.0,
+                    "min": 0.1,
+                    "max": 10.0,
+                    "step": 0.1
+                }),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "blur"
+
+    CATEGORY = "image/postprocessing"
+
+    def blur(self, image: torch.Tensor, blur_radius: int, sigma: float):
+        if blur_radius == 0:
+            return (image,)
+
+        batch_size, height, width, channels = image.shape
+
+        kernel_size = blur_radius * 2 + 1
+        kernel = gaussian_kernel(kernel_size, sigma, device=image.device).repeat(channels, 1, 1).unsqueeze(1)
+
+        image = image.permute(0, 3, 1, 2) # Torch wants (B, C, H, W) we use (B, H, W, C)
+        padded_image = F.pad(image, (blur_radius,blur_radius,blur_radius,blur_radius), 'reflect')
+        blurred = F.conv2d(padded_image, kernel, padding=kernel_size // 2, groups=channels)[:,:,blur_radius:-blur_radius, blur_radius:-blur_radius]
+        blurred = blurred.permute(0, 2, 3, 1)
+
+        return (blurred,)
+
+class Quantize:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "image": ("IMAGE",),
+                "colors": ("INT", {
+                    "default": 256,
+                    "min": 1,
+                    "max": 256,
+                    "step": 1
+                }),
+                "dither": (["none", "floyd-steinberg", "bayer-2", "bayer-4", "bayer-8", "bayer-16"],),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "quantize"
+
+    CATEGORY = "image/postprocessing"
+
+    def bayer(im, pal_im, order):
+        def normalized_bayer_matrix(n):
+            if n == 0:
+                return np.zeros((1,1), "float32")
+            else:
+                q = 4 ** n
+                m = q * normalized_bayer_matrix(n - 1)
+                return np.bmat(((m-1.5, m+0.5), (m+1.5, m-0.5))) / q
+
+        num_colors = len(pal_im.getpalette()) // 3
+        spread = 2 * 256 / num_colors
+        bayer_n = int(math.log2(order))
+        bayer_matrix = torch.from_numpy(spread * normalized_bayer_matrix(bayer_n) + 0.5)
+
+        result = torch.from_numpy(np.array(im).astype(np.float32))
+        tw = math.ceil(result.shape[0] / bayer_matrix.shape[0])
+        th = math.ceil(result.shape[1] / bayer_matrix.shape[1])
+        tiled_matrix = bayer_matrix.tile(tw, th).unsqueeze(-1)
+        result.add_(tiled_matrix[:result.shape[0],:result.shape[1]]).clamp_(0, 255)
+        result = result.to(dtype=torch.uint8)
+
+        im = Image.fromarray(result.cpu().numpy())
+        im = im.quantize(palette=pal_im, dither=Image.Dither.NONE)
+        return im
+
+    def quantize(self, image: torch.Tensor, colors: int, dither: str):
+        batch_size, height, width, _ = image.shape
+        result = torch.zeros_like(image)
+
+        for b in range(batch_size):
+            im = Image.fromarray((image[b] * 255).to(torch.uint8).numpy(), mode='RGB')
+
+            pal_im = im.quantize(colors=colors) # Required as described in https://github.com/python-pillow/Pillow/issues/5836
+
+            if dither == "none":
+                quantized_image = im.quantize(palette=pal_im, dither=Image.Dither.NONE)
+            elif dither == "floyd-steinberg":
+                quantized_image = im.quantize(palette=pal_im, dither=Image.Dither.FLOYDSTEINBERG)
+            elif dither.startswith("bayer"):
+                order = int(dither.split('-')[-1])
+                quantized_image = Quantize.bayer(im, pal_im, order)
+
+            quantized_array = torch.tensor(np.array(quantized_image.convert("RGB"))).float() / 255
+            result[b] = quantized_array
+
+        return (result,)
+
+class Sharpen:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "image": ("IMAGE",),
+                "sharpen_radius": ("INT", {
+                    "default": 1,
+                    "min": 1,
+                    "max": 31,
+                    "step": 1
+                }),
+                "sigma": ("FLOAT", {
+                    "default": 1.0,
+                    "min": 0.1,
+                    "max": 10.0,
+                    "step": 0.1
+                }),
+                "alpha": ("FLOAT", {
+                    "default": 1.0,
+                    "min": 0.0,
+                    "max": 5.0,
+                    "step": 0.1
+                }),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "sharpen"
+
+    CATEGORY = "image/postprocessing"
+
+    def sharpen(self, image: torch.Tensor, sharpen_radius: int, sigma:float, alpha: float):
+        if sharpen_radius == 0:
+            return (image,)
+
+        batch_size, height, width, channels = image.shape
+
+        kernel_size = sharpen_radius * 2 + 1
+        kernel = gaussian_kernel(kernel_size, sigma, device=image.device) * -(alpha*10)
+        center = kernel_size // 2
+        kernel[center, center] = kernel[center, center] - kernel.sum() + 1.0
+        kernel = kernel.repeat(channels, 1, 1).unsqueeze(1)
+
+        tensor_image = image.permute(0, 3, 1, 2) # Torch wants (B, C, H, W) we use (B, H, W, C)
+        tensor_image = F.pad(tensor_image, (sharpen_radius,sharpen_radius,sharpen_radius,sharpen_radius), 'reflect')
+        sharpened = F.conv2d(tensor_image, kernel, padding=center, groups=channels)[:,:,sharpen_radius:-sharpen_radius, sharpen_radius:-sharpen_radius]
+        sharpened = sharpened.permute(0, 2, 3, 1)
+
+        result = torch.clamp(sharpened, 0, 1)
+
+        return (result,)
+
+class ImageScaleToTotalPixels:
+    upscale_methods = ["nearest-exact", "bilinear", "area", "bicubic", "lanczos"]
+    crop_methods = ["disabled", "center"]
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "image": ("IMAGE",), "upscale_method": (s.upscale_methods,),
+                              "megapixels": ("FLOAT", {"default": 1.0, "min": 0.01, "max": 16.0, "step": 0.01}),
+                            }}
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "upscale"
+
+    CATEGORY = "image/upscaling"
+
+    def upscale(self, image, upscale_method, megapixels):
+        samples = image.movedim(-1,1)
+        total = int(megapixels * 1024 * 1024)
+
+        scale_by = math.sqrt(total / (samples.shape[3] * samples.shape[2]))
+        width = round(samples.shape[3] * scale_by)
+        height = round(samples.shape[2] * scale_by)
+
+        s = ldm_patched.modules.utils.common_upscale(samples, width, height, upscale_method, "disabled")
+        s = s.movedim(1,-1)
+        return (s,)
+
+NODE_CLASS_MAPPINGS = {
+    "ImageBlend": Blend,
+    "ImageBlur": Blur,
+    "ImageQuantize": Quantize,
+    "ImageSharpen": Sharpen,
+    "ImageScaleToTotalPixels": ImageScaleToTotalPixels,
+}

ldm_patched/contrib/external_rebatch.py

@@ -0,0 +1,110 @@
+# https://github.com/comfyanonymous/ComfyUI/blob/master/nodes.py 
+
+import torch
+
+class LatentRebatch:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "latents": ("LATENT",),
+                              "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
+                              }}
+    RETURN_TYPES = ("LATENT",)
+    INPUT_IS_LIST = True
+    OUTPUT_IS_LIST = (True, )
+
+    FUNCTION = "rebatch"
+
+    CATEGORY = "latent/batch"
+
+    @staticmethod
+    def get_batch(latents, list_ind, offset):
+        '''prepare a batch out of the list of latents'''
+        samples = latents[list_ind]['samples']
+        shape = samples.shape
+        mask = latents[list_ind]['noise_mask'] if 'noise_mask' in latents[list_ind] else torch.ones((shape[0], 1, shape[2]*8, shape[3]*8), device='cpu')
+        if mask.shape[-1] != shape[-1] * 8 or mask.shape[-2] != shape[-2]:
+            torch.nn.functional.interpolate(mask.reshape((-1, 1, mask.shape[-2], mask.shape[-1])), size=(shape[-2]*8, shape[-1]*8), mode="bilinear")
+        if mask.shape[0] < samples.shape[0]:
+            mask = mask.repeat((shape[0] - 1) // mask.shape[0] + 1, 1, 1, 1)[:shape[0]]
+        if 'batch_index' in latents[list_ind]:
+            batch_inds = latents[list_ind]['batch_index']
+        else:
+            batch_inds = [x+offset for x in range(shape[0])]
+        return samples, mask, batch_inds
+
+    @staticmethod
+    def get_slices(indexable, num, batch_size):
+        '''divides an indexable object into num slices of length batch_size, and a remainder'''
+        slices = []
+        for i in range(num):
+            slices.append(indexable[i*batch_size:(i+1)*batch_size])
+        if num * batch_size < len(indexable):
+            return slices, indexable[num * batch_size:]
+        else:
+            return slices, None
+    
+    @staticmethod
+    def slice_batch(batch, num, batch_size):
+        result = [LatentRebatch.get_slices(x, num, batch_size) for x in batch]
+        return list(zip(*result))
+
+    @staticmethod
+    def cat_batch(batch1, batch2):
+        if batch1[0] is None:
+            return batch2
+        result = [torch.cat((b1, b2)) if torch.is_tensor(b1) else b1 + b2 for b1, b2 in zip(batch1, batch2)]
+        return result
+
+    def rebatch(self, latents, batch_size):
+        batch_size = batch_size[0]
+
+        output_list = []
+        current_batch = (None, None, None)
+        processed = 0
+
+        for i in range(len(latents)):
+            # fetch new entry of list
+            #samples, masks, indices = self.get_batch(latents, i)
+            next_batch = self.get_batch(latents, i, processed)
+            processed += len(next_batch[2])
+            # set to current if current is None
+            if current_batch[0] is None:
+                current_batch = next_batch
+            # add previous to list if dimensions do not match
+            elif next_batch[0].shape[-1] != current_batch[0].shape[-1] or next_batch[0].shape[-2] != current_batch[0].shape[-2]:
+                sliced, _ = self.slice_batch(current_batch, 1, batch_size)
+                output_list.append({'samples': sliced[0][0], 'noise_mask': sliced[1][0], 'batch_index': sliced[2][0]})
+                current_batch = next_batch
+            # cat if everything checks out
+            else:
+                current_batch = self.cat_batch(current_batch, next_batch)
+
+            # add to list if dimensions gone above target batch size
+            if current_batch[0].shape[0] > batch_size:
+                num = current_batch[0].shape[0] // batch_size
+                sliced, remainder = self.slice_batch(current_batch, num, batch_size)
+                
+                for i in range(num):
+                    output_list.append({'samples': sliced[0][i], 'noise_mask': sliced[1][i], 'batch_index': sliced[2][i]})
+
+                current_batch = remainder
+
+        #add remainder
+        if current_batch[0] is not None:
+            sliced, _ = self.slice_batch(current_batch, 1, batch_size)
+            output_list.append({'samples': sliced[0][0], 'noise_mask': sliced[1][0], 'batch_index': sliced[2][0]})
+
+        #get rid of empty masks
+        for s in output_list:
+            if s['noise_mask'].mean() == 1.0:
+                del s['noise_mask']
+
+        return (output_list,)
+
+NODE_CLASS_MAPPINGS = {
+    "RebatchLatents": LatentRebatch,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "RebatchLatents": "Rebatch Latents",
+}

ldm_patched/contrib/external_tomesd.py

@@ -0,0 +1,179 @@
+# https://github.com/comfyanonymous/ComfyUI/blob/master/nodes.py 
+
+#Taken from: https://github.com/dbolya/tomesd
+
+import torch
+from typing import Tuple, Callable
+import math
+
+def do_nothing(x: torch.Tensor, mode:str=None):
+    return x
+
+
+def mps_gather_workaround(input, dim, index):
+    if input.shape[-1] == 1:
+        return torch.gather(
+            input.unsqueeze(-1),
+            dim - 1 if dim < 0 else dim,
+            index.unsqueeze(-1)
+        ).squeeze(-1)
+    else:
+        return torch.gather(input, dim, index)
+
+
+def bipartite_soft_matching_random2d(metric: torch.Tensor,
+                                     w: int, h: int, sx: int, sy: int, r: int,
+                                     no_rand: bool = False) -> Tuple[Callable, Callable]:
+    """
+    Partitions the tokens into src and dst and merges r tokens from src to dst.
+    Dst tokens are partitioned by choosing one randomy in each (sx, sy) region.
+    Args:
+     - metric [B, N, C]: metric to use for similarity
+     - w: image width in tokens
+     - h: image height in tokens
+     - sx: stride in the x dimension for dst, must divide w
+     - sy: stride in the y dimension for dst, must divide h
+     - r: number of tokens to remove (by merging)
+     - no_rand: if true, disable randomness (use top left corner only)
+    """
+    B, N, _ = metric.shape
+
+    if r <= 0 or w == 1 or h == 1:
+        return do_nothing, do_nothing
+
+    gather = mps_gather_workaround if metric.device.type == "mps" else torch.gather
+    
+    with torch.no_grad():
+        
+        hsy, wsx = h // sy, w // sx
+
+        # For each sy by sx kernel, randomly assign one token to be dst and the rest src
+        if no_rand:
+            rand_idx = torch.zeros(hsy, wsx, 1, device=metric.device, dtype=torch.int64)
+        else:
+            rand_idx = torch.randint(sy*sx, size=(hsy, wsx, 1), device=metric.device)
+        
+        # The image might not divide sx and sy, so we need to work on a view of the top left if the idx buffer instead
+        idx_buffer_view = torch.zeros(hsy, wsx, sy*sx, device=metric.device, dtype=torch.int64)
+        idx_buffer_view.scatter_(dim=2, index=rand_idx, src=-torch.ones_like(rand_idx, dtype=rand_idx.dtype))
+        idx_buffer_view = idx_buffer_view.view(hsy, wsx, sy, sx).transpose(1, 2).reshape(hsy * sy, wsx * sx)
+
+        # Image is not divisible by sx or sy so we need to move it into a new buffer
+        if (hsy * sy) < h or (wsx * sx) < w:
+            idx_buffer = torch.zeros(h, w, device=metric.device, dtype=torch.int64)
+            idx_buffer[:(hsy * sy), :(wsx * sx)] = idx_buffer_view
+        else:
+            idx_buffer = idx_buffer_view
+
+        # We set dst tokens to be -1 and src to be 0, so an argsort gives us dst|src indices
+        rand_idx = idx_buffer.reshape(1, -1, 1).argsort(dim=1)
+
+        # We're finished with these
+        del idx_buffer, idx_buffer_view
+
+        # rand_idx is currently dst|src, so split them
+        num_dst = hsy * wsx
+        a_idx = rand_idx[:, num_dst:, :] # src
+        b_idx = rand_idx[:, :num_dst, :] # dst
+
+        def split(x):
+            C = x.shape[-1]
+            src = gather(x, dim=1, index=a_idx.expand(B, N - num_dst, C))
+            dst = gather(x, dim=1, index=b_idx.expand(B, num_dst, C))
+            return src, dst
+
+        # Cosine similarity between A and B
+        metric = metric / metric.norm(dim=-1, keepdim=True)
+        a, b = split(metric)
+        scores = a @ b.transpose(-1, -2)
+
+        # Can't reduce more than the # tokens in src
+        r = min(a.shape[1], r)
+
+        # Find the most similar greedily
+        node_max, node_idx = scores.max(dim=-1)
+        edge_idx = node_max.argsort(dim=-1, descending=True)[..., None]
+
+        unm_idx = edge_idx[..., r:, :]  # Unmerged Tokens
+        src_idx = edge_idx[..., :r, :]  # Merged Tokens
+        dst_idx = gather(node_idx[..., None], dim=-2, index=src_idx)
+
+    def merge(x: torch.Tensor, mode="mean") -> torch.Tensor:
+        src, dst = split(x)
+        n, t1, c = src.shape
+        
+        unm = gather(src, dim=-2, index=unm_idx.expand(n, t1 - r, c))
+        src = gather(src, dim=-2, index=src_idx.expand(n, r, c))
+        dst = dst.scatter_reduce(-2, dst_idx.expand(n, r, c), src, reduce=mode)
+
+        return torch.cat([unm, dst], dim=1)
+
+    def unmerge(x: torch.Tensor) -> torch.Tensor:
+        unm_len = unm_idx.shape[1]
+        unm, dst = x[..., :unm_len, :], x[..., unm_len:, :]
+        _, _, c = unm.shape
+
+        src = gather(dst, dim=-2, index=dst_idx.expand(B, r, c))
+
+        # Combine back to the original shape
+        out = torch.zeros(B, N, c, device=x.device, dtype=x.dtype)
+        out.scatter_(dim=-2, index=b_idx.expand(B, num_dst, c), src=dst)
+        out.scatter_(dim=-2, index=gather(a_idx.expand(B, a_idx.shape[1], 1), dim=1, index=unm_idx).expand(B, unm_len, c), src=unm)
+        out.scatter_(dim=-2, index=gather(a_idx.expand(B, a_idx.shape[1], 1), dim=1, index=src_idx).expand(B, r, c), src=src)
+
+        return out
+
+    return merge, unmerge
+
+
+def get_functions(x, ratio, original_shape):
+    b, c, original_h, original_w = original_shape
+    original_tokens = original_h * original_w
+    downsample = int(math.ceil(math.sqrt(original_tokens // x.shape[1])))
+    stride_x = 2
+    stride_y = 2
+    max_downsample = 1
+
+    if downsample <= max_downsample:
+        w = int(math.ceil(original_w / downsample))
+        h = int(math.ceil(original_h / downsample))
+        r = int(x.shape[1] * ratio)
+        no_rand = False
+        m, u = bipartite_soft_matching_random2d(x, w, h, stride_x, stride_y, r, no_rand)
+        return m, u
+
+    nothing = lambda y: y
+    return nothing, nothing
+
+
+
+class TomePatchModel:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model": ("MODEL",),
+                              "ratio": ("FLOAT", {"default": 0.3, "min": 0.0, "max": 1.0, "step": 0.01}),
+                              }}
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "patch"
+
+    CATEGORY = "_for_testing"
+
+    def patch(self, model, ratio):
+        self.u = None
+        def tomesd_m(q, k, v, extra_options):
+            #NOTE: In the reference code get_functions takes x (input of the transformer block) as the argument instead of q
+            #however from my basic testing it seems that using q instead gives better results
+            m, self.u = get_functions(q, ratio, extra_options["original_shape"])
+            return m(q), k, v
+        def tomesd_u(n, extra_options):
+            return self.u(n)
+
+        m = model.clone()
+        m.set_model_attn1_patch(tomesd_m)
+        m.set_model_attn1_output_patch(tomesd_u)
+        return (m, )
+
+
+NODE_CLASS_MAPPINGS = {
+    "TomePatchModel": TomePatchModel,
+}

ldm_patched/contrib/external_upscale_model.py

@@ -0,0 +1,68 @@
+# https://github.com/comfyanonymous/ComfyUI/blob/master/nodes.py 
+
+import os
+from ldm_patched.pfn import model_loading
+from ldm_patched.modules import model_management
+import torch
+import ldm_patched.modules.utils
+import ldm_patched.utils.path_utils
+
+class UpscaleModelLoader:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model_name": (ldm_patched.utils.path_utils.get_filename_list("upscale_models"), ),
+                             }}
+    RETURN_TYPES = ("UPSCALE_MODEL",)
+    FUNCTION = "load_model"
+
+    CATEGORY = "loaders"
+
+    def load_model(self, model_name):
+        model_path = ldm_patched.utils.path_utils.get_full_path("upscale_models", model_name)
+        sd = ldm_patched.modules.utils.load_torch_file(model_path, safe_load=True)
+        if "module.layers.0.residual_group.blocks.0.norm1.weight" in sd:
+            sd = ldm_patched.modules.utils.state_dict_prefix_replace(sd, {"module.":""})
+        out = model_loading.load_state_dict(sd).eval()
+        return (out, )
+
+
+class ImageUpscaleWithModel:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "upscale_model": ("UPSCALE_MODEL",),
+                              "image": ("IMAGE",),
+                              }}
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "upscale"
+
+    CATEGORY = "image/upscaling"
+
+    def upscale(self, upscale_model, image):
+        device = model_management.get_torch_device()
+        upscale_model.to(device)
+        in_img = image.movedim(-1,-3).to(device)
+        free_memory = model_management.get_free_memory(device)
+
+        tile = 512
+        overlap = 32
+
+        oom = True
+        while oom:
+            try:
+                steps = in_img.shape[0] * ldm_patched.modules.utils.get_tiled_scale_steps(in_img.shape[3], in_img.shape[2], tile_x=tile, tile_y=tile, overlap=overlap)
+                pbar = ldm_patched.modules.utils.ProgressBar(steps)
+                s = ldm_patched.modules.utils.tiled_scale(in_img, lambda a: upscale_model(a), tile_x=tile, tile_y=tile, overlap=overlap, upscale_amount=upscale_model.scale, pbar=pbar)
+                oom = False
+            except model_management.OOM_EXCEPTION as e:
+                tile //= 2
+                if tile < 128:
+                    raise e
+
+        upscale_model.cpu()
+        s = torch.clamp(s.movedim(-3,-1), min=0, max=1.0)
+        return (s,)
+
+NODE_CLASS_MAPPINGS = {
+    "UpscaleModelLoader": UpscaleModelLoader,
+    "ImageUpscaleWithModel": ImageUpscaleWithModel
+}

ldm_patched/contrib/external_video_model.py

@@ -0,0 +1,91 @@
+# https://github.com/comfyanonymous/ComfyUI/blob/master/nodes.py 
+
+import ldm_patched.contrib.external
+import torch
+import ldm_patched.modules.utils
+import ldm_patched.modules.sd
+import ldm_patched.utils.path_utils
+
+
+class ImageOnlyCheckpointLoader:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "ckpt_name": (ldm_patched.utils.path_utils.get_filename_list("checkpoints"), ),
+                             }}
+    RETURN_TYPES = ("MODEL", "CLIP_VISION", "VAE")
+    FUNCTION = "load_checkpoint"
+
+    CATEGORY = "loaders/video_models"
+
+    def load_checkpoint(self, ckpt_name, output_vae=True, output_clip=True):
+        ckpt_path = ldm_patched.utils.path_utils.get_full_path("checkpoints", ckpt_name)
+        out = ldm_patched.modules.sd.load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=False, output_clipvision=True, embedding_directory=ldm_patched.utils.path_utils.get_folder_paths("embeddings"))
+        return (out[0], out[3], out[2])
+
+
+class SVD_img2vid_Conditioning:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "clip_vision": ("CLIP_VISION",),
+                              "init_image": ("IMAGE",),
+                              "vae": ("VAE",),
+                              "width": ("INT", {"default": 1024, "min": 16, "max": ldm_patched.contrib.external.MAX_RESOLUTION, "step": 8}),
+                              "height": ("INT", {"default": 576, "min": 16, "max": ldm_patched.contrib.external.MAX_RESOLUTION, "step": 8}),
+                              "video_frames": ("INT", {"default": 14, "min": 1, "max": 4096}),
+                              "motion_bucket_id": ("INT", {"default": 127, "min": 1, "max": 1023}),
+                              "fps": ("INT", {"default": 6, "min": 1, "max": 1024}),
+                              "augmentation_level": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 10.0, "step": 0.01})
+                             }}
+    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT")
+    RETURN_NAMES = ("positive", "negative", "latent")
+
+    FUNCTION = "encode"
+
+    CATEGORY = "conditioning/video_models"
+
+    def encode(self, clip_vision, init_image, vae, width, height, video_frames, motion_bucket_id, fps, augmentation_level):
+        output = clip_vision.encode_image(init_image)
+        pooled = output.image_embeds.unsqueeze(0)
+        pixels = ldm_patched.modules.utils.common_upscale(init_image.movedim(-1,1), width, height, "bilinear", "center").movedim(1,-1)
+        encode_pixels = pixels[:,:,:,:3]
+        if augmentation_level > 0:
+            encode_pixels += torch.randn_like(pixels) * augmentation_level
+        t = vae.encode(encode_pixels)
+        positive = [[pooled, {"motion_bucket_id": motion_bucket_id, "fps": fps, "augmentation_level": augmentation_level, "concat_latent_image": t}]]
+        negative = [[torch.zeros_like(pooled), {"motion_bucket_id": motion_bucket_id, "fps": fps, "augmentation_level": augmentation_level, "concat_latent_image": torch.zeros_like(t)}]]
+        latent = torch.zeros([video_frames, 4, height // 8, width // 8])
+        return (positive, negative, {"samples":latent})
+
+class VideoLinearCFGGuidance:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model": ("MODEL",),
+                              "min_cfg": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.5, "round": 0.01}),
+                              }}
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "patch"
+
+    CATEGORY = "sampling/video_models"
+
+    def patch(self, model, min_cfg):
+        def linear_cfg(args):
+            cond = args["cond"]
+            uncond = args["uncond"]
+            cond_scale = args["cond_scale"]
+
+            scale = torch.linspace(min_cfg, cond_scale, cond.shape[0], device=cond.device).reshape((cond.shape[0], 1, 1, 1))
+            return uncond + scale * (cond - uncond)
+
+        m = model.clone()
+        m.set_model_sampler_cfg_function(linear_cfg)
+        return (m, )
+
+NODE_CLASS_MAPPINGS = {
+    "ImageOnlyCheckpointLoader": ImageOnlyCheckpointLoader,
+    "SVD_img2vid_Conditioning": SVD_img2vid_Conditioning,
+    "VideoLinearCFGGuidance": VideoLinearCFGGuidance,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "ImageOnlyCheckpointLoader": "Image Only Checkpoint Loader (img2vid model)",
+}

ldm_patched/controlnet/cldm.py

@@ -0,0 +1,312 @@
+#taken from: https://github.com/lllyasviel/ControlNet
+#and modified
+
+import torch
+import torch as th
+import torch.nn as nn
+
+from ldm_patched.ldm.modules.diffusionmodules.util import (
+    zero_module,
+    timestep_embedding,
+)
+
+from ldm_patched.ldm.modules.attention import SpatialTransformer
+from ldm_patched.ldm.modules.diffusionmodules.openaimodel import UNetModel, TimestepEmbedSequential, ResBlock, Downsample
+from ldm_patched.ldm.util import exists
+import ldm_patched.modules.ops
+
+class ControlledUnetModel(UNetModel):
+    #implemented in the ldm unet
+    pass
+
+class ControlNet(nn.Module):
+    def __init__(
+        self,
+        image_size,
+        in_channels,
+        model_channels,
+        hint_channels,
+        num_res_blocks,
+        dropout=0,
+        channel_mult=(1, 2, 4, 8),
+        conv_resample=True,
+        dims=2,
+        num_classes=None,
+        use_checkpoint=False,
+        dtype=torch.float32,
+        num_heads=-1,
+        num_head_channels=-1,
+        num_heads_upsample=-1,
+        use_scale_shift_norm=False,
+        resblock_updown=False,
+        use_new_attention_order=False,
+        use_spatial_transformer=False,    # custom transformer support
+        transformer_depth=1,              # custom transformer support
+        context_dim=None,                 # custom transformer support
+        n_embed=None,                     # custom support for prediction of discrete ids into codebook of first stage vq model
+        legacy=True,
+        disable_self_attentions=None,
+        num_attention_blocks=None,
+        disable_middle_self_attn=False,
+        use_linear_in_transformer=False,
+        adm_in_channels=None,
+        transformer_depth_middle=None,
+        transformer_depth_output=None,
+        device=None,
+        operations=ldm_patched.modules.ops.disable_weight_init,
+        **kwargs,
+    ):
+        super().__init__()
+        assert use_spatial_transformer == True, "use_spatial_transformer has to be true"
+        if use_spatial_transformer:
+            assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
+
+        if context_dim is not None:
+            assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
+            # from omegaconf.listconfig import ListConfig
+            # if type(context_dim) == ListConfig:
+            #     context_dim = list(context_dim)
+
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+
+        if num_heads == -1:
+            assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
+
+        if num_head_channels == -1:
+            assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
+
+        self.dims = dims
+        self.image_size = image_size
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+
+        if isinstance(num_res_blocks, int):
+            self.num_res_blocks = len(channel_mult) * [num_res_blocks]
+        else:
+            if len(num_res_blocks) != len(channel_mult):
+                raise ValueError("provide num_res_blocks either as an int (globally constant) or "
+                                 "as a list/tuple (per-level) with the same length as channel_mult")
+            self.num_res_blocks = num_res_blocks
+
+        if disable_self_attentions is not None:
+            # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not
+            assert len(disable_self_attentions) == len(channel_mult)
+        if num_attention_blocks is not None:
+            assert len(num_attention_blocks) == len(self.num_res_blocks)
+            assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(len(num_attention_blocks))))
+
+        transformer_depth = transformer_depth[:]
+
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.num_classes = num_classes
+        self.use_checkpoint = use_checkpoint
+        self.dtype = dtype
+        self.num_heads = num_heads
+        self.num_head_channels = num_head_channels
+        self.num_heads_upsample = num_heads_upsample
+        self.predict_codebook_ids = n_embed is not None
+
+        time_embed_dim = model_channels * 4
+        self.time_embed = nn.Sequential(
+            operations.Linear(model_channels, time_embed_dim, dtype=self.dtype, device=device),
+            nn.SiLU(),
+            operations.Linear(time_embed_dim, time_embed_dim, dtype=self.dtype, device=device),
+        )
+
+        if self.num_classes is not None:
+            if isinstance(self.num_classes, int):
+                self.label_emb = nn.Embedding(num_classes, time_embed_dim)
+            elif self.num_classes == "continuous":
+                print("setting up linear c_adm embedding layer")
+                self.label_emb = nn.Linear(1, time_embed_dim)
+            elif self.num_classes == "sequential":
+                assert adm_in_channels is not None
+                self.label_emb = nn.Sequential(
+                    nn.Sequential(
+                        operations.Linear(adm_in_channels, time_embed_dim, dtype=self.dtype, device=device),
+                        nn.SiLU(),
+                        operations.Linear(time_embed_dim, time_embed_dim, dtype=self.dtype, device=device),
+                    )
+                )
+            else:
+                raise ValueError()
+
+        self.input_blocks = nn.ModuleList(
+            [
+                TimestepEmbedSequential(
+                    operations.conv_nd(dims, in_channels, model_channels, 3, padding=1, dtype=self.dtype, device=device)
+                )
+            ]
+        )
+        self.zero_convs = nn.ModuleList([self.make_zero_conv(model_channels, operations=operations, dtype=self.dtype, device=device)])
+
+        self.input_hint_block = TimestepEmbedSequential(
+                    operations.conv_nd(dims, hint_channels, 16, 3, padding=1, dtype=self.dtype, device=device),
+                    nn.SiLU(),
+                    operations.conv_nd(dims, 16, 16, 3, padding=1, dtype=self.dtype, device=device),
+                    nn.SiLU(),
+                    operations.conv_nd(dims, 16, 32, 3, padding=1, stride=2, dtype=self.dtype, device=device),
+                    nn.SiLU(),
+                    operations.conv_nd(dims, 32, 32, 3, padding=1, dtype=self.dtype, device=device),
+                    nn.SiLU(),
+                    operations.conv_nd(dims, 32, 96, 3, padding=1, stride=2, dtype=self.dtype, device=device),
+                    nn.SiLU(),
+                    operations.conv_nd(dims, 96, 96, 3, padding=1, dtype=self.dtype, device=device),
+                    nn.SiLU(),
+                    operations.conv_nd(dims, 96, 256, 3, padding=1, stride=2, dtype=self.dtype, device=device),
+                    nn.SiLU(),
+                    operations.conv_nd(dims, 256, model_channels, 3, padding=1, dtype=self.dtype, device=device)
+        )
+
+        self._feature_size = model_channels
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for nr in range(self.num_res_blocks[level]):
+                layers = [
+                    ResBlock(
+                        ch,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=mult * model_channels,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                        dtype=self.dtype,
+                        device=device,
+                        operations=operations,
+                    )
+                ]
+                ch = mult * model_channels
+                num_transformers = transformer_depth.pop(0)
+                if num_transformers > 0:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        #num_heads = 1
+                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    if exists(disable_self_attentions):
+                        disabled_sa = disable_self_attentions[level]
+                    else:
+                        disabled_sa = False
+
+                    if not exists(num_attention_blocks) or nr < num_attention_blocks[level]:
+                        layers.append(
+                            SpatialTransformer(
+                                ch, num_heads, dim_head, depth=num_transformers, context_dim=context_dim,
+                                disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer,
+                                use_checkpoint=use_checkpoint, dtype=self.dtype, device=device, operations=operations
+                            )
+                        )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                self.zero_convs.append(self.make_zero_conv(ch, operations=operations, dtype=self.dtype, device=device))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True,
+                            dtype=self.dtype,
+                            device=device,
+                            operations=operations
+                        )
+                        if resblock_updown
+                        else Downsample(
+                            ch, conv_resample, dims=dims, out_channels=out_ch, dtype=self.dtype, device=device, operations=operations
+                        )
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                self.zero_convs.append(self.make_zero_conv(ch, operations=operations, dtype=self.dtype, device=device))
+                ds *= 2
+                self._feature_size += ch
+
+        if num_head_channels == -1:
+            dim_head = ch // num_heads
+        else:
+            num_heads = ch // num_head_channels
+            dim_head = num_head_channels
+        if legacy:
+            #num_heads = 1
+            dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+        mid_block = [
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+                dtype=self.dtype,
+                device=device,
+                operations=operations
+            )]
+        if transformer_depth_middle >= 0:
+            mid_block += [SpatialTransformer(  # always uses a self-attn
+                            ch, num_heads, dim_head, depth=transformer_depth_middle, context_dim=context_dim,
+                            disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer,
+                            use_checkpoint=use_checkpoint, dtype=self.dtype, device=device, operations=operations
+                        ),
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+                dtype=self.dtype,
+                device=device,
+                operations=operations
+            )]
+        self.middle_block = TimestepEmbedSequential(*mid_block)
+        self.middle_block_out = self.make_zero_conv(ch, operations=operations, dtype=self.dtype, device=device)
+        self._feature_size += ch
+
+    def make_zero_conv(self, channels, operations=None, dtype=None, device=None):
+        return TimestepEmbedSequential(operations.conv_nd(self.dims, channels, channels, 1, padding=0, dtype=dtype, device=device))
+
+    def forward(self, x, hint, timesteps, context, y=None, **kwargs):
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).to(x.dtype)
+        emb = self.time_embed(t_emb)
+
+        guided_hint = self.input_hint_block(hint, emb, context)
+
+        outs = []
+
+        hs = []
+        if self.num_classes is not None:
+            assert y.shape[0] == x.shape[0]
+            emb = emb + self.label_emb(y)
+
+        h = x
+        for module, zero_conv in zip(self.input_blocks, self.zero_convs):
+            if guided_hint is not None:
+                h = module(h, emb, context)
+                h += guided_hint
+                guided_hint = None
+            else:
+                h = module(h, emb, context)
+            outs.append(zero_conv(h, emb, context))
+
+        h = self.middle_block(h, emb, context)
+        outs.append(self.middle_block_out(h, emb, context))
+
+        return outs
+

ldm_patched/k_diffusion/sampling.py

@@ -0,0 +1,810 @@
+import math
+
+from scipy import integrate
+import torch
+from torch import nn
+import torchsde
+from tqdm.auto import trange, tqdm
+
+from . import utils
+
+
+def append_zero(x):
+    return torch.cat([x, x.new_zeros([1])])
+
+
+def get_sigmas_karras(n, sigma_min, sigma_max, rho=7., device='cpu'):
+    """Constructs the noise schedule of Karras et al. (2022)."""
+    ramp = torch.linspace(0, 1, n, device=device)
+    min_inv_rho = sigma_min ** (1 / rho)
+    max_inv_rho = sigma_max ** (1 / rho)
+    sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
+    return append_zero(sigmas).to(device)
+
+
+def get_sigmas_exponential(n, sigma_min, sigma_max, device='cpu'):
+    """Constructs an exponential noise schedule."""
+    sigmas = torch.linspace(math.log(sigma_max), math.log(sigma_min), n, device=device).exp()
+    return append_zero(sigmas)
+
+
+def get_sigmas_polyexponential(n, sigma_min, sigma_max, rho=1., device='cpu'):
+    """Constructs an polynomial in log sigma noise schedule."""
+    ramp = torch.linspace(1, 0, n, device=device) ** rho
+    sigmas = torch.exp(ramp * (math.log(sigma_max) - math.log(sigma_min)) + math.log(sigma_min))
+    return append_zero(sigmas)
+
+
+def get_sigmas_vp(n, beta_d=19.9, beta_min=0.1, eps_s=1e-3, device='cpu'):
+    """Constructs a continuous VP noise schedule."""
+    t = torch.linspace(1, eps_s, n, device=device)
+    sigmas = torch.sqrt(torch.exp(beta_d * t ** 2 / 2 + beta_min * t) - 1)
+    return append_zero(sigmas)
+
+
+def to_d(x, sigma, denoised):
+    """Converts a denoiser output to a Karras ODE derivative."""
+    return (x - denoised) / utils.append_dims(sigma, x.ndim)
+
+
+def get_ancestral_step(sigma_from, sigma_to, eta=1.):
+    """Calculates the noise level (sigma_down) to step down to and the amount
+    of noise to add (sigma_up) when doing an ancestral sampling step."""
+    if not eta:
+        return sigma_to, 0.
+    sigma_up = min(sigma_to, eta * (sigma_to ** 2 * (sigma_from ** 2 - sigma_to ** 2) / sigma_from ** 2) ** 0.5)
+    sigma_down = (sigma_to ** 2 - sigma_up ** 2) ** 0.5
+    return sigma_down, sigma_up
+
+
+def default_noise_sampler(x):
+    return lambda sigma, sigma_next: torch.randn_like(x)
+
+
+class BatchedBrownianTree:
+    """A wrapper around torchsde.BrownianTree that enables batches of entropy."""
+
+    def __init__(self, x, t0, t1, seed=None, **kwargs):
+        self.cpu_tree = True
+        if "cpu" in kwargs:
+            self.cpu_tree = kwargs.pop("cpu")
+        t0, t1, self.sign = self.sort(t0, t1)
+        w0 = kwargs.get('w0', torch.zeros_like(x))
+        if seed is None:
+            seed = torch.randint(0, 2 ** 63 - 1, []).item()
+        self.batched = True
+        try:
+            assert len(seed) == x.shape[0]
+            w0 = w0[0]
+        except TypeError:
+            seed = [seed]
+            self.batched = False
+        if self.cpu_tree:
+            self.trees = [torchsde.BrownianTree(t0.cpu(), w0.cpu(), t1.cpu(), entropy=s, **kwargs) for s in seed]
+        else:
+            self.trees = [torchsde.BrownianTree(t0, w0, t1, entropy=s, **kwargs) for s in seed]
+
+    @staticmethod
+    def sort(a, b):
+        return (a, b, 1) if a < b else (b, a, -1)
+
+    def __call__(self, t0, t1):
+        t0, t1, sign = self.sort(t0, t1)
+        if self.cpu_tree:
+            w = torch.stack([tree(t0.cpu().float(), t1.cpu().float()).to(t0.dtype).to(t0.device) for tree in self.trees]) * (self.sign * sign)
+        else:
+            w = torch.stack([tree(t0, t1) for tree in self.trees]) * (self.sign * sign)
+
+        return w if self.batched else w[0]
+
+
+class BrownianTreeNoiseSampler:
+    """A noise sampler backed by a torchsde.BrownianTree.
+
+    Args:
+        x (Tensor): The tensor whose shape, device and dtype to use to generate
+            random samples.
+        sigma_min (float): The low end of the valid interval.
+        sigma_max (float): The high end of the valid interval.
+        seed (int or List[int]): The random seed. If a list of seeds is
+            supplied instead of a single integer, then the noise sampler will
+            use one BrownianTree per batch item, each with its own seed.
+        transform (callable): A function that maps sigma to the sampler's
+            internal timestep.
+    """
+
+    def __init__(self, x, sigma_min, sigma_max, seed=None, transform=lambda x: x, cpu=False):
+        self.transform = transform
+        t0, t1 = self.transform(torch.as_tensor(sigma_min)), self.transform(torch.as_tensor(sigma_max))
+        self.tree = BatchedBrownianTree(x, t0, t1, seed, cpu=cpu)
+
+    def __call__(self, sigma, sigma_next):
+        t0, t1 = self.transform(torch.as_tensor(sigma)), self.transform(torch.as_tensor(sigma_next))
+        return self.tree(t0, t1) / (t1 - t0).abs().sqrt()
+
+
+@torch.no_grad()
+def sample_euler(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.):
+    """Implements Algorithm 2 (Euler steps) from Karras et al. (2022)."""
+    extra_args = {} if extra_args is None else extra_args
+    s_in = x.new_ones([x.shape[0]])
+    for i in trange(len(sigmas) - 1, disable=disable):
+        gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.
+        sigma_hat = sigmas[i] * (gamma + 1)
+        if gamma > 0:
+            eps = torch.randn_like(x) * s_noise
+            x = x + eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5
+        denoised = model(x, sigma_hat * s_in, **extra_args)
+        d = to_d(x, sigma_hat, denoised)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
+        dt = sigmas[i + 1] - sigma_hat
+        # Euler method
+        x = x + d * dt
+    return x
+
+
+@torch.no_grad()
+def sample_euler_ancestral(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
+    """Ancestral sampling with Euler method steps."""
+    extra_args = {} if extra_args is None else extra_args
+    noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
+    s_in = x.new_ones([x.shape[0]])
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        d = to_d(x, sigmas[i], denoised)
+        # Euler method
+        dt = sigma_down - sigmas[i]
+        x = x + d * dt
+        if sigmas[i + 1] > 0:
+            x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
+    return x
+
+
+@torch.no_grad()
+def sample_heun(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.):
+    """Implements Algorithm 2 (Heun steps) from Karras et al. (2022)."""
+    extra_args = {} if extra_args is None else extra_args
+    s_in = x.new_ones([x.shape[0]])
+    for i in trange(len(sigmas) - 1, disable=disable):
+        gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.
+        sigma_hat = sigmas[i] * (gamma + 1)
+        if gamma > 0:
+            eps = torch.randn_like(x) * s_noise
+            x = x + eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5
+        denoised = model(x, sigma_hat * s_in, **extra_args)
+        d = to_d(x, sigma_hat, denoised)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
+        dt = sigmas[i + 1] - sigma_hat
+        if sigmas[i + 1] == 0:
+            # Euler method
+            x = x + d * dt
+        else:
+            # Heun's method
+            x_2 = x + d * dt
+            denoised_2 = model(x_2, sigmas[i + 1] * s_in, **extra_args)
+            d_2 = to_d(x_2, sigmas[i + 1], denoised_2)
+            d_prime = (d + d_2) / 2
+            x = x + d_prime * dt
+    return x
+
+
+@torch.no_grad()
+def sample_dpm_2(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.):
+    """A sampler inspired by DPM-Solver-2 and Algorithm 2 from Karras et al. (2022)."""
+    extra_args = {} if extra_args is None else extra_args
+    s_in = x.new_ones([x.shape[0]])
+    for i in trange(len(sigmas) - 1, disable=disable):
+        gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.
+        sigma_hat = sigmas[i] * (gamma + 1)
+        if gamma > 0:
+            eps = torch.randn_like(x) * s_noise
+            x = x + eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5
+        denoised = model(x, sigma_hat * s_in, **extra_args)
+        d = to_d(x, sigma_hat, denoised)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
+        if sigmas[i + 1] == 0:
+            # Euler method
+            dt = sigmas[i + 1] - sigma_hat
+            x = x + d * dt
+        else:
+            # DPM-Solver-2
+            sigma_mid = sigma_hat.log().lerp(sigmas[i + 1].log(), 0.5).exp()
+            dt_1 = sigma_mid - sigma_hat
+            dt_2 = sigmas[i + 1] - sigma_hat
+            x_2 = x + d * dt_1
+            denoised_2 = model(x_2, sigma_mid * s_in, **extra_args)
+            d_2 = to_d(x_2, sigma_mid, denoised_2)
+            x = x + d_2 * dt_2
+    return x
+
+
+@torch.no_grad()
+def sample_dpm_2_ancestral(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
+    """Ancestral sampling with DPM-Solver second-order steps."""
+    extra_args = {} if extra_args is None else extra_args
+    noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
+    s_in = x.new_ones([x.shape[0]])
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        d = to_d(x, sigmas[i], denoised)
+        if sigma_down == 0:
+            # Euler method
+            dt = sigma_down - sigmas[i]
+            x = x + d * dt
+        else:
+            # DPM-Solver-2
+            sigma_mid = sigmas[i].log().lerp(sigma_down.log(), 0.5).exp()
+            dt_1 = sigma_mid - sigmas[i]
+            dt_2 = sigma_down - sigmas[i]
+            x_2 = x + d * dt_1
+            denoised_2 = model(x_2, sigma_mid * s_in, **extra_args)
+            d_2 = to_d(x_2, sigma_mid, denoised_2)
+            x = x + d_2 * dt_2
+            x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
+    return x
+
+
+def linear_multistep_coeff(order, t, i, j):
+    if order - 1 > i:
+        raise ValueError(f'Order {order} too high for step {i}')
+    def fn(tau):
+        prod = 1.
+        for k in range(order):
+            if j == k:
+                continue
+            prod *= (tau - t[i - k]) / (t[i - j] - t[i - k])
+        return prod
+    return integrate.quad(fn, t[i], t[i + 1], epsrel=1e-4)[0]
+
+
+@torch.no_grad()
+def sample_lms(model, x, sigmas, extra_args=None, callback=None, disable=None, order=4):
+    extra_args = {} if extra_args is None else extra_args
+    s_in = x.new_ones([x.shape[0]])
+    sigmas_cpu = sigmas.detach().cpu().numpy()
+    ds = []
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        d = to_d(x, sigmas[i], denoised)
+        ds.append(d)
+        if len(ds) > order:
+            ds.pop(0)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        cur_order = min(i + 1, order)
+        coeffs = [linear_multistep_coeff(cur_order, sigmas_cpu, i, j) for j in range(cur_order)]
+        x = x + sum(coeff * d for coeff, d in zip(coeffs, reversed(ds)))
+    return x
+
+
+class PIDStepSizeController:
+    """A PID controller for ODE adaptive step size control."""
+    def __init__(self, h, pcoeff, icoeff, dcoeff, order=1, accept_safety=0.81, eps=1e-8):
+        self.h = h
+        self.b1 = (pcoeff + icoeff + dcoeff) / order
+        self.b2 = -(pcoeff + 2 * dcoeff) / order
+        self.b3 = dcoeff / order
+        self.accept_safety = accept_safety
+        self.eps = eps
+        self.errs = []
+
+    def limiter(self, x):
+        return 1 + math.atan(x - 1)
+
+    def propose_step(self, error):
+        inv_error = 1 / (float(error) + self.eps)
+        if not self.errs:
+            self.errs = [inv_error, inv_error, inv_error]
+        self.errs[0] = inv_error
+        factor = self.errs[0] ** self.b1 * self.errs[1] ** self.b2 * self.errs[2] ** self.b3
+        factor = self.limiter(factor)
+        accept = factor >= self.accept_safety
+        if accept:
+            self.errs[2] = self.errs[1]
+            self.errs[1] = self.errs[0]
+        self.h *= factor
+        return accept
+
+
+class DPMSolver(nn.Module):
+    """DPM-Solver. See https://arxiv.org/abs/2206.00927."""
+
+    def __init__(self, model, extra_args=None, eps_callback=None, info_callback=None):
+        super().__init__()
+        self.model = model
+        self.extra_args = {} if extra_args is None else extra_args
+        self.eps_callback = eps_callback
+        self.info_callback = info_callback
+
+    def t(self, sigma):
+        return -sigma.log()
+
+    def sigma(self, t):
+        return t.neg().exp()
+
+    def eps(self, eps_cache, key, x, t, *args, **kwargs):
+        if key in eps_cache:
+            return eps_cache[key], eps_cache
+        sigma = self.sigma(t) * x.new_ones([x.shape[0]])
+        eps = (x - self.model(x, sigma, *args, **self.extra_args, **kwargs)) / self.sigma(t)
+        if self.eps_callback is not None:
+            self.eps_callback()
+        return eps, {key: eps, **eps_cache}
+
+    def dpm_solver_1_step(self, x, t, t_next, eps_cache=None):
+        eps_cache = {} if eps_cache is None else eps_cache
+        h = t_next - t
+        eps, eps_cache = self.eps(eps_cache, 'eps', x, t)
+        x_1 = x - self.sigma(t_next) * h.expm1() * eps
+        return x_1, eps_cache
+
+    def dpm_solver_2_step(self, x, t, t_next, r1=1 / 2, eps_cache=None):
+        eps_cache = {} if eps_cache is None else eps_cache
+        h = t_next - t
+        eps, eps_cache = self.eps(eps_cache, 'eps', x, t)
+        s1 = t + r1 * h
+        u1 = x - self.sigma(s1) * (r1 * h).expm1() * eps
+        eps_r1, eps_cache = self.eps(eps_cache, 'eps_r1', u1, s1)
+        x_2 = x - self.sigma(t_next) * h.expm1() * eps - self.sigma(t_next) / (2 * r1) * h.expm1() * (eps_r1 - eps)
+        return x_2, eps_cache
+
+    def dpm_solver_3_step(self, x, t, t_next, r1=1 / 3, r2=2 / 3, eps_cache=None):
+        eps_cache = {} if eps_cache is None else eps_cache
+        h = t_next - t
+        eps, eps_cache = self.eps(eps_cache, 'eps', x, t)
+        s1 = t + r1 * h
+        s2 = t + r2 * h
+        u1 = x - self.sigma(s1) * (r1 * h).expm1() * eps
+        eps_r1, eps_cache = self.eps(eps_cache, 'eps_r1', u1, s1)
+        u2 = x - self.sigma(s2) * (r2 * h).expm1() * eps - self.sigma(s2) * (r2 / r1) * ((r2 * h).expm1() / (r2 * h) - 1) * (eps_r1 - eps)
+        eps_r2, eps_cache = self.eps(eps_cache, 'eps_r2', u2, s2)
+        x_3 = x - self.sigma(t_next) * h.expm1() * eps - self.sigma(t_next) / r2 * (h.expm1() / h - 1) * (eps_r2 - eps)
+        return x_3, eps_cache
+
+    def dpm_solver_fast(self, x, t_start, t_end, nfe, eta=0., s_noise=1., noise_sampler=None):
+        noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
+        if not t_end > t_start and eta:
+            raise ValueError('eta must be 0 for reverse sampling')
+
+        m = math.floor(nfe / 3) + 1
+        ts = torch.linspace(t_start, t_end, m + 1, device=x.device)
+
+        if nfe % 3 == 0:
+            orders = [3] * (m - 2) + [2, 1]
+        else:
+            orders = [3] * (m - 1) + [nfe % 3]
+
+        for i in range(len(orders)):
+            eps_cache = {}
+            t, t_next = ts[i], ts[i + 1]
+            if eta:
+                sd, su = get_ancestral_step(self.sigma(t), self.sigma(t_next), eta)
+                t_next_ = torch.minimum(t_end, self.t(sd))
+                su = (self.sigma(t_next) ** 2 - self.sigma(t_next_) ** 2) ** 0.5
+            else:
+                t_next_, su = t_next, 0.
+
+            eps, eps_cache = self.eps(eps_cache, 'eps', x, t)
+            denoised = x - self.sigma(t) * eps
+            if self.info_callback is not None:
+                self.info_callback({'x': x, 'i': i, 't': ts[i], 't_up': t, 'denoised': denoised})
+
+            if orders[i] == 1:
+                x, eps_cache = self.dpm_solver_1_step(x, t, t_next_, eps_cache=eps_cache)
+            elif orders[i] == 2:
+                x, eps_cache = self.dpm_solver_2_step(x, t, t_next_, eps_cache=eps_cache)
+            else:
+                x, eps_cache = self.dpm_solver_3_step(x, t, t_next_, eps_cache=eps_cache)
+
+            x = x + su * s_noise * noise_sampler(self.sigma(t), self.sigma(t_next))
+
+        return x
+
+    def dpm_solver_adaptive(self, x, t_start, t_end, order=3, rtol=0.05, atol=0.0078, h_init=0.05, pcoeff=0., icoeff=1., dcoeff=0., accept_safety=0.81, eta=0., s_noise=1., noise_sampler=None):
+        noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
+        if order not in {2, 3}:
+            raise ValueError('order should be 2 or 3')
+        forward = t_end > t_start
+        if not forward and eta:
+            raise ValueError('eta must be 0 for reverse sampling')
+        h_init = abs(h_init) * (1 if forward else -1)
+        atol = torch.tensor(atol)
+        rtol = torch.tensor(rtol)
+        s = t_start
+        x_prev = x
+        accept = True
+        pid = PIDStepSizeController(h_init, pcoeff, icoeff, dcoeff, 1.5 if eta else order, accept_safety)
+        info = {'steps': 0, 'nfe': 0, 'n_accept': 0, 'n_reject': 0}
+
+        while s < t_end - 1e-5 if forward else s > t_end + 1e-5:
+            eps_cache = {}
+            t = torch.minimum(t_end, s + pid.h) if forward else torch.maximum(t_end, s + pid.h)
+            if eta:
+                sd, su = get_ancestral_step(self.sigma(s), self.sigma(t), eta)
+                t_ = torch.minimum(t_end, self.t(sd))
+                su = (self.sigma(t) ** 2 - self.sigma(t_) ** 2) ** 0.5
+            else:
+                t_, su = t, 0.
+
+            eps, eps_cache = self.eps(eps_cache, 'eps', x, s)
+            denoised = x - self.sigma(s) * eps
+
+            if order == 2:
+                x_low, eps_cache = self.dpm_solver_1_step(x, s, t_, eps_cache=eps_cache)
+                x_high, eps_cache = self.dpm_solver_2_step(x, s, t_, eps_cache=eps_cache)
+            else:
+                x_low, eps_cache = self.dpm_solver_2_step(x, s, t_, r1=1 / 3, eps_cache=eps_cache)
+                x_high, eps_cache = self.dpm_solver_3_step(x, s, t_, eps_cache=eps_cache)
+            delta = torch.maximum(atol, rtol * torch.maximum(x_low.abs(), x_prev.abs()))
+            error = torch.linalg.norm((x_low - x_high) / delta) / x.numel() ** 0.5
+            accept = pid.propose_step(error)
+            if accept:
+                x_prev = x_low
+                x = x_high + su * s_noise * noise_sampler(self.sigma(s), self.sigma(t))
+                s = t
+                info['n_accept'] += 1
+            else:
+                info['n_reject'] += 1
+            info['nfe'] += order
+            info['steps'] += 1
+
+            if self.info_callback is not None:
+                self.info_callback({'x': x, 'i': info['steps'] - 1, 't': s, 't_up': s, 'denoised': denoised, 'error': error, 'h': pid.h, **info})
+
+        return x, info
+
+
+@torch.no_grad()
+def sample_dpm_fast(model, x, sigma_min, sigma_max, n, extra_args=None, callback=None, disable=None, eta=0., s_noise=1., noise_sampler=None):
+    """DPM-Solver-Fast (fixed step size). See https://arxiv.org/abs/2206.00927."""
+    if sigma_min <= 0 or sigma_max <= 0:
+        raise ValueError('sigma_min and sigma_max must not be 0')
+    with tqdm(total=n, disable=disable) as pbar:
+        dpm_solver = DPMSolver(model, extra_args, eps_callback=pbar.update)
+        if callback is not None:
+            dpm_solver.info_callback = lambda info: callback({'sigma': dpm_solver.sigma(info['t']), 'sigma_hat': dpm_solver.sigma(info['t_up']), **info})
+        return dpm_solver.dpm_solver_fast(x, dpm_solver.t(torch.tensor(sigma_max)), dpm_solver.t(torch.tensor(sigma_min)), n, eta, s_noise, noise_sampler)
+
+
+@torch.no_grad()
+def sample_dpm_adaptive(model, x, sigma_min, sigma_max, extra_args=None, callback=None, disable=None, order=3, rtol=0.05, atol=0.0078, h_init=0.05, pcoeff=0., icoeff=1., dcoeff=0., accept_safety=0.81, eta=0., s_noise=1., noise_sampler=None, return_info=False):
+    """DPM-Solver-12 and 23 (adaptive step size). See https://arxiv.org/abs/2206.00927."""
+    if sigma_min <= 0 or sigma_max <= 0:
+        raise ValueError('sigma_min and sigma_max must not be 0')
+    with tqdm(disable=disable) as pbar:
+        dpm_solver = DPMSolver(model, extra_args, eps_callback=pbar.update)
+        if callback is not None:
+            dpm_solver.info_callback = lambda info: callback({'sigma': dpm_solver.sigma(info['t']), 'sigma_hat': dpm_solver.sigma(info['t_up']), **info})
+        x, info = dpm_solver.dpm_solver_adaptive(x, dpm_solver.t(torch.tensor(sigma_max)), dpm_solver.t(torch.tensor(sigma_min)), order, rtol, atol, h_init, pcoeff, icoeff, dcoeff, accept_safety, eta, s_noise, noise_sampler)
+    if return_info:
+        return x, info
+    return x
+
+
+@torch.no_grad()
+def sample_dpmpp_2s_ancestral(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
+    """Ancestral sampling with DPM-Solver++(2S) second-order steps."""
+    extra_args = {} if extra_args is None else extra_args
+    noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
+    s_in = x.new_ones([x.shape[0]])
+    sigma_fn = lambda t: t.neg().exp()
+    t_fn = lambda sigma: sigma.log().neg()
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        if sigma_down == 0:
+            # Euler method
+            d = to_d(x, sigmas[i], denoised)
+            dt = sigma_down - sigmas[i]
+            x = x + d * dt
+        else:
+            # DPM-Solver++(2S)
+            t, t_next = t_fn(sigmas[i]), t_fn(sigma_down)
+            r = 1 / 2
+            h = t_next - t
+            s = t + r * h
+            x_2 = (sigma_fn(s) / sigma_fn(t)) * x - (-h * r).expm1() * denoised
+            denoised_2 = model(x_2, sigma_fn(s) * s_in, **extra_args)
+            x = (sigma_fn(t_next) / sigma_fn(t)) * x - (-h).expm1() * denoised_2
+        # Noise addition
+        if sigmas[i + 1] > 0:
+            x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
+    return x
+
+
+@torch.no_grad()
+def sample_dpmpp_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=1 / 2):
+    """DPM-Solver++ (stochastic)."""
+    sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
+    seed = extra_args.get("seed", None)
+    noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) if noise_sampler is None else noise_sampler
+    extra_args = {} if extra_args is None else extra_args
+    s_in = x.new_ones([x.shape[0]])
+    sigma_fn = lambda t: t.neg().exp()
+    t_fn = lambda sigma: sigma.log().neg()
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        if sigmas[i + 1] == 0:
+            # Euler method
+            d = to_d(x, sigmas[i], denoised)
+            dt = sigmas[i + 1] - sigmas[i]
+            x = x + d * dt
+        else:
+            # DPM-Solver++
+            t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1])
+            h = t_next - t
+            s = t + h * r
+            fac = 1 / (2 * r)
+
+            # Step 1
+            sd, su = get_ancestral_step(sigma_fn(t), sigma_fn(s), eta)
+            s_ = t_fn(sd)
+            x_2 = (sigma_fn(s_) / sigma_fn(t)) * x - (t - s_).expm1() * denoised
+            x_2 = x_2 + noise_sampler(sigma_fn(t), sigma_fn(s)) * s_noise * su
+            denoised_2 = model(x_2, sigma_fn(s) * s_in, **extra_args)
+
+            # Step 2
+            sd, su = get_ancestral_step(sigma_fn(t), sigma_fn(t_next), eta)
+            t_next_ = t_fn(sd)
+            denoised_d = (1 - fac) * denoised + fac * denoised_2
+            x = (sigma_fn(t_next_) / sigma_fn(t)) * x - (t - t_next_).expm1() * denoised_d
+            x = x + noise_sampler(sigma_fn(t), sigma_fn(t_next)) * s_noise * su
+    return x
+
+
+@torch.no_grad()
+def sample_dpmpp_2m(model, x, sigmas, extra_args=None, callback=None, disable=None):
+    """DPM-Solver++(2M)."""
+    extra_args = {} if extra_args is None else extra_args
+    s_in = x.new_ones([x.shape[0]])
+    sigma_fn = lambda t: t.neg().exp()
+    t_fn = lambda sigma: sigma.log().neg()
+    old_denoised = None
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1])
+        h = t_next - t
+        if old_denoised is None or sigmas[i + 1] == 0:
+            x = (sigma_fn(t_next) / sigma_fn(t)) * x - (-h).expm1() * denoised
+        else:
+            h_last = t - t_fn(sigmas[i - 1])
+            r = h_last / h
+            denoised_d = (1 + 1 / (2 * r)) * denoised - (1 / (2 * r)) * old_denoised
+            x = (sigma_fn(t_next) / sigma_fn(t)) * x - (-h).expm1() * denoised_d
+        old_denoised = denoised
+    return x
+
+@torch.no_grad()
+def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='midpoint'):
+    """DPM-Solver++(2M) SDE."""
+
+    if solver_type not in {'heun', 'midpoint'}:
+        raise ValueError('solver_type must be \'heun\' or \'midpoint\'')
+
+    seed = extra_args.get("seed", None)
+    sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
+    noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) if noise_sampler is None else noise_sampler
+    extra_args = {} if extra_args is None else extra_args
+    s_in = x.new_ones([x.shape[0]])
+
+    old_denoised = None
+    h_last = None
+    h = None
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        if sigmas[i + 1] == 0:
+            # Denoising step
+            x = denoised
+        else:
+            # DPM-Solver++(2M) SDE
+            t, s = -sigmas[i].log(), -sigmas[i + 1].log()
+            h = s - t
+            eta_h = eta * h
+
+            x = sigmas[i + 1] / sigmas[i] * (-eta_h).exp() * x + (-h - eta_h).expm1().neg() * denoised
+
+            if old_denoised is not None:
+                r = h_last / h
+                if solver_type == 'heun':
+                    x = x + ((-h - eta_h).expm1().neg() / (-h - eta_h) + 1) * (1 / r) * (denoised - old_denoised)
+                elif solver_type == 'midpoint':
+                    x = x + 0.5 * (-h - eta_h).expm1().neg() * (1 / r) * (denoised - old_denoised)
+
+            if eta:
+                x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * eta_h).expm1().neg().sqrt() * s_noise
+
+        old_denoised = denoised
+        h_last = h
+    return x
+
+@torch.no_grad()
+def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
+    """DPM-Solver++(3M) SDE."""
+
+    seed = extra_args.get("seed", None)
+    sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
+    noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) if noise_sampler is None else noise_sampler
+    extra_args = {} if extra_args is None else extra_args
+    s_in = x.new_ones([x.shape[0]])
+
+    denoised_1, denoised_2 = None, None
+    h, h_1, h_2 = None, None, None
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        if sigmas[i + 1] == 0:
+            # Denoising step
+            x = denoised
+        else:
+            t, s = -sigmas[i].log(), -sigmas[i + 1].log()
+            h = s - t
+            h_eta = h * (eta + 1)
+
+            x = torch.exp(-h_eta) * x + (-h_eta).expm1().neg() * denoised
+
+            if h_2 is not None:
+                r0 = h_1 / h
+                r1 = h_2 / h
+                d1_0 = (denoised - denoised_1) / r0
+                d1_1 = (denoised_1 - denoised_2) / r1
+                d1 = d1_0 + (d1_0 - d1_1) * r0 / (r0 + r1)
+                d2 = (d1_0 - d1_1) / (r0 + r1)
+                phi_2 = h_eta.neg().expm1() / h_eta + 1
+                phi_3 = phi_2 / h_eta - 0.5
+                x = x + phi_2 * d1 - phi_3 * d2
+            elif h_1 is not None:
+                r = h_1 / h
+                d = (denoised - denoised_1) / r
+                phi_2 = h_eta.neg().expm1() / h_eta + 1
+                x = x + phi_2 * d
+
+            if eta:
+                x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * h * eta).expm1().neg().sqrt() * s_noise
+
+        denoised_1, denoised_2 = denoised, denoised_1
+        h_1, h_2 = h, h_1
+    return x
+
+@torch.no_grad()
+def sample_dpmpp_3m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
+    sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
+    noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
+    return sample_dpmpp_3m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler)
+
+@torch.no_grad()
+def sample_dpmpp_2m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='midpoint'):
+    sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
+    noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
+    return sample_dpmpp_2m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, solver_type=solver_type)
+
+@torch.no_grad()
+def sample_dpmpp_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=1 / 2):
+    sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
+    noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
+    return sample_dpmpp_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, r=r)
+
+
+def DDPMSampler_step(x, sigma, sigma_prev, noise, noise_sampler):
+    alpha_cumprod = 1 / ((sigma * sigma) + 1)
+    alpha_cumprod_prev = 1 / ((sigma_prev * sigma_prev) + 1)
+    alpha = (alpha_cumprod / alpha_cumprod_prev)
+
+    mu = (1.0 / alpha).sqrt() * (x - (1 - alpha) * noise / (1 - alpha_cumprod).sqrt())
+    if sigma_prev > 0:
+        mu += ((1 - alpha) * (1. - alpha_cumprod_prev) / (1. - alpha_cumprod)).sqrt() * noise_sampler(sigma, sigma_prev)
+    return mu
+
+def generic_step_sampler(model, x, sigmas, extra_args=None, callback=None, disable=None, noise_sampler=None, step_function=None):
+    extra_args = {} if extra_args is None else extra_args
+    noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
+    s_in = x.new_ones([x.shape[0]])
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        x = step_function(x / torch.sqrt(1.0 + sigmas[i] ** 2.0), sigmas[i], sigmas[i + 1], (x - denoised) / sigmas[i], noise_sampler)
+        if sigmas[i + 1] != 0:
+            x *= torch.sqrt(1.0 + sigmas[i + 1] ** 2.0)
+    return x
+
+
+@torch.no_grad()
+def sample_ddpm(model, x, sigmas, extra_args=None, callback=None, disable=None, noise_sampler=None):
+    return generic_step_sampler(model, x, sigmas, extra_args, callback, disable, noise_sampler, DDPMSampler_step)
+
+@torch.no_grad()
+def sample_lcm(model, x, sigmas, extra_args=None, callback=None, disable=None, noise_sampler=None):
+    extra_args = {} if extra_args is None else extra_args
+    noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
+    s_in = x.new_ones([x.shape[0]])
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+
+        x = denoised
+        if sigmas[i + 1] > 0:
+            x += sigmas[i + 1] * noise_sampler(sigmas[i], sigmas[i + 1])
+    return x
+
+
+
+@torch.no_grad()
+def sample_heunpp2(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.):
+    # From MIT licensed: https://github.com/Carzit/sd-webui-samplers-scheduler/
+    extra_args = {} if extra_args is None else extra_args
+    s_in = x.new_ones([x.shape[0]])
+    s_end = sigmas[-1]
+    for i in trange(len(sigmas) - 1, disable=disable):
+        gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.
+        eps = torch.randn_like(x) * s_noise
+        sigma_hat = sigmas[i] * (gamma + 1)
+        if gamma > 0:
+            x = x + eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5
+        denoised = model(x, sigma_hat * s_in, **extra_args)
+        d = to_d(x, sigma_hat, denoised)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
+        dt = sigmas[i + 1] - sigma_hat
+        if sigmas[i + 1] == s_end:
+            # Euler method
+            x = x + d * dt
+        elif sigmas[i + 2] == s_end:
+
+            # Heun's method
+            x_2 = x + d * dt
+            denoised_2 = model(x_2, sigmas[i + 1] * s_in, **extra_args)
+            d_2 = to_d(x_2, sigmas[i + 1], denoised_2)
+
+            w = 2 * sigmas[0]
+            w2 = sigmas[i+1]/w
+            w1 = 1 - w2
+
+            d_prime = d * w1 + d_2 * w2
+
+
+            x = x + d_prime * dt
+
+        else:
+            # Heun++
+            x_2 = x + d * dt
+            denoised_2 = model(x_2, sigmas[i + 1] * s_in, **extra_args)
+            d_2 = to_d(x_2, sigmas[i + 1], denoised_2)
+            dt_2 = sigmas[i + 2] - sigmas[i + 1]
+
+            x_3 = x_2 + d_2 * dt_2
+            denoised_3 = model(x_3, sigmas[i + 2] * s_in, **extra_args)
+            d_3 = to_d(x_3, sigmas[i + 2], denoised_3)
+
+            w = 3 * sigmas[0]
+            w2 = sigmas[i + 1] / w
+            w3 = sigmas[i + 2] / w
+            w1 = 1 - w2 - w3
+
+            d_prime = w1 * d + w2 * d_2 + w3 * d_3
+            x = x + d_prime * dt
+    return x

ldm_patched/k_diffusion/utils.py

@@ -0,0 +1,313 @@
+from contextlib import contextmanager
+import hashlib
+import math
+from pathlib import Path
+import shutil
+import urllib
+import warnings
+
+from PIL import Image
+import torch
+from torch import nn, optim
+from torch.utils import data
+
+
+def hf_datasets_augs_helper(examples, transform, image_key, mode='RGB'):
+    """Apply passed in transforms for HuggingFace Datasets."""
+    images = [transform(image.convert(mode)) for image in examples[image_key]]
+    return {image_key: images}
+
+
+def append_dims(x, target_dims):
+    """Appends dimensions to the end of a tensor until it has target_dims dimensions."""
+    dims_to_append = target_dims - x.ndim
+    if dims_to_append < 0:
+        raise ValueError(f'input has {x.ndim} dims but target_dims is {target_dims}, which is less')
+    expanded = x[(...,) + (None,) * dims_to_append]
+    # MPS will get inf values if it tries to index into the new axes, but detaching fixes this.
+    # https://github.com/pytorch/pytorch/issues/84364
+    return expanded.detach().clone() if expanded.device.type == 'mps' else expanded
+
+
+def n_params(module):
+    """Returns the number of trainable parameters in a module."""
+    return sum(p.numel() for p in module.parameters())
+
+
+def download_file(path, url, digest=None):
+    """Downloads a file if it does not exist, optionally checking its SHA-256 hash."""
+    path = Path(path)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    if not path.exists():
+        with urllib.request.urlopen(url) as response, open(path, 'wb') as f:
+            shutil.copyfileobj(response, f)
+    if digest is not None:
+        file_digest = hashlib.sha256(open(path, 'rb').read()).hexdigest()
+        if digest != file_digest:
+            raise OSError(f'hash of {path} (url: {url}) failed to validate')
+    return path
+
+
+@contextmanager
+def train_mode(model, mode=True):
+    """A context manager that places a model into training mode and restores
+    the previous mode on exit."""
+    modes = [module.training for module in model.modules()]
+    try:
+        yield model.train(mode)
+    finally:
+        for i, module in enumerate(model.modules()):
+            module.training = modes[i]
+
+
+def eval_mode(model):
+    """A context manager that places a model into evaluation mode and restores
+    the previous mode on exit."""
+    return train_mode(model, False)
+
+
+@torch.no_grad()
+def ema_update(model, averaged_model, decay):
+    """Incorporates updated model parameters into an exponential moving averaged
+    version of a model. It should be called after each optimizer step."""
+    model_params = dict(model.named_parameters())
+    averaged_params = dict(averaged_model.named_parameters())
+    assert model_params.keys() == averaged_params.keys()
+
+    for name, param in model_params.items():
+        averaged_params[name].mul_(decay).add_(param, alpha=1 - decay)
+
+    model_buffers = dict(model.named_buffers())
+    averaged_buffers = dict(averaged_model.named_buffers())
+    assert model_buffers.keys() == averaged_buffers.keys()
+
+    for name, buf in model_buffers.items():
+        averaged_buffers[name].copy_(buf)
+
+
+class EMAWarmup:
+    """Implements an EMA warmup using an inverse decay schedule.
+    If inv_gamma=1 and power=1, implements a simple average. inv_gamma=1, power=2/3 are
+    good values for models you plan to train for a million or more steps (reaches decay
+    factor 0.999 at 31.6K steps, 0.9999 at 1M steps), inv_gamma=1, power=3/4 for models
+    you plan to train for less (reaches decay factor 0.999 at 10K steps, 0.9999 at
+    215.4k steps).
+    Args:
+        inv_gamma (float): Inverse multiplicative factor of EMA warmup. Default: 1.
+        power (float): Exponential factor of EMA warmup. Default: 1.
+        min_value (float): The minimum EMA decay rate. Default: 0.
+        max_value (float): The maximum EMA decay rate. Default: 1.
+        start_at (int): The epoch to start averaging at. Default: 0.
+        last_epoch (int): The index of last epoch. Default: 0.
+    """
+
+    def __init__(self, inv_gamma=1., power=1., min_value=0., max_value=1., start_at=0,
+                 last_epoch=0):
+        self.inv_gamma = inv_gamma
+        self.power = power
+        self.min_value = min_value
+        self.max_value = max_value
+        self.start_at = start_at
+        self.last_epoch = last_epoch
+
+    def state_dict(self):
+        """Returns the state of the class as a :class:`dict`."""
+        return dict(self.__dict__.items())
+
+    def load_state_dict(self, state_dict):
+        """Loads the class's state.
+        Args:
+            state_dict (dict): scaler state. Should be an object returned
+                from a call to :meth:`state_dict`.
+        """
+        self.__dict__.update(state_dict)
+
+    def get_value(self):
+        """Gets the current EMA decay rate."""
+        epoch = max(0, self.last_epoch - self.start_at)
+        value = 1 - (1 + epoch / self.inv_gamma) ** -self.power
+        return 0. if epoch < 0 else min(self.max_value, max(self.min_value, value))
+
+    def step(self):
+        """Updates the step count."""
+        self.last_epoch += 1
+
+
+class InverseLR(optim.lr_scheduler._LRScheduler):
+    """Implements an inverse decay learning rate schedule with an optional exponential
+    warmup. When last_epoch=-1, sets initial lr as lr.
+    inv_gamma is the number of steps/epochs required for the learning rate to decay to
+    (1 / 2)**power of its original value.
+    Args:
+        optimizer (Optimizer): Wrapped optimizer.
+        inv_gamma (float): Inverse multiplicative factor of learning rate decay. Default: 1.
+        power (float): Exponential factor of learning rate decay. Default: 1.
+        warmup (float): Exponential warmup factor (0 <= warmup < 1, 0 to disable)
+            Default: 0.
+        min_lr (float): The minimum learning rate. Default: 0.
+        last_epoch (int): The index of last epoch. Default: -1.
+        verbose (bool): If ``True``, prints a message to stdout for
+            each update. Default: ``False``.
+    """
+
+    def __init__(self, optimizer, inv_gamma=1., power=1., warmup=0., min_lr=0.,
+                 last_epoch=-1, verbose=False):
+        self.inv_gamma = inv_gamma
+        self.power = power
+        if not 0. <= warmup < 1:
+            raise ValueError('Invalid value for warmup')
+        self.warmup = warmup
+        self.min_lr = min_lr
+        super().__init__(optimizer, last_epoch, verbose)
+
+    def get_lr(self):
+        if not self._get_lr_called_within_step:
+            warnings.warn("To get the last learning rate computed by the scheduler, "
+                          "please use `get_last_lr()`.")
+
+        return self._get_closed_form_lr()
+
+    def _get_closed_form_lr(self):
+        warmup = 1 - self.warmup ** (self.last_epoch + 1)
+        lr_mult = (1 + self.last_epoch / self.inv_gamma) ** -self.power
+        return [warmup * max(self.min_lr, base_lr * lr_mult)
+                for base_lr in self.base_lrs]
+
+
+class ExponentialLR(optim.lr_scheduler._LRScheduler):
+    """Implements an exponential learning rate schedule with an optional exponential
+    warmup. When last_epoch=-1, sets initial lr as lr. Decays the learning rate
+    continuously by decay (default 0.5) every num_steps steps.
+    Args:
+        optimizer (Optimizer): Wrapped optimizer.
+        num_steps (float): The number of steps to decay the learning rate by decay in.
+        decay (float): The factor by which to decay the learning rate every num_steps
+            steps. Default: 0.5.
+        warmup (float): Exponential warmup factor (0 <= warmup < 1, 0 to disable)
+            Default: 0.
+        min_lr (float): The minimum learning rate. Default: 0.
+        last_epoch (int): The index of last epoch. Default: -1.
+        verbose (bool): If ``True``, prints a message to stdout for
+            each update. Default: ``False``.
+    """
+
+    def __init__(self, optimizer, num_steps, decay=0.5, warmup=0., min_lr=0.,
+                 last_epoch=-1, verbose=False):
+        self.num_steps = num_steps
+        self.decay = decay
+        if not 0. <= warmup < 1:
+            raise ValueError('Invalid value for warmup')
+        self.warmup = warmup
+        self.min_lr = min_lr
+        super().__init__(optimizer, last_epoch, verbose)
+
+    def get_lr(self):
+        if not self._get_lr_called_within_step:
+            warnings.warn("To get the last learning rate computed by the scheduler, "
+                          "please use `get_last_lr()`.")
+
+        return self._get_closed_form_lr()
+
+    def _get_closed_form_lr(self):
+        warmup = 1 - self.warmup ** (self.last_epoch + 1)
+        lr_mult = (self.decay ** (1 / self.num_steps)) ** self.last_epoch
+        return [warmup * max(self.min_lr, base_lr * lr_mult)
+                for base_lr in self.base_lrs]
+
+
+def rand_log_normal(shape, loc=0., scale=1., device='cpu', dtype=torch.float32):
+    """Draws samples from an lognormal distribution."""
+    return (torch.randn(shape, device=device, dtype=dtype) * scale + loc).exp()
+
+
+def rand_log_logistic(shape, loc=0., scale=1., min_value=0., max_value=float('inf'), device='cpu', dtype=torch.float32):
+    """Draws samples from an optionally truncated log-logistic distribution."""
+    min_value = torch.as_tensor(min_value, device=device, dtype=torch.float64)
+    max_value = torch.as_tensor(max_value, device=device, dtype=torch.float64)
+    min_cdf = min_value.log().sub(loc).div(scale).sigmoid()
+    max_cdf = max_value.log().sub(loc).div(scale).sigmoid()
+    u = torch.rand(shape, device=device, dtype=torch.float64) * (max_cdf - min_cdf) + min_cdf
+    return u.logit().mul(scale).add(loc).exp().to(dtype)
+
+
+def rand_log_uniform(shape, min_value, max_value, device='cpu', dtype=torch.float32):
+    """Draws samples from an log-uniform distribution."""
+    min_value = math.log(min_value)
+    max_value = math.log(max_value)
+    return (torch.rand(shape, device=device, dtype=dtype) * (max_value - min_value) + min_value).exp()
+
+
+def rand_v_diffusion(shape, sigma_data=1., min_value=0., max_value=float('inf'), device='cpu', dtype=torch.float32):
+    """Draws samples from a truncated v-diffusion training timestep distribution."""
+    min_cdf = math.atan(min_value / sigma_data) * 2 / math.pi
+    max_cdf = math.atan(max_value / sigma_data) * 2 / math.pi
+    u = torch.rand(shape, device=device, dtype=dtype) * (max_cdf - min_cdf) + min_cdf
+    return torch.tan(u * math.pi / 2) * sigma_data
+
+
+def rand_split_log_normal(shape, loc, scale_1, scale_2, device='cpu', dtype=torch.float32):
+    """Draws samples from a split lognormal distribution."""
+    n = torch.randn(shape, device=device, dtype=dtype).abs()
+    u = torch.rand(shape, device=device, dtype=dtype)
+    n_left = n * -scale_1 + loc
+    n_right = n * scale_2 + loc
+    ratio = scale_1 / (scale_1 + scale_2)
+    return torch.where(u < ratio, n_left, n_right).exp()
+
+
+class FolderOfImages(data.Dataset):
+    """Recursively finds all images in a directory. It does not support
+    classes/targets."""
+
+    IMG_EXTENSIONS = {'.jpg', '.jpeg', '.png', '.ppm', '.bmp', '.pgm', '.tif', '.tiff', '.webp'}
+
+    def __init__(self, root, transform=None):
+        super().__init__()
+        self.root = Path(root)
+        self.transform = nn.Identity() if transform is None else transform
+        self.paths = sorted(path for path in self.root.rglob('*') if path.suffix.lower() in self.IMG_EXTENSIONS)
+
+    def __repr__(self):
+        return f'FolderOfImages(root="{self.root}", len: {len(self)})'
+
+    def __len__(self):
+        return len(self.paths)
+
+    def __getitem__(self, key):
+        path = self.paths[key]
+        with open(path, 'rb') as f:
+            image = Image.open(f).convert('RGB')
+        image = self.transform(image)
+        return image,
+
+
+class CSVLogger:
+    def __init__(self, filename, columns):
+        self.filename = Path(filename)
+        self.columns = columns
+        if self.filename.exists():
+            self.file = open(self.filename, 'a')
+        else:
+            self.file = open(self.filename, 'w')
+            self.write(*self.columns)
+
+    def write(self, *args):
+        print(*args, sep=',', file=self.file, flush=True)
+
+
+@contextmanager
+def tf32_mode(cudnn=None, matmul=None):
+    """A context manager that sets whether TF32 is allowed on cuDNN or matmul."""
+    cudnn_old = torch.backends.cudnn.allow_tf32
+    matmul_old = torch.backends.cuda.matmul.allow_tf32
+    try:
+        if cudnn is not None:
+            torch.backends.cudnn.allow_tf32 = cudnn
+        if matmul is not None:
+            torch.backends.cuda.matmul.allow_tf32 = matmul
+        yield
+    finally:
+        if cudnn is not None:
+            torch.backends.cudnn.allow_tf32 = cudnn_old
+        if matmul is not None:
+            torch.backends.cuda.matmul.allow_tf32 = matmul_old

ldm_patched/ldm/models/autoencoder.py

@@ -0,0 +1,227 @@
+import torch
+# import pytorch_lightning as pl
+import torch.nn.functional as F
+from contextlib import contextmanager
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+from ldm_patched.ldm.modules.distributions.distributions import DiagonalGaussianDistribution
+
+from ldm_patched.ldm.util import instantiate_from_config
+from ldm_patched.ldm.modules.ema import LitEma
+
+class DiagonalGaussianRegularizer(torch.nn.Module):
+    def __init__(self, sample: bool = True):
+        super().__init__()
+        self.sample = sample
+
+    def get_trainable_parameters(self) -> Any:
+        yield from ()
+
+    def forward(self, z: torch.Tensor) -> Tuple[torch.Tensor, dict]:
+        log = dict()
+        posterior = DiagonalGaussianDistribution(z)
+        if self.sample:
+            z = posterior.sample()
+        else:
+            z = posterior.mode()
+        kl_loss = posterior.kl()
+        kl_loss = torch.sum(kl_loss) / kl_loss.shape[0]
+        log["kl_loss"] = kl_loss
+        return z, log
+
+
+class AbstractAutoencoder(torch.nn.Module):
+    """
+    This is the base class for all autoencoders, including image autoencoders, image autoencoders with discriminators,
+    unCLIP models, etc. Hence, it is fairly general, and specific features
+    (e.g. discriminator training, encoding, decoding) must be implemented in subclasses.
+    """
+
+    def __init__(
+        self,
+        ema_decay: Union[None, float] = None,
+        monitor: Union[None, str] = None,
+        input_key: str = "jpg",
+        **kwargs,
+    ):
+        super().__init__()
+
+        self.input_key = input_key
+        self.use_ema = ema_decay is not None
+        if monitor is not None:
+            self.monitor = monitor
+
+        if self.use_ema:
+            self.model_ema = LitEma(self, decay=ema_decay)
+            logpy.info(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
+
+    def get_input(self, batch) -> Any:
+        raise NotImplementedError()
+
+    def on_train_batch_end(self, *args, **kwargs):
+        # for EMA computation
+        if self.use_ema:
+            self.model_ema(self)
+
+    @contextmanager
+    def ema_scope(self, context=None):
+        if self.use_ema:
+            self.model_ema.store(self.parameters())
+            self.model_ema.copy_to(self)
+            if context is not None:
+                logpy.info(f"{context}: Switched to EMA weights")
+        try:
+            yield None
+        finally:
+            if self.use_ema:
+                self.model_ema.restore(self.parameters())
+                if context is not None:
+                    logpy.info(f"{context}: Restored training weights")
+
+    def encode(self, *args, **kwargs) -> torch.Tensor:
+        raise NotImplementedError("encode()-method of abstract base class called")
+
+    def decode(self, *args, **kwargs) -> torch.Tensor:
+        raise NotImplementedError("decode()-method of abstract base class called")
+
+    def instantiate_optimizer_from_config(self, params, lr, cfg):
+        logpy.info(f"loading >>> {cfg['target']} <<< optimizer from config")
+        return get_obj_from_str(cfg["target"])(
+            params, lr=lr, **cfg.get("params", dict())
+        )
+
+    def configure_optimizers(self) -> Any:
+        raise NotImplementedError()
+
+
+class AutoencodingEngine(AbstractAutoencoder):
+    """
+    Base class for all image autoencoders that we train, like VQGAN or AutoencoderKL
+    (we also restore them explicitly as special cases for legacy reasons).
+    Regularizations such as KL or VQ are moved to the regularizer class.
+    """
+
+    def __init__(
+        self,
+        *args,
+        encoder_config: Dict,
+        decoder_config: Dict,
+        regularizer_config: Dict,
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+
+        self.encoder: torch.nn.Module = instantiate_from_config(encoder_config)
+        self.decoder: torch.nn.Module = instantiate_from_config(decoder_config)
+        self.regularization: AbstractRegularizer = instantiate_from_config(
+            regularizer_config
+        )
+
+    def get_last_layer(self):
+        return self.decoder.get_last_layer()
+
+    def encode(
+        self,
+        x: torch.Tensor,
+        return_reg_log: bool = False,
+        unregularized: bool = False,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, dict]]:
+        z = self.encoder(x)
+        if unregularized:
+            return z, dict()
+        z, reg_log = self.regularization(z)
+        if return_reg_log:
+            return z, reg_log
+        return z
+
+    def decode(self, z: torch.Tensor, **kwargs) -> torch.Tensor:
+        x = self.decoder(z, **kwargs)
+        return x
+
+    def forward(
+        self, x: torch.Tensor, **additional_decode_kwargs
+    ) -> Tuple[torch.Tensor, torch.Tensor, dict]:
+        z, reg_log = self.encode(x, return_reg_log=True)
+        dec = self.decode(z, **additional_decode_kwargs)
+        return z, dec, reg_log
+
+
+class AutoencodingEngineLegacy(AutoencodingEngine):
+    def __init__(self, embed_dim: int, **kwargs):
+        self.max_batch_size = kwargs.pop("max_batch_size", None)
+        ddconfig = kwargs.pop("ddconfig")
+        super().__init__(
+            encoder_config={
+                "target": "ldm_patched.ldm.modules.diffusionmodules.model.Encoder",
+                "params": ddconfig,
+            },
+            decoder_config={
+                "target": "ldm_patched.ldm.modules.diffusionmodules.model.Decoder",
+                "params": ddconfig,
+            },
+            **kwargs,
+        )
+        self.quant_conv = torch.nn.Conv2d(
+            (1 + ddconfig["double_z"]) * ddconfig["z_channels"],
+            (1 + ddconfig["double_z"]) * embed_dim,
+            1,
+        )
+        self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
+        self.embed_dim = embed_dim
+
+    def get_autoencoder_params(self) -> list:
+        params = super().get_autoencoder_params()
+        return params
+
+    def encode(
+        self, x: torch.Tensor, return_reg_log: bool = False
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, dict]]:
+        if self.max_batch_size is None:
+            z = self.encoder(x)
+            z = self.quant_conv(z)
+        else:
+            N = x.shape[0]
+            bs = self.max_batch_size
+            n_batches = int(math.ceil(N / bs))
+            z = list()
+            for i_batch in range(n_batches):
+                z_batch = self.encoder(x[i_batch * bs : (i_batch + 1) * bs])
+                z_batch = self.quant_conv(z_batch)
+                z.append(z_batch)
+            z = torch.cat(z, 0)
+
+        z, reg_log = self.regularization(z)
+        if return_reg_log:
+            return z, reg_log
+        return z
+
+    def decode(self, z: torch.Tensor, **decoder_kwargs) -> torch.Tensor:
+        if self.max_batch_size is None:
+            dec = self.post_quant_conv(z)
+            dec = self.decoder(dec, **decoder_kwargs)
+        else:
+            N = z.shape[0]
+            bs = self.max_batch_size
+            n_batches = int(math.ceil(N / bs))
+            dec = list()
+            for i_batch in range(n_batches):
+                dec_batch = self.post_quant_conv(z[i_batch * bs : (i_batch + 1) * bs])
+                dec_batch = self.decoder(dec_batch, **decoder_kwargs)
+                dec.append(dec_batch)
+            dec = torch.cat(dec, 0)
+
+        return dec
+
+
+class AutoencoderKL(AutoencodingEngineLegacy):
+    def __init__(self, **kwargs):
+        if "lossconfig" in kwargs:
+            kwargs["loss_config"] = kwargs.pop("lossconfig")
+        super().__init__(
+            regularizer_config={
+                "target": (
+                    "ldm_patched.ldm.models.autoencoder.DiagonalGaussianRegularizer"
+                )
+            },
+            **kwargs,
+        )

ldm_patched/ldm/modules/attention.py

@@ -0,0 +1,774 @@
+from inspect import isfunction
+import math
+import torch
+import torch.nn.functional as F
+from torch import nn, einsum
+from einops import rearrange, repeat
+from typing import Optional, Any
+from functools import partial
+
+
+from .diffusionmodules.util import checkpoint, AlphaBlender, timestep_embedding
+from .sub_quadratic_attention import efficient_dot_product_attention
+
+from ldm_patched.modules import model_management
+
+if model_management.xformers_enabled():
+    import xformers
+    import xformers.ops
+
+from ldm_patched.modules.args_parser import args
+import ldm_patched.modules.ops
+ops = ldm_patched.modules.ops.disable_weight_init
+
+# CrossAttn precision handling
+if args.disable_attention_upcast:
+    print("disabling upcasting of attention")
+    _ATTN_PRECISION = "fp16"
+else:
+    _ATTN_PRECISION = "fp32"
+
+
+def exists(val):
+    return val is not None
+
+
+def uniq(arr):
+    return{el: True for el in arr}.keys()
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d
+
+
+def max_neg_value(t):
+    return -torch.finfo(t.dtype).max
+
+
+def init_(tensor):
+    dim = tensor.shape[-1]
+    std = 1 / math.sqrt(dim)
+    tensor.uniform_(-std, std)
+    return tensor
+
+
+# feedforward
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out, dtype=None, device=None, operations=ops):
+        super().__init__()
+        self.proj = operations.Linear(dim_in, dim_out * 2, dtype=dtype, device=device)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=-1)
+        return x * F.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0., dtype=None, device=None, operations=ops):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = nn.Sequential(
+            operations.Linear(dim, inner_dim, dtype=dtype, device=device),
+            nn.GELU()
+        ) if not glu else GEGLU(dim, inner_dim, dtype=dtype, device=device, operations=operations)
+
+        self.net = nn.Sequential(
+            project_in,
+            nn.Dropout(dropout),
+            operations.Linear(inner_dim, dim_out, dtype=dtype, device=device)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+def Normalize(in_channels, dtype=None, device=None):
+    return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True, dtype=dtype, device=device)
+
+def attention_basic(q, k, v, heads, mask=None):
+    b, _, dim_head = q.shape
+    dim_head //= heads
+    scale = dim_head ** -0.5
+
+    h = heads
+    q, k, v = map(
+        lambda t: t.unsqueeze(3)
+        .reshape(b, -1, heads, dim_head)
+        .permute(0, 2, 1, 3)
+        .reshape(b * heads, -1, dim_head)
+        .contiguous(),
+        (q, k, v),
+    )
+
+    # force cast to fp32 to avoid overflowing
+    if _ATTN_PRECISION =="fp32":
+        with torch.autocast(enabled=False, device_type = 'cuda'):
+            q, k = q.float(), k.float()
+            sim = einsum('b i d, b j d -> b i j', q, k) * scale
+    else:
+        sim = einsum('b i d, b j d -> b i j', q, k) * scale
+
+    del q, k
+
+    if exists(mask):
+        if mask.dtype == torch.bool:
+            mask = rearrange(mask, 'b ... -> b (...)') #TODO: check if this bool part matches pytorch attention
+            max_neg_value = -torch.finfo(sim.dtype).max
+            mask = repeat(mask, 'b j -> (b h) () j', h=h)
+            sim.masked_fill_(~mask, max_neg_value)
+        else:
+            sim += mask
+
+    # attention, what we cannot get enough of
+    sim = sim.softmax(dim=-1)
+
+    out = einsum('b i j, b j d -> b i d', sim.to(v.dtype), v)
+    out = (
+        out.unsqueeze(0)
+        .reshape(b, heads, -1, dim_head)
+        .permute(0, 2, 1, 3)
+        .reshape(b, -1, heads * dim_head)
+    )
+    return out
+
+
+def attention_sub_quad(query, key, value, heads, mask=None):
+    b, _, dim_head = query.shape
+    dim_head //= heads
+
+    scale = dim_head ** -0.5
+    query = query.unsqueeze(3).reshape(b, -1, heads, dim_head).permute(0, 2, 1, 3).reshape(b * heads, -1, dim_head)
+    value = value.unsqueeze(3).reshape(b, -1, heads, dim_head).permute(0, 2, 1, 3).reshape(b * heads, -1, dim_head)
+
+    key = key.unsqueeze(3).reshape(b, -1, heads, dim_head).permute(0, 2, 3, 1).reshape(b * heads, dim_head, -1)
+
+    dtype = query.dtype
+    upcast_attention = _ATTN_PRECISION =="fp32" and query.dtype != torch.float32
+    if upcast_attention:
+        bytes_per_token = torch.finfo(torch.float32).bits//8
+    else:
+        bytes_per_token = torch.finfo(query.dtype).bits//8
+    batch_x_heads, q_tokens, _ = query.shape
+    _, _, k_tokens = key.shape
+    qk_matmul_size_bytes = batch_x_heads * bytes_per_token * q_tokens * k_tokens
+
+    mem_free_total, mem_free_torch = model_management.get_free_memory(query.device, True)
+
+    kv_chunk_size_min = None
+    kv_chunk_size = None
+    query_chunk_size = None
+
+    for x in [4096, 2048, 1024, 512, 256]:
+        count = mem_free_total / (batch_x_heads * bytes_per_token * x * 4.0)
+        if count >= k_tokens:
+            kv_chunk_size = k_tokens
+            query_chunk_size = x
+            break
+
+    if query_chunk_size is None:
+        query_chunk_size = 512
+
+    hidden_states = efficient_dot_product_attention(
+        query,
+        key,
+        value,
+        query_chunk_size=query_chunk_size,
+        kv_chunk_size=kv_chunk_size,
+        kv_chunk_size_min=kv_chunk_size_min,
+        use_checkpoint=False,
+        upcast_attention=upcast_attention,
+    )
+
+    hidden_states = hidden_states.to(dtype)
+
+    hidden_states = hidden_states.unflatten(0, (-1, heads)).transpose(1,2).flatten(start_dim=2)
+    return hidden_states
+
+def attention_split(q, k, v, heads, mask=None):
+    b, _, dim_head = q.shape
+    dim_head //= heads
+    scale = dim_head ** -0.5
+
+    h = heads
+    q, k, v = map(
+        lambda t: t.unsqueeze(3)
+        .reshape(b, -1, heads, dim_head)
+        .permute(0, 2, 1, 3)
+        .reshape(b * heads, -1, dim_head)
+        .contiguous(),
+        (q, k, v),
+    )
+
+    r1 = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device, dtype=q.dtype)
+
+    mem_free_total = model_management.get_free_memory(q.device)
+
+    if _ATTN_PRECISION =="fp32":
+        element_size = 4
+    else:
+        element_size = q.element_size()
+
+    gb = 1024 ** 3
+    tensor_size = q.shape[0] * q.shape[1] * k.shape[1] * element_size
+    modifier = 3
+    mem_required = tensor_size * modifier
+    steps = 1
+
+
+    if mem_required > mem_free_total:
+        steps = 2**(math.ceil(math.log(mem_required / mem_free_total, 2)))
+        # print(f"Expected tensor size:{tensor_size/gb:0.1f}GB, cuda free:{mem_free_cuda/gb:0.1f}GB "
+        #      f"torch free:{mem_free_torch/gb:0.1f} total:{mem_free_total/gb:0.1f} steps:{steps}")
+
+    if steps > 64:
+        max_res = math.floor(math.sqrt(math.sqrt(mem_free_total / 2.5)) / 8) * 64
+        raise RuntimeError(f'Not enough memory, use lower resolution (max approx. {max_res}x{max_res}). '
+                            f'Need: {mem_required/64/gb:0.1f}GB free, Have:{mem_free_total/gb:0.1f}GB free')
+
+    # print("steps", steps, mem_required, mem_free_total, modifier, q.element_size(), tensor_size)
+    first_op_done = False
+    cleared_cache = False
+    while True:
+        try:
+            slice_size = q.shape[1] // steps if (q.shape[1] % steps) == 0 else q.shape[1]
+            for i in range(0, q.shape[1], slice_size):
+                end = i + slice_size
+                if _ATTN_PRECISION =="fp32":
+                    with torch.autocast(enabled=False, device_type = 'cuda'):
+                        s1 = einsum('b i d, b j d -> b i j', q[:, i:end].float(), k.float()) * scale
+                else:
+                    s1 = einsum('b i d, b j d -> b i j', q[:, i:end], k) * scale
+
+                s2 = s1.softmax(dim=-1).to(v.dtype)
+                del s1
+                first_op_done = True
+
+                r1[:, i:end] = einsum('b i j, b j d -> b i d', s2, v)
+                del s2
+            break
+        except model_management.OOM_EXCEPTION as e:
+            if first_op_done == False:
+                model_management.soft_empty_cache(True)
+                if cleared_cache == False:
+                    cleared_cache = True
+                    print("out of memory error, emptying cache and trying again")
+                    continue
+                steps *= 2
+                if steps > 64:
+                    raise e
+                print("out of memory error, increasing steps and trying again", steps)
+            else:
+                raise e
+
+    del q, k, v
+
+    r1 = (
+        r1.unsqueeze(0)
+        .reshape(b, heads, -1, dim_head)
+        .permute(0, 2, 1, 3)
+        .reshape(b, -1, heads * dim_head)
+    )
+    return r1
+
+BROKEN_XFORMERS = False
+try:
+    x_vers = xformers.__version__
+    #I think 0.0.23 is also broken (q with bs bigger than 65535 gives CUDA error)
+    BROKEN_XFORMERS = x_vers.startswith("0.0.21") or x_vers.startswith("0.0.22") or x_vers.startswith("0.0.23")
+except:
+    pass
+
+def attention_xformers(q, k, v, heads, mask=None):
+    b, _, dim_head = q.shape
+    dim_head //= heads
+    if BROKEN_XFORMERS:
+        if b * heads > 65535:
+            return attention_pytorch(q, k, v, heads, mask)
+
+    q, k, v = map(
+        lambda t: t.unsqueeze(3)
+        .reshape(b, -1, heads, dim_head)
+        .permute(0, 2, 1, 3)
+        .reshape(b * heads, -1, dim_head)
+        .contiguous(),
+        (q, k, v),
+    )
+
+    # actually compute the attention, what we cannot get enough of
+    out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None)
+
+    if exists(mask):
+        raise NotImplementedError
+    out = (
+        out.unsqueeze(0)
+        .reshape(b, heads, -1, dim_head)
+        .permute(0, 2, 1, 3)
+        .reshape(b, -1, heads * dim_head)
+    )
+    return out
+
+def attention_pytorch(q, k, v, heads, mask=None):
+    b, _, dim_head = q.shape
+    dim_head //= heads
+    q, k, v = map(
+        lambda t: t.view(b, -1, heads, dim_head).transpose(1, 2),
+        (q, k, v),
+    )
+
+    out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
+    out = (
+        out.transpose(1, 2).reshape(b, -1, heads * dim_head)
+    )
+    return out
+
+
+optimized_attention = attention_basic
+optimized_attention_masked = attention_basic
+
+if model_management.xformers_enabled():
+    print("Using xformers cross attention")
+    optimized_attention = attention_xformers
+elif model_management.pytorch_attention_enabled():
+    print("Using pytorch cross attention")
+    optimized_attention = attention_pytorch
+else:
+    if args.attention_split:
+        print("Using split optimization for cross attention")
+        optimized_attention = attention_split
+    else:
+        print("Using sub quadratic optimization for cross attention, if you have memory or speed issues try using: --attention-split")
+        optimized_attention = attention_sub_quad
+
+if model_management.pytorch_attention_enabled():
+    optimized_attention_masked = attention_pytorch
+
+def optimized_attention_for_device(device, mask=False):
+    if device == torch.device("cpu"): #TODO
+        if model_management.pytorch_attention_enabled():
+            return attention_pytorch
+        else:
+            return attention_basic
+    if mask:
+        return optimized_attention_masked
+
+    return optimized_attention
+
+
+class CrossAttention(nn.Module):
+    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., dtype=None, device=None, operations=ops):
+        super().__init__()
+        inner_dim = dim_head * heads
+        context_dim = default(context_dim, query_dim)
+
+        self.heads = heads
+        self.dim_head = dim_head
+
+        self.to_q = operations.Linear(query_dim, inner_dim, bias=False, dtype=dtype, device=device)
+        self.to_k = operations.Linear(context_dim, inner_dim, bias=False, dtype=dtype, device=device)
+        self.to_v = operations.Linear(context_dim, inner_dim, bias=False, dtype=dtype, device=device)
+
+        self.to_out = nn.Sequential(operations.Linear(inner_dim, query_dim, dtype=dtype, device=device), nn.Dropout(dropout))
+
+    def forward(self, x, context=None, value=None, mask=None):
+        q = self.to_q(x)
+        context = default(context, x)
+        k = self.to_k(context)
+        if value is not None:
+            v = self.to_v(value)
+            del value
+        else:
+            v = self.to_v(context)
+
+        if mask is None:
+            out = optimized_attention(q, k, v, self.heads)
+        else:
+            out = optimized_attention_masked(q, k, v, self.heads, mask)
+        return self.to_out(out)
+
+
+class BasicTransformerBlock(nn.Module):
+    def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True, ff_in=False, inner_dim=None,
+                 disable_self_attn=False, disable_temporal_crossattention=False, switch_temporal_ca_to_sa=False, dtype=None, device=None, operations=ops):
+        super().__init__()
+
+        self.ff_in = ff_in or inner_dim is not None
+        if inner_dim is None:
+            inner_dim = dim
+
+        self.is_res = inner_dim == dim
+
+        if self.ff_in:
+            self.norm_in = operations.LayerNorm(dim, dtype=dtype, device=device)
+            self.ff_in = FeedForward(dim, dim_out=inner_dim, dropout=dropout, glu=gated_ff, dtype=dtype, device=device, operations=operations)
+
+        self.disable_self_attn = disable_self_attn
+        self.attn1 = CrossAttention(query_dim=inner_dim, heads=n_heads, dim_head=d_head, dropout=dropout,
+                              context_dim=context_dim if self.disable_self_attn else None, dtype=dtype, device=device, operations=operations)  # is a self-attention if not self.disable_self_attn
+        self.ff = FeedForward(inner_dim, dim_out=dim, dropout=dropout, glu=gated_ff, dtype=dtype, device=device, operations=operations)
+
+        if disable_temporal_crossattention:
+            if switch_temporal_ca_to_sa:
+                raise ValueError
+            else:
+                self.attn2 = None
+        else:
+            context_dim_attn2 = None
+            if not switch_temporal_ca_to_sa:
+                context_dim_attn2 = context_dim
+
+            self.attn2 = CrossAttention(query_dim=inner_dim, context_dim=context_dim_attn2,
+                                heads=n_heads, dim_head=d_head, dropout=dropout, dtype=dtype, device=device, operations=operations)  # is self-attn if context is none
+            self.norm2 = operations.LayerNorm(inner_dim, dtype=dtype, device=device)
+
+        self.norm1 = operations.LayerNorm(inner_dim, dtype=dtype, device=device)
+        self.norm3 = operations.LayerNorm(inner_dim, dtype=dtype, device=device)
+        self.checkpoint = checkpoint
+        self.n_heads = n_heads
+        self.d_head = d_head
+        self.switch_temporal_ca_to_sa = switch_temporal_ca_to_sa
+
+    def forward(self, x, context=None, transformer_options={}):
+        return checkpoint(self._forward, (x, context, transformer_options), self.parameters(), self.checkpoint)
+
+    def _forward(self, x, context=None, transformer_options={}):
+        extra_options = {}
+        block = transformer_options.get("block", None)
+        block_index = transformer_options.get("block_index", 0)
+        transformer_patches = {}
+        transformer_patches_replace = {}
+
+        for k in transformer_options:
+            if k == "patches":
+                transformer_patches = transformer_options[k]
+            elif k == "patches_replace":
+                transformer_patches_replace = transformer_options[k]
+            else:
+                extra_options[k] = transformer_options[k]
+
+        extra_options["n_heads"] = self.n_heads
+        extra_options["dim_head"] = self.d_head
+
+        if self.ff_in:
+            x_skip = x
+            x = self.ff_in(self.norm_in(x))
+            if self.is_res:
+                x += x_skip
+
+        n = self.norm1(x)
+        if self.disable_self_attn:
+            context_attn1 = context
+        else:
+            context_attn1 = None
+        value_attn1 = None
+
+        if "attn1_patch" in transformer_patches:
+            patch = transformer_patches["attn1_patch"]
+            if context_attn1 is None:
+                context_attn1 = n
+            value_attn1 = context_attn1
+            for p in patch:
+                n, context_attn1, value_attn1 = p(n, context_attn1, value_attn1, extra_options)
+
+        if block is not None:
+            transformer_block = (block[0], block[1], block_index)
+        else:
+            transformer_block = None
+        attn1_replace_patch = transformer_patches_replace.get("attn1", {})
+        block_attn1 = transformer_block
+        if block_attn1 not in attn1_replace_patch:
+            block_attn1 = block
+
+        if block_attn1 in attn1_replace_patch:
+            if context_attn1 is None:
+                context_attn1 = n
+                value_attn1 = n
+            n = self.attn1.to_q(n)
+            context_attn1 = self.attn1.to_k(context_attn1)
+            value_attn1 = self.attn1.to_v(value_attn1)
+            n = attn1_replace_patch[block_attn1](n, context_attn1, value_attn1, extra_options)
+            n = self.attn1.to_out(n)
+        else:
+            n = self.attn1(n, context=context_attn1, value=value_attn1)
+
+        if "attn1_output_patch" in transformer_patches:
+            patch = transformer_patches["attn1_output_patch"]
+            for p in patch:
+                n = p(n, extra_options)
+
+        x += n
+        if "middle_patch" in transformer_patches:
+            patch = transformer_patches["middle_patch"]
+            for p in patch:
+                x = p(x, extra_options)
+
+        if self.attn2 is not None:
+            n = self.norm2(x)
+            if self.switch_temporal_ca_to_sa:
+                context_attn2 = n
+            else:
+                context_attn2 = context
+            value_attn2 = None
+            if "attn2_patch" in transformer_patches:
+                patch = transformer_patches["attn2_patch"]
+                value_attn2 = context_attn2
+                for p in patch:
+                    n, context_attn2, value_attn2 = p(n, context_attn2, value_attn2, extra_options)
+
+            attn2_replace_patch = transformer_patches_replace.get("attn2", {})
+            block_attn2 = transformer_block
+            if block_attn2 not in attn2_replace_patch:
+                block_attn2 = block
+
+            if block_attn2 in attn2_replace_patch:
+                if value_attn2 is None:
+                    value_attn2 = context_attn2
+                n = self.attn2.to_q(n)
+                context_attn2 = self.attn2.to_k(context_attn2)
+                value_attn2 = self.attn2.to_v(value_attn2)
+                n = attn2_replace_patch[block_attn2](n, context_attn2, value_attn2, extra_options)
+                n = self.attn2.to_out(n)
+            else:
+                n = self.attn2(n, context=context_attn2, value=value_attn2)
+
+        if "attn2_output_patch" in transformer_patches:
+            patch = transformer_patches["attn2_output_patch"]
+            for p in patch:
+                n = p(n, extra_options)
+
+        x += n
+        if self.is_res:
+            x_skip = x
+        x = self.ff(self.norm3(x))
+        if self.is_res:
+            x += x_skip
+
+        return x
+
+
+class SpatialTransformer(nn.Module):
+    """
+    Transformer block for image-like data.
+    First, project the input (aka embedding)
+    and reshape to b, t, d.
+    Then apply standard transformer action.
+    Finally, reshape to image
+    NEW: use_linear for more efficiency instead of the 1x1 convs
+    """
+    def __init__(self, in_channels, n_heads, d_head,
+                 depth=1, dropout=0., context_dim=None,
+                 disable_self_attn=False, use_linear=False,
+                 use_checkpoint=True, dtype=None, device=None, operations=ops):
+        super().__init__()
+        if exists(context_dim) and not isinstance(context_dim, list):
+            context_dim = [context_dim] * depth
+        self.in_channels = in_channels
+        inner_dim = n_heads * d_head
+        self.norm = operations.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True, dtype=dtype, device=device)
+        if not use_linear:
+            self.proj_in = operations.Conv2d(in_channels,
+                                     inner_dim,
+                                     kernel_size=1,
+                                     stride=1,
+                                     padding=0, dtype=dtype, device=device)
+        else:
+            self.proj_in = operations.Linear(in_channels, inner_dim, dtype=dtype, device=device)
+
+        self.transformer_blocks = nn.ModuleList(
+            [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim[d],
+                                   disable_self_attn=disable_self_attn, checkpoint=use_checkpoint, dtype=dtype, device=device, operations=operations)
+                for d in range(depth)]
+        )
+        if not use_linear:
+            self.proj_out = operations.Conv2d(inner_dim,in_channels,
+                                                  kernel_size=1,
+                                                  stride=1,
+                                                  padding=0, dtype=dtype, device=device)
+        else:
+            self.proj_out = operations.Linear(in_channels, inner_dim, dtype=dtype, device=device)
+        self.use_linear = use_linear
+
+    def forward(self, x, context=None, transformer_options={}):
+        # note: if no context is given, cross-attention defaults to self-attention
+        if not isinstance(context, list):
+            context = [context] * len(self.transformer_blocks)
+        b, c, h, w = x.shape
+        x_in = x
+        x = self.norm(x)
+        if not self.use_linear:
+            x = self.proj_in(x)
+        x = rearrange(x, 'b c h w -> b (h w) c').contiguous()
+        if self.use_linear:
+            x = self.proj_in(x)
+        for i, block in enumerate(self.transformer_blocks):
+            transformer_options["block_index"] = i
+            x = block(x, context=context[i], transformer_options=transformer_options)
+        if self.use_linear:
+            x = self.proj_out(x)
+        x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous()
+        if not self.use_linear:
+            x = self.proj_out(x)
+        return x + x_in
+
+
+class SpatialVideoTransformer(SpatialTransformer):
+    def __init__(
+        self,
+        in_channels,
+        n_heads,
+        d_head,
+        depth=1,
+        dropout=0.0,
+        use_linear=False,
+        context_dim=None,
+        use_spatial_context=False,
+        timesteps=None,
+        merge_strategy: str = "fixed",
+        merge_factor: float = 0.5,
+        time_context_dim=None,
+        ff_in=False,
+        checkpoint=False,
+        time_depth=1,
+        disable_self_attn=False,
+        disable_temporal_crossattention=False,
+        max_time_embed_period: int = 10000,
+        dtype=None, device=None, operations=ops
+    ):
+        super().__init__(
+            in_channels,
+            n_heads,
+            d_head,
+            depth=depth,
+            dropout=dropout,
+            use_checkpoint=checkpoint,
+            context_dim=context_dim,
+            use_linear=use_linear,
+            disable_self_attn=disable_self_attn,
+            dtype=dtype, device=device, operations=operations
+        )
+        self.time_depth = time_depth
+        self.depth = depth
+        self.max_time_embed_period = max_time_embed_period
+
+        time_mix_d_head = d_head
+        n_time_mix_heads = n_heads
+
+        time_mix_inner_dim = int(time_mix_d_head * n_time_mix_heads)
+
+        inner_dim = n_heads * d_head
+        if use_spatial_context:
+            time_context_dim = context_dim
+
+        self.time_stack = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim,
+                    n_time_mix_heads,
+                    time_mix_d_head,
+                    dropout=dropout,
+                    context_dim=time_context_dim,
+                    # timesteps=timesteps,
+                    checkpoint=checkpoint,
+                    ff_in=ff_in,
+                    inner_dim=time_mix_inner_dim,
+                    disable_self_attn=disable_self_attn,
+                    disable_temporal_crossattention=disable_temporal_crossattention,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for _ in range(self.depth)
+            ]
+        )
+
+        assert len(self.time_stack) == len(self.transformer_blocks)
+
+        self.use_spatial_context = use_spatial_context
+        self.in_channels = in_channels
+
+        time_embed_dim = self.in_channels * 4
+        self.time_pos_embed = nn.Sequential(
+            operations.Linear(self.in_channels, time_embed_dim, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Linear(time_embed_dim, self.in_channels, dtype=dtype, device=device),
+        )
+
+        self.time_mixer = AlphaBlender(
+            alpha=merge_factor, merge_strategy=merge_strategy
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        context: Optional[torch.Tensor] = None,
+        time_context: Optional[torch.Tensor] = None,
+        timesteps: Optional[int] = None,
+        image_only_indicator: Optional[torch.Tensor] = None,
+        transformer_options={}
+    ) -> torch.Tensor:
+        _, _, h, w = x.shape
+        x_in = x
+        spatial_context = None
+        if exists(context):
+            spatial_context = context
+
+        if self.use_spatial_context:
+            assert (
+                context.ndim == 3
+            ), f"n dims of spatial context should be 3 but are {context.ndim}"
+
+            if time_context is None:
+                time_context = context
+            time_context_first_timestep = time_context[::timesteps]
+            time_context = repeat(
+                time_context_first_timestep, "b ... -> (b n) ...", n=h * w
+            )
+        elif time_context is not None and not self.use_spatial_context:
+            time_context = repeat(time_context, "b ... -> (b n) ...", n=h * w)
+            if time_context.ndim == 2:
+                time_context = rearrange(time_context, "b c -> b 1 c")
+
+        x = self.norm(x)
+        if not self.use_linear:
+            x = self.proj_in(x)
+        x = rearrange(x, "b c h w -> b (h w) c")
+        if self.use_linear:
+            x = self.proj_in(x)
+
+        num_frames = torch.arange(timesteps, device=x.device)
+        num_frames = repeat(num_frames, "t -> b t", b=x.shape[0] // timesteps)
+        num_frames = rearrange(num_frames, "b t -> (b t)")
+        t_emb = timestep_embedding(num_frames, self.in_channels, repeat_only=False, max_period=self.max_time_embed_period).to(x.dtype)
+        emb = self.time_pos_embed(t_emb)
+        emb = emb[:, None, :]
+
+        for it_, (block, mix_block) in enumerate(
+            zip(self.transformer_blocks, self.time_stack)
+        ):
+            transformer_options["block_index"] = it_
+            x = block(
+                x,
+                context=spatial_context,
+                transformer_options=transformer_options,
+            )
+
+            x_mix = x
+            x_mix = x_mix + emb
+
+            B, S, C = x_mix.shape
+            x_mix = rearrange(x_mix, "(b t) s c -> (b s) t c", t=timesteps)
+            x_mix = mix_block(x_mix, context=time_context) #TODO: transformer_options
+            x_mix = rearrange(
+                x_mix, "(b s) t c -> (b t) s c", s=S, b=B // timesteps, c=C, t=timesteps
+            )
+
+            x = self.time_mixer(x_spatial=x, x_temporal=x_mix, image_only_indicator=image_only_indicator)
+
+        if self.use_linear:
+            x = self.proj_out(x)
+        x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w)
+        if not self.use_linear:
+            x = self.proj_out(x)
+        out = x + x_in
+        return out
+
+

ldm_patched/ldm/modules/diffusionmodules/model.py

@@ -0,0 +1,650 @@
+# pytorch_diffusion + derived encoder decoder
+import math
+import torch
+import torch.nn as nn
+import numpy as np
+from einops import rearrange
+from typing import Optional, Any
+
+from ldm_patched.modules import model_management
+import ldm_patched.modules.ops
+ops = ldm_patched.modules.ops.disable_weight_init
+
+if model_management.xformers_enabled_vae():
+    import xformers
+    import xformers.ops
+
+def get_timestep_embedding(timesteps, embedding_dim):
+    """
+    This matches the implementation in Denoising Diffusion Probabilistic Models:
+    From Fairseq.
+    Build sinusoidal embeddings.
+    This matches the implementation in tensor2tensor, but differs slightly
+    from the description in Section 3.5 of "Attention Is All You Need".
+    """
+    assert len(timesteps.shape) == 1
+
+    half_dim = embedding_dim // 2
+    emb = math.log(10000) / (half_dim - 1)
+    emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
+    emb = emb.to(device=timesteps.device)
+    emb = timesteps.float()[:, None] * emb[None, :]
+    emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+    if embedding_dim % 2 == 1:  # zero pad
+        emb = torch.nn.functional.pad(emb, (0,1,0,0))
+    return emb
+
+
+def nonlinearity(x):
+    # swish
+    return x*torch.sigmoid(x)
+
+
+def Normalize(in_channels, num_groups=32):
+    return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+
+
+class Upsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            self.conv = ops.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        try:
+            x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+        except: #operation not implemented for bf16
+            b, c, h, w = x.shape
+            out = torch.empty((b, c, h*2, w*2), dtype=x.dtype, layout=x.layout, device=x.device)
+            split = 8
+            l = out.shape[1] // split
+            for i in range(0, out.shape[1], l):
+                out[:,i:i+l] = torch.nn.functional.interpolate(x[:,i:i+l].to(torch.float32), scale_factor=2.0, mode="nearest").to(x.dtype)
+            del x
+            x = out
+
+        if self.with_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = ops.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=2,
+                                        padding=0)
+
+    def forward(self, x):
+        if self.with_conv:
+            pad = (0,1,0,1)
+            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
+            x = self.conv(x)
+        else:
+            x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
+        return x
+
+
+class ResnetBlock(nn.Module):
+    def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False,
+                 dropout, temb_channels=512):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+
+        self.swish = torch.nn.SiLU(inplace=True)
+        self.norm1 = Normalize(in_channels)
+        self.conv1 = ops.Conv2d(in_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if temb_channels > 0:
+            self.temb_proj = ops.Linear(temb_channels,
+                                             out_channels)
+        self.norm2 = Normalize(out_channels)
+        self.dropout = torch.nn.Dropout(dropout, inplace=True)
+        self.conv2 = ops.Conv2d(out_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = ops.Conv2d(in_channels,
+                                                     out_channels,
+                                                     kernel_size=3,
+                                                     stride=1,
+                                                     padding=1)
+            else:
+                self.nin_shortcut = ops.Conv2d(in_channels,
+                                                    out_channels,
+                                                    kernel_size=1,
+                                                    stride=1,
+                                                    padding=0)
+
+    def forward(self, x, temb):
+        h = x
+        h = self.norm1(h)
+        h = self.swish(h)
+        h = self.conv1(h)
+
+        if temb is not None:
+            h = h + self.temb_proj(self.swish(temb))[:,:,None,None]
+
+        h = self.norm2(h)
+        h = self.swish(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return x+h
+
+def slice_attention(q, k, v):
+    r1 = torch.zeros_like(k, device=q.device)
+    scale = (int(q.shape[-1])**(-0.5))
+
+    mem_free_total = model_management.get_free_memory(q.device)
+
+    gb = 1024 ** 3
+    tensor_size = q.shape[0] * q.shape[1] * k.shape[2] * q.element_size()
+    modifier = 3 if q.element_size() == 2 else 2.5
+    mem_required = tensor_size * modifier
+    steps = 1
+
+    if mem_required > mem_free_total:
+        steps = 2**(math.ceil(math.log(mem_required / mem_free_total, 2)))
+
+    while True:
+        try:
+            slice_size = q.shape[1] // steps if (q.shape[1] % steps) == 0 else q.shape[1]
+            for i in range(0, q.shape[1], slice_size):
+                end = i + slice_size
+                s1 = torch.bmm(q[:, i:end], k) * scale
+
+                s2 = torch.nn.functional.softmax(s1, dim=2).permute(0,2,1)
+                del s1
+
+                r1[:, :, i:end] = torch.bmm(v, s2)
+                del s2
+            break
+        except model_management.OOM_EXCEPTION as e:
+            model_management.soft_empty_cache(True)
+            steps *= 2
+            if steps > 128:
+                raise e
+            print("out of memory error, increasing steps and trying again", steps)
+
+    return r1
+
+def normal_attention(q, k, v):
+    # compute attention
+    b,c,h,w = q.shape
+
+    q = q.reshape(b,c,h*w)
+    q = q.permute(0,2,1)   # b,hw,c
+    k = k.reshape(b,c,h*w) # b,c,hw
+    v = v.reshape(b,c,h*w)
+
+    r1 = slice_attention(q, k, v)
+    h_ = r1.reshape(b,c,h,w)
+    del r1
+    return h_
+
+def xformers_attention(q, k, v):
+    # compute attention
+    B, C, H, W = q.shape
+    q, k, v = map(
+        lambda t: t.view(B, C, -1).transpose(1, 2).contiguous(),
+        (q, k, v),
+    )
+
+    try:
+        out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None)
+        out = out.transpose(1, 2).reshape(B, C, H, W)
+    except NotImplementedError as e:
+        out = slice_attention(q.view(B, -1, C), k.view(B, -1, C).transpose(1, 2), v.view(B, -1, C).transpose(1, 2)).reshape(B, C, H, W)
+    return out
+
+def pytorch_attention(q, k, v):
+    # compute attention
+    B, C, H, W = q.shape
+    q, k, v = map(
+        lambda t: t.view(B, 1, C, -1).transpose(2, 3).contiguous(),
+        (q, k, v),
+    )
+
+    try:
+        out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=False)
+        out = out.transpose(2, 3).reshape(B, C, H, W)
+    except model_management.OOM_EXCEPTION as e:
+        print("scaled_dot_product_attention OOMed: switched to slice attention")
+        out = slice_attention(q.view(B, -1, C), k.view(B, -1, C).transpose(1, 2), v.view(B, -1, C).transpose(1, 2)).reshape(B, C, H, W)
+    return out
+
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = ops.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = ops.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = ops.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = ops.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+
+        if model_management.xformers_enabled_vae():
+            print("Using xformers attention in VAE")
+            self.optimized_attention = xformers_attention
+        elif model_management.pytorch_attention_enabled():
+            print("Using pytorch attention in VAE")
+            self.optimized_attention = pytorch_attention
+        else:
+            print("Using split attention in VAE")
+            self.optimized_attention = normal_attention
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        h_ = self.optimized_attention(q, k, v)
+
+        h_ = self.proj_out(h_)
+
+        return x+h_
+
+
+def make_attn(in_channels, attn_type="vanilla", attn_kwargs=None):
+    return AttnBlock(in_channels)
+
+
+class Model(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = self.ch*4
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        self.use_timestep = use_timestep
+        if self.use_timestep:
+            # timestep embedding
+            self.temb = nn.Module()
+            self.temb.dense = nn.ModuleList([
+                ops.Linear(self.ch,
+                                self.temb_ch),
+                ops.Linear(self.temb_ch,
+                                self.temb_ch),
+            ])
+
+        # downsampling
+        self.conv_in = ops.Conv2d(in_channels,
+                                       self.ch,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,)+tuple(ch_mult)
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch*in_ch_mult[i_level]
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions-1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch*ch_mult[i_level]
+            skip_in = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks+1):
+                if i_block == self.num_res_blocks:
+                    skip_in = ch*in_ch_mult[i_level]
+                block.append(ResnetBlock(in_channels=block_in+skip_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up) # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = ops.Conv2d(block_in,
+                                        out_ch,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x, t=None, context=None):
+        #assert x.shape[2] == x.shape[3] == self.resolution
+        if context is not None:
+            # assume aligned context, cat along channel axis
+            x = torch.cat((x, context), dim=1)
+        if self.use_timestep:
+            # timestep embedding
+            assert t is not None
+            temb = get_timestep_embedding(t, self.ch)
+            temb = self.temb.dense[0](temb)
+            temb = nonlinearity(temb)
+            temb = self.temb.dense[1](temb)
+        else:
+            temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions-1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks+1):
+                h = self.up[i_level].block[i_block](
+                    torch.cat([h, hs.pop()], dim=1), temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+    def get_last_layer(self):
+        return self.conv_out.weight
+
+
+class Encoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, double_z=True, use_linear_attn=False, attn_type="vanilla",
+                 **ignore_kwargs):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        # downsampling
+        self.conv_in = ops.Conv2d(in_channels,
+                                       self.ch,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,)+tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch*in_ch_mult[i_level]
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions-1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = ops.Conv2d(block_in,
+                                        2*z_channels if double_z else z_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        # timestep embedding
+        temb = None
+        # downsampling
+        h = self.conv_in(x)
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](h, temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+            if i_level != self.num_resolutions-1:
+                h = self.down[i_level].downsample(h)
+
+        # middle
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class Decoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False,
+                 conv_out_op=ops.Conv2d,
+                 resnet_op=ResnetBlock,
+                 attn_op=AttnBlock,
+                **ignorekwargs):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.give_pre_end = give_pre_end
+        self.tanh_out = tanh_out
+
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        in_ch_mult = (1,)+tuple(ch_mult)
+        block_in = ch*ch_mult[self.num_resolutions-1]
+        curr_res = resolution // 2**(self.num_resolutions-1)
+        self.z_shape = (1,z_channels,curr_res,curr_res)
+        print("Working with z of shape {} = {} dimensions.".format(
+            self.z_shape, np.prod(self.z_shape)))
+
+        # z to block_in
+        self.conv_in = ops.Conv2d(z_channels,
+                                       block_in,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = resnet_op(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = attn_op(block_in)
+        self.mid.block_2 = resnet_op(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks+1):
+                block.append(resnet_op(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(attn_op(block_in))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up) # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = conv_out_op(block_in,
+                                        out_ch,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, z, **kwargs):
+        #assert z.shape[1:] == self.z_shape[1:]
+        self.last_z_shape = z.shape
+
+        # timestep embedding
+        temb = None
+
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        h = self.mid.block_1(h, temb, **kwargs)
+        h = self.mid.attn_1(h, **kwargs)
+        h = self.mid.block_2(h, temb, **kwargs)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks+1):
+                h = self.up[i_level].block[i_block](h, temb, **kwargs)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h, **kwargs)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        if self.give_pre_end:
+            return h
+
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h, **kwargs)
+        if self.tanh_out:
+            h = torch.tanh(h)
+        return h

ldm_patched/ldm/modules/diffusionmodules/openaimodel.py

@@ -0,0 +1,893 @@
+from abc import abstractmethod
+import math
+
+import numpy as np
+import torch as th
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+from functools import partial
+
+from .util import (
+    checkpoint,
+    avg_pool_nd,
+    zero_module,
+    timestep_embedding,
+    AlphaBlender,
+)
+from ..attention import SpatialTransformer, SpatialVideoTransformer, default
+from ldm_patched.ldm.util import exists
+import ldm_patched.modules.ops
+ops = ldm_patched.modules.ops.disable_weight_init
+
+class TimestepBlock(nn.Module):
+    """
+    Any module where forward() takes timestep embeddings as a second argument.
+    """
+
+    @abstractmethod
+    def forward(self, x, emb):
+        """
+        Apply the module to `x` given `emb` timestep embeddings.
+        """
+
+#This is needed because accelerate makes a copy of transformer_options which breaks "transformer_index"
+def forward_timestep_embed(ts, x, emb, context=None, transformer_options={}, output_shape=None, time_context=None, num_video_frames=None, image_only_indicator=None):
+    for layer in ts:
+        if isinstance(layer, VideoResBlock):
+            x = layer(x, emb, num_video_frames, image_only_indicator)
+        elif isinstance(layer, TimestepBlock):
+            x = layer(x, emb)
+        elif isinstance(layer, SpatialVideoTransformer):
+            x = layer(x, context, time_context, num_video_frames, image_only_indicator, transformer_options)
+            if "transformer_index" in transformer_options:
+                transformer_options["transformer_index"] += 1
+        elif isinstance(layer, SpatialTransformer):
+            x = layer(x, context, transformer_options)
+            if "transformer_index" in transformer_options:
+                transformer_options["transformer_index"] += 1
+        elif isinstance(layer, Upsample):
+            x = layer(x, output_shape=output_shape)
+        else:
+            x = layer(x)
+    return x
+
+class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
+    """
+    A sequential module that passes timestep embeddings to the children that
+    support it as an extra input.
+    """
+
+    def forward(self, *args, **kwargs):
+        return forward_timestep_embed(self, *args, **kwargs)
+
+class Upsample(nn.Module):
+    """
+    An upsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 upsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1, dtype=None, device=None, operations=ops):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        if use_conv:
+            self.conv = operations.conv_nd(dims, self.channels, self.out_channels, 3, padding=padding, dtype=dtype, device=device)
+
+    def forward(self, x, output_shape=None):
+        assert x.shape[1] == self.channels
+        if self.dims == 3:
+            shape = [x.shape[2], x.shape[3] * 2, x.shape[4] * 2]
+            if output_shape is not None:
+                shape[1] = output_shape[3]
+                shape[2] = output_shape[4]
+        else:
+            shape = [x.shape[2] * 2, x.shape[3] * 2]
+            if output_shape is not None:
+                shape[0] = output_shape[2]
+                shape[1] = output_shape[3]
+
+        x = F.interpolate(x, size=shape, mode="nearest")
+        if self.use_conv:
+            x = self.conv(x)
+        return x
+
+class Downsample(nn.Module):
+    """
+    A downsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 downsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1, dtype=None, device=None, operations=ops):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        stride = 2 if dims != 3 else (1, 2, 2)
+        if use_conv:
+            self.op = operations.conv_nd(
+                dims, self.channels, self.out_channels, 3, stride=stride, padding=padding, dtype=dtype, device=device
+            )
+        else:
+            assert self.channels == self.out_channels
+            self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        return self.op(x)
+
+
+class ResBlock(TimestepBlock):
+    """
+    A residual block that can optionally change the number of channels.
+    :param channels: the number of input channels.
+    :param emb_channels: the number of timestep embedding channels.
+    :param dropout: the rate of dropout.
+    :param out_channels: if specified, the number of out channels.
+    :param use_conv: if True and out_channels is specified, use a spatial
+        convolution instead of a smaller 1x1 convolution to change the
+        channels in the skip connection.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param use_checkpoint: if True, use gradient checkpointing on this module.
+    :param up: if True, use this block for upsampling.
+    :param down: if True, use this block for downsampling.
+    """
+
+    def __init__(
+        self,
+        channels,
+        emb_channels,
+        dropout,
+        out_channels=None,
+        use_conv=False,
+        use_scale_shift_norm=False,
+        dims=2,
+        use_checkpoint=False,
+        up=False,
+        down=False,
+        kernel_size=3,
+        exchange_temb_dims=False,
+        skip_t_emb=False,
+        dtype=None,
+        device=None,
+        operations=ops
+    ):
+        super().__init__()
+        self.channels = channels
+        self.emb_channels = emb_channels
+        self.dropout = dropout
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_checkpoint = use_checkpoint
+        self.use_scale_shift_norm = use_scale_shift_norm
+        self.exchange_temb_dims = exchange_temb_dims
+
+        if isinstance(kernel_size, list):
+            padding = [k // 2 for k in kernel_size]
+        else:
+            padding = kernel_size // 2
+
+        self.in_layers = nn.Sequential(
+            operations.GroupNorm(32, channels, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.conv_nd(dims, channels, self.out_channels, kernel_size, padding=padding, dtype=dtype, device=device),
+        )
+
+        self.updown = up or down
+
+        if up:
+            self.h_upd = Upsample(channels, False, dims, dtype=dtype, device=device)
+            self.x_upd = Upsample(channels, False, dims, dtype=dtype, device=device)
+        elif down:
+            self.h_upd = Downsample(channels, False, dims, dtype=dtype, device=device)
+            self.x_upd = Downsample(channels, False, dims, dtype=dtype, device=device)
+        else:
+            self.h_upd = self.x_upd = nn.Identity()
+
+        self.skip_t_emb = skip_t_emb
+        if self.skip_t_emb:
+            self.emb_layers = None
+            self.exchange_temb_dims = False
+        else:
+            self.emb_layers = nn.Sequential(
+                nn.SiLU(),
+                operations.Linear(
+                    emb_channels,
+                    2 * self.out_channels if use_scale_shift_norm else self.out_channels, dtype=dtype, device=device
+                ),
+            )
+        self.out_layers = nn.Sequential(
+            operations.GroupNorm(32, self.out_channels, dtype=dtype, device=device),
+            nn.SiLU(),
+            nn.Dropout(p=dropout),
+            operations.conv_nd(dims, self.out_channels, self.out_channels, kernel_size, padding=padding, dtype=dtype, device=device)
+            ,
+        )
+
+        if self.out_channels == channels:
+            self.skip_connection = nn.Identity()
+        elif use_conv:
+            self.skip_connection = operations.conv_nd(
+                dims, channels, self.out_channels, kernel_size, padding=padding, dtype=dtype, device=device
+            )
+        else:
+            self.skip_connection = operations.conv_nd(dims, channels, self.out_channels, 1, dtype=dtype, device=device)
+
+    def forward(self, x, emb):
+        """
+        Apply the block to a Tensor, conditioned on a timestep embedding.
+        :param x: an [N x C x ...] Tensor of features.
+        :param emb: an [N x emb_channels] Tensor of timestep embeddings.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        return checkpoint(
+            self._forward, (x, emb), self.parameters(), self.use_checkpoint
+        )
+
+
+    def _forward(self, x, emb):
+        if self.updown:
+            in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
+            h = in_rest(x)
+            h = self.h_upd(h)
+            x = self.x_upd(x)
+            h = in_conv(h)
+        else:
+            h = self.in_layers(x)
+
+        emb_out = None
+        if not self.skip_t_emb:
+            emb_out = self.emb_layers(emb).type(h.dtype)
+            while len(emb_out.shape) < len(h.shape):
+                emb_out = emb_out[..., None]
+        if self.use_scale_shift_norm:
+            out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
+            h = out_norm(h)
+            if emb_out is not None:
+                scale, shift = th.chunk(emb_out, 2, dim=1)
+                h *= (1 + scale)
+                h += shift
+            h = out_rest(h)
+        else:
+            if emb_out is not None:
+                if self.exchange_temb_dims:
+                    emb_out = rearrange(emb_out, "b t c ... -> b c t ...")
+                h = h + emb_out
+            h = self.out_layers(h)
+        return self.skip_connection(x) + h
+
+
+class VideoResBlock(ResBlock):
+    def __init__(
+        self,
+        channels: int,
+        emb_channels: int,
+        dropout: float,
+        video_kernel_size=3,
+        merge_strategy: str = "fixed",
+        merge_factor: float = 0.5,
+        out_channels=None,
+        use_conv: bool = False,
+        use_scale_shift_norm: bool = False,
+        dims: int = 2,
+        use_checkpoint: bool = False,
+        up: bool = False,
+        down: bool = False,
+        dtype=None,
+        device=None,
+        operations=ops
+    ):
+        super().__init__(
+            channels,
+            emb_channels,
+            dropout,
+            out_channels=out_channels,
+            use_conv=use_conv,
+            use_scale_shift_norm=use_scale_shift_norm,
+            dims=dims,
+            use_checkpoint=use_checkpoint,
+            up=up,
+            down=down,
+            dtype=dtype,
+            device=device,
+            operations=operations
+        )
+
+        self.time_stack = ResBlock(
+            default(out_channels, channels),
+            emb_channels,
+            dropout=dropout,
+            dims=3,
+            out_channels=default(out_channels, channels),
+            use_scale_shift_norm=False,
+            use_conv=False,
+            up=False,
+            down=False,
+            kernel_size=video_kernel_size,
+            use_checkpoint=use_checkpoint,
+            exchange_temb_dims=True,
+            dtype=dtype,
+            device=device,
+            operations=operations
+        )
+        self.time_mixer = AlphaBlender(
+            alpha=merge_factor,
+            merge_strategy=merge_strategy,
+            rearrange_pattern="b t -> b 1 t 1 1",
+        )
+
+    def forward(
+        self,
+        x: th.Tensor,
+        emb: th.Tensor,
+        num_video_frames: int,
+        image_only_indicator = None,
+    ) -> th.Tensor:
+        x = super().forward(x, emb)
+
+        x_mix = rearrange(x, "(b t) c h w -> b c t h w", t=num_video_frames)
+        x = rearrange(x, "(b t) c h w -> b c t h w", t=num_video_frames)
+
+        x = self.time_stack(
+            x, rearrange(emb, "(b t) ... -> b t ...", t=num_video_frames)
+        )
+        x = self.time_mixer(
+            x_spatial=x_mix, x_temporal=x, image_only_indicator=image_only_indicator
+        )
+        x = rearrange(x, "b c t h w -> (b t) c h w")
+        return x
+
+
+class Timestep(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.dim = dim
+
+    def forward(self, t):
+        return timestep_embedding(t, self.dim)
+
+def apply_control(h, control, name):
+    if control is not None and name in control and len(control[name]) > 0:
+        ctrl = control[name].pop()
+        if ctrl is not None:
+            try:
+                h += ctrl
+            except:
+                print("warning control could not be applied", h.shape, ctrl.shape)
+    return h
+
+class UNetModel(nn.Module):
+    """
+    The full UNet model with attention and timestep embedding.
+    :param in_channels: channels in the input Tensor.
+    :param model_channels: base channel count for the model.
+    :param out_channels: channels in the output Tensor.
+    :param num_res_blocks: number of residual blocks per downsample.
+    :param dropout: the dropout probability.
+    :param channel_mult: channel multiplier for each level of the UNet.
+    :param conv_resample: if True, use learned convolutions for upsampling and
+        downsampling.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param num_classes: if specified (as an int), then this model will be
+        class-conditional with `num_classes` classes.
+    :param use_checkpoint: use gradient checkpointing to reduce memory usage.
+    :param num_heads: the number of attention heads in each attention layer.
+    :param num_heads_channels: if specified, ignore num_heads and instead use
+                               a fixed channel width per attention head.
+    :param num_heads_upsample: works with num_heads to set a different number
+                               of heads for upsampling. Deprecated.
+    :param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
+    :param resblock_updown: use residual blocks for up/downsampling.
+    :param use_new_attention_order: use a different attention pattern for potentially
+                                    increased efficiency.
+    """
+
+    def __init__(
+        self,
+        image_size,
+        in_channels,
+        model_channels,
+        out_channels,
+        num_res_blocks,
+        dropout=0,
+        channel_mult=(1, 2, 4, 8),
+        conv_resample=True,
+        dims=2,
+        num_classes=None,
+        use_checkpoint=False,
+        dtype=th.float32,
+        num_heads=-1,
+        num_head_channels=-1,
+        num_heads_upsample=-1,
+        use_scale_shift_norm=False,
+        resblock_updown=False,
+        use_new_attention_order=False,
+        use_spatial_transformer=False,    # custom transformer support
+        transformer_depth=1,              # custom transformer support
+        context_dim=None,                 # custom transformer support
+        n_embed=None,                     # custom support for prediction of discrete ids into codebook of first stage vq model
+        legacy=True,
+        disable_self_attentions=None,
+        num_attention_blocks=None,
+        disable_middle_self_attn=False,
+        use_linear_in_transformer=False,
+        adm_in_channels=None,
+        transformer_depth_middle=None,
+        transformer_depth_output=None,
+        use_temporal_resblock=False,
+        use_temporal_attention=False,
+        time_context_dim=None,
+        extra_ff_mix_layer=False,
+        use_spatial_context=False,
+        merge_strategy=None,
+        merge_factor=0.0,
+        video_kernel_size=None,
+        disable_temporal_crossattention=False,
+        max_ddpm_temb_period=10000,
+        device=None,
+        operations=ops,
+    ):
+        super().__init__()
+        assert use_spatial_transformer == True, "use_spatial_transformer has to be true"
+        if use_spatial_transformer:
+            assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
+
+        if context_dim is not None:
+            assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
+            # from omegaconf.listconfig import ListConfig
+            # if type(context_dim) == ListConfig:
+            #     context_dim = list(context_dim)
+
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+
+        if num_heads == -1:
+            assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
+
+        if num_head_channels == -1:
+            assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
+
+        self.image_size = image_size
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+
+        if isinstance(num_res_blocks, int):
+            self.num_res_blocks = len(channel_mult) * [num_res_blocks]
+        else:
+            if len(num_res_blocks) != len(channel_mult):
+                raise ValueError("provide num_res_blocks either as an int (globally constant) or "
+                                 "as a list/tuple (per-level) with the same length as channel_mult")
+            self.num_res_blocks = num_res_blocks
+
+        if disable_self_attentions is not None:
+            # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not
+            assert len(disable_self_attentions) == len(channel_mult)
+        if num_attention_blocks is not None:
+            assert len(num_attention_blocks) == len(self.num_res_blocks)
+
+        transformer_depth = transformer_depth[:]
+        transformer_depth_output = transformer_depth_output[:]
+
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.num_classes = num_classes
+        self.use_checkpoint = use_checkpoint
+        self.dtype = dtype
+        self.num_heads = num_heads
+        self.num_head_channels = num_head_channels
+        self.num_heads_upsample = num_heads_upsample
+        self.use_temporal_resblocks = use_temporal_resblock
+        self.predict_codebook_ids = n_embed is not None
+
+        self.default_num_video_frames = None
+        self.default_image_only_indicator = None
+
+        time_embed_dim = model_channels * 4
+        self.time_embed = nn.Sequential(
+            operations.Linear(model_channels, time_embed_dim, dtype=self.dtype, device=device),
+            nn.SiLU(),
+            operations.Linear(time_embed_dim, time_embed_dim, dtype=self.dtype, device=device),
+        )
+
+        if self.num_classes is not None:
+            if isinstance(self.num_classes, int):
+                self.label_emb = nn.Embedding(num_classes, time_embed_dim)
+            elif self.num_classes == "continuous":
+                print("setting up linear c_adm embedding layer")
+                self.label_emb = nn.Linear(1, time_embed_dim)
+            elif self.num_classes == "sequential":
+                assert adm_in_channels is not None
+                self.label_emb = nn.Sequential(
+                    nn.Sequential(
+                        operations.Linear(adm_in_channels, time_embed_dim, dtype=self.dtype, device=device),
+                        nn.SiLU(),
+                        operations.Linear(time_embed_dim, time_embed_dim, dtype=self.dtype, device=device),
+                    )
+                )
+            else:
+                raise ValueError()
+
+        self.input_blocks = nn.ModuleList(
+            [
+                TimestepEmbedSequential(
+                    operations.conv_nd(dims, in_channels, model_channels, 3, padding=1, dtype=self.dtype, device=device)
+                )
+            ]
+        )
+        self._feature_size = model_channels
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+
+        def get_attention_layer(
+            ch,
+            num_heads,
+            dim_head,
+            depth=1,
+            context_dim=None,
+            use_checkpoint=False,
+            disable_self_attn=False,
+        ):
+            if use_temporal_attention:
+                return SpatialVideoTransformer(
+                    ch,
+                    num_heads,
+                    dim_head,
+                    depth=depth,
+                    context_dim=context_dim,
+                    time_context_dim=time_context_dim,
+                    dropout=dropout,
+                    ff_in=extra_ff_mix_layer,
+                    use_spatial_context=use_spatial_context,
+                    merge_strategy=merge_strategy,
+                    merge_factor=merge_factor,
+                    checkpoint=use_checkpoint,
+                    use_linear=use_linear_in_transformer,
+                    disable_self_attn=disable_self_attn,
+                    disable_temporal_crossattention=disable_temporal_crossattention,
+                    max_time_embed_period=max_ddpm_temb_period,
+                    dtype=self.dtype, device=device, operations=operations
+                )
+            else:
+                return SpatialTransformer(
+                                ch, num_heads, dim_head, depth=depth, context_dim=context_dim,
+                                disable_self_attn=disable_self_attn, use_linear=use_linear_in_transformer,
+                                use_checkpoint=use_checkpoint, dtype=self.dtype, device=device, operations=operations
+                            )
+
+        def get_resblock(
+            merge_factor,
+            merge_strategy,
+            video_kernel_size,
+            ch,
+            time_embed_dim,
+            dropout,
+            out_channels,
+            dims,
+            use_checkpoint,
+            use_scale_shift_norm,
+            down=False,
+            up=False,
+            dtype=None,
+            device=None,
+            operations=ops
+        ):
+            if self.use_temporal_resblocks:
+                return VideoResBlock(
+                    merge_factor=merge_factor,
+                    merge_strategy=merge_strategy,
+                    video_kernel_size=video_kernel_size,
+                    channels=ch,
+                    emb_channels=time_embed_dim,
+                    dropout=dropout,
+                    out_channels=out_channels,
+                    dims=dims,
+                    use_checkpoint=use_checkpoint,
+                    use_scale_shift_norm=use_scale_shift_norm,
+                    down=down,
+                    up=up,
+                    dtype=dtype,
+                    device=device,
+                    operations=operations
+                )
+            else:
+                return ResBlock(
+                    channels=ch,
+                    emb_channels=time_embed_dim,
+                    dropout=dropout,
+                    out_channels=out_channels,
+                    use_checkpoint=use_checkpoint,
+                    dims=dims,
+                    use_scale_shift_norm=use_scale_shift_norm,
+                    down=down,
+                    up=up,
+                    dtype=dtype,
+                    device=device,
+                    operations=operations
+                )
+
+        for level, mult in enumerate(channel_mult):
+            for nr in range(self.num_res_blocks[level]):
+                layers = [
+                    get_resblock(
+                        merge_factor=merge_factor,
+                        merge_strategy=merge_strategy,
+                        video_kernel_size=video_kernel_size,
+                        ch=ch,
+                        time_embed_dim=time_embed_dim,
+                        dropout=dropout,
+                        out_channels=mult * model_channels,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                        dtype=self.dtype,
+                        device=device,
+                        operations=operations,
+                    )
+                ]
+                ch = mult * model_channels
+                num_transformers = transformer_depth.pop(0)
+                if num_transformers > 0:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        #num_heads = 1
+                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    if exists(disable_self_attentions):
+                        disabled_sa = disable_self_attentions[level]
+                    else:
+                        disabled_sa = False
+
+                    if not exists(num_attention_blocks) or nr < num_attention_blocks[level]:
+                        layers.append(get_attention_layer(
+                                ch, num_heads, dim_head, depth=num_transformers, context_dim=context_dim,
+                                disable_self_attn=disabled_sa, use_checkpoint=use_checkpoint)
+                        )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        get_resblock(
+                            merge_factor=merge_factor,
+                            merge_strategy=merge_strategy,
+                            video_kernel_size=video_kernel_size,
+                            ch=ch,
+                            time_embed_dim=time_embed_dim,
+                            dropout=dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True,
+                            dtype=self.dtype,
+                            device=device,
+                            operations=operations
+                        )
+                        if resblock_updown
+                        else Downsample(
+                            ch, conv_resample, dims=dims, out_channels=out_ch, dtype=self.dtype, device=device, operations=operations
+                        )
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+                self._feature_size += ch
+
+        if num_head_channels == -1:
+            dim_head = ch // num_heads
+        else:
+            num_heads = ch // num_head_channels
+            dim_head = num_head_channels
+        if legacy:
+            #num_heads = 1
+            dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+        mid_block = [
+            get_resblock(
+                merge_factor=merge_factor,
+                merge_strategy=merge_strategy,
+                video_kernel_size=video_kernel_size,
+                ch=ch,
+                time_embed_dim=time_embed_dim,
+                dropout=dropout,
+                out_channels=None,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+                dtype=self.dtype,
+                device=device,
+                operations=operations
+            )]
+        if transformer_depth_middle >= 0:
+            mid_block += [get_attention_layer(  # always uses a self-attn
+                            ch, num_heads, dim_head, depth=transformer_depth_middle, context_dim=context_dim,
+                            disable_self_attn=disable_middle_self_attn, use_checkpoint=use_checkpoint
+                        ),
+            get_resblock(
+                merge_factor=merge_factor,
+                merge_strategy=merge_strategy,
+                video_kernel_size=video_kernel_size,
+                ch=ch,
+                time_embed_dim=time_embed_dim,
+                dropout=dropout,
+                out_channels=None,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+                dtype=self.dtype,
+                device=device,
+                operations=operations
+            )]
+        self.middle_block = TimestepEmbedSequential(*mid_block)
+        self._feature_size += ch
+
+        self.output_blocks = nn.ModuleList([])
+        for level, mult in list(enumerate(channel_mult))[::-1]:
+            for i in range(self.num_res_blocks[level] + 1):
+                ich = input_block_chans.pop()
+                layers = [
+                    get_resblock(
+                        merge_factor=merge_factor,
+                        merge_strategy=merge_strategy,
+                        video_kernel_size=video_kernel_size,
+                        ch=ch + ich,
+                        time_embed_dim=time_embed_dim,
+                        dropout=dropout,
+                        out_channels=model_channels * mult,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                        dtype=self.dtype,
+                        device=device,
+                        operations=operations
+                    )
+                ]
+                ch = model_channels * mult
+                num_transformers = transformer_depth_output.pop()
+                if num_transformers > 0:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        #num_heads = 1
+                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    if exists(disable_self_attentions):
+                        disabled_sa = disable_self_attentions[level]
+                    else:
+                        disabled_sa = False
+
+                    if not exists(num_attention_blocks) or i < num_attention_blocks[level]:
+                        layers.append(
+                            get_attention_layer(
+                                ch, num_heads, dim_head, depth=num_transformers, context_dim=context_dim,
+                                disable_self_attn=disabled_sa, use_checkpoint=use_checkpoint
+                            )
+                        )
+                if level and i == self.num_res_blocks[level]:
+                    out_ch = ch
+                    layers.append(
+                        get_resblock(
+                            merge_factor=merge_factor,
+                            merge_strategy=merge_strategy,
+                            video_kernel_size=video_kernel_size,
+                            ch=ch,
+                            time_embed_dim=time_embed_dim,
+                            dropout=dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            up=True,
+                            dtype=self.dtype,
+                            device=device,
+                            operations=operations
+                        )
+                        if resblock_updown
+                        else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch, dtype=self.dtype, device=device, operations=operations)
+                    )
+                    ds //= 2
+                self.output_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+
+        self.out = nn.Sequential(
+            operations.GroupNorm(32, ch, dtype=self.dtype, device=device),
+            nn.SiLU(),
+            zero_module(operations.conv_nd(dims, model_channels, out_channels, 3, padding=1, dtype=self.dtype, device=device)),
+        )
+        if self.predict_codebook_ids:
+            self.id_predictor = nn.Sequential(
+            operations.GroupNorm(32, ch, dtype=self.dtype, device=device),
+            operations.conv_nd(dims, model_channels, n_embed, 1, dtype=self.dtype, device=device),
+            #nn.LogSoftmax(dim=1)  # change to cross_entropy and produce non-normalized logits
+        )
+
+    def forward(self, x, timesteps=None, context=None, y=None, control=None, transformer_options={}, **kwargs):
+        """
+        Apply the model to an input batch.
+        :param x: an [N x C x ...] Tensor of inputs.
+        :param timesteps: a 1-D batch of timesteps.
+        :param context: conditioning plugged in via crossattn
+        :param y: an [N] Tensor of labels, if class-conditional.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        transformer_options["original_shape"] = list(x.shape)
+        transformer_options["transformer_index"] = 0
+        transformer_patches = transformer_options.get("patches", {})
+
+        num_video_frames = kwargs.get("num_video_frames", self.default_num_video_frames)
+        image_only_indicator = kwargs.get("image_only_indicator", self.default_image_only_indicator)
+        time_context = kwargs.get("time_context", None)
+
+        assert (y is not None) == (
+            self.num_classes is not None
+        ), "must specify y if and only if the model is class-conditional"
+        hs = []
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).to(x.dtype)
+        emb = self.time_embed(t_emb)
+
+        if self.num_classes is not None:
+            assert y.shape[0] == x.shape[0]
+            emb = emb + self.label_emb(y)
+
+        h = x
+        for id, module in enumerate(self.input_blocks):
+            transformer_options["block"] = ("input", id)
+            h = forward_timestep_embed(module, h, emb, context, transformer_options, time_context=time_context, num_video_frames=num_video_frames, image_only_indicator=image_only_indicator)
+            h = apply_control(h, control, 'input')
+            if "input_block_patch" in transformer_patches:
+                patch = transformer_patches["input_block_patch"]
+                for p in patch:
+                    h = p(h, transformer_options)
+
+            hs.append(h)
+            if "input_block_patch_after_skip" in transformer_patches:
+                patch = transformer_patches["input_block_patch_after_skip"]
+                for p in patch:
+                    h = p(h, transformer_options)
+
+        transformer_options["block"] = ("middle", 0)
+        h = forward_timestep_embed(self.middle_block, h, emb, context, transformer_options, time_context=time_context, num_video_frames=num_video_frames, image_only_indicator=image_only_indicator)
+        h = apply_control(h, control, 'middle')
+
+
+        for id, module in enumerate(self.output_blocks):
+            transformer_options["block"] = ("output", id)
+            hsp = hs.pop()
+            hsp = apply_control(hsp, control, 'output')
+
+            if "output_block_patch" in transformer_patches:
+                patch = transformer_patches["output_block_patch"]
+                for p in patch:
+                    h, hsp = p(h, hsp, transformer_options)
+
+            h = th.cat([h, hsp], dim=1)
+            del hsp
+            if len(hs) > 0:
+                output_shape = hs[-1].shape
+            else:
+                output_shape = None
+            h = forward_timestep_embed(module, h, emb, context, transformer_options, output_shape, time_context=time_context, num_video_frames=num_video_frames, image_only_indicator=image_only_indicator)
+        h = h.type(x.dtype)
+        if self.predict_codebook_ids:
+            return self.id_predictor(h)
+        else:
+            return self.out(h)

ldm_patched/ldm/modules/diffusionmodules/upscaling.py

@@ -0,0 +1,81 @@
+import torch
+import torch.nn as nn
+import numpy as np
+from functools import partial
+
+from .util import extract_into_tensor, make_beta_schedule
+from ldm_patched.ldm.util import default
+
+
+class AbstractLowScaleModel(nn.Module):
+    # for concatenating a downsampled image to the latent representation
+    def __init__(self, noise_schedule_config=None):
+        super(AbstractLowScaleModel, self).__init__()
+        if noise_schedule_config is not None:
+            self.register_schedule(**noise_schedule_config)
+
+    def register_schedule(self, beta_schedule="linear", timesteps=1000,
+                          linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+        betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end,
+                                   cosine_s=cosine_s)
+        alphas = 1. - betas
+        alphas_cumprod = np.cumprod(alphas, axis=0)
+        alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
+
+        timesteps, = betas.shape
+        self.num_timesteps = int(timesteps)
+        self.linear_start = linear_start
+        self.linear_end = linear_end
+        assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep'
+
+        to_torch = partial(torch.tensor, dtype=torch.float32)
+
+        self.register_buffer('betas', to_torch(betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
+        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
+
+    def q_sample(self, x_start, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
+                extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
+
+    def forward(self, x):
+        return x, None
+
+    def decode(self, x):
+        return x
+
+
+class SimpleImageConcat(AbstractLowScaleModel):
+    # no noise level conditioning
+    def __init__(self):
+        super(SimpleImageConcat, self).__init__(noise_schedule_config=None)
+        self.max_noise_level = 0
+
+    def forward(self, x):
+        # fix to constant noise level
+        return x, torch.zeros(x.shape[0], device=x.device).long()
+
+
+class ImageConcatWithNoiseAugmentation(AbstractLowScaleModel):
+    def __init__(self, noise_schedule_config, max_noise_level=1000, to_cuda=False):
+        super().__init__(noise_schedule_config=noise_schedule_config)
+        self.max_noise_level = max_noise_level
+
+    def forward(self, x, noise_level=None):
+        if noise_level is None:
+            noise_level = torch.randint(0, self.max_noise_level, (x.shape[0],), device=x.device).long()
+        else:
+            assert isinstance(noise_level, torch.Tensor)
+        z = self.q_sample(x, noise_level)
+        return z, noise_level
+
+
+

ldm_patched/ldm/modules/diffusionmodules/util.py

@@ -0,0 +1,304 @@
+# adopted from
+# https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py
+# and
+# https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
+# and
+# https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py
+#
+# thanks!
+
+
+import os
+import math
+import torch
+import torch.nn as nn
+import numpy as np
+from einops import repeat, rearrange
+
+from ldm_patched.ldm.util import instantiate_from_config
+
+class AlphaBlender(nn.Module):
+    strategies = ["learned", "fixed", "learned_with_images"]
+
+    def __init__(
+        self,
+        alpha: float,
+        merge_strategy: str = "learned_with_images",
+        rearrange_pattern: str = "b t -> (b t) 1 1",
+    ):
+        super().__init__()
+        self.merge_strategy = merge_strategy
+        self.rearrange_pattern = rearrange_pattern
+
+        assert (
+            merge_strategy in self.strategies
+        ), f"merge_strategy needs to be in {self.strategies}"
+
+        if self.merge_strategy == "fixed":
+            self.register_buffer("mix_factor", torch.Tensor([alpha]))
+        elif (
+            self.merge_strategy == "learned"
+            or self.merge_strategy == "learned_with_images"
+        ):
+            self.register_parameter(
+                "mix_factor", torch.nn.Parameter(torch.Tensor([alpha]))
+            )
+        else:
+            raise ValueError(f"unknown merge strategy {self.merge_strategy}")
+
+    def get_alpha(self, image_only_indicator: torch.Tensor) -> torch.Tensor:
+        # skip_time_mix = rearrange(repeat(skip_time_mix, 'b -> (b t) () () ()', t=t), '(b t) 1 ... -> b 1 t ...', t=t)
+        if self.merge_strategy == "fixed":
+            # make shape compatible
+            # alpha = repeat(self.mix_factor, '1 -> b () t  () ()', t=t, b=bs)
+            alpha = self.mix_factor
+        elif self.merge_strategy == "learned":
+            alpha = torch.sigmoid(self.mix_factor)
+            # make shape compatible
+            # alpha = repeat(alpha, '1 -> s () ()', s = t * bs)
+        elif self.merge_strategy == "learned_with_images":
+            assert image_only_indicator is not None, "need image_only_indicator ..."
+            alpha = torch.where(
+                image_only_indicator.bool(),
+                torch.ones(1, 1, device=image_only_indicator.device),
+                rearrange(torch.sigmoid(self.mix_factor), "... -> ... 1"),
+            )
+            alpha = rearrange(alpha, self.rearrange_pattern)
+            # make shape compatible
+            # alpha = repeat(alpha, '1 -> s () ()', s = t * bs)
+        else:
+            raise NotImplementedError()
+        return alpha
+
+    def forward(
+        self,
+        x_spatial,
+        x_temporal,
+        image_only_indicator=None,
+    ) -> torch.Tensor:
+        alpha = self.get_alpha(image_only_indicator)
+        x = (
+            alpha.to(x_spatial.dtype) * x_spatial
+            + (1.0 - alpha).to(x_spatial.dtype) * x_temporal
+        )
+        return x
+
+
+def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+    if schedule == "linear":
+        betas = (
+                torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2
+        )
+
+    elif schedule == "cosine":
+        timesteps = (
+                torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s
+        )
+        alphas = timesteps / (1 + cosine_s) * np.pi / 2
+        alphas = torch.cos(alphas).pow(2)
+        alphas = alphas / alphas[0]
+        betas = 1 - alphas[1:] / alphas[:-1]
+        betas = np.clip(betas, a_min=0, a_max=0.999)
+
+    elif schedule == "squaredcos_cap_v2":  # used for karlo prior
+        # return early
+        return betas_for_alpha_bar(
+            n_timestep,
+            lambda t: math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2,
+        )
+
+    elif schedule == "sqrt_linear":
+        betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64)
+    elif schedule == "sqrt":
+        betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5
+    else:
+        raise ValueError(f"schedule '{schedule}' unknown.")
+    return betas.numpy()
+
+
+def make_ddim_timesteps(ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True):
+    if ddim_discr_method == 'uniform':
+        c = num_ddpm_timesteps // num_ddim_timesteps
+        ddim_timesteps = np.asarray(list(range(0, num_ddpm_timesteps, c)))
+    elif ddim_discr_method == 'quad':
+        ddim_timesteps = ((np.linspace(0, np.sqrt(num_ddpm_timesteps * .8), num_ddim_timesteps)) ** 2).astype(int)
+    else:
+        raise NotImplementedError(f'There is no ddim discretization method called "{ddim_discr_method}"')
+
+    # assert ddim_timesteps.shape[0] == num_ddim_timesteps
+    # add one to get the final alpha values right (the ones from first scale to data during sampling)
+    steps_out = ddim_timesteps + 1
+    if verbose:
+        print(f'Selected timesteps for ddim sampler: {steps_out}')
+    return steps_out
+
+
+def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True):
+    # select alphas for computing the variance schedule
+    alphas = alphacums[ddim_timesteps]
+    alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist())
+
+    # according the the formula provided in https://arxiv.org/abs/2010.02502
+    sigmas = eta * np.sqrt((1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev))
+    if verbose:
+        print(f'Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}')
+        print(f'For the chosen value of eta, which is {eta}, '
+              f'this results in the following sigma_t schedule for ddim sampler {sigmas}')
+    return sigmas, alphas, alphas_prev
+
+
+def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999):
+    """
+    Create a beta schedule that discretizes the given alpha_t_bar function,
+    which defines the cumulative product of (1-beta) over time from t = [0,1].
+    :param num_diffusion_timesteps: the number of betas to produce.
+    :param alpha_bar: a lambda that takes an argument t from 0 to 1 and
+                      produces the cumulative product of (1-beta) up to that
+                      part of the diffusion process.
+    :param max_beta: the maximum beta to use; use values lower than 1 to
+                     prevent singularities.
+    """
+    betas = []
+    for i in range(num_diffusion_timesteps):
+        t1 = i / num_diffusion_timesteps
+        t2 = (i + 1) / num_diffusion_timesteps
+        betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
+    return np.array(betas)
+
+
+def extract_into_tensor(a, t, x_shape):
+    b, *_ = t.shape
+    out = a.gather(-1, t)
+    return out.reshape(b, *((1,) * (len(x_shape) - 1)))
+
+
+def checkpoint(func, inputs, params, flag):
+    """
+    Evaluate a function without caching intermediate activations, allowing for
+    reduced memory at the expense of extra compute in the backward pass.
+    :param func: the function to evaluate.
+    :param inputs: the argument sequence to pass to `func`.
+    :param params: a sequence of parameters `func` depends on but does not
+                   explicitly take as arguments.
+    :param flag: if False, disable gradient checkpointing.
+    """
+    if flag:
+        args = tuple(inputs) + tuple(params)
+        return CheckpointFunction.apply(func, len(inputs), *args)
+    else:
+        return func(*inputs)
+
+
+class CheckpointFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, run_function, length, *args):
+        ctx.run_function = run_function
+        ctx.input_tensors = list(args[:length])
+        ctx.input_params = list(args[length:])
+        ctx.gpu_autocast_kwargs = {"enabled": torch.is_autocast_enabled(),
+                                   "dtype": torch.get_autocast_gpu_dtype(),
+                                   "cache_enabled": torch.is_autocast_cache_enabled()}
+        with torch.no_grad():
+            output_tensors = ctx.run_function(*ctx.input_tensors)
+        return output_tensors
+
+    @staticmethod
+    def backward(ctx, *output_grads):
+        ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors]
+        with torch.enable_grad(), \
+                torch.cuda.amp.autocast(**ctx.gpu_autocast_kwargs):
+            # Fixes a bug where the first op in run_function modifies the
+            # Tensor storage in place, which is not allowed for detach()'d
+            # Tensors.
+            shallow_copies = [x.view_as(x) for x in ctx.input_tensors]
+            output_tensors = ctx.run_function(*shallow_copies)
+        input_grads = torch.autograd.grad(
+            output_tensors,
+            ctx.input_tensors + ctx.input_params,
+            output_grads,
+            allow_unused=True,
+        )
+        del ctx.input_tensors
+        del ctx.input_params
+        del output_tensors
+        return (None, None) + input_grads
+
+
+def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False):
+    """
+    Create sinusoidal timestep embeddings.
+    :param timesteps: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an [N x dim] Tensor of positional embeddings.
+    """
+    if not repeat_only:
+        half = dim // 2
+        freqs = torch.exp(
+            -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32, device=timesteps.device) / half
+        )
+        args = timesteps[:, None].float() * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    else:
+        embedding = repeat(timesteps, 'b -> b d', d=dim)
+    return embedding
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def scale_module(module, scale):
+    """
+    Scale the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().mul_(scale)
+    return module
+
+
+def mean_flat(tensor):
+    """
+    Take the mean over all non-batch dimensions.
+    """
+    return tensor.mean(dim=list(range(1, len(tensor.shape))))
+
+
+def avg_pool_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D average pooling module.
+    """
+    if dims == 1:
+        return nn.AvgPool1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.AvgPool2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.AvgPool3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+class HybridConditioner(nn.Module):
+
+    def __init__(self, c_concat_config, c_crossattn_config):
+        super().__init__()
+        self.concat_conditioner = instantiate_from_config(c_concat_config)
+        self.crossattn_conditioner = instantiate_from_config(c_crossattn_config)
+
+    def forward(self, c_concat, c_crossattn):
+        c_concat = self.concat_conditioner(c_concat)
+        c_crossattn = self.crossattn_conditioner(c_crossattn)
+        return {'c_concat': [c_concat], 'c_crossattn': [c_crossattn]}
+
+
+def noise_like(shape, device, repeat=False):
+    repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1)))
+    noise = lambda: torch.randn(shape, device=device)
+    return repeat_noise() if repeat else noise()

ldm_patched/ldm/modules/distributions/distributions.py

@@ -0,0 +1,92 @@
+import torch
+import numpy as np
+
+
+class AbstractDistribution:
+    def sample(self):
+        raise NotImplementedError()
+
+    def mode(self):
+        raise NotImplementedError()
+
+
+class DiracDistribution(AbstractDistribution):
+    def __init__(self, value):
+        self.value = value
+
+    def sample(self):
+        return self.value
+
+    def mode(self):
+        return self.value
+
+
+class DiagonalGaussianDistribution(object):
+    def __init__(self, parameters, deterministic=False):
+        self.parameters = parameters
+        self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
+        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
+        self.deterministic = deterministic
+        self.std = torch.exp(0.5 * self.logvar)
+        self.var = torch.exp(self.logvar)
+        if self.deterministic:
+            self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device)
+
+    def sample(self):
+        x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device)
+        return x
+
+    def kl(self, other=None):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        else:
+            if other is None:
+                return 0.5 * torch.sum(torch.pow(self.mean, 2)
+                                       + self.var - 1.0 - self.logvar,
+                                       dim=[1, 2, 3])
+            else:
+                return 0.5 * torch.sum(
+                    torch.pow(self.mean - other.mean, 2) / other.var
+                    + self.var / other.var - 1.0 - self.logvar + other.logvar,
+                    dim=[1, 2, 3])
+
+    def nll(self, sample, dims=[1,2,3]):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        logtwopi = np.log(2.0 * np.pi)
+        return 0.5 * torch.sum(
+            logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
+            dim=dims)
+
+    def mode(self):
+        return self.mean
+
+
+def normal_kl(mean1, logvar1, mean2, logvar2):
+    """
+    source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
+    Compute the KL divergence between two gaussians.
+    Shapes are automatically broadcasted, so batches can be compared to
+    scalars, among other use cases.
+    """
+    tensor = None
+    for obj in (mean1, logvar1, mean2, logvar2):
+        if isinstance(obj, torch.Tensor):
+            tensor = obj
+            break
+    assert tensor is not None, "at least one argument must be a Tensor"
+
+    # Force variances to be Tensors. Broadcasting helps convert scalars to
+    # Tensors, but it does not work for torch.exp().
+    logvar1, logvar2 = [
+        x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor)
+        for x in (logvar1, logvar2)
+    ]
+
+    return 0.5 * (
+        -1.0
+        + logvar2
+        - logvar1
+        + torch.exp(logvar1 - logvar2)
+        + ((mean1 - mean2) ** 2) * torch.exp(-logvar2)
+    )

ldm_patched/ldm/modules/ema.py

@@ -0,0 +1,80 @@
+import torch
+from torch import nn
+
+
+class LitEma(nn.Module):
+    def __init__(self, model, decay=0.9999, use_num_upates=True):
+        super().__init__()
+        if decay < 0.0 or decay > 1.0:
+            raise ValueError('Decay must be between 0 and 1')
+
+        self.m_name2s_name = {}
+        self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32))
+        self.register_buffer('num_updates', torch.tensor(0, dtype=torch.int) if use_num_upates
+        else torch.tensor(-1, dtype=torch.int))
+
+        for name, p in model.named_parameters():
+            if p.requires_grad:
+                # remove as '.'-character is not allowed in buffers
+                s_name = name.replace('.', '')
+                self.m_name2s_name.update({name: s_name})
+                self.register_buffer(s_name, p.clone().detach().data)
+
+        self.collected_params = []
+
+    def reset_num_updates(self):
+        del self.num_updates
+        self.register_buffer('num_updates', torch.tensor(0, dtype=torch.int))
+
+    def forward(self, model):
+        decay = self.decay
+
+        if self.num_updates >= 0:
+            self.num_updates += 1
+            decay = min(self.decay, (1 + self.num_updates) / (10 + self.num_updates))
+
+        one_minus_decay = 1.0 - decay
+
+        with torch.no_grad():
+            m_param = dict(model.named_parameters())
+            shadow_params = dict(self.named_buffers())
+
+            for key in m_param:
+                if m_param[key].requires_grad:
+                    sname = self.m_name2s_name[key]
+                    shadow_params[sname] = shadow_params[sname].type_as(m_param[key])
+                    shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key]))
+                else:
+                    assert not key in self.m_name2s_name
+
+    def copy_to(self, model):
+        m_param = dict(model.named_parameters())
+        shadow_params = dict(self.named_buffers())
+        for key in m_param:
+            if m_param[key].requires_grad:
+                m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data)
+            else:
+                assert not key in self.m_name2s_name
+
+    def store(self, parameters):
+        """
+        Save the current parameters for restoring later.
+        Args:
+          parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+            temporarily stored.
+        """
+        self.collected_params = [param.clone() for param in parameters]
+
+    def restore(self, parameters):
+        """
+        Restore the parameters stored with the `store` method.
+        Useful to validate the model with EMA parameters without affecting the
+        original optimization process. Store the parameters before the
+        `copy_to` method. After validation (or model saving), use this to
+        restore the former parameters.
+        Args:
+          parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+            updated with the stored parameters.
+        """
+        for c_param, param in zip(self.collected_params, parameters):
+            param.data.copy_(c_param.data)

ldm_patched/ldm/modules/encoders/noise_aug_modules.py

@@ -0,0 +1,35 @@
+from ..diffusionmodules.upscaling import ImageConcatWithNoiseAugmentation
+from ..diffusionmodules.openaimodel import Timestep
+import torch
+
+class CLIPEmbeddingNoiseAugmentation(ImageConcatWithNoiseAugmentation):
+    def __init__(self, *args, clip_stats_path=None, timestep_dim=256, **kwargs):
+        super().__init__(*args, **kwargs)
+        if clip_stats_path is None:
+            clip_mean, clip_std = torch.zeros(timestep_dim), torch.ones(timestep_dim)
+        else:
+            clip_mean, clip_std = torch.load(clip_stats_path, map_location="cpu")
+        self.register_buffer("data_mean", clip_mean[None, :], persistent=False)
+        self.register_buffer("data_std", clip_std[None, :], persistent=False)
+        self.time_embed = Timestep(timestep_dim)
+
+    def scale(self, x):
+        # re-normalize to centered mean and unit variance
+        x = (x - self.data_mean) * 1. / self.data_std
+        return x
+
+    def unscale(self, x):
+        # back to original data stats
+        x = (x * self.data_std) + self.data_mean
+        return x
+
+    def forward(self, x, noise_level=None):
+        if noise_level is None:
+            noise_level = torch.randint(0, self.max_noise_level, (x.shape[0],), device=x.device).long()
+        else:
+            assert isinstance(noise_level, torch.Tensor)
+        x = self.scale(x)
+        z = self.q_sample(x, noise_level)
+        z = self.unscale(z)
+        noise_level = self.time_embed(noise_level)
+        return z, noise_level

ldm_patched/ldm/modules/sub_quadratic_attention.py

@@ -0,0 +1,251 @@
+# original source:
+#   https://github.com/AminRezaei0x443/memory-efficient-attention/blob/1bc0d9e6ac5f82ea43a375135c4e1d3896ee1694/memory_efficient_attention/attention_torch.py
+# license:
+#   MIT
+# credit:
+#   Amin Rezaei (original author)
+#   Alex Birch (optimized algorithm for 3D tensors, at the expense of removing bias, masking and callbacks)
+# implementation of:
+#   Self-attention Does Not Need O(n2) Memory":
+#   https://arxiv.org/abs/2112.05682v2
+
+from functools import partial
+import torch
+from torch import Tensor
+from torch.utils.checkpoint import checkpoint
+import math
+
+try:
+	from typing import Optional, NamedTuple, List, Protocol
+except ImportError:
+	from typing import Optional, NamedTuple, List
+	from typing_extensions import Protocol
+
+from torch import Tensor
+from typing import List
+
+from ldm_patched.modules import model_management
+
+def dynamic_slice(
+    x: Tensor,
+    starts: List[int],
+    sizes: List[int],
+) -> Tensor:
+    slicing = [slice(start, start + size) for start, size in zip(starts, sizes)]
+    return x[slicing]
+
+class AttnChunk(NamedTuple):
+    exp_values: Tensor
+    exp_weights_sum: Tensor
+    max_score: Tensor
+
+class SummarizeChunk(Protocol):
+    @staticmethod
+    def __call__(
+        query: Tensor,
+        key_t: Tensor,
+        value: Tensor,
+    ) -> AttnChunk: ...
+
+class ComputeQueryChunkAttn(Protocol):
+    @staticmethod
+    def __call__(
+        query: Tensor,
+        key_t: Tensor,
+        value: Tensor,
+    ) -> Tensor: ...
+
+def _summarize_chunk(
+    query: Tensor,
+    key_t: Tensor,
+    value: Tensor,
+    scale: float,
+    upcast_attention: bool,
+) -> AttnChunk:
+    if upcast_attention:
+        with torch.autocast(enabled=False, device_type = 'cuda'):
+            query = query.float()
+            key_t = key_t.float()
+            attn_weights = torch.baddbmm(
+                torch.empty(1, 1, 1, device=query.device, dtype=query.dtype),
+                query,
+                key_t,
+                alpha=scale,
+                beta=0,
+            )
+    else:
+        attn_weights = torch.baddbmm(
+            torch.empty(1, 1, 1, device=query.device, dtype=query.dtype),
+            query,
+            key_t,
+            alpha=scale,
+            beta=0,
+        )
+    max_score, _ = torch.max(attn_weights, -1, keepdim=True)
+    max_score = max_score.detach()
+    attn_weights -= max_score
+    torch.exp(attn_weights, out=attn_weights)
+    exp_weights = attn_weights.to(value.dtype)
+    exp_values = torch.bmm(exp_weights, value)
+    max_score = max_score.squeeze(-1)
+    return AttnChunk(exp_values, exp_weights.sum(dim=-1), max_score)
+
+def _query_chunk_attention(
+    query: Tensor,
+    key_t: Tensor,
+    value: Tensor,
+    summarize_chunk: SummarizeChunk,
+    kv_chunk_size: int,
+) -> Tensor:
+    batch_x_heads, k_channels_per_head, k_tokens = key_t.shape
+    _, _, v_channels_per_head = value.shape
+
+    def chunk_scanner(chunk_idx: int) -> AttnChunk:
+        key_chunk = dynamic_slice(
+            key_t,
+            (0, 0, chunk_idx),
+            (batch_x_heads, k_channels_per_head, kv_chunk_size)
+        )
+        value_chunk = dynamic_slice(
+            value,
+            (0, chunk_idx, 0),
+            (batch_x_heads, kv_chunk_size, v_channels_per_head)
+        )
+        return summarize_chunk(query, key_chunk, value_chunk)
+
+    chunks: List[AttnChunk] = [
+        chunk_scanner(chunk) for chunk in torch.arange(0, k_tokens, kv_chunk_size)
+    ]
+    acc_chunk = AttnChunk(*map(torch.stack, zip(*chunks)))
+    chunk_values, chunk_weights, chunk_max = acc_chunk
+
+    global_max, _ = torch.max(chunk_max, 0, keepdim=True)
+    max_diffs = torch.exp(chunk_max - global_max)
+    chunk_values *= torch.unsqueeze(max_diffs, -1)
+    chunk_weights *= max_diffs
+
+    all_values = chunk_values.sum(dim=0)
+    all_weights = torch.unsqueeze(chunk_weights, -1).sum(dim=0)
+    return all_values / all_weights
+
+# TODO: refactor CrossAttention#get_attention_scores to share code with this
+def _get_attention_scores_no_kv_chunking(
+    query: Tensor,
+    key_t: Tensor,
+    value: Tensor,
+    scale: float,
+    upcast_attention: bool,
+) -> Tensor:
+    if upcast_attention:
+        with torch.autocast(enabled=False, device_type = 'cuda'):
+            query = query.float()
+            key_t = key_t.float()
+            attn_scores = torch.baddbmm(
+                torch.empty(1, 1, 1, device=query.device, dtype=query.dtype),
+                query,
+                key_t,
+                alpha=scale,
+                beta=0,
+            )
+    else:
+        attn_scores = torch.baddbmm(
+            torch.empty(1, 1, 1, device=query.device, dtype=query.dtype),
+            query,
+            key_t,
+            alpha=scale,
+            beta=0,
+        )
+
+    try:
+        attn_probs = attn_scores.softmax(dim=-1)
+        del attn_scores
+    except model_management.OOM_EXCEPTION:
+        print("ran out of memory while running softmax in  _get_attention_scores_no_kv_chunking, trying slower in place softmax instead")
+        attn_scores -= attn_scores.max(dim=-1, keepdim=True).values
+        torch.exp(attn_scores, out=attn_scores)
+        summed = torch.sum(attn_scores, dim=-1, keepdim=True)
+        attn_scores /= summed
+        attn_probs = attn_scores
+
+    hidden_states_slice = torch.bmm(attn_probs.to(value.dtype), value)
+    return hidden_states_slice
+
+class ScannedChunk(NamedTuple):
+    chunk_idx: int
+    attn_chunk: AttnChunk
+
+def efficient_dot_product_attention(
+    query: Tensor,
+    key_t: Tensor,
+    value: Tensor,
+    query_chunk_size=1024,
+    kv_chunk_size: Optional[int] = None,
+    kv_chunk_size_min: Optional[int] = None,
+    use_checkpoint=True,
+    upcast_attention=False,
+):
+    """Computes efficient dot-product attention given query, transposed key, and value.
+      This is efficient version of attention presented in
+      https://arxiv.org/abs/2112.05682v2 which comes with O(sqrt(n)) memory requirements.
+      Args:
+        query: queries for calculating attention with shape of
+          `[batch * num_heads, tokens, channels_per_head]`.
+        key_t: keys for calculating attention with shape of
+          `[batch * num_heads, channels_per_head, tokens]`.
+        value: values to be used in attention with shape of
+          `[batch * num_heads, tokens, channels_per_head]`.
+        query_chunk_size: int: query chunks size
+        kv_chunk_size: Optional[int]: key/value chunks size. if None: defaults to sqrt(key_tokens)
+        kv_chunk_size_min: Optional[int]: key/value minimum chunk size. only considered when kv_chunk_size is None. changes `sqrt(key_tokens)` into `max(sqrt(key_tokens), kv_chunk_size_min)`, to ensure our chunk sizes don't get too small (smaller chunks = more chunks = less concurrent work done).
+        use_checkpoint: bool: whether to use checkpointing (recommended True for training, False for inference)
+      Returns:
+        Output of shape `[batch * num_heads, query_tokens, channels_per_head]`.
+      """
+    batch_x_heads, q_tokens, q_channels_per_head = query.shape
+    _, _, k_tokens = key_t.shape
+    scale = q_channels_per_head ** -0.5
+
+    kv_chunk_size = min(kv_chunk_size or int(math.sqrt(k_tokens)), k_tokens)
+    if kv_chunk_size_min is not None:
+        kv_chunk_size = max(kv_chunk_size, kv_chunk_size_min)
+
+    def get_query_chunk(chunk_idx: int) -> Tensor:
+        return dynamic_slice(
+            query,
+            (0, chunk_idx, 0),
+            (batch_x_heads, min(query_chunk_size, q_tokens), q_channels_per_head)
+        )
+    
+    summarize_chunk: SummarizeChunk = partial(_summarize_chunk, scale=scale, upcast_attention=upcast_attention)
+    summarize_chunk: SummarizeChunk = partial(checkpoint, summarize_chunk) if use_checkpoint else summarize_chunk
+    compute_query_chunk_attn: ComputeQueryChunkAttn = partial(
+        _get_attention_scores_no_kv_chunking,
+        scale=scale,
+        upcast_attention=upcast_attention
+    ) if k_tokens <= kv_chunk_size else (
+        # fast-path for when there's just 1 key-value chunk per query chunk (this is just sliced attention btw)
+        partial(
+            _query_chunk_attention,
+            kv_chunk_size=kv_chunk_size,
+            summarize_chunk=summarize_chunk,
+        )
+    )
+
+    if q_tokens <= query_chunk_size:
+        # fast-path for when there's just 1 query chunk
+        return compute_query_chunk_attn(
+            query=query,
+            key_t=key_t,
+            value=value,
+        )
+    
+    # TODO: maybe we should use torch.empty_like(query) to allocate storage in-advance,
+    # and pass slices to be mutated, instead of torch.cat()ing the returned slices
+    res = torch.cat([
+        compute_query_chunk_attn(
+            query=get_query_chunk(i * query_chunk_size),
+            key_t=key_t,
+            value=value,
+        ) for i in range(math.ceil(q_tokens / query_chunk_size))
+    ], dim=1)
+    return res

ldm_patched/ldm/modules/temporal_ae.py

@@ -0,0 +1,245 @@
+import functools
+from typing import Callable, Iterable, Union
+
+import torch
+from einops import rearrange, repeat
+
+import ldm_patched.modules.ops
+ops = ldm_patched.modules.ops.disable_weight_init
+
+from .diffusionmodules.model import (
+    AttnBlock,
+    Decoder,
+    ResnetBlock,
+)
+from .diffusionmodules.openaimodel import ResBlock, timestep_embedding
+from .attention import BasicTransformerBlock
+
+def partialclass(cls, *args, **kwargs):
+    class NewCls(cls):
+        __init__ = functools.partialmethod(cls.__init__, *args, **kwargs)
+
+    return NewCls
+
+
+class VideoResBlock(ResnetBlock):
+    def __init__(
+        self,
+        out_channels,
+        *args,
+        dropout=0.0,
+        video_kernel_size=3,
+        alpha=0.0,
+        merge_strategy="learned",
+        **kwargs,
+    ):
+        super().__init__(out_channels=out_channels, dropout=dropout, *args, **kwargs)
+        if video_kernel_size is None:
+            video_kernel_size = [3, 1, 1]
+        self.time_stack = ResBlock(
+            channels=out_channels,
+            emb_channels=0,
+            dropout=dropout,
+            dims=3,
+            use_scale_shift_norm=False,
+            use_conv=False,
+            up=False,
+            down=False,
+            kernel_size=video_kernel_size,
+            use_checkpoint=False,
+            skip_t_emb=True,
+        )
+
+        self.merge_strategy = merge_strategy
+        if self.merge_strategy == "fixed":
+            self.register_buffer("mix_factor", torch.Tensor([alpha]))
+        elif self.merge_strategy == "learned":
+            self.register_parameter(
+                "mix_factor", torch.nn.Parameter(torch.Tensor([alpha]))
+            )
+        else:
+            raise ValueError(f"unknown merge strategy {self.merge_strategy}")
+
+    def get_alpha(self, bs):
+        if self.merge_strategy == "fixed":
+            return self.mix_factor
+        elif self.merge_strategy == "learned":
+            return torch.sigmoid(self.mix_factor)
+        else:
+            raise NotImplementedError()
+
+    def forward(self, x, temb, skip_video=False, timesteps=None):
+        b, c, h, w = x.shape
+        if timesteps is None:
+            timesteps = b
+
+        x = super().forward(x, temb)
+
+        if not skip_video:
+            x_mix = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps)
+
+            x = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps)
+
+            x = self.time_stack(x, temb)
+
+            alpha = self.get_alpha(bs=b // timesteps)
+            x = alpha * x + (1.0 - alpha) * x_mix
+
+            x = rearrange(x, "b c t h w -> (b t) c h w")
+        return x
+
+
+class AE3DConv(torch.nn.Conv2d):
+    def __init__(self, in_channels, out_channels, video_kernel_size=3, *args, **kwargs):
+        super().__init__(in_channels, out_channels, *args, **kwargs)
+        if isinstance(video_kernel_size, Iterable):
+            padding = [int(k // 2) for k in video_kernel_size]
+        else:
+            padding = int(video_kernel_size // 2)
+
+        self.time_mix_conv = torch.nn.Conv3d(
+            in_channels=out_channels,
+            out_channels=out_channels,
+            kernel_size=video_kernel_size,
+            padding=padding,
+        )
+
+    def forward(self, input, timesteps=None, skip_video=False):
+        if timesteps is None:
+            timesteps = input.shape[0]
+        x = super().forward(input)
+        if skip_video:
+            return x
+        x = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps)
+        x = self.time_mix_conv(x)
+        return rearrange(x, "b c t h w -> (b t) c h w")
+
+
+class AttnVideoBlock(AttnBlock):
+    def __init__(
+        self, in_channels: int, alpha: float = 0, merge_strategy: str = "learned"
+    ):
+        super().__init__(in_channels)
+        # no context, single headed, as in base class
+        self.time_mix_block = BasicTransformerBlock(
+            dim=in_channels,
+            n_heads=1,
+            d_head=in_channels,
+            checkpoint=False,
+            ff_in=True,
+        )
+
+        time_embed_dim = self.in_channels * 4
+        self.video_time_embed = torch.nn.Sequential(
+            ops.Linear(self.in_channels, time_embed_dim),
+            torch.nn.SiLU(),
+            ops.Linear(time_embed_dim, self.in_channels),
+        )
+
+        self.merge_strategy = merge_strategy
+        if self.merge_strategy == "fixed":
+            self.register_buffer("mix_factor", torch.Tensor([alpha]))
+        elif self.merge_strategy == "learned":
+            self.register_parameter(
+                "mix_factor", torch.nn.Parameter(torch.Tensor([alpha]))
+            )
+        else:
+            raise ValueError(f"unknown merge strategy {self.merge_strategy}")
+
+    def forward(self, x, timesteps=None, skip_time_block=False):
+        if skip_time_block:
+            return super().forward(x)
+
+        if timesteps is None:
+            timesteps = x.shape[0]
+
+        x_in = x
+        x = self.attention(x)
+        h, w = x.shape[2:]
+        x = rearrange(x, "b c h w -> b (h w) c")
+
+        x_mix = x
+        num_frames = torch.arange(timesteps, device=x.device)
+        num_frames = repeat(num_frames, "t -> b t", b=x.shape[0] // timesteps)
+        num_frames = rearrange(num_frames, "b t -> (b t)")
+        t_emb = timestep_embedding(num_frames, self.in_channels, repeat_only=False)
+        emb = self.video_time_embed(t_emb)  # b, n_channels
+        emb = emb[:, None, :]
+        x_mix = x_mix + emb
+
+        alpha = self.get_alpha()
+        x_mix = self.time_mix_block(x_mix, timesteps=timesteps)
+        x = alpha * x + (1.0 - alpha) * x_mix  # alpha merge
+
+        x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w)
+        x = self.proj_out(x)
+
+        return x_in + x
+
+    def get_alpha(
+        self,
+    ):
+        if self.merge_strategy == "fixed":
+            return self.mix_factor
+        elif self.merge_strategy == "learned":
+            return torch.sigmoid(self.mix_factor)
+        else:
+            raise NotImplementedError(f"unknown merge strategy {self.merge_strategy}")
+
+
+
+def make_time_attn(
+    in_channels,
+    attn_type="vanilla",
+    attn_kwargs=None,
+    alpha: float = 0,
+    merge_strategy: str = "learned",
+):
+    return partialclass(
+        AttnVideoBlock, in_channels, alpha=alpha, merge_strategy=merge_strategy
+    )
+
+
+class Conv2DWrapper(torch.nn.Conv2d):
+    def forward(self, input: torch.Tensor, **kwargs) -> torch.Tensor:
+        return super().forward(input)
+
+
+class VideoDecoder(Decoder):
+    available_time_modes = ["all", "conv-only", "attn-only"]
+
+    def __init__(
+        self,
+        *args,
+        video_kernel_size: Union[int, list] = 3,
+        alpha: float = 0.0,
+        merge_strategy: str = "learned",
+        time_mode: str = "conv-only",
+        **kwargs,
+    ):
+        self.video_kernel_size = video_kernel_size
+        self.alpha = alpha
+        self.merge_strategy = merge_strategy
+        self.time_mode = time_mode
+        assert (
+            self.time_mode in self.available_time_modes
+        ), f"time_mode parameter has to be in {self.available_time_modes}"
+
+        if self.time_mode != "attn-only":
+            kwargs["conv_out_op"] = partialclass(AE3DConv, video_kernel_size=self.video_kernel_size)
+        if self.time_mode not in ["conv-only", "only-last-conv"]:
+            kwargs["attn_op"] = partialclass(make_time_attn, alpha=self.alpha, merge_strategy=self.merge_strategy)
+        if self.time_mode not in ["attn-only", "only-last-conv"]:
+            kwargs["resnet_op"] = partialclass(VideoResBlock, video_kernel_size=self.video_kernel_size, alpha=self.alpha, merge_strategy=self.merge_strategy)
+
+        super().__init__(*args, **kwargs)
+
+    def get_last_layer(self, skip_time_mix=False, **kwargs):
+        if self.time_mode == "attn-only":
+            raise NotImplementedError("TODO")
+        else:
+            return (
+                self.conv_out.time_mix_conv.weight
+                if not skip_time_mix
+                else self.conv_out.weight
+            )

ldm_patched/ldm/util.py

@@ -0,0 +1,197 @@
+import importlib
+
+import torch
+from torch import optim
+import numpy as np
+
+from inspect import isfunction
+from PIL import Image, ImageDraw, ImageFont
+
+
+def log_txt_as_img(wh, xc, size=10):
+    # wh a tuple of (width, height)
+    # xc a list of captions to plot
+    b = len(xc)
+    txts = list()
+    for bi in range(b):
+        txt = Image.new("RGB", wh, color="white")
+        draw = ImageDraw.Draw(txt)
+        font = ImageFont.truetype('data/DejaVuSans.ttf', size=size)
+        nc = int(40 * (wh[0] / 256))
+        lines = "\n".join(xc[bi][start:start + nc] for start in range(0, len(xc[bi]), nc))
+
+        try:
+            draw.text((0, 0), lines, fill="black", font=font)
+        except UnicodeEncodeError:
+            print("Cant encode string for logging. Skipping.")
+
+        txt = np.array(txt).transpose(2, 0, 1) / 127.5 - 1.0
+        txts.append(txt)
+    txts = np.stack(txts)
+    txts = torch.tensor(txts)
+    return txts
+
+
+def ismap(x):
+    if not isinstance(x, torch.Tensor):
+        return False
+    return (len(x.shape) == 4) and (x.shape[1] > 3)
+
+
+def isimage(x):
+    if not isinstance(x,torch.Tensor):
+        return False
+    return (len(x.shape) == 4) and (x.shape[1] == 3 or x.shape[1] == 1)
+
+
+def exists(x):
+    return x is not None
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+def mean_flat(tensor):
+    """
+    https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/nn.py#L86
+    Take the mean over all non-batch dimensions.
+    """
+    return tensor.mean(dim=list(range(1, len(tensor.shape))))
+
+
+def count_params(model, verbose=False):
+    total_params = sum(p.numel() for p in model.parameters())
+    if verbose:
+        print(f"{model.__class__.__name__} has {total_params*1.e-6:.2f} M params.")
+    return total_params
+
+
+def instantiate_from_config(config):
+    if not "target" in config:
+        if config == '__is_first_stage__':
+            return None
+        elif config == "__is_unconditional__":
+            return None
+        raise KeyError("Expected key `target` to instantiate.")
+    return get_obj_from_str(config["target"])(**config.get("params", dict()))
+
+
+def get_obj_from_str(string, reload=False):
+    module, cls = string.rsplit(".", 1)
+    if reload:
+        module_imp = importlib.import_module(module)
+        importlib.reload(module_imp)
+    return getattr(importlib.import_module(module, package=None), cls)
+
+
+class AdamWwithEMAandWings(optim.Optimizer):
+    # credit to https://gist.github.com/crowsonkb/65f7265353f403714fce3b2595e0b298
+    def __init__(self, params, lr=1.e-3, betas=(0.9, 0.999), eps=1.e-8,  # TODO: check hyperparameters before using
+                 weight_decay=1.e-2, amsgrad=False, ema_decay=0.9999,   # ema decay to match previous code
+                 ema_power=1., param_names=()):
+        """AdamW that saves EMA versions of the parameters."""
+        if not 0.0 <= lr:
+            raise ValueError("Invalid learning rate: {}".format(lr))
+        if not 0.0 <= eps:
+            raise ValueError("Invalid epsilon value: {}".format(eps))
+        if not 0.0 <= betas[0] < 1.0:
+            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
+        if not 0.0 <= betas[1] < 1.0:
+            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
+        if not 0.0 <= weight_decay:
+            raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
+        if not 0.0 <= ema_decay <= 1.0:
+            raise ValueError("Invalid ema_decay value: {}".format(ema_decay))
+        defaults = dict(lr=lr, betas=betas, eps=eps,
+                        weight_decay=weight_decay, amsgrad=amsgrad, ema_decay=ema_decay,
+                        ema_power=ema_power, param_names=param_names)
+        super().__init__(params, defaults)
+
+    def __setstate__(self, state):
+        super().__setstate__(state)
+        for group in self.param_groups:
+            group.setdefault('amsgrad', False)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+        Args:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            params_with_grad = []
+            grads = []
+            exp_avgs = []
+            exp_avg_sqs = []
+            ema_params_with_grad = []
+            state_sums = []
+            max_exp_avg_sqs = []
+            state_steps = []
+            amsgrad = group['amsgrad']
+            beta1, beta2 = group['betas']
+            ema_decay = group['ema_decay']
+            ema_power = group['ema_power']
+
+            for p in group['params']:
+                if p.grad is None:
+                    continue
+                params_with_grad.append(p)
+                if p.grad.is_sparse:
+                    raise RuntimeError('AdamW does not support sparse gradients')
+                grads.append(p.grad)
+
+                state = self.state[p]
+
+                # State initialization
+                if len(state) == 0:
+                    state['step'] = 0
+                    # Exponential moving average of gradient values
+                    state['exp_avg'] = torch.zeros_like(p, memory_format=torch.preserve_format)
+                    # Exponential moving average of squared gradient values
+                    state['exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)
+                    if amsgrad:
+                        # Maintains max of all exp. moving avg. of sq. grad. values
+                        state['max_exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)
+                    # Exponential moving average of parameter values
+                    state['param_exp_avg'] = p.detach().float().clone()
+
+                exp_avgs.append(state['exp_avg'])
+                exp_avg_sqs.append(state['exp_avg_sq'])
+                ema_params_with_grad.append(state['param_exp_avg'])
+
+                if amsgrad:
+                    max_exp_avg_sqs.append(state['max_exp_avg_sq'])
+
+                # update the steps for each param group update
+                state['step'] += 1
+                # record the step after step update
+                state_steps.append(state['step'])
+
+            optim._functional.adamw(params_with_grad,
+                    grads,
+                    exp_avgs,
+                    exp_avg_sqs,
+                    max_exp_avg_sqs,
+                    state_steps,
+                    amsgrad=amsgrad,
+                    beta1=beta1,
+                    beta2=beta2,
+                    lr=group['lr'],
+                    weight_decay=group['weight_decay'],
+                    eps=group['eps'],
+                    maximize=False)
+
+            cur_ema_decay = min(ema_decay, 1 - state['step'] ** -ema_power)
+            for param, ema_param in zip(params_with_grad, ema_params_with_grad):
+                ema_param.mul_(cur_ema_decay).add_(param.float(), alpha=1 - cur_ema_decay)
+
+        return loss

ldm_patched/modules/args_parser.py

@@ -0,0 +1,122 @@
+import argparse
+import enum
+import ldm_patched.modules.options
+
+class EnumAction(argparse.Action):
+    """
+    Argparse action for handling Enums
+    """
+    def __init__(self, **kwargs):
+        # Pop off the type value
+        enum_type = kwargs.pop("type", None)
+
+        # Ensure an Enum subclass is provided
+        if enum_type is None:
+            raise ValueError("type must be assigned an Enum when using EnumAction")
+        if not issubclass(enum_type, enum.Enum):
+            raise TypeError("type must be an Enum when using EnumAction")
+
+        # Generate choices from the Enum
+        choices = tuple(e.value for e in enum_type)
+        kwargs.setdefault("choices", choices)
+        kwargs.setdefault("metavar", f"[{','.join(list(choices))}]")
+
+        super(EnumAction, self).__init__(**kwargs)
+
+        self._enum = enum_type
+
+    def __call__(self, parser, namespace, values, option_string=None):
+        # Convert value back into an Enum
+        value = self._enum(values)
+        setattr(namespace, self.dest, value)
+
+
+parser = argparse.ArgumentParser()
+
+parser.add_argument("--listen", type=str, default="127.0.0.1", metavar="IP", nargs="?", const="0.0.0.0")
+parser.add_argument("--port", type=int, default=8188)
+parser.add_argument("--disable-header-check", type=str, default=None, metavar="ORIGIN", nargs="?", const="*")
+parser.add_argument("--web-upload-size", type=float, default=100)
+
+parser.add_argument("--external-working-path", type=str, default=None, metavar="PATH", nargs='+', action='append')
+parser.add_argument("--output-path", type=str, default=None)
+parser.add_argument("--temp-path", type=str, default=None)
+parser.add_argument("--cache-path", type=str, default=None)
+parser.add_argument("--in-browser", action="store_true")
+parser.add_argument("--disable-in-browser", action="store_true")
+parser.add_argument("--gpu-device-id", type=int, default=None, metavar="DEVICE_ID")
+cm_group = parser.add_mutually_exclusive_group()
+cm_group.add_argument("--async-cuda-allocation", action="store_true")
+cm_group.add_argument("--disable-async-cuda-allocation", action="store_true")
+
+parser.add_argument("--disable-attention-upcast", action="store_true")
+
+fp_group = parser.add_mutually_exclusive_group()
+fp_group.add_argument("--all-in-fp32", action="store_true")
+fp_group.add_argument("--all-in-fp16", action="store_true")
+
+fpunet_group = parser.add_mutually_exclusive_group()
+fpunet_group.add_argument("--unet-in-bf16", action="store_true")
+fpunet_group.add_argument("--unet-in-fp16", action="store_true")
+fpunet_group.add_argument("--unet-in-fp8-e4m3fn", action="store_true")
+fpunet_group.add_argument("--unet-in-fp8-e5m2", action="store_true")
+
+fpvae_group = parser.add_mutually_exclusive_group()
+fpvae_group.add_argument("--vae-in-fp16", action="store_true")
+fpvae_group.add_argument("--vae-in-fp32", action="store_true")
+fpvae_group.add_argument("--vae-in-bf16", action="store_true")
+
+fpte_group = parser.add_mutually_exclusive_group()
+fpte_group.add_argument("--clip-in-fp8-e4m3fn", action="store_true")
+fpte_group.add_argument("--clip-in-fp8-e5m2", action="store_true")
+fpte_group.add_argument("--clip-in-fp16", action="store_true")
+fpte_group.add_argument("--clip-in-fp32", action="store_true")
+
+
+parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE", const=-1)
+
+parser.add_argument("--disable-ipex-hijack", action="store_true")
+
+class LatentPreviewMethod(enum.Enum):
+    NoPreviews = "none"
+    Auto = "auto"
+    Latent2RGB = "fast"
+    TAESD = "taesd"
+
+parser.add_argument("--preview-option", type=LatentPreviewMethod, default=LatentPreviewMethod.NoPreviews, action=EnumAction)
+
+attn_group = parser.add_mutually_exclusive_group()
+attn_group.add_argument("--attention-split", action="store_true")
+attn_group.add_argument("--attention-quad", action="store_true")
+attn_group.add_argument("--attention-pytorch", action="store_true")
+
+parser.add_argument("--disable-xformers", action="store_true")
+
+vram_group = parser.add_mutually_exclusive_group()
+vram_group.add_argument("--always-gpu", action="store_true")
+vram_group.add_argument("--always-high-vram", action="store_true")
+vram_group.add_argument("--always-normal-vram", action="store_true")
+vram_group.add_argument("--always-low-vram", action="store_true")
+vram_group.add_argument("--always-no-vram", action="store_true")
+vram_group.add_argument("--always-cpu", action="store_true")
+
+
+parser.add_argument("--always-offload-from-vram", action="store_true")
+
+
+parser.add_argument("--disable-server-log", action="store_true")
+parser.add_argument("--debug-mode", action="store_true")
+parser.add_argument("--is-windows-embedded-python", action="store_true")
+
+parser.add_argument("--disable-server-info", action="store_true")
+
+if ldm_patched.modules.options.args_parsing:
+    args = parser.parse_args([])
+else:
+    args = parser.parse_args([])
+
+if args.is_windows_embedded_python:
+    args.in_browser = True
+
+if args.disable_in_browser:
+    args.in_browser = False

ldm_patched/modules/checkpoint_pickle.py

@@ -0,0 +1,13 @@
+import pickle
+
+load = pickle.load
+
+class Empty:
+    pass
+
+class Unpickler(pickle.Unpickler):
+    def find_class(self, module, name):
+        #TODO: safe unpickle
+        if module.startswith("pytorch_lightning"):
+            return Empty
+        return super().find_class(module, name)

ldm_patched/modules/clip_config_bigg.json

@@ -0,0 +1,23 @@
+{
+  "architectures": [
+    "CLIPTextModel"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "dropout": 0.0,
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_size": 1280,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 5120,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 77,
+  "model_type": "clip_text_model",
+  "num_attention_heads": 20,
+  "num_hidden_layers": 32,
+  "pad_token_id": 1,
+  "projection_dim": 1280,
+  "torch_dtype": "float32",
+  "vocab_size": 49408
+}

ldm_patched/modules/clip_model.py

@@ -0,0 +1,188 @@
+import torch
+from ldm_patched.ldm.modules.attention import optimized_attention_for_device
+
+class CLIPAttention(torch.nn.Module):
+    def __init__(self, embed_dim, heads, dtype, device, operations):
+        super().__init__()
+
+        self.heads = heads
+        self.q_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device)
+        self.k_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device)
+        self.v_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device)
+
+        self.out_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device)
+
+    def forward(self, x, mask=None, optimized_attention=None):
+        q = self.q_proj(x)
+        k = self.k_proj(x)
+        v = self.v_proj(x)
+
+        out = optimized_attention(q, k, v, self.heads, mask)
+        return self.out_proj(out)
+
+ACTIVATIONS = {"quick_gelu": lambda a: a * torch.sigmoid(1.702 * a),
+               "gelu": torch.nn.functional.gelu,
+}
+
+class CLIPMLP(torch.nn.Module):
+    def __init__(self, embed_dim, intermediate_size, activation, dtype, device, operations):
+        super().__init__()
+        self.fc1 = operations.Linear(embed_dim, intermediate_size, bias=True, dtype=dtype, device=device)
+        self.activation = ACTIVATIONS[activation]
+        self.fc2 = operations.Linear(intermediate_size, embed_dim, bias=True, dtype=dtype, device=device)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.activation(x)
+        x = self.fc2(x)
+        return x
+
+class CLIPLayer(torch.nn.Module):
+    def __init__(self, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations):
+        super().__init__()
+        self.layer_norm1 = operations.LayerNorm(embed_dim, dtype=dtype, device=device)
+        self.self_attn = CLIPAttention(embed_dim, heads, dtype, device, operations)
+        self.layer_norm2 = operations.LayerNorm(embed_dim, dtype=dtype, device=device)
+        self.mlp = CLIPMLP(embed_dim, intermediate_size, intermediate_activation, dtype, device, operations)
+
+    def forward(self, x, mask=None, optimized_attention=None):
+        x += self.self_attn(self.layer_norm1(x), mask, optimized_attention)
+        x += self.mlp(self.layer_norm2(x))
+        return x
+
+
+class CLIPEncoder(torch.nn.Module):
+    def __init__(self, num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations):
+        super().__init__()
+        self.layers = torch.nn.ModuleList([CLIPLayer(embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations) for i in range(num_layers)])
+
+    def forward(self, x, mask=None, intermediate_output=None):
+        optimized_attention = optimized_attention_for_device(x.device, mask=mask is not None)
+
+        if intermediate_output is not None:
+            if intermediate_output < 0:
+                intermediate_output = len(self.layers) + intermediate_output
+
+        intermediate = None
+        for i, l in enumerate(self.layers):
+            x = l(x, mask, optimized_attention)
+            if i == intermediate_output:
+                intermediate = x.clone()
+        return x, intermediate
+
+class CLIPEmbeddings(torch.nn.Module):
+    def __init__(self, embed_dim, vocab_size=49408, num_positions=77, dtype=None, device=None):
+        super().__init__()
+        self.token_embedding = torch.nn.Embedding(vocab_size, embed_dim, dtype=dtype, device=device)
+        self.position_embedding = torch.nn.Embedding(num_positions, embed_dim, dtype=dtype, device=device)
+
+    def forward(self, input_tokens):
+        return self.token_embedding(input_tokens) + self.position_embedding.weight
+
+
+class CLIPTextModel_(torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        num_layers = config_dict["num_hidden_layers"]
+        embed_dim = config_dict["hidden_size"]
+        heads = config_dict["num_attention_heads"]
+        intermediate_size = config_dict["intermediate_size"]
+        intermediate_activation = config_dict["hidden_act"]
+
+        super().__init__()
+        self.embeddings = CLIPEmbeddings(embed_dim, dtype=torch.float32, device=device)
+        self.encoder = CLIPEncoder(num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations)
+        self.final_layer_norm = operations.LayerNorm(embed_dim, dtype=dtype, device=device)
+
+    def forward(self, input_tokens, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True):
+        x = self.embeddings(input_tokens)
+        mask = None
+        if attention_mask is not None:
+            mask = 1.0 - attention_mask.to(x.dtype).unsqueeze(1).unsqueeze(1).expand(attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1])
+            mask = mask.masked_fill(mask.to(torch.bool), float("-inf"))
+
+        causal_mask = torch.empty(x.shape[1], x.shape[1], dtype=x.dtype, device=x.device).fill_(float("-inf")).triu_(1)
+        if mask is not None:
+            mask += causal_mask
+        else:
+            mask = causal_mask
+
+        x, i = self.encoder(x, mask=mask, intermediate_output=intermediate_output)
+        x = self.final_layer_norm(x)
+        if i is not None and final_layer_norm_intermediate:
+            i = self.final_layer_norm(i)
+
+        pooled_output = x[torch.arange(x.shape[0], device=x.device), input_tokens.to(dtype=torch.int, device=x.device).argmax(dim=-1),]
+        return x, i, pooled_output
+
+class CLIPTextModel(torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        self.num_layers = config_dict["num_hidden_layers"]
+        self.text_model = CLIPTextModel_(config_dict, dtype, device, operations)
+        self.dtype = dtype
+
+    def get_input_embeddings(self):
+        return self.text_model.embeddings.token_embedding
+
+    def set_input_embeddings(self, embeddings):
+        self.text_model.embeddings.token_embedding = embeddings
+
+    def forward(self, *args, **kwargs):
+        return self.text_model(*args, **kwargs)
+
+class CLIPVisionEmbeddings(torch.nn.Module):
+    def __init__(self, embed_dim, num_channels=3, patch_size=14, image_size=224, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.class_embedding = torch.nn.Parameter(torch.empty(embed_dim, dtype=dtype, device=device))
+
+        self.patch_embedding = operations.Conv2d(
+            in_channels=num_channels,
+            out_channels=embed_dim,
+            kernel_size=patch_size,
+            stride=patch_size,
+            bias=False,
+            dtype=dtype,
+            device=device
+        )
+
+        num_patches = (image_size // patch_size) ** 2
+        num_positions = num_patches + 1
+        self.position_embedding = torch.nn.Embedding(num_positions, embed_dim, dtype=dtype, device=device)
+
+    def forward(self, pixel_values):
+        embeds = self.patch_embedding(pixel_values).flatten(2).transpose(1, 2)
+        return torch.cat([self.class_embedding.expand(pixel_values.shape[0], 1, -1), embeds], dim=1) + self.position_embedding.weight
+
+
+class CLIPVision(torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        num_layers = config_dict["num_hidden_layers"]
+        embed_dim = config_dict["hidden_size"]
+        heads = config_dict["num_attention_heads"]
+        intermediate_size = config_dict["intermediate_size"]
+        intermediate_activation = config_dict["hidden_act"]
+
+        self.embeddings = CLIPVisionEmbeddings(embed_dim, config_dict["num_channels"], config_dict["patch_size"], config_dict["image_size"], dtype=torch.float32, device=device, operations=operations)
+        self.pre_layrnorm = operations.LayerNorm(embed_dim)
+        self.encoder = CLIPEncoder(num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations)
+        self.post_layernorm = operations.LayerNorm(embed_dim)
+
+    def forward(self, pixel_values, attention_mask=None, intermediate_output=None):
+        x = self.embeddings(pixel_values)
+        x = self.pre_layrnorm(x)
+        #TODO: attention_mask?
+        x, i = self.encoder(x, mask=None, intermediate_output=intermediate_output)
+        pooled_output = self.post_layernorm(x[:, 0, :])
+        return x, i, pooled_output
+
+class CLIPVisionModelProjection(torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        self.vision_model = CLIPVision(config_dict, dtype, device, operations)
+        self.visual_projection = operations.Linear(config_dict["hidden_size"], config_dict["projection_dim"], bias=False)
+
+    def forward(self, *args, **kwargs):
+        x = self.vision_model(*args, **kwargs)
+        out = self.visual_projection(x[2])
+        return (x[0], x[1], out)

ldm_patched/modules/clip_vision.py

@@ -0,0 +1,117 @@
+from .utils import load_torch_file, transformers_convert, common_upscale
+import os
+import torch
+import contextlib
+import json
+
+import ldm_patched.modules.ops
+import ldm_patched.modules.model_patcher
+import ldm_patched.modules.model_management
+import ldm_patched.modules.utils
+import ldm_patched.modules.clip_model
+
+class Output:
+    def __getitem__(self, key):
+        return getattr(self, key)
+    def __setitem__(self, key, item):
+        setattr(self, key, item)
+
+def clip_preprocess(image, size=224):
+    mean = torch.tensor([ 0.48145466,0.4578275,0.40821073], device=image.device, dtype=image.dtype)
+    std = torch.tensor([0.26862954,0.26130258,0.27577711], device=image.device, dtype=image.dtype)
+    scale = (size / min(image.shape[1], image.shape[2]))
+    image = torch.nn.functional.interpolate(image.movedim(-1, 1), size=(round(scale * image.shape[1]), round(scale * image.shape[2])), mode="bicubic", antialias=True)
+    h = (image.shape[2] - size)//2
+    w = (image.shape[3] - size)//2
+    image = image[:,:,h:h+size,w:w+size]
+    image = torch.clip((255. * image), 0, 255).round() / 255.0
+    return (image - mean.view([3,1,1])) / std.view([3,1,1])
+
+class ClipVisionModel():
+    def __init__(self, json_config):
+        with open(json_config) as f:
+            config = json.load(f)
+
+        self.load_device = ldm_patched.modules.model_management.text_encoder_device()
+        offload_device = ldm_patched.modules.model_management.text_encoder_offload_device()
+        self.dtype = torch.float32
+        if ldm_patched.modules.model_management.should_use_fp16(self.load_device, prioritize_performance=False):
+            self.dtype = torch.float16
+
+        self.model = ldm_patched.modules.clip_model.CLIPVisionModelProjection(config, self.dtype, offload_device, ldm_patched.modules.ops.disable_weight_init)
+
+        self.patcher = ldm_patched.modules.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+    def load_sd(self, sd):
+        return self.model.load_state_dict(sd, strict=False)
+
+    def encode_image(self, image):
+        ldm_patched.modules.model_management.load_model_gpu(self.patcher)
+        pixel_values = clip_preprocess(image.to(self.load_device))
+
+        if self.dtype != torch.float32:
+            precision_scope = torch.autocast
+        else:
+            precision_scope = lambda a, b: contextlib.nullcontext(a)
+
+        with precision_scope(ldm_patched.modules.model_management.get_autocast_device(self.load_device), torch.float32):
+            out = self.model(pixel_values=pixel_values, intermediate_output=-2)
+
+        outputs = Output()
+        outputs["last_hidden_state"] = out[0].to(ldm_patched.modules.model_management.intermediate_device())
+        outputs["image_embeds"] = out[2].to(ldm_patched.modules.model_management.intermediate_device())
+        outputs["penultimate_hidden_states"] = out[1].to(ldm_patched.modules.model_management.intermediate_device())
+        return outputs
+
+def convert_to_transformers(sd, prefix):
+    sd_k = sd.keys()
+    if "{}transformer.resblocks.0.attn.in_proj_weight".format(prefix) in sd_k:
+        keys_to_replace = {
+            "{}class_embedding".format(prefix): "vision_model.embeddings.class_embedding",
+            "{}conv1.weight".format(prefix): "vision_model.embeddings.patch_embedding.weight",
+            "{}positional_embedding".format(prefix): "vision_model.embeddings.position_embedding.weight",
+            "{}ln_post.bias".format(prefix): "vision_model.post_layernorm.bias",
+            "{}ln_post.weight".format(prefix): "vision_model.post_layernorm.weight",
+            "{}ln_pre.bias".format(prefix): "vision_model.pre_layrnorm.bias",
+            "{}ln_pre.weight".format(prefix): "vision_model.pre_layrnorm.weight",
+        }
+
+        for x in keys_to_replace:
+            if x in sd_k:
+                sd[keys_to_replace[x]] = sd.pop(x)
+
+        if "{}proj".format(prefix) in sd_k:
+            sd['visual_projection.weight'] = sd.pop("{}proj".format(prefix)).transpose(0, 1)
+
+        sd = transformers_convert(sd, prefix, "vision_model.", 48)
+    return sd
+
+def load_clipvision_from_sd(sd, prefix="", convert_keys=False):
+    if convert_keys:
+        sd = convert_to_transformers(sd, prefix)
+    if "vision_model.encoder.layers.47.layer_norm1.weight" in sd:
+        json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_g.json")
+    elif "vision_model.encoder.layers.30.layer_norm1.weight" in sd:
+        json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_h.json")
+    elif "vision_model.encoder.layers.22.layer_norm1.weight" in sd:
+        json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl.json")
+    else:
+        return None
+
+    clip = ClipVisionModel(json_config)
+    m, u = clip.load_sd(sd)
+    if len(m) > 0:
+        print("extra clip vision:", m)
+    u = set(u)
+    keys = list(sd.keys())
+    for k in keys:
+        if k not in u:
+            t = sd.pop(k)
+            del t
+    return clip
+
+def load(ckpt_path):
+    sd = load_torch_file(ckpt_path)
+    if "visual.transformer.resblocks.0.attn.in_proj_weight" in sd:
+        return load_clipvision_from_sd(sd, prefix="visual.", convert_keys=True)
+    else:
+        return load_clipvision_from_sd(sd)

ldm_patched/modules/clip_vision_config_g.json

@@ -0,0 +1,18 @@
+{
+  "attention_dropout": 0.0,
+  "dropout": 0.0,
+  "hidden_act": "gelu",
+  "hidden_size": 1664,
+  "image_size": 224,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 8192,
+  "layer_norm_eps": 1e-05,
+  "model_type": "clip_vision_model",
+  "num_attention_heads": 16,
+  "num_channels": 3,
+  "num_hidden_layers": 48,
+  "patch_size": 14,
+  "projection_dim": 1280,
+  "torch_dtype": "float32"
+}

ldm_patched/modules/clip_vision_config_h.json

@@ -0,0 +1,18 @@
+{
+  "attention_dropout": 0.0,
+  "dropout": 0.0,
+  "hidden_act": "gelu",
+  "hidden_size": 1280,
+  "image_size": 224,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 5120,
+  "layer_norm_eps": 1e-05,
+  "model_type": "clip_vision_model",
+  "num_attention_heads": 16,
+  "num_channels": 3,
+  "num_hidden_layers": 32,
+  "patch_size": 14,
+  "projection_dim": 1024,
+  "torch_dtype": "float32"
+}

ldm_patched/modules/clip_vision_config_vitl.json

@@ -0,0 +1,18 @@
+{
+  "attention_dropout": 0.0,
+  "dropout": 0.0,
+  "hidden_act": "quick_gelu",
+  "hidden_size": 1024,
+  "image_size": 224,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "layer_norm_eps": 1e-05,
+  "model_type": "clip_vision_model",
+  "num_attention_heads": 16,
+  "num_channels": 3,
+  "num_hidden_layers": 24,
+  "patch_size": 14,
+  "projection_dim": 768,
+  "torch_dtype": "float32"
+}

ldm_patched/modules/conds.py

@@ -0,0 +1,79 @@
+import enum
+import torch
+import math
+import ldm_patched.modules.utils
+
+
+def lcm(a, b): #TODO: eventually replace by math.lcm (added in python3.9)
+    return abs(a*b) // math.gcd(a, b)
+
+class CONDRegular:
+    def __init__(self, cond):
+        self.cond = cond
+
+    def _copy_with(self, cond):
+        return self.__class__(cond)
+
+    def process_cond(self, batch_size, device, **kwargs):
+        return self._copy_with(ldm_patched.modules.utils.repeat_to_batch_size(self.cond, batch_size).to(device))
+
+    def can_concat(self, other):
+        if self.cond.shape != other.cond.shape:
+            return False
+        return True
+
+    def concat(self, others):
+        conds = [self.cond]
+        for x in others:
+            conds.append(x.cond)
+        return torch.cat(conds)
+
+class CONDNoiseShape(CONDRegular):
+    def process_cond(self, batch_size, device, area, **kwargs):
+        data = self.cond[:,:,area[2]:area[0] + area[2],area[3]:area[1] + area[3]]
+        return self._copy_with(ldm_patched.modules.utils.repeat_to_batch_size(data, batch_size).to(device))
+
+
+class CONDCrossAttn(CONDRegular):
+    def can_concat(self, other):
+        s1 = self.cond.shape
+        s2 = other.cond.shape
+        if s1 != s2:
+            if s1[0] != s2[0] or s1[2] != s2[2]: #these 2 cases should not happen
+                return False
+
+            mult_min = lcm(s1[1], s2[1])
+            diff = mult_min // min(s1[1], s2[1])
+            if diff > 4: #arbitrary limit on the padding because it's probably going to impact performance negatively if it's too much
+                return False
+        return True
+
+    def concat(self, others):
+        conds = [self.cond]
+        crossattn_max_len = self.cond.shape[1]
+        for x in others:
+            c = x.cond
+            crossattn_max_len = lcm(crossattn_max_len, c.shape[1])
+            conds.append(c)
+
+        out = []
+        for c in conds:
+            if c.shape[1] < crossattn_max_len:
+                c = c.repeat(1, crossattn_max_len // c.shape[1], 1) #padding with repeat doesn't change result
+            out.append(c)
+        return torch.cat(out)
+
+class CONDConstant(CONDRegular):
+    def __init__(self, cond):
+        self.cond = cond
+
+    def process_cond(self, batch_size, device, **kwargs):
+        return self._copy_with(self.cond)
+
+    def can_concat(self, other):
+        if self.cond != other.cond:
+            return False
+        return True
+
+    def concat(self, others):
+        return self.cond

ldm_patched/modules/controlnet.py

@@ -0,0 +1,514 @@
+import torch
+import math
+import os
+import contextlib
+import ldm_patched.modules.utils
+import ldm_patched.modules.model_management
+import ldm_patched.modules.model_detection
+import ldm_patched.modules.model_patcher
+import ldm_patched.modules.ops
+
+import ldm_patched.controlnet.cldm
+import ldm_patched.t2ia.adapter
+
+
+def broadcast_image_to(tensor, target_batch_size, batched_number):
+    current_batch_size = tensor.shape[0]
+    #print(current_batch_size, target_batch_size)
+    if current_batch_size == 1:
+        return tensor
+
+    per_batch = target_batch_size // batched_number
+    tensor = tensor[:per_batch]
+
+    if per_batch > tensor.shape[0]:
+        tensor = torch.cat([tensor] * (per_batch // tensor.shape[0]) + [tensor[:(per_batch % tensor.shape[0])]], dim=0)
+
+    current_batch_size = tensor.shape[0]
+    if current_batch_size == target_batch_size:
+        return tensor
+    else:
+        return torch.cat([tensor] * batched_number, dim=0)
+
+class ControlBase:
+    def __init__(self, device=None):
+        self.cond_hint_original = None
+        self.cond_hint = None
+        self.strength = 1.0
+        self.timestep_percent_range = (0.0, 1.0)
+        self.global_average_pooling = False
+        self.timestep_range = None
+
+        if device is None:
+            device = ldm_patched.modules.model_management.get_torch_device()
+        self.device = device
+        self.previous_controlnet = None
+
+    def set_cond_hint(self, cond_hint, strength=1.0, timestep_percent_range=(0.0, 1.0)):
+        self.cond_hint_original = cond_hint
+        self.strength = strength
+        self.timestep_percent_range = timestep_percent_range
+        return self
+
+    def pre_run(self, model, percent_to_timestep_function):
+        self.timestep_range = (percent_to_timestep_function(self.timestep_percent_range[0]), percent_to_timestep_function(self.timestep_percent_range[1]))
+        if self.previous_controlnet is not None:
+            self.previous_controlnet.pre_run(model, percent_to_timestep_function)
+
+    def set_previous_controlnet(self, controlnet):
+        self.previous_controlnet = controlnet
+        return self
+
+    def cleanup(self):
+        if self.previous_controlnet is not None:
+            self.previous_controlnet.cleanup()
+        if self.cond_hint is not None:
+            del self.cond_hint
+            self.cond_hint = None
+        self.timestep_range = None
+
+    def get_models(self):
+        out = []
+        if self.previous_controlnet is not None:
+            out += self.previous_controlnet.get_models()
+        return out
+
+    def copy_to(self, c):
+        c.cond_hint_original = self.cond_hint_original
+        c.strength = self.strength
+        c.timestep_percent_range = self.timestep_percent_range
+        c.global_average_pooling = self.global_average_pooling
+
+    def inference_memory_requirements(self, dtype):
+        if self.previous_controlnet is not None:
+            return self.previous_controlnet.inference_memory_requirements(dtype)
+        return 0
+
+    def control_merge(self, control_input, control_output, control_prev, output_dtype):
+        out = {'input':[], 'middle':[], 'output': []}
+
+        if control_input is not None:
+            for i in range(len(control_input)):
+                key = 'input'
+                x = control_input[i]
+                if x is not None:
+                    x *= self.strength
+                    if x.dtype != output_dtype:
+                        x = x.to(output_dtype)
+                out[key].insert(0, x)
+
+        if control_output is not None:
+            for i in range(len(control_output)):
+                if i == (len(control_output) - 1):
+                    key = 'middle'
+                    index = 0
+                else:
+                    key = 'output'
+                    index = i
+                x = control_output[i]
+                if x is not None:
+                    if self.global_average_pooling:
+                        x = torch.mean(x, dim=(2, 3), keepdim=True).repeat(1, 1, x.shape[2], x.shape[3])
+
+                    x *= self.strength
+                    if x.dtype != output_dtype:
+                        x = x.to(output_dtype)
+
+                out[key].append(x)
+        if control_prev is not None:
+            for x in ['input', 'middle', 'output']:
+                o = out[x]
+                for i in range(len(control_prev[x])):
+                    prev_val = control_prev[x][i]
+                    if i >= len(o):
+                        o.append(prev_val)
+                    elif prev_val is not None:
+                        if o[i] is None:
+                            o[i] = prev_val
+                        else:
+                            o[i] += prev_val
+        return out
+
+class ControlNet(ControlBase):
+    def __init__(self, control_model, global_average_pooling=False, device=None, load_device=None, manual_cast_dtype=None):
+        super().__init__(device)
+        self.control_model = control_model
+        self.load_device = load_device
+        self.control_model_wrapped = ldm_patched.modules.model_patcher.ModelPatcher(self.control_model, load_device=load_device, offload_device=ldm_patched.modules.model_management.unet_offload_device())
+        self.global_average_pooling = global_average_pooling
+        self.model_sampling_current = None
+        self.manual_cast_dtype = manual_cast_dtype
+
+    def get_control(self, x_noisy, t, cond, batched_number):
+        control_prev = None
+        if self.previous_controlnet is not None:
+            control_prev = self.previous_controlnet.get_control(x_noisy, t, cond, batched_number)
+
+        if self.timestep_range is not None:
+            if t[0] > self.timestep_range[0] or t[0] < self.timestep_range[1]:
+                if control_prev is not None:
+                    return control_prev
+                else:
+                    return None
+
+        dtype = self.control_model.dtype
+        if self.manual_cast_dtype is not None:
+            dtype = self.manual_cast_dtype
+
+        output_dtype = x_noisy.dtype
+        if self.cond_hint is None or x_noisy.shape[2] * 8 != self.cond_hint.shape[2] or x_noisy.shape[3] * 8 != self.cond_hint.shape[3]:
+            if self.cond_hint is not None:
+                del self.cond_hint
+            self.cond_hint = None
+            self.cond_hint = ldm_patched.modules.utils.common_upscale(self.cond_hint_original, x_noisy.shape[3] * 8, x_noisy.shape[2] * 8, 'nearest-exact', "center").to(dtype).to(self.device)
+        if x_noisy.shape[0] != self.cond_hint.shape[0]:
+            self.cond_hint = broadcast_image_to(self.cond_hint, x_noisy.shape[0], batched_number)
+
+        context = cond['c_crossattn']
+        y = cond.get('y', None)
+        if y is not None:
+            y = y.to(dtype)
+        timestep = self.model_sampling_current.timestep(t)
+        x_noisy = self.model_sampling_current.calculate_input(t, x_noisy)
+
+        control = self.control_model(x=x_noisy.to(dtype), hint=self.cond_hint, timesteps=timestep.float(), context=context.to(dtype), y=y)
+        return self.control_merge(None, control, control_prev, output_dtype)
+
+    def copy(self):
+        c = ControlNet(self.control_model, global_average_pooling=self.global_average_pooling, load_device=self.load_device, manual_cast_dtype=self.manual_cast_dtype)
+        self.copy_to(c)
+        return c
+
+    def get_models(self):
+        out = super().get_models()
+        out.append(self.control_model_wrapped)
+        return out
+
+    def pre_run(self, model, percent_to_timestep_function):
+        super().pre_run(model, percent_to_timestep_function)
+        self.model_sampling_current = model.model_sampling
+
+    def cleanup(self):
+        self.model_sampling_current = None
+        super().cleanup()
+
+class ControlLoraOps:
+    class Linear(torch.nn.Module):
+        def __init__(self, in_features: int, out_features: int, bias: bool = True,
+                    device=None, dtype=None) -> None:
+            factory_kwargs = {'device': device, 'dtype': dtype}
+            super().__init__()
+            self.in_features = in_features
+            self.out_features = out_features
+            self.weight = None
+            self.up = None
+            self.down = None
+            self.bias = None
+
+        def forward(self, input):
+            weight, bias = ldm_patched.modules.ops.cast_bias_weight(self, input)
+            if self.up is not None:
+                return torch.nn.functional.linear(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias)
+            else:
+                return torch.nn.functional.linear(input, weight, bias)
+
+    class Conv2d(torch.nn.Module):
+        def __init__(
+            self,
+            in_channels,
+            out_channels,
+            kernel_size,
+            stride=1,
+            padding=0,
+            dilation=1,
+            groups=1,
+            bias=True,
+            padding_mode='zeros',
+            device=None,
+            dtype=None
+        ):
+            super().__init__()
+            self.in_channels = in_channels
+            self.out_channels = out_channels
+            self.kernel_size = kernel_size
+            self.stride = stride
+            self.padding = padding
+            self.dilation = dilation
+            self.transposed = False
+            self.output_padding = 0
+            self.groups = groups
+            self.padding_mode = padding_mode
+
+            self.weight = None
+            self.bias = None
+            self.up = None
+            self.down = None
+
+
+        def forward(self, input):
+            weight, bias = ldm_patched.modules.ops.cast_bias_weight(self, input)
+            if self.up is not None:
+                return torch.nn.functional.conv2d(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias, self.stride, self.padding, self.dilation, self.groups)
+            else:
+                return torch.nn.functional.conv2d(input, weight, bias, self.stride, self.padding, self.dilation, self.groups)
+
+
+class ControlLora(ControlNet):
+    def __init__(self, control_weights, global_average_pooling=False, device=None):
+        ControlBase.__init__(self, device)
+        self.control_weights = control_weights
+        self.global_average_pooling = global_average_pooling
+
+    def pre_run(self, model, percent_to_timestep_function):
+        super().pre_run(model, percent_to_timestep_function)
+        controlnet_config = model.model_config.unet_config.copy()
+        controlnet_config.pop("out_channels")
+        controlnet_config["hint_channels"] = self.control_weights["input_hint_block.0.weight"].shape[1]
+        self.manual_cast_dtype = model.manual_cast_dtype
+        dtype = model.get_dtype()
+        if self.manual_cast_dtype is None:
+            class control_lora_ops(ControlLoraOps, ldm_patched.modules.ops.disable_weight_init):
+                pass
+        else:
+            class control_lora_ops(ControlLoraOps, ldm_patched.modules.ops.manual_cast):
+                pass
+            dtype = self.manual_cast_dtype
+
+        controlnet_config["operations"] = control_lora_ops
+        controlnet_config["dtype"] = dtype
+        self.control_model = ldm_patched.controlnet.cldm.ControlNet(**controlnet_config)
+        self.control_model.to(ldm_patched.modules.model_management.get_torch_device())
+        diffusion_model = model.diffusion_model
+        sd = diffusion_model.state_dict()
+        cm = self.control_model.state_dict()
+
+        for k in sd:
+            weight = ldm_patched.modules.model_management.resolve_lowvram_weight(sd[k], diffusion_model, k)
+            try:
+                ldm_patched.modules.utils.set_attr(self.control_model, k, weight)
+            except:
+                pass
+
+        for k in self.control_weights:
+            if k not in {"lora_controlnet"}:
+                ldm_patched.modules.utils.set_attr(self.control_model, k, self.control_weights[k].to(dtype).to(ldm_patched.modules.model_management.get_torch_device()))
+
+    def copy(self):
+        c = ControlLora(self.control_weights, global_average_pooling=self.global_average_pooling)
+        self.copy_to(c)
+        return c
+
+    def cleanup(self):
+        del self.control_model
+        self.control_model = None
+        super().cleanup()
+
+    def get_models(self):
+        out = ControlBase.get_models(self)
+        return out
+
+    def inference_memory_requirements(self, dtype):
+        return ldm_patched.modules.utils.calculate_parameters(self.control_weights) * ldm_patched.modules.model_management.dtype_size(dtype) + ControlBase.inference_memory_requirements(self, dtype)
+
+def load_controlnet(ckpt_path, model=None):
+    controlnet_data = ldm_patched.modules.utils.load_torch_file(ckpt_path, safe_load=True)
+    if "lora_controlnet" in controlnet_data:
+        return ControlLora(controlnet_data)
+
+    controlnet_config = None
+    if "controlnet_cond_embedding.conv_in.weight" in controlnet_data: #diffusers format
+        unet_dtype = ldm_patched.modules.model_management.unet_dtype()
+        controlnet_config = ldm_patched.modules.model_detection.unet_config_from_diffusers_unet(controlnet_data, unet_dtype)
+        diffusers_keys = ldm_patched.modules.utils.unet_to_diffusers(controlnet_config)
+        diffusers_keys["controlnet_mid_block.weight"] = "middle_block_out.0.weight"
+        diffusers_keys["controlnet_mid_block.bias"] = "middle_block_out.0.bias"
+
+        count = 0
+        loop = True
+        while loop:
+            suffix = [".weight", ".bias"]
+            for s in suffix:
+                k_in = "controlnet_down_blocks.{}{}".format(count, s)
+                k_out = "zero_convs.{}.0{}".format(count, s)
+                if k_in not in controlnet_data:
+                    loop = False
+                    break
+                diffusers_keys[k_in] = k_out
+            count += 1
+
+        count = 0
+        loop = True
+        while loop:
+            suffix = [".weight", ".bias"]
+            for s in suffix:
+                if count == 0:
+                    k_in = "controlnet_cond_embedding.conv_in{}".format(s)
+                else:
+                    k_in = "controlnet_cond_embedding.blocks.{}{}".format(count - 1, s)
+                k_out = "input_hint_block.{}{}".format(count * 2, s)
+                if k_in not in controlnet_data:
+                    k_in = "controlnet_cond_embedding.conv_out{}".format(s)
+                    loop = False
+                diffusers_keys[k_in] = k_out
+            count += 1
+
+        new_sd = {}
+        for k in diffusers_keys:
+            if k in controlnet_data:
+                new_sd[diffusers_keys[k]] = controlnet_data.pop(k)
+
+        leftover_keys = controlnet_data.keys()
+        if len(leftover_keys) > 0:
+            print("leftover keys:", leftover_keys)
+        controlnet_data = new_sd
+
+    pth_key = 'control_model.zero_convs.0.0.weight'
+    pth = False
+    key = 'zero_convs.0.0.weight'
+    if pth_key in controlnet_data:
+        pth = True
+        key = pth_key
+        prefix = "control_model."
+    elif key in controlnet_data:
+        prefix = ""
+    else:
+        net = load_t2i_adapter(controlnet_data)
+        if net is None:
+            print("error checkpoint does not contain controlnet or t2i adapter data", ckpt_path)
+        return net
+
+    if controlnet_config is None:
+        unet_dtype = ldm_patched.modules.model_management.unet_dtype()
+        controlnet_config = ldm_patched.modules.model_detection.model_config_from_unet(controlnet_data, prefix, unet_dtype, True).unet_config
+    load_device = ldm_patched.modules.model_management.get_torch_device()
+    manual_cast_dtype = ldm_patched.modules.model_management.unet_manual_cast(unet_dtype, load_device)
+    if manual_cast_dtype is not None:
+        controlnet_config["operations"] = ldm_patched.modules.ops.manual_cast
+    controlnet_config.pop("out_channels")
+    controlnet_config["hint_channels"] = controlnet_data["{}input_hint_block.0.weight".format(prefix)].shape[1]
+    control_model = ldm_patched.controlnet.cldm.ControlNet(**controlnet_config)
+
+    if pth:
+        if 'difference' in controlnet_data:
+            if model is not None:
+                ldm_patched.modules.model_management.load_models_gpu([model])
+                model_sd = model.model_state_dict()
+                for x in controlnet_data:
+                    c_m = "control_model."
+                    if x.startswith(c_m):
+                        sd_key = "diffusion_model.{}".format(x[len(c_m):])
+                        if sd_key in model_sd:
+                            cd = controlnet_data[x]
+                            cd += model_sd[sd_key].type(cd.dtype).to(cd.device)
+            else:
+                print("WARNING: Loaded a diff controlnet without a model. It will very likely not work.")
+
+        class WeightsLoader(torch.nn.Module):
+            pass
+        w = WeightsLoader()
+        w.control_model = control_model
+        missing, unexpected = w.load_state_dict(controlnet_data, strict=False)
+    else:
+        missing, unexpected = control_model.load_state_dict(controlnet_data, strict=False)
+    print(missing, unexpected)
+
+    global_average_pooling = False
+    filename = os.path.splitext(ckpt_path)[0]
+    if filename.endswith("_shuffle") or filename.endswith("_shuffle_fp16"): #TODO: smarter way of enabling global_average_pooling
+        global_average_pooling = True
+
+    control = ControlNet(control_model, global_average_pooling=global_average_pooling, load_device=load_device, manual_cast_dtype=manual_cast_dtype)
+    return control
+
+class T2IAdapter(ControlBase):
+    def __init__(self, t2i_model, channels_in, device=None):
+        super().__init__(device)
+        self.t2i_model = t2i_model
+        self.channels_in = channels_in
+        self.control_input = None
+
+    def scale_image_to(self, width, height):
+        unshuffle_amount = self.t2i_model.unshuffle_amount
+        width = math.ceil(width / unshuffle_amount) * unshuffle_amount
+        height = math.ceil(height / unshuffle_amount) * unshuffle_amount
+        return width, height
+
+    def get_control(self, x_noisy, t, cond, batched_number):
+        control_prev = None
+        if self.previous_controlnet is not None:
+            control_prev = self.previous_controlnet.get_control(x_noisy, t, cond, batched_number)
+
+        if self.timestep_range is not None:
+            if t[0] > self.timestep_range[0] or t[0] < self.timestep_range[1]:
+                if control_prev is not None:
+                    return control_prev
+                else:
+                    return None
+
+        if self.cond_hint is None or x_noisy.shape[2] * 8 != self.cond_hint.shape[2] or x_noisy.shape[3] * 8 != self.cond_hint.shape[3]:
+            if self.cond_hint is not None:
+                del self.cond_hint
+            self.control_input = None
+            self.cond_hint = None
+            width, height = self.scale_image_to(x_noisy.shape[3] * 8, x_noisy.shape[2] * 8)
+            self.cond_hint = ldm_patched.modules.utils.common_upscale(self.cond_hint_original, width, height, 'nearest-exact', "center").float().to(self.device)
+            if self.channels_in == 1 and self.cond_hint.shape[1] > 1:
+                self.cond_hint = torch.mean(self.cond_hint, 1, keepdim=True)
+        if x_noisy.shape[0] != self.cond_hint.shape[0]:
+            self.cond_hint = broadcast_image_to(self.cond_hint, x_noisy.shape[0], batched_number)
+        if self.control_input is None:
+            self.t2i_model.to(x_noisy.dtype)
+            self.t2i_model.to(self.device)
+            self.control_input = self.t2i_model(self.cond_hint.to(x_noisy.dtype))
+            self.t2i_model.cpu()
+
+        control_input = list(map(lambda a: None if a is None else a.clone(), self.control_input))
+        mid = None
+        if self.t2i_model.xl == True:
+            mid = control_input[-1:]
+            control_input = control_input[:-1]
+        return self.control_merge(control_input, mid, control_prev, x_noisy.dtype)
+
+    def copy(self):
+        c = T2IAdapter(self.t2i_model, self.channels_in)
+        self.copy_to(c)
+        return c
+
+def load_t2i_adapter(t2i_data):
+    if 'adapter' in t2i_data:
+        t2i_data = t2i_data['adapter']
+    if 'adapter.body.0.resnets.0.block1.weight' in t2i_data: #diffusers format
+        prefix_replace = {}
+        for i in range(4):
+            for j in range(2):
+                prefix_replace["adapter.body.{}.resnets.{}.".format(i, j)] = "body.{}.".format(i * 2 + j)
+            prefix_replace["adapter.body.{}.".format(i, j)] = "body.{}.".format(i * 2)
+        prefix_replace["adapter."] = ""
+        t2i_data = ldm_patched.modules.utils.state_dict_prefix_replace(t2i_data, prefix_replace)
+    keys = t2i_data.keys()
+
+    if "body.0.in_conv.weight" in keys:
+        cin = t2i_data['body.0.in_conv.weight'].shape[1]
+        model_ad = ldm_patched.t2ia.adapter.Adapter_light(cin=cin, channels=[320, 640, 1280, 1280], nums_rb=4)
+    elif 'conv_in.weight' in keys:
+        cin = t2i_data['conv_in.weight'].shape[1]
+        channel = t2i_data['conv_in.weight'].shape[0]
+        ksize = t2i_data['body.0.block2.weight'].shape[2]
+        use_conv = False
+        down_opts = list(filter(lambda a: a.endswith("down_opt.op.weight"), keys))
+        if len(down_opts) > 0:
+            use_conv = True
+        xl = False
+        if cin == 256 or cin == 768:
+            xl = True
+        model_ad = ldm_patched.t2ia.adapter.Adapter(cin=cin, channels=[channel, channel*2, channel*4, channel*4][:4], nums_rb=2, ksize=ksize, sk=True, use_conv=use_conv, xl=xl)
+    else:
+        return None
+    missing, unexpected = model_ad.load_state_dict(t2i_data)
+    if len(missing) > 0:
+        print("t2i missing", missing)
+
+    if len(unexpected) > 0:
+        print("t2i unexpected", unexpected)
+
+    return T2IAdapter(model_ad, model_ad.input_channels)

ldm_patched/modules/diffusers_convert.py

@@ -0,0 +1,261 @@
+import re
+import torch
+
+# conversion code from https://github.com/huggingface/diffusers/blob/main/scripts/convert_diffusers_to_original_stable_diffusion.py
+
+# =================#
+# UNet Conversion #
+# =================#
+
+unet_conversion_map = [
+    # (stable-diffusion, HF Diffusers)
+    ("time_embed.0.weight", "time_embedding.linear_1.weight"),
+    ("time_embed.0.bias", "time_embedding.linear_1.bias"),
+    ("time_embed.2.weight", "time_embedding.linear_2.weight"),
+    ("time_embed.2.bias", "time_embedding.linear_2.bias"),
+    ("input_blocks.0.0.weight", "conv_in.weight"),
+    ("input_blocks.0.0.bias", "conv_in.bias"),
+    ("out.0.weight", "conv_norm_out.weight"),
+    ("out.0.bias", "conv_norm_out.bias"),
+    ("out.2.weight", "conv_out.weight"),
+    ("out.2.bias", "conv_out.bias"),
+]
+
+unet_conversion_map_resnet = [
+    # (stable-diffusion, HF Diffusers)
+    ("in_layers.0", "norm1"),
+    ("in_layers.2", "conv1"),
+    ("out_layers.0", "norm2"),
+    ("out_layers.3", "conv2"),
+    ("emb_layers.1", "time_emb_proj"),
+    ("skip_connection", "conv_shortcut"),
+]
+
+unet_conversion_map_layer = []
+# hardcoded number of downblocks and resnets/attentions...
+# would need smarter logic for other networks.
+for i in range(4):
+    # loop over downblocks/upblocks
+
+    for j in range(2):
+        # loop over resnets/attentions for downblocks
+        hf_down_res_prefix = f"down_blocks.{i}.resnets.{j}."
+        sd_down_res_prefix = f"input_blocks.{3 * i + j + 1}.0."
+        unet_conversion_map_layer.append((sd_down_res_prefix, hf_down_res_prefix))
+
+        if i < 3:
+            # no attention layers in down_blocks.3
+            hf_down_atn_prefix = f"down_blocks.{i}.attentions.{j}."
+            sd_down_atn_prefix = f"input_blocks.{3 * i + j + 1}.1."
+            unet_conversion_map_layer.append((sd_down_atn_prefix, hf_down_atn_prefix))
+
+    for j in range(3):
+        # loop over resnets/attentions for upblocks
+        hf_up_res_prefix = f"up_blocks.{i}.resnets.{j}."
+        sd_up_res_prefix = f"output_blocks.{3 * i + j}.0."
+        unet_conversion_map_layer.append((sd_up_res_prefix, hf_up_res_prefix))
+
+        if i > 0:
+            # no attention layers in up_blocks.0
+            hf_up_atn_prefix = f"up_blocks.{i}.attentions.{j}."
+            sd_up_atn_prefix = f"output_blocks.{3 * i + j}.1."
+            unet_conversion_map_layer.append((sd_up_atn_prefix, hf_up_atn_prefix))
+
+    if i < 3:
+        # no downsample in down_blocks.3
+        hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0.conv."
+        sd_downsample_prefix = f"input_blocks.{3 * (i + 1)}.0.op."
+        unet_conversion_map_layer.append((sd_downsample_prefix, hf_downsample_prefix))
+
+        # no upsample in up_blocks.3
+        hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."
+        sd_upsample_prefix = f"output_blocks.{3 * i + 2}.{1 if i == 0 else 2}."
+        unet_conversion_map_layer.append((sd_upsample_prefix, hf_upsample_prefix))
+
+hf_mid_atn_prefix = "mid_block.attentions.0."
+sd_mid_atn_prefix = "middle_block.1."
+unet_conversion_map_layer.append((sd_mid_atn_prefix, hf_mid_atn_prefix))
+
+for j in range(2):
+    hf_mid_res_prefix = f"mid_block.resnets.{j}."
+    sd_mid_res_prefix = f"middle_block.{2 * j}."
+    unet_conversion_map_layer.append((sd_mid_res_prefix, hf_mid_res_prefix))
+
+
+def convert_unet_state_dict(unet_state_dict):
+    # buyer beware: this is a *brittle* function,
+    # and correct output requires that all of these pieces interact in
+    # the exact order in which I have arranged them.
+    mapping = {k: k for k in unet_state_dict.keys()}
+    for sd_name, hf_name in unet_conversion_map:
+        mapping[hf_name] = sd_name
+    for k, v in mapping.items():
+        if "resnets" in k:
+            for sd_part, hf_part in unet_conversion_map_resnet:
+                v = v.replace(hf_part, sd_part)
+            mapping[k] = v
+    for k, v in mapping.items():
+        for sd_part, hf_part in unet_conversion_map_layer:
+            v = v.replace(hf_part, sd_part)
+        mapping[k] = v
+    new_state_dict = {v: unet_state_dict[k] for k, v in mapping.items()}
+    return new_state_dict
+
+
+# ================#
+# VAE Conversion #
+# ================#
+
+vae_conversion_map = [
+    # (stable-diffusion, HF Diffusers)
+    ("nin_shortcut", "conv_shortcut"),
+    ("norm_out", "conv_norm_out"),
+    ("mid.attn_1.", "mid_block.attentions.0."),
+]
+
+for i in range(4):
+    # down_blocks have two resnets
+    for j in range(2):
+        hf_down_prefix = f"encoder.down_blocks.{i}.resnets.{j}."
+        sd_down_prefix = f"encoder.down.{i}.block.{j}."
+        vae_conversion_map.append((sd_down_prefix, hf_down_prefix))
+
+    if i < 3:
+        hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0."
+        sd_downsample_prefix = f"down.{i}.downsample."
+        vae_conversion_map.append((sd_downsample_prefix, hf_downsample_prefix))
+
+        hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."
+        sd_upsample_prefix = f"up.{3 - i}.upsample."
+        vae_conversion_map.append((sd_upsample_prefix, hf_upsample_prefix))
+
+    # up_blocks have three resnets
+    # also, up blocks in hf are numbered in reverse from sd
+    for j in range(3):
+        hf_up_prefix = f"decoder.up_blocks.{i}.resnets.{j}."
+        sd_up_prefix = f"decoder.up.{3 - i}.block.{j}."
+        vae_conversion_map.append((sd_up_prefix, hf_up_prefix))
+
+# this part accounts for mid blocks in both the encoder and the decoder
+for i in range(2):
+    hf_mid_res_prefix = f"mid_block.resnets.{i}."
+    sd_mid_res_prefix = f"mid.block_{i + 1}."
+    vae_conversion_map.append((sd_mid_res_prefix, hf_mid_res_prefix))
+
+vae_conversion_map_attn = [
+    # (stable-diffusion, HF Diffusers)
+    ("norm.", "group_norm."),
+    ("q.", "query."),
+    ("k.", "key."),
+    ("v.", "value."),
+    ("q.", "to_q."),
+    ("k.", "to_k."),
+    ("v.", "to_v."),
+    ("proj_out.", "to_out.0."),
+    ("proj_out.", "proj_attn."),
+]
+
+
+def reshape_weight_for_sd(w):
+    # convert HF linear weights to SD conv2d weights
+    return w.reshape(*w.shape, 1, 1)
+
+
+def convert_vae_state_dict(vae_state_dict):
+    mapping = {k: k for k in vae_state_dict.keys()}
+    for k, v in mapping.items():
+        for sd_part, hf_part in vae_conversion_map:
+            v = v.replace(hf_part, sd_part)
+        mapping[k] = v
+    for k, v in mapping.items():
+        if "attentions" in k:
+            for sd_part, hf_part in vae_conversion_map_attn:
+                v = v.replace(hf_part, sd_part)
+            mapping[k] = v
+    new_state_dict = {v: vae_state_dict[k] for k, v in mapping.items()}
+    weights_to_convert = ["q", "k", "v", "proj_out"]
+    for k, v in new_state_dict.items():
+        for weight_name in weights_to_convert:
+            if f"mid.attn_1.{weight_name}.weight" in k:
+                print(f"Reshaping {k} for SD format")
+                new_state_dict[k] = reshape_weight_for_sd(v)
+    return new_state_dict
+
+
+# =========================#
+# Text Encoder Conversion #
+# =========================#
+
+
+textenc_conversion_lst = [
+    # (stable-diffusion, HF Diffusers)
+    ("resblocks.", "text_model.encoder.layers."),
+    ("ln_1", "layer_norm1"),
+    ("ln_2", "layer_norm2"),
+    (".c_fc.", ".fc1."),
+    (".c_proj.", ".fc2."),
+    (".attn", ".self_attn"),
+    ("ln_final.", "transformer.text_model.final_layer_norm."),
+    ("token_embedding.weight", "transformer.text_model.embeddings.token_embedding.weight"),
+    ("positional_embedding", "transformer.text_model.embeddings.position_embedding.weight"),
+]
+protected = {re.escape(x[1]): x[0] for x in textenc_conversion_lst}
+textenc_pattern = re.compile("|".join(protected.keys()))
+
+# Ordering is from https://github.com/pytorch/pytorch/blob/master/test/cpp/api/modules.cpp
+code2idx = {"q": 0, "k": 1, "v": 2}
+
+
+def convert_text_enc_state_dict_v20(text_enc_dict, prefix=""):
+    new_state_dict = {}
+    capture_qkv_weight = {}
+    capture_qkv_bias = {}
+    for k, v in text_enc_dict.items():
+        if not k.startswith(prefix):
+            continue
+        if (
+                k.endswith(".self_attn.q_proj.weight")
+                or k.endswith(".self_attn.k_proj.weight")
+                or k.endswith(".self_attn.v_proj.weight")
+        ):
+            k_pre = k[: -len(".q_proj.weight")]
+            k_code = k[-len("q_proj.weight")]
+            if k_pre not in capture_qkv_weight:
+                capture_qkv_weight[k_pre] = [None, None, None]
+            capture_qkv_weight[k_pre][code2idx[k_code]] = v
+            continue
+
+        if (
+                k.endswith(".self_attn.q_proj.bias")
+                or k.endswith(".self_attn.k_proj.bias")
+                or k.endswith(".self_attn.v_proj.bias")
+        ):
+            k_pre = k[: -len(".q_proj.bias")]
+            k_code = k[-len("q_proj.bias")]
+            if k_pre not in capture_qkv_bias:
+                capture_qkv_bias[k_pre] = [None, None, None]
+            capture_qkv_bias[k_pre][code2idx[k_code]] = v
+            continue
+
+        relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k)
+        new_state_dict[relabelled_key] = v
+
+    for k_pre, tensors in capture_qkv_weight.items():
+        if None in tensors:
+            raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing")
+        relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre)
+        new_state_dict[relabelled_key + ".in_proj_weight"] = torch.cat(tensors)
+
+    for k_pre, tensors in capture_qkv_bias.items():
+        if None in tensors:
+            raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing")
+        relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre)
+        new_state_dict[relabelled_key + ".in_proj_bias"] = torch.cat(tensors)
+
+    return new_state_dict
+
+
+def convert_text_enc_state_dict(text_enc_dict):
+    return text_enc_dict
+
+

ldm_patched/modules/diffusers_load.py

@@ -0,0 +1,37 @@
+import json
+import os
+
+import ldm_patched.modules.sd
+
+def first_file(path, filenames):
+    for f in filenames:
+        p = os.path.join(path, f)
+        if os.path.exists(p):
+            return p
+    return None
+
+def load_diffusers(model_path, output_vae=True, output_clip=True, embedding_directory=None):
+    diffusion_model_names = ["diffusion_pytorch_model.fp16.safetensors", "diffusion_pytorch_model.safetensors", "diffusion_pytorch_model.fp16.bin", "diffusion_pytorch_model.bin"]
+    unet_path = first_file(os.path.join(model_path, "unet"), diffusion_model_names)
+    vae_path = first_file(os.path.join(model_path, "vae"), diffusion_model_names)
+
+    text_encoder_model_names = ["model.fp16.safetensors", "model.safetensors", "pytorch_model.fp16.bin", "pytorch_model.bin"]
+    text_encoder1_path = first_file(os.path.join(model_path, "text_encoder"), text_encoder_model_names)
+    text_encoder2_path = first_file(os.path.join(model_path, "text_encoder_2"), text_encoder_model_names)
+
+    text_encoder_paths = [text_encoder1_path]
+    if text_encoder2_path is not None:
+        text_encoder_paths.append(text_encoder2_path)
+
+    unet = ldm_patched.modules.sd.load_unet(unet_path)
+
+    clip = None
+    if output_clip:
+        clip = ldm_patched.modules.sd.load_clip(text_encoder_paths, embedding_directory=embedding_directory)
+
+    vae = None
+    if output_vae:
+        sd = ldm_patched.modules.utils.load_torch_file(vae_path)
+        vae = ldm_patched.modules.sd.VAE(sd=sd)
+
+    return (unet, clip, vae)

ldm_patched/modules/gligen.py

@@ -0,0 +1,341 @@
+import torch
+from torch import nn, einsum
+from ldm_patched.ldm.modules.attention import CrossAttention
+from inspect import isfunction
+
+
+def exists(val):
+    return val is not None
+
+
+def uniq(arr):
+    return{el: True for el in arr}.keys()
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+# feedforward
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=-1)
+        return x * torch.nn.functional.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = nn.Sequential(
+            nn.Linear(dim, inner_dim),
+            nn.GELU()
+        ) if not glu else GEGLU(dim, inner_dim)
+
+        self.net = nn.Sequential(
+            project_in,
+            nn.Dropout(dropout),
+            nn.Linear(inner_dim, dim_out)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+class GatedCrossAttentionDense(nn.Module):
+    def __init__(self, query_dim, context_dim, n_heads, d_head):
+        super().__init__()
+
+        self.attn = CrossAttention(
+            query_dim=query_dim,
+            context_dim=context_dim,
+            heads=n_heads,
+            dim_head=d_head)
+        self.ff = FeedForward(query_dim, glu=True)
+
+        self.norm1 = nn.LayerNorm(query_dim)
+        self.norm2 = nn.LayerNorm(query_dim)
+
+        self.register_parameter('alpha_attn', nn.Parameter(torch.tensor(0.)))
+        self.register_parameter('alpha_dense', nn.Parameter(torch.tensor(0.)))
+
+        # this can be useful: we can externally change magnitude of tanh(alpha)
+        # for example, when it is set to 0, then the entire model is same as
+        # original one
+        self.scale = 1
+
+    def forward(self, x, objs):
+
+        x = x + self.scale * \
+            torch.tanh(self.alpha_attn) * self.attn(self.norm1(x), objs, objs)
+        x = x + self.scale * \
+            torch.tanh(self.alpha_dense) * self.ff(self.norm2(x))
+
+        return x
+
+
+class GatedSelfAttentionDense(nn.Module):
+    def __init__(self, query_dim, context_dim, n_heads, d_head):
+        super().__init__()
+
+        # we need a linear projection since we need cat visual feature and obj
+        # feature
+        self.linear = nn.Linear(context_dim, query_dim)
+
+        self.attn = CrossAttention(
+            query_dim=query_dim,
+            context_dim=query_dim,
+            heads=n_heads,
+            dim_head=d_head)
+        self.ff = FeedForward(query_dim, glu=True)
+
+        self.norm1 = nn.LayerNorm(query_dim)
+        self.norm2 = nn.LayerNorm(query_dim)
+
+        self.register_parameter('alpha_attn', nn.Parameter(torch.tensor(0.)))
+        self.register_parameter('alpha_dense', nn.Parameter(torch.tensor(0.)))
+
+        # this can be useful: we can externally change magnitude of tanh(alpha)
+        # for example, when it is set to 0, then the entire model is same as
+        # original one
+        self.scale = 1
+
+    def forward(self, x, objs):
+
+        N_visual = x.shape[1]
+        objs = self.linear(objs)
+
+        x = x + self.scale * torch.tanh(self.alpha_attn) * self.attn(
+            self.norm1(torch.cat([x, objs], dim=1)))[:, 0:N_visual, :]
+        x = x + self.scale * \
+            torch.tanh(self.alpha_dense) * self.ff(self.norm2(x))
+
+        return x
+
+
+class GatedSelfAttentionDense2(nn.Module):
+    def __init__(self, query_dim, context_dim, n_heads, d_head):
+        super().__init__()
+
+        # we need a linear projection since we need cat visual feature and obj
+        # feature
+        self.linear = nn.Linear(context_dim, query_dim)
+
+        self.attn = CrossAttention(
+            query_dim=query_dim, context_dim=query_dim, dim_head=d_head)
+        self.ff = FeedForward(query_dim, glu=True)
+
+        self.norm1 = nn.LayerNorm(query_dim)
+        self.norm2 = nn.LayerNorm(query_dim)
+
+        self.register_parameter('alpha_attn', nn.Parameter(torch.tensor(0.)))
+        self.register_parameter('alpha_dense', nn.Parameter(torch.tensor(0.)))
+
+        # this can be useful: we can externally change magnitude of tanh(alpha)
+        # for example, when it is set to 0, then the entire model is same as
+        # original one
+        self.scale = 1
+
+    def forward(self, x, objs):
+
+        B, N_visual, _ = x.shape
+        B, N_ground, _ = objs.shape
+
+        objs = self.linear(objs)
+
+        # sanity check
+        size_v = math.sqrt(N_visual)
+        size_g = math.sqrt(N_ground)
+        assert int(size_v) == size_v, "Visual tokens must be square rootable"
+        assert int(size_g) == size_g, "Grounding tokens must be square rootable"
+        size_v = int(size_v)
+        size_g = int(size_g)
+
+        # select grounding token and resize it to visual token size as residual
+        out = self.attn(self.norm1(torch.cat([x, objs], dim=1)))[
+            :, N_visual:, :]
+        out = out.permute(0, 2, 1).reshape(B, -1, size_g, size_g)
+        out = torch.nn.functional.interpolate(
+            out, (size_v, size_v), mode='bicubic')
+        residual = out.reshape(B, -1, N_visual).permute(0, 2, 1)
+
+        # add residual to visual feature
+        x = x + self.scale * torch.tanh(self.alpha_attn) * residual
+        x = x + self.scale * \
+            torch.tanh(self.alpha_dense) * self.ff(self.norm2(x))
+
+        return x
+
+
+class FourierEmbedder():
+    def __init__(self, num_freqs=64, temperature=100):
+
+        self.num_freqs = num_freqs
+        self.temperature = temperature
+        self.freq_bands = temperature ** (torch.arange(num_freqs) / num_freqs)
+
+    @torch.no_grad()
+    def __call__(self, x, cat_dim=-1):
+        "x: arbitrary shape of tensor. dim: cat dim"
+        out = []
+        for freq in self.freq_bands:
+            out.append(torch.sin(freq * x))
+            out.append(torch.cos(freq * x))
+        return torch.cat(out, cat_dim)
+
+
+class PositionNet(nn.Module):
+    def __init__(self, in_dim, out_dim, fourier_freqs=8):
+        super().__init__()
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+
+        self.fourier_embedder = FourierEmbedder(num_freqs=fourier_freqs)
+        self.position_dim = fourier_freqs * 2 * 4  # 2 is sin&cos, 4 is xyxy
+
+        self.linears = nn.Sequential(
+            nn.Linear(self.in_dim + self.position_dim, 512),
+            nn.SiLU(),
+            nn.Linear(512, 512),
+            nn.SiLU(),
+            nn.Linear(512, out_dim),
+        )
+
+        self.null_positive_feature = torch.nn.Parameter(
+            torch.zeros([self.in_dim]))
+        self.null_position_feature = torch.nn.Parameter(
+            torch.zeros([self.position_dim]))
+
+    def forward(self, boxes, masks, positive_embeddings):
+        B, N, _ = boxes.shape
+        dtype = self.linears[0].weight.dtype
+        masks = masks.unsqueeze(-1).to(dtype)
+        positive_embeddings = positive_embeddings.to(dtype)
+
+        # embedding position (it may includes padding as placeholder)
+        xyxy_embedding = self.fourier_embedder(boxes.to(dtype))  # B*N*4 --> B*N*C
+
+        # learnable null embedding
+        positive_null = self.null_positive_feature.view(1, 1, -1)
+        xyxy_null = self.null_position_feature.view(1, 1, -1)
+
+        # replace padding with learnable null embedding
+        positive_embeddings = positive_embeddings * \
+            masks + (1 - masks) * positive_null
+        xyxy_embedding = xyxy_embedding * masks + (1 - masks) * xyxy_null
+
+        objs = self.linears(
+            torch.cat([positive_embeddings, xyxy_embedding], dim=-1))
+        assert objs.shape == torch.Size([B, N, self.out_dim])
+        return objs
+
+
+class Gligen(nn.Module):
+    def __init__(self, modules, position_net, key_dim):
+        super().__init__()
+        self.module_list = nn.ModuleList(modules)
+        self.position_net = position_net
+        self.key_dim = key_dim
+        self.max_objs = 30
+        self.current_device = torch.device("cpu")
+
+    def _set_position(self, boxes, masks, positive_embeddings):
+        objs = self.position_net(boxes, masks, positive_embeddings)
+        def func(x, extra_options):
+            key = extra_options["transformer_index"]
+            module = self.module_list[key]
+            return module(x, objs)
+        return func
+
+    def set_position(self, latent_image_shape, position_params, device):
+        batch, c, h, w = latent_image_shape
+        masks = torch.zeros([self.max_objs], device="cpu")
+        boxes = []
+        positive_embeddings = []
+        for p in position_params:
+            x1 = (p[4]) / w
+            y1 = (p[3]) / h
+            x2 = (p[4] + p[2]) / w
+            y2 = (p[3] + p[1]) / h
+            masks[len(boxes)] = 1.0
+            boxes += [torch.tensor((x1, y1, x2, y2)).unsqueeze(0)]
+            positive_embeddings += [p[0]]
+        append_boxes = []
+        append_conds = []
+        if len(boxes) < self.max_objs:
+            append_boxes = [torch.zeros(
+                [self.max_objs - len(boxes), 4], device="cpu")]
+            append_conds = [torch.zeros(
+                [self.max_objs - len(boxes), self.key_dim], device="cpu")]
+
+        box_out = torch.cat(
+            boxes + append_boxes).unsqueeze(0).repeat(batch, 1, 1)
+        masks = masks.unsqueeze(0).repeat(batch, 1)
+        conds = torch.cat(positive_embeddings +
+                          append_conds).unsqueeze(0).repeat(batch, 1, 1)
+        return self._set_position(
+            box_out.to(device),
+            masks.to(device),
+            conds.to(device))
+
+    def set_empty(self, latent_image_shape, device):
+        batch, c, h, w = latent_image_shape
+        masks = torch.zeros([self.max_objs], device="cpu").repeat(batch, 1)
+        box_out = torch.zeros([self.max_objs, 4],
+                              device="cpu").repeat(batch, 1, 1)
+        conds = torch.zeros([self.max_objs, self.key_dim],
+                            device="cpu").repeat(batch, 1, 1)
+        return self._set_position(
+            box_out.to(device),
+            masks.to(device),
+            conds.to(device))
+
+
+def load_gligen(sd):
+    sd_k = sd.keys()
+    output_list = []
+    key_dim = 768
+    for a in ["input_blocks", "middle_block", "output_blocks"]:
+        for b in range(20):
+            k_temp = filter(lambda k: "{}.{}.".format(a, b)
+                            in k and ".fuser." in k, sd_k)
+            k_temp = map(lambda k: (k, k.split(".fuser.")[-1]), k_temp)
+
+            n_sd = {}
+            for k in k_temp:
+                n_sd[k[1]] = sd[k[0]]
+            if len(n_sd) > 0:
+                query_dim = n_sd["linear.weight"].shape[0]
+                key_dim = n_sd["linear.weight"].shape[1]
+
+                if key_dim == 768:  # SD1.x
+                    n_heads = 8
+                    d_head = query_dim // n_heads
+                else:
+                    d_head = 64
+                    n_heads = query_dim // d_head
+
+                gated = GatedSelfAttentionDense(
+                    query_dim, key_dim, n_heads, d_head)
+                gated.load_state_dict(n_sd, strict=False)
+                output_list.append(gated)
+
+    if "position_net.null_positive_feature" in sd_k:
+        in_dim = sd["position_net.null_positive_feature"].shape[0]
+        out_dim = sd["position_net.linears.4.weight"].shape[0]
+
+        class WeightsLoader(torch.nn.Module):
+            pass
+        w = WeightsLoader()
+        w.position_net = PositionNet(in_dim, out_dim)
+        w.load_state_dict(sd, strict=False)
+
+    gligen = Gligen(output_list, w.position_net, key_dim)
+    return gligen

ldm_patched/modules/latent_formats.py

@@ -0,0 +1,35 @@
+
+class LatentFormat:
+    scale_factor = 1.0
+    latent_rgb_factors = None
+    taesd_decoder_name = None
+
+    def process_in(self, latent):
+        return latent * self.scale_factor
+
+    def process_out(self, latent):
+        return latent / self.scale_factor
+
+class SD15(LatentFormat):
+    def __init__(self, scale_factor=0.18215):
+        self.scale_factor = scale_factor
+        self.latent_rgb_factors = [
+                    #   R        G        B
+                    [ 0.3512,  0.2297,  0.3227],
+                    [ 0.3250,  0.4974,  0.2350],
+                    [-0.2829,  0.1762,  0.2721],
+                    [-0.2120, -0.2616, -0.7177]
+                ]
+        self.taesd_decoder_name = "taesd_decoder"
+
+class SDXL(LatentFormat):
+    def __init__(self):
+        self.scale_factor = 0.13025
+        self.latent_rgb_factors = [
+                    #   R        G        B
+                    [ 0.3920,  0.4054,  0.4549],
+                    [-0.2634, -0.0196,  0.0653],
+                    [ 0.0568,  0.1687, -0.0755],
+                    [-0.3112, -0.2359, -0.2076]
+                ]
+        self.taesd_decoder_name = "taesdxl_decoder"

ldm_patched/modules/lora.py

@@ -0,0 +1,224 @@
+import ldm_patched.modules.utils
+
+LORA_CLIP_MAP = {
+    "mlp.fc1": "mlp_fc1",
+    "mlp.fc2": "mlp_fc2",
+    "self_attn.k_proj": "self_attn_k_proj",
+    "self_attn.q_proj": "self_attn_q_proj",
+    "self_attn.v_proj": "self_attn_v_proj",
+    "self_attn.out_proj": "self_attn_out_proj",
+}
+
+
+def load_lora(lora, to_load):
+    patch_dict = {}
+    loaded_keys = set()
+    for x in to_load:
+        alpha_name = "{}.alpha".format(x)
+        alpha = None
+        if alpha_name in lora.keys():
+            alpha = lora[alpha_name].item()
+            loaded_keys.add(alpha_name)
+
+        regular_lora = "{}.lora_up.weight".format(x)
+        diffusers_lora = "{}_lora.up.weight".format(x)
+        transformers_lora = "{}.lora_linear_layer.up.weight".format(x)
+        A_name = None
+
+        if regular_lora in lora.keys():
+            A_name = regular_lora
+            B_name = "{}.lora_down.weight".format(x)
+            mid_name = "{}.lora_mid.weight".format(x)
+        elif diffusers_lora in lora.keys():
+            A_name = diffusers_lora
+            B_name = "{}_lora.down.weight".format(x)
+            mid_name = None
+        elif transformers_lora in lora.keys():
+            A_name = transformers_lora
+            B_name ="{}.lora_linear_layer.down.weight".format(x)
+            mid_name = None
+
+        if A_name is not None:
+            mid = None
+            if mid_name is not None and mid_name in lora.keys():
+                mid = lora[mid_name]
+                loaded_keys.add(mid_name)
+            patch_dict[to_load[x]] = ("lora", (lora[A_name], lora[B_name], alpha, mid))
+            loaded_keys.add(A_name)
+            loaded_keys.add(B_name)
+
+
+        ######## loha
+        hada_w1_a_name = "{}.hada_w1_a".format(x)
+        hada_w1_b_name = "{}.hada_w1_b".format(x)
+        hada_w2_a_name = "{}.hada_w2_a".format(x)
+        hada_w2_b_name = "{}.hada_w2_b".format(x)
+        hada_t1_name = "{}.hada_t1".format(x)
+        hada_t2_name = "{}.hada_t2".format(x)
+        if hada_w1_a_name in lora.keys():
+            hada_t1 = None
+            hada_t2 = None
+            if hada_t1_name in lora.keys():
+                hada_t1 = lora[hada_t1_name]
+                hada_t2 = lora[hada_t2_name]
+                loaded_keys.add(hada_t1_name)
+                loaded_keys.add(hada_t2_name)
+
+            patch_dict[to_load[x]] = ("loha", (lora[hada_w1_a_name], lora[hada_w1_b_name], alpha, lora[hada_w2_a_name], lora[hada_w2_b_name], hada_t1, hada_t2))
+            loaded_keys.add(hada_w1_a_name)
+            loaded_keys.add(hada_w1_b_name)
+            loaded_keys.add(hada_w2_a_name)
+            loaded_keys.add(hada_w2_b_name)
+
+
+        ######## lokr
+        lokr_w1_name = "{}.lokr_w1".format(x)
+        lokr_w2_name = "{}.lokr_w2".format(x)
+        lokr_w1_a_name = "{}.lokr_w1_a".format(x)
+        lokr_w1_b_name = "{}.lokr_w1_b".format(x)
+        lokr_t2_name = "{}.lokr_t2".format(x)
+        lokr_w2_a_name = "{}.lokr_w2_a".format(x)
+        lokr_w2_b_name = "{}.lokr_w2_b".format(x)
+
+        lokr_w1 = None
+        if lokr_w1_name in lora.keys():
+            lokr_w1 = lora[lokr_w1_name]
+            loaded_keys.add(lokr_w1_name)
+
+        lokr_w2 = None
+        if lokr_w2_name in lora.keys():
+            lokr_w2 = lora[lokr_w2_name]
+            loaded_keys.add(lokr_w2_name)
+
+        lokr_w1_a = None
+        if lokr_w1_a_name in lora.keys():
+            lokr_w1_a = lora[lokr_w1_a_name]
+            loaded_keys.add(lokr_w1_a_name)
+
+        lokr_w1_b = None
+        if lokr_w1_b_name in lora.keys():
+            lokr_w1_b = lora[lokr_w1_b_name]
+            loaded_keys.add(lokr_w1_b_name)
+
+        lokr_w2_a = None
+        if lokr_w2_a_name in lora.keys():
+            lokr_w2_a = lora[lokr_w2_a_name]
+            loaded_keys.add(lokr_w2_a_name)
+
+        lokr_w2_b = None
+        if lokr_w2_b_name in lora.keys():
+            lokr_w2_b = lora[lokr_w2_b_name]
+            loaded_keys.add(lokr_w2_b_name)
+
+        lokr_t2 = None
+        if lokr_t2_name in lora.keys():
+            lokr_t2 = lora[lokr_t2_name]
+            loaded_keys.add(lokr_t2_name)
+
+        if (lokr_w1 is not None) or (lokr_w2 is not None) or (lokr_w1_a is not None) or (lokr_w2_a is not None):
+            patch_dict[to_load[x]] = ("lokr", (lokr_w1, lokr_w2, alpha, lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2))
+
+        #glora
+        a1_name = "{}.a1.weight".format(x)
+        a2_name = "{}.a2.weight".format(x)
+        b1_name = "{}.b1.weight".format(x)
+        b2_name = "{}.b2.weight".format(x)
+        if a1_name in lora:
+            patch_dict[to_load[x]] = ("glora", (lora[a1_name], lora[a2_name], lora[b1_name], lora[b2_name], alpha))
+            loaded_keys.add(a1_name)
+            loaded_keys.add(a2_name)
+            loaded_keys.add(b1_name)
+            loaded_keys.add(b2_name)
+
+        w_norm_name = "{}.w_norm".format(x)
+        b_norm_name = "{}.b_norm".format(x)
+        w_norm = lora.get(w_norm_name, None)
+        b_norm = lora.get(b_norm_name, None)
+
+        if w_norm is not None:
+            loaded_keys.add(w_norm_name)
+            patch_dict[to_load[x]] = ("diff", (w_norm,))
+            if b_norm is not None:
+                loaded_keys.add(b_norm_name)
+                patch_dict["{}.bias".format(to_load[x][:-len(".weight")])] = ("diff", (b_norm,))
+
+        diff_name = "{}.diff".format(x)
+        diff_weight = lora.get(diff_name, None)
+        if diff_weight is not None:
+            patch_dict[to_load[x]] = ("diff", (diff_weight,))
+            loaded_keys.add(diff_name)
+
+        diff_bias_name = "{}.diff_b".format(x)
+        diff_bias = lora.get(diff_bias_name, None)
+        if diff_bias is not None:
+            patch_dict["{}.bias".format(to_load[x][:-len(".weight")])] = ("diff", (diff_bias,))
+            loaded_keys.add(diff_bias_name)
+
+    for x in lora.keys():
+        if x not in loaded_keys:
+            print("lora key not loaded", x)
+    return patch_dict
+
+def model_lora_keys_clip(model, key_map={}):
+    sdk = model.state_dict().keys()
+
+    text_model_lora_key = "lora_te_text_model_encoder_layers_{}_{}"
+    clip_l_present = False
+    for b in range(32): #TODO: clean up
+        for c in LORA_CLIP_MAP:
+            k = "clip_h.transformer.text_model.encoder.layers.{}.{}.weight".format(b, c)
+            if k in sdk:
+                lora_key = text_model_lora_key.format(b, LORA_CLIP_MAP[c])
+                key_map[lora_key] = k
+                lora_key = "lora_te1_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c])
+                key_map[lora_key] = k
+                lora_key = "text_encoder.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
+                key_map[lora_key] = k
+
+            k = "clip_l.transformer.text_model.encoder.layers.{}.{}.weight".format(b, c)
+            if k in sdk:
+                lora_key = text_model_lora_key.format(b, LORA_CLIP_MAP[c])
+                key_map[lora_key] = k
+                lora_key = "lora_te1_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #SDXL base
+                key_map[lora_key] = k
+                clip_l_present = True
+                lora_key = "text_encoder.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
+                key_map[lora_key] = k
+
+            k = "clip_g.transformer.text_model.encoder.layers.{}.{}.weight".format(b, c)
+            if k in sdk:
+                if clip_l_present:
+                    lora_key = "lora_te2_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #SDXL base
+                    key_map[lora_key] = k
+                    lora_key = "text_encoder_2.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
+                    key_map[lora_key] = k
+                else:
+                    lora_key = "lora_te_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #TODO: test if this is correct for SDXL-Refiner
+                    key_map[lora_key] = k
+                    lora_key = "text_encoder.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
+                    key_map[lora_key] = k
+
+    return key_map
+
+def model_lora_keys_unet(model, key_map={}):
+    sdk = model.state_dict().keys()
+
+    for k in sdk:
+        if k.startswith("diffusion_model.") and k.endswith(".weight"):
+            key_lora = k[len("diffusion_model."):-len(".weight")].replace(".", "_")
+            key_map["lora_unet_{}".format(key_lora)] = k
+
+    diffusers_keys = ldm_patched.modules.utils.unet_to_diffusers(model.model_config.unet_config)
+    for k in diffusers_keys:
+        if k.endswith(".weight"):
+            unet_key = "diffusion_model.{}".format(diffusers_keys[k])
+            key_lora = k[:-len(".weight")].replace(".", "_")
+            key_map["lora_unet_{}".format(key_lora)] = unet_key
+
+            diffusers_lora_prefix = ["", "unet."]
+            for p in diffusers_lora_prefix:
+                diffusers_lora_key = "{}{}".format(p, k[:-len(".weight")].replace(".to_", ".processor.to_"))
+                if diffusers_lora_key.endswith(".to_out.0"):
+                    diffusers_lora_key = diffusers_lora_key[:-2]
+                key_map[diffusers_lora_key] = unet_key
+    return key_map

ldm_patched/modules/model_base.py

@@ -0,0 +1,330 @@
+import torch
+from ldm_patched.ldm.modules.diffusionmodules.openaimodel import UNetModel
+from ldm_patched.ldm.modules.encoders.noise_aug_modules import CLIPEmbeddingNoiseAugmentation
+from ldm_patched.ldm.modules.diffusionmodules.openaimodel import Timestep
+import ldm_patched.modules.model_management
+import ldm_patched.modules.conds
+import ldm_patched.modules.ops
+from enum import Enum
+import contextlib
+from . import utils
+
+class ModelType(Enum):
+    EPS = 1
+    V_PREDICTION = 2
+    V_PREDICTION_EDM = 3
+
+
+from ldm_patched.modules.model_sampling import EPS, V_PREDICTION, ModelSamplingDiscrete, ModelSamplingContinuousEDM
+
+
+def model_sampling(model_config, model_type):
+    s = ModelSamplingDiscrete
+
+    if model_type == ModelType.EPS:
+        c = EPS
+    elif model_type == ModelType.V_PREDICTION:
+        c = V_PREDICTION
+    elif model_type == ModelType.V_PREDICTION_EDM:
+        c = V_PREDICTION
+        s = ModelSamplingContinuousEDM
+
+    class ModelSampling(s, c):
+        pass
+
+    return ModelSampling(model_config)
+
+
+class BaseModel(torch.nn.Module):
+    def __init__(self, model_config, model_type=ModelType.EPS, device=None):
+        super().__init__()
+
+        unet_config = model_config.unet_config
+        self.latent_format = model_config.latent_format
+        self.model_config = model_config
+        self.manual_cast_dtype = model_config.manual_cast_dtype
+
+        if not unet_config.get("disable_unet_model_creation", False):
+            if self.manual_cast_dtype is not None:
+                operations = ldm_patched.modules.ops.manual_cast
+            else:
+                operations = ldm_patched.modules.ops.disable_weight_init
+            self.diffusion_model = UNetModel(**unet_config, device=device, operations=operations)
+        self.model_type = model_type
+        self.model_sampling = model_sampling(model_config, model_type)
+
+        self.adm_channels = unet_config.get("adm_in_channels", None)
+        if self.adm_channels is None:
+            self.adm_channels = 0
+        self.inpaint_model = False
+        print("model_type", model_type.name)
+        print("UNet ADM Dimension", self.adm_channels)
+
+    def apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, **kwargs):
+        sigma = t
+        xc = self.model_sampling.calculate_input(sigma, x)
+        if c_concat is not None:
+            xc = torch.cat([xc] + [c_concat], dim=1)
+
+        context = c_crossattn
+        dtype = self.get_dtype()
+
+        if self.manual_cast_dtype is not None:
+            dtype = self.manual_cast_dtype
+
+        xc = xc.to(dtype)
+        t = self.model_sampling.timestep(t).float()
+        context = context.to(dtype)
+        extra_conds = {}
+        for o in kwargs:
+            extra = kwargs[o]
+            if hasattr(extra, "to"):
+                extra = extra.to(dtype)
+            extra_conds[o] = extra
+
+        model_output = self.diffusion_model(xc, t, context=context, control=control, transformer_options=transformer_options, **extra_conds).float()
+        return self.model_sampling.calculate_denoised(sigma, model_output, x)
+
+    def get_dtype(self):
+        return self.diffusion_model.dtype
+
+    def is_adm(self):
+        return self.adm_channels > 0
+
+    def encode_adm(self, **kwargs):
+        return None
+
+    def extra_conds(self, **kwargs):
+        out = {}
+        if self.inpaint_model:
+            concat_keys = ("mask", "masked_image")
+            cond_concat = []
+            denoise_mask = kwargs.get("denoise_mask", None)
+            latent_image = kwargs.get("latent_image", None)
+            noise = kwargs.get("noise", None)
+            device = kwargs["device"]
+
+            def blank_inpaint_image_like(latent_image):
+                blank_image = torch.ones_like(latent_image)
+                # these are the values for "zero" in pixel space translated to latent space
+                blank_image[:,0] *= 0.8223
+                blank_image[:,1] *= -0.6876
+                blank_image[:,2] *= 0.6364
+                blank_image[:,3] *= 0.1380
+                return blank_image
+
+            for ck in concat_keys:
+                if denoise_mask is not None:
+                    if ck == "mask":
+                        cond_concat.append(denoise_mask[:,:1].to(device))
+                    elif ck == "masked_image":
+                        cond_concat.append(latent_image.to(device)) #NOTE: the latent_image should be masked by the mask in pixel space
+                else:
+                    if ck == "mask":
+                        cond_concat.append(torch.ones_like(noise)[:,:1])
+                    elif ck == "masked_image":
+                        cond_concat.append(blank_inpaint_image_like(noise))
+            data = torch.cat(cond_concat, dim=1)
+            out['c_concat'] = ldm_patched.modules.conds.CONDNoiseShape(data)
+        adm = self.encode_adm(**kwargs)
+        if adm is not None:
+            out['y'] = ldm_patched.modules.conds.CONDRegular(adm)
+        return out
+
+    def load_model_weights(self, sd, unet_prefix=""):
+        to_load = {}
+        keys = list(sd.keys())
+        for k in keys:
+            if k.startswith(unet_prefix):
+                to_load[k[len(unet_prefix):]] = sd.pop(k)
+
+        to_load = self.model_config.process_unet_state_dict(to_load)
+        m, u = self.diffusion_model.load_state_dict(to_load, strict=False)
+        if len(m) > 0:
+            print("unet missing:", m)
+
+        if len(u) > 0:
+            print("unet unexpected:", u)
+        del to_load
+        return self
+
+    def process_latent_in(self, latent):
+        return self.latent_format.process_in(latent)
+
+    def process_latent_out(self, latent):
+        return self.latent_format.process_out(latent)
+
+    def state_dict_for_saving(self, clip_state_dict, vae_state_dict):
+        clip_state_dict = self.model_config.process_clip_state_dict_for_saving(clip_state_dict)
+        unet_sd = self.diffusion_model.state_dict()
+        unet_state_dict = {}
+        for k in unet_sd:
+            unet_state_dict[k] = ldm_patched.modules.model_management.resolve_lowvram_weight(unet_sd[k], self.diffusion_model, k)
+
+        unet_state_dict = self.model_config.process_unet_state_dict_for_saving(unet_state_dict)
+        vae_state_dict = self.model_config.process_vae_state_dict_for_saving(vae_state_dict)
+        if self.get_dtype() == torch.float16:
+            clip_state_dict = utils.convert_sd_to(clip_state_dict, torch.float16)
+            vae_state_dict = utils.convert_sd_to(vae_state_dict, torch.float16)
+
+        if self.model_type == ModelType.V_PREDICTION:
+            unet_state_dict["v_pred"] = torch.tensor([])
+
+        return {**unet_state_dict, **vae_state_dict, **clip_state_dict}
+
+    def set_inpaint(self):
+        self.inpaint_model = True
+
+    def memory_required(self, input_shape):
+        if ldm_patched.modules.model_management.xformers_enabled() or ldm_patched.modules.model_management.pytorch_attention_flash_attention():
+            dtype = self.get_dtype()
+            if self.manual_cast_dtype is not None:
+                dtype = self.manual_cast_dtype
+            #TODO: this needs to be tweaked
+            area = input_shape[0] * input_shape[2] * input_shape[3]
+            return (area * ldm_patched.modules.model_management.dtype_size(dtype) / 50) * (1024 * 1024)
+        else:
+            #TODO: this formula might be too aggressive since I tweaked the sub-quad and split algorithms to use less memory.
+            area = input_shape[0] * input_shape[2] * input_shape[3]
+            return (((area * 0.6) / 0.9) + 1024) * (1024 * 1024)
+
+
+def unclip_adm(unclip_conditioning, device, noise_augmentor, noise_augment_merge=0.0):
+    adm_inputs = []
+    weights = []
+    noise_aug = []
+    for unclip_cond in unclip_conditioning:
+        for adm_cond in unclip_cond["clip_vision_output"].image_embeds:
+            weight = unclip_cond["strength"]
+            noise_augment = unclip_cond["noise_augmentation"]
+            noise_level = round((noise_augmentor.max_noise_level - 1) * noise_augment)
+            c_adm, noise_level_emb = noise_augmentor(adm_cond.to(device), noise_level=torch.tensor([noise_level], device=device))
+            adm_out = torch.cat((c_adm, noise_level_emb), 1) * weight
+            weights.append(weight)
+            noise_aug.append(noise_augment)
+            adm_inputs.append(adm_out)
+
+    if len(noise_aug) > 1:
+        adm_out = torch.stack(adm_inputs).sum(0)
+        noise_augment = noise_augment_merge
+        noise_level = round((noise_augmentor.max_noise_level - 1) * noise_augment)
+        c_adm, noise_level_emb = noise_augmentor(adm_out[:, :noise_augmentor.time_embed.dim], noise_level=torch.tensor([noise_level], device=device))
+        adm_out = torch.cat((c_adm, noise_level_emb), 1)
+
+    return adm_out
+
+class SD21UNCLIP(BaseModel):
+    def __init__(self, model_config, noise_aug_config, model_type=ModelType.V_PREDICTION, device=None):
+        super().__init__(model_config, model_type, device=device)
+        self.noise_augmentor = CLIPEmbeddingNoiseAugmentation(**noise_aug_config)
+
+    def encode_adm(self, **kwargs):
+        unclip_conditioning = kwargs.get("unclip_conditioning", None)
+        device = kwargs["device"]
+        if unclip_conditioning is None:
+            return torch.zeros((1, self.adm_channels))
+        else:
+            return unclip_adm(unclip_conditioning, device, self.noise_augmentor, kwargs.get("unclip_noise_augment_merge", 0.05))
+
+def sdxl_pooled(args, noise_augmentor):
+    if "unclip_conditioning" in args:
+        return unclip_adm(args.get("unclip_conditioning", None), args["device"], noise_augmentor)[:,:1280]
+    else:
+        return args["pooled_output"]
+
+class SDXLRefiner(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.EPS, device=None):
+        super().__init__(model_config, model_type, device=device)
+        self.embedder = Timestep(256)
+        self.noise_augmentor = CLIPEmbeddingNoiseAugmentation(**{"noise_schedule_config": {"timesteps": 1000, "beta_schedule": "squaredcos_cap_v2"}, "timestep_dim": 1280})
+
+    def encode_adm(self, **kwargs):
+        clip_pooled = sdxl_pooled(kwargs, self.noise_augmentor)
+        width = kwargs.get("width", 768)
+        height = kwargs.get("height", 768)
+        crop_w = kwargs.get("crop_w", 0)
+        crop_h = kwargs.get("crop_h", 0)
+
+        if kwargs.get("prompt_type", "") == "negative":
+            aesthetic_score = kwargs.get("aesthetic_score", 2.5)
+        else:
+            aesthetic_score = kwargs.get("aesthetic_score", 6)
+
+        out = []
+        out.append(self.embedder(torch.Tensor([height])))
+        out.append(self.embedder(torch.Tensor([width])))
+        out.append(self.embedder(torch.Tensor([crop_h])))
+        out.append(self.embedder(torch.Tensor([crop_w])))
+        out.append(self.embedder(torch.Tensor([aesthetic_score])))
+        flat = torch.flatten(torch.cat(out)).unsqueeze(dim=0).repeat(clip_pooled.shape[0], 1)
+        return torch.cat((clip_pooled.to(flat.device), flat), dim=1)
+
+class SDXL(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.EPS, device=None):
+        super().__init__(model_config, model_type, device=device)
+        self.embedder = Timestep(256)
+        self.noise_augmentor = CLIPEmbeddingNoiseAugmentation(**{"noise_schedule_config": {"timesteps": 1000, "beta_schedule": "squaredcos_cap_v2"}, "timestep_dim": 1280})
+
+    def encode_adm(self, **kwargs):
+        clip_pooled = sdxl_pooled(kwargs, self.noise_augmentor)
+        width = kwargs.get("width", 768)
+        height = kwargs.get("height", 768)
+        crop_w = kwargs.get("crop_w", 0)
+        crop_h = kwargs.get("crop_h", 0)
+        target_width = kwargs.get("target_width", width)
+        target_height = kwargs.get("target_height", height)
+
+        out = []
+        out.append(self.embedder(torch.Tensor([height])))
+        out.append(self.embedder(torch.Tensor([width])))
+        out.append(self.embedder(torch.Tensor([crop_h])))
+        out.append(self.embedder(torch.Tensor([crop_w])))
+        out.append(self.embedder(torch.Tensor([target_height])))
+        out.append(self.embedder(torch.Tensor([target_width])))
+        flat = torch.flatten(torch.cat(out)).unsqueeze(dim=0).repeat(clip_pooled.shape[0], 1)
+        return torch.cat((clip_pooled.to(flat.device), flat), dim=1)
+
+class SVD_img2vid(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.V_PREDICTION_EDM, device=None):
+        super().__init__(model_config, model_type, device=device)
+        self.embedder = Timestep(256)
+
+    def encode_adm(self, **kwargs):
+        fps_id = kwargs.get("fps", 6) - 1
+        motion_bucket_id = kwargs.get("motion_bucket_id", 127)
+        augmentation = kwargs.get("augmentation_level", 0)
+
+        out = []
+        out.append(self.embedder(torch.Tensor([fps_id])))
+        out.append(self.embedder(torch.Tensor([motion_bucket_id])))
+        out.append(self.embedder(torch.Tensor([augmentation])))
+
+        flat = torch.flatten(torch.cat(out)).unsqueeze(dim=0)
+        return flat
+
+    def extra_conds(self, **kwargs):
+        out = {}
+        adm = self.encode_adm(**kwargs)
+        if adm is not None:
+            out['y'] = ldm_patched.modules.conds.CONDRegular(adm)
+
+        latent_image = kwargs.get("concat_latent_image", None)
+        noise = kwargs.get("noise", None)
+        device = kwargs["device"]
+
+        if latent_image is None:
+            latent_image = torch.zeros_like(noise)
+
+        if latent_image.shape[1:] != noise.shape[1:]:
+            latent_image = utils.common_upscale(latent_image, noise.shape[-1], noise.shape[-2], "bilinear", "center")
+
+        latent_image = utils.resize_to_batch_size(latent_image, noise.shape[0])
+
+        out['c_concat'] = ldm_patched.modules.conds.CONDNoiseShape(latent_image)
+
+        if "time_conditioning" in kwargs:
+            out["time_context"] = ldm_patched.modules.conds.CONDCrossAttn(kwargs["time_conditioning"])
+
+        out['image_only_indicator'] = ldm_patched.modules.conds.CONDConstant(torch.zeros((1,), device=device))
+        out['num_video_frames'] = ldm_patched.modules.conds.CONDConstant(noise.shape[0])
+        return out

ldm_patched/modules/model_detection.py

@@ -0,0 +1,308 @@
+import ldm_patched.modules.supported_models
+import ldm_patched.modules.supported_models_base
+
+def count_blocks(state_dict_keys, prefix_string):
+    count = 0
+    while True:
+        c = False
+        for k in state_dict_keys:
+            if k.startswith(prefix_string.format(count)):
+                c = True
+                break
+        if c == False:
+            break
+        count += 1
+    return count
+
+def calculate_transformer_depth(prefix, state_dict_keys, state_dict):
+    context_dim = None
+    use_linear_in_transformer = False
+
+    transformer_prefix = prefix + "1.transformer_blocks."
+    transformer_keys = sorted(list(filter(lambda a: a.startswith(transformer_prefix), state_dict_keys)))
+    if len(transformer_keys) > 0:
+        last_transformer_depth = count_blocks(state_dict_keys, transformer_prefix + '{}')
+        context_dim = state_dict['{}0.attn2.to_k.weight'.format(transformer_prefix)].shape[1]
+        use_linear_in_transformer = len(state_dict['{}1.proj_in.weight'.format(prefix)].shape) == 2
+        time_stack = '{}1.time_stack.0.attn1.to_q.weight'.format(prefix) in state_dict or '{}1.time_mix_blocks.0.attn1.to_q.weight'.format(prefix) in state_dict
+        return last_transformer_depth, context_dim, use_linear_in_transformer, time_stack
+    return None
+
+def detect_unet_config(state_dict, key_prefix, dtype):
+    state_dict_keys = list(state_dict.keys())
+
+    unet_config = {
+        "use_checkpoint": False,
+        "image_size": 32,
+        "out_channels": 4,
+        "use_spatial_transformer": True,
+        "legacy": False
+    }
+
+    y_input = '{}label_emb.0.0.weight'.format(key_prefix)
+    if y_input in state_dict_keys:
+        unet_config["num_classes"] = "sequential"
+        unet_config["adm_in_channels"] = state_dict[y_input].shape[1]
+    else:
+        unet_config["adm_in_channels"] = None
+
+    unet_config["dtype"] = dtype
+    model_channels = state_dict['{}input_blocks.0.0.weight'.format(key_prefix)].shape[0]
+    in_channels = state_dict['{}input_blocks.0.0.weight'.format(key_prefix)].shape[1]
+
+    num_res_blocks = []
+    channel_mult = []
+    attention_resolutions = []
+    transformer_depth = []
+    transformer_depth_output = []
+    context_dim = None
+    use_linear_in_transformer = False
+
+    video_model = False
+
+    current_res = 1
+    count = 0
+
+    last_res_blocks = 0
+    last_channel_mult = 0
+
+    input_block_count = count_blocks(state_dict_keys, '{}input_blocks'.format(key_prefix) + '.{}.')
+    for count in range(input_block_count):
+        prefix = '{}input_blocks.{}.'.format(key_prefix, count)
+        prefix_output = '{}output_blocks.{}.'.format(key_prefix, input_block_count - count - 1)
+
+        block_keys = sorted(list(filter(lambda a: a.startswith(prefix), state_dict_keys)))
+        if len(block_keys) == 0:
+            break
+
+        block_keys_output = sorted(list(filter(lambda a: a.startswith(prefix_output), state_dict_keys)))
+
+        if "{}0.op.weight".format(prefix) in block_keys: #new layer
+            num_res_blocks.append(last_res_blocks)
+            channel_mult.append(last_channel_mult)
+
+            current_res *= 2
+            last_res_blocks = 0
+            last_channel_mult = 0
+            out = calculate_transformer_depth(prefix_output, state_dict_keys, state_dict)
+            if out is not None:
+                transformer_depth_output.append(out[0])
+            else:
+                transformer_depth_output.append(0)
+        else:
+            res_block_prefix = "{}0.in_layers.0.weight".format(prefix)
+            if res_block_prefix in block_keys:
+                last_res_blocks += 1
+                last_channel_mult = state_dict["{}0.out_layers.3.weight".format(prefix)].shape[0] // model_channels
+
+                out = calculate_transformer_depth(prefix, state_dict_keys, state_dict)
+                if out is not None:
+                    transformer_depth.append(out[0])
+                    if context_dim is None:
+                        context_dim = out[1]
+                        use_linear_in_transformer = out[2]
+                        video_model = out[3]
+                else:
+                    transformer_depth.append(0)
+
+            res_block_prefix = "{}0.in_layers.0.weight".format(prefix_output)
+            if res_block_prefix in block_keys_output:
+                out = calculate_transformer_depth(prefix_output, state_dict_keys, state_dict)
+                if out is not None:
+                    transformer_depth_output.append(out[0])
+                else:
+                    transformer_depth_output.append(0)
+
+
+    num_res_blocks.append(last_res_blocks)
+    channel_mult.append(last_channel_mult)
+    if "{}middle_block.1.proj_in.weight".format(key_prefix) in state_dict_keys:
+        transformer_depth_middle = count_blocks(state_dict_keys, '{}middle_block.1.transformer_blocks.'.format(key_prefix) + '{}')
+    else:
+        transformer_depth_middle = -1
+
+    unet_config["in_channels"] = in_channels
+    unet_config["model_channels"] = model_channels
+    unet_config["num_res_blocks"] = num_res_blocks
+    unet_config["transformer_depth"] = transformer_depth
+    unet_config["transformer_depth_output"] = transformer_depth_output
+    unet_config["channel_mult"] = channel_mult
+    unet_config["transformer_depth_middle"] = transformer_depth_middle
+    unet_config['use_linear_in_transformer'] = use_linear_in_transformer
+    unet_config["context_dim"] = context_dim
+
+    if video_model:
+        unet_config["extra_ff_mix_layer"] = True
+        unet_config["use_spatial_context"] = True
+        unet_config["merge_strategy"] = "learned_with_images"
+        unet_config["merge_factor"] = 0.0
+        unet_config["video_kernel_size"] = [3, 1, 1]
+        unet_config["use_temporal_resblock"] = True
+        unet_config["use_temporal_attention"] = True
+    else:
+        unet_config["use_temporal_resblock"] = False
+        unet_config["use_temporal_attention"] = False
+
+    return unet_config
+
+def model_config_from_unet_config(unet_config):
+    for model_config in ldm_patched.modules.supported_models.models:
+        if model_config.matches(unet_config):
+            return model_config(unet_config)
+
+    print("no match", unet_config)
+    return None
+
+def model_config_from_unet(state_dict, unet_key_prefix, dtype, use_base_if_no_match=False):
+    unet_config = detect_unet_config(state_dict, unet_key_prefix, dtype)
+    model_config = model_config_from_unet_config(unet_config)
+    if model_config is None and use_base_if_no_match:
+        return ldm_patched.modules.supported_models_base.BASE(unet_config)
+    else:
+        return model_config
+
+def convert_config(unet_config):
+    new_config = unet_config.copy()
+    num_res_blocks = new_config.get("num_res_blocks", None)
+    channel_mult = new_config.get("channel_mult", None)
+
+    if isinstance(num_res_blocks, int):
+        num_res_blocks = len(channel_mult) * [num_res_blocks]
+
+    if "attention_resolutions" in new_config:
+        attention_resolutions = new_config.pop("attention_resolutions")
+        transformer_depth = new_config.get("transformer_depth", None)
+        transformer_depth_middle = new_config.get("transformer_depth_middle", None)
+
+        if isinstance(transformer_depth, int):
+            transformer_depth = len(channel_mult) * [transformer_depth]
+        if transformer_depth_middle is None:
+            transformer_depth_middle =  transformer_depth[-1]
+        t_in = []
+        t_out = []
+        s = 1
+        for i in range(len(num_res_blocks)):
+            res = num_res_blocks[i]
+            d = 0
+            if s in attention_resolutions:
+                d = transformer_depth[i]
+
+            t_in += [d] * res
+            t_out += [d] * (res + 1)
+            s *= 2
+        transformer_depth = t_in
+        transformer_depth_output = t_out
+        new_config["transformer_depth"] = t_in
+        new_config["transformer_depth_output"] = t_out
+        new_config["transformer_depth_middle"] = transformer_depth_middle
+
+    new_config["num_res_blocks"] = num_res_blocks
+    return new_config
+
+
+def unet_config_from_diffusers_unet(state_dict, dtype):
+    match = {}
+    transformer_depth = []
+
+    attn_res = 1
+    down_blocks = count_blocks(state_dict, "down_blocks.{}")
+    for i in range(down_blocks):
+        attn_blocks = count_blocks(state_dict, "down_blocks.{}.attentions.".format(i) + '{}')
+        for ab in range(attn_blocks):
+            transformer_count = count_blocks(state_dict, "down_blocks.{}.attentions.{}.transformer_blocks.".format(i, ab) + '{}')
+            transformer_depth.append(transformer_count)
+            if transformer_count > 0:
+                match["context_dim"] = state_dict["down_blocks.{}.attentions.{}.transformer_blocks.0.attn2.to_k.weight".format(i, ab)].shape[1]
+
+        attn_res *= 2
+        if attn_blocks == 0:
+            transformer_depth.append(0)
+            transformer_depth.append(0)
+
+    match["transformer_depth"] = transformer_depth
+
+    match["model_channels"] = state_dict["conv_in.weight"].shape[0]
+    match["in_channels"] = state_dict["conv_in.weight"].shape[1]
+    match["adm_in_channels"] = None
+    if "class_embedding.linear_1.weight" in state_dict:
+        match["adm_in_channels"] = state_dict["class_embedding.linear_1.weight"].shape[1]
+    elif "add_embedding.linear_1.weight" in state_dict:
+        match["adm_in_channels"] = state_dict["add_embedding.linear_1.weight"].shape[1]
+
+    SDXL = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+            'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320,
+            'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 2, 2, 10, 10], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 10,
+            'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 2, 2, 2, 10, 10, 10],
+            'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    SDXL_refiner = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+                    'num_classes': 'sequential', 'adm_in_channels': 2560, 'dtype': dtype, 'in_channels': 4, 'model_channels': 384,
+                    'num_res_blocks': [2, 2, 2, 2], 'transformer_depth': [0, 0, 4, 4, 4, 4, 0, 0], 'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 4,
+                    'use_linear_in_transformer': True, 'context_dim': 1280, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 4, 4, 4, 4, 4, 4, 0, 0, 0],
+                    'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    SD21 = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+            'adm_in_channels': None, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 'num_res_blocks': [2, 2, 2, 2],
+            'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0], 'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1, 'use_linear_in_transformer': True,
+            'context_dim': 1024, 'num_head_channels': 64, 'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
+            'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    SD21_uncliph = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+                    'num_classes': 'sequential', 'adm_in_channels': 2048, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320,
+                    'num_res_blocks': [2, 2, 2, 2], 'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0], 'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1,
+                    'use_linear_in_transformer': True, 'context_dim': 1024, 'num_head_channels': 64, 'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
+                    'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    SD21_unclipl = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+                    'num_classes': 'sequential', 'adm_in_channels': 1536, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320,
+                    'num_res_blocks': [2, 2, 2, 2], 'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0], 'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1,
+                    'use_linear_in_transformer': True, 'context_dim': 1024, 'num_head_channels': 64, 'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
+                    'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    SD15 = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, 'adm_in_channels': None,
+            'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 'num_res_blocks': [2, 2, 2, 2], 'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0],
+            'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1, 'use_linear_in_transformer': False, 'context_dim': 768, 'num_heads': 8,
+            'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
+            'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    SDXL_mid_cnet = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+                     'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320,
+                     'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 0, 0, 1, 1], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 1,
+                     'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 0, 0, 0, 1, 1, 1],
+                     'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    SDXL_small_cnet = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+                       'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320,
+                       'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 0, 0, 0, 0], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 0,
+                       'use_linear_in_transformer': True, 'num_head_channels': 64, 'context_dim': 1, 'transformer_depth_output': [0, 0, 0, 0, 0, 0, 0, 0, 0],
+                       'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    SDXL_diffusers_inpaint = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+                              'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 9, 'model_channels': 320,
+                              'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 2, 2, 10, 10], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 10,
+                              'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 2, 2, 2, 10, 10, 10],
+                              'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    SSD_1B = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+              'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320,
+              'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 2, 2, 4, 4], 'transformer_depth_output': [0, 0, 0, 1, 1, 2, 10, 4, 4],
+              'channel_mult': [1, 2, 4], 'transformer_depth_middle': -1, 'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64,
+              'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    supported_models = [SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet, SDXL_diffusers_inpaint, SSD_1B]
+
+    for unet_config in supported_models:
+        matches = True
+        for k in match:
+            if match[k] != unet_config[k]:
+                matches = False
+                break
+        if matches:
+            return convert_config(unet_config)
+    return None
+
+def model_config_from_diffusers_unet(state_dict, dtype):
+    unet_config = unet_config_from_diffusers_unet(state_dict, dtype)
+    if unet_config is not None:
+        return model_config_from_unet_config(unet_config)
+    return None

ldm_patched/modules/model_management.py

@@ -0,0 +1,775 @@
+import psutil
+from enum import Enum
+from ldm_patched.modules.args_parser import args
+import ldm_patched.modules.utils
+import torch
+import sys
+
+class VRAMState(Enum):
+    DISABLED = 0    #No vram present: no need to move models to vram
+    NO_VRAM = 1     #Very low vram: enable all the options to save vram
+    LOW_VRAM = 2
+    NORMAL_VRAM = 3
+    HIGH_VRAM = 4
+    SHARED = 5      #No dedicated vram: memory shared between CPU and GPU but models still need to be moved between both.
+
+class CPUState(Enum):
+    GPU = 0
+    CPU = 1
+    MPS = 2
+
+# Determine VRAM State
+vram_state = VRAMState.NORMAL_VRAM
+set_vram_to = VRAMState.NORMAL_VRAM
+cpu_state = CPUState.GPU
+
+total_vram = 0
+
+lowvram_available = True
+xpu_available = False
+
+directml_enabled = False
+if args.directml is not None:
+    import torch_directml
+    directml_enabled = True
+    device_index = args.directml
+    if device_index < 0:
+        directml_device = torch_directml.device()
+    else:
+        directml_device = torch_directml.device(device_index)
+    print("Using directml with device:", torch_directml.device_name(device_index))
+    # torch_directml.disable_tiled_resources(True)
+    lowvram_available = False #TODO: need to find a way to get free memory in directml before this can be enabled by default.
+
+try:
+    import intel_extension_for_pytorch as ipex
+    if torch.xpu.is_available():
+        xpu_available = True
+except:
+    pass
+
+try:
+    if torch.backends.mps.is_available():
+        cpu_state = CPUState.MPS
+        import torch.mps
+except:
+    pass
+
+if args.always_cpu:
+    cpu_state = CPUState.CPU
+
+def is_intel_xpu():
+    global cpu_state
+    global xpu_available
+    if cpu_state == CPUState.GPU:
+        if xpu_available:
+            return True
+    return False
+
+def get_torch_device():
+    global directml_enabled
+    global cpu_state
+    if directml_enabled:
+        global directml_device
+        return directml_device
+    if cpu_state == CPUState.MPS:
+        return torch.device("mps")
+    if cpu_state == CPUState.CPU:
+        return torch.device("cpu")
+    else:
+        if is_intel_xpu():
+            return torch.device("xpu")
+        else:
+            return torch.device(torch.cuda.current_device())
+
+def get_total_memory(dev=None, torch_total_too=False):
+    global directml_enabled
+    if dev is None:
+        dev = get_torch_device()
+
+    if hasattr(dev, 'type') and (dev.type == 'cpu' or dev.type == 'mps'):
+        mem_total = psutil.virtual_memory().total
+        mem_total_torch = mem_total
+    else:
+        if directml_enabled:
+            mem_total = 1024 * 1024 * 1024 #TODO
+            mem_total_torch = mem_total
+        elif is_intel_xpu():
+            stats = torch.xpu.memory_stats(dev)
+            mem_reserved = stats['reserved_bytes.all.current']
+            mem_total = torch.xpu.get_device_properties(dev).total_memory
+            mem_total_torch = mem_reserved
+        else:
+            stats = torch.cuda.memory_stats(dev)
+            mem_reserved = stats['reserved_bytes.all.current']
+            _, mem_total_cuda = torch.cuda.mem_get_info(dev)
+            mem_total_torch = mem_reserved
+            mem_total = mem_total_cuda
+
+    if torch_total_too:
+        return (mem_total, mem_total_torch)
+    else:
+        return mem_total
+
+total_vram = get_total_memory(get_torch_device()) / (1024 * 1024)
+total_ram = psutil.virtual_memory().total / (1024 * 1024)
+print("Total VRAM {:0.0f} MB, total RAM {:0.0f} MB".format(total_vram, total_ram))
+if not args.always_normal_vram and not args.always_cpu:
+    if lowvram_available and total_vram <= 4096:
+        print("Trying to enable lowvram mode because your GPU seems to have 4GB or less. If you don't want this use: --always-normal-vram")
+        set_vram_to = VRAMState.LOW_VRAM
+
+try:
+    OOM_EXCEPTION = torch.cuda.OutOfMemoryError
+except:
+    OOM_EXCEPTION = Exception
+
+XFORMERS_VERSION = ""
+XFORMERS_ENABLED_VAE = True
+if args.disable_xformers:
+    XFORMERS_IS_AVAILABLE = False
+else:
+    try:
+        import xformers
+        import xformers.ops
+        XFORMERS_IS_AVAILABLE = True
+        try:
+            XFORMERS_IS_AVAILABLE = xformers._has_cpp_library
+        except:
+            pass
+        try:
+            XFORMERS_VERSION = xformers.version.__version__
+            print("xformers version:", XFORMERS_VERSION)
+            if XFORMERS_VERSION.startswith("0.0.18"):
+                print()
+                print("WARNING: This version of xformers has a major bug where you will get black images when generating high resolution images.")
+                print("Please downgrade or upgrade xformers to a different version.")
+                print()
+                XFORMERS_ENABLED_VAE = False
+        except:
+            pass
+    except:
+        XFORMERS_IS_AVAILABLE = False
+
+def is_nvidia():
+    global cpu_state
+    if cpu_state == CPUState.GPU:
+        if torch.version.cuda:
+            return True
+    return False
+
+ENABLE_PYTORCH_ATTENTION = False
+if args.attention_pytorch:
+    ENABLE_PYTORCH_ATTENTION = True
+    XFORMERS_IS_AVAILABLE = False
+
+VAE_DTYPE = torch.float32
+
+try:
+    if is_nvidia():
+        torch_version = torch.version.__version__
+        if int(torch_version[0]) >= 2:
+            if ENABLE_PYTORCH_ATTENTION == False and args.attention_split == False and args.attention_quad == False:
+                ENABLE_PYTORCH_ATTENTION = True
+            if torch.cuda.is_bf16_supported():
+                VAE_DTYPE = torch.bfloat16
+    if is_intel_xpu():
+        if args.attention_split == False and args.attention_quad == False:
+            ENABLE_PYTORCH_ATTENTION = True
+except:
+    pass
+
+if is_intel_xpu():
+    VAE_DTYPE = torch.bfloat16
+
+if args.vae_in_fp16:
+    VAE_DTYPE = torch.float16
+elif args.vae_in_bf16:
+    VAE_DTYPE = torch.bfloat16
+elif args.vae_in_fp32:
+    VAE_DTYPE = torch.float32
+
+
+if ENABLE_PYTORCH_ATTENTION:
+    torch.backends.cuda.enable_math_sdp(True)
+    torch.backends.cuda.enable_flash_sdp(True)
+    torch.backends.cuda.enable_mem_efficient_sdp(True)
+
+if args.always_low_vram:
+    set_vram_to = VRAMState.LOW_VRAM
+    lowvram_available = True
+elif args.always_no_vram:
+    set_vram_to = VRAMState.NO_VRAM
+elif args.always_high_vram or args.always_gpu:
+    vram_state = VRAMState.HIGH_VRAM
+
+FORCE_FP32 = False
+FORCE_FP16 = False
+if args.all_in_fp32:
+    print("Forcing FP32, if this improves things please report it.")
+    FORCE_FP32 = True
+
+if args.all_in_fp16:
+    print("Forcing FP16.")
+    FORCE_FP16 = True
+
+if lowvram_available:
+    try:
+        import accelerate
+        if set_vram_to in (VRAMState.LOW_VRAM, VRAMState.NO_VRAM):
+            vram_state = set_vram_to
+    except Exception as e:
+        import traceback
+        print(traceback.format_exc())
+        print("ERROR: LOW VRAM MODE NEEDS accelerate.")
+        lowvram_available = False
+
+
+if cpu_state != CPUState.GPU:
+    vram_state = VRAMState.DISABLED
+
+if cpu_state == CPUState.MPS:
+    vram_state = VRAMState.SHARED
+
+print(f"Set vram state to: {vram_state.name}")
+
+ALWAYS_VRAM_OFFLOAD = args.always_offload_from_vram
+
+if ALWAYS_VRAM_OFFLOAD:
+    print("Always offload VRAM")
+
+def get_torch_device_name(device):
+    if hasattr(device, 'type'):
+        if device.type == "cuda":
+            try:
+                allocator_backend = torch.cuda.get_allocator_backend()
+            except:
+                allocator_backend = ""
+            return "{} {} : {}".format(device, torch.cuda.get_device_name(device), allocator_backend)
+        else:
+            return "{}".format(device.type)
+    elif is_intel_xpu():
+        return "{} {}".format(device, torch.xpu.get_device_name(device))
+    else:
+        return "CUDA {}: {}".format(device, torch.cuda.get_device_name(device))
+
+try:
+    print("Device:", get_torch_device_name(get_torch_device()))
+except:
+    print("Could not pick default device.")
+
+print("VAE dtype:", VAE_DTYPE)
+
+current_loaded_models = []
+
+class LoadedModel:
+    def __init__(self, model):
+        self.model = model
+        self.model_accelerated = False
+        self.device = model.load_device
+
+    def model_memory(self):
+        return self.model.model_size()
+
+    def model_memory_required(self, device):
+        if device == self.model.current_device:
+            return 0
+        else:
+            return self.model_memory()
+
+    def model_load(self, lowvram_model_memory=0):
+        patch_model_to = None
+        if lowvram_model_memory == 0:
+            patch_model_to = self.device
+
+        self.model.model_patches_to(self.device)
+        self.model.model_patches_to(self.model.model_dtype())
+
+        try:
+            self.real_model = self.model.patch_model(device_to=patch_model_to) #TODO: do something with loras and offloading to CPU
+        except Exception as e:
+            self.model.unpatch_model(self.model.offload_device)
+            self.model_unload()
+            raise e
+
+        if lowvram_model_memory > 0:
+            print("loading in lowvram mode", lowvram_model_memory/(1024 * 1024))
+            device_map = accelerate.infer_auto_device_map(self.real_model, max_memory={0: "{}MiB".format(lowvram_model_memory // (1024 * 1024)), "cpu": "16GiB"})
+            accelerate.dispatch_model(self.real_model, device_map=device_map, main_device=self.device)
+            self.model_accelerated = True
+
+        if is_intel_xpu() and not args.disable_ipex_hijack:
+            self.real_model = torch.xpu.optimize(self.real_model.eval(), inplace=True, auto_kernel_selection=True, graph_mode=True)
+
+        return self.real_model
+
+    def model_unload(self):
+        if self.model_accelerated:
+            accelerate.hooks.remove_hook_from_submodules(self.real_model)
+            self.model_accelerated = False
+
+        self.model.unpatch_model(self.model.offload_device)
+        self.model.model_patches_to(self.model.offload_device)
+
+    def __eq__(self, other):
+        return self.model is other.model
+
+def minimum_inference_memory():
+    return (1024 * 1024 * 1024)
+
+def unload_model_clones(model):
+    to_unload = []
+    for i in range(len(current_loaded_models)):
+        if model.is_clone(current_loaded_models[i].model):
+            to_unload = [i] + to_unload
+
+    for i in to_unload:
+        print("unload clone", i)
+        current_loaded_models.pop(i).model_unload()
+
+def free_memory(memory_required, device, keep_loaded=[]):
+    unloaded_model = False
+    for i in range(len(current_loaded_models) -1, -1, -1):
+        if not ALWAYS_VRAM_OFFLOAD:
+            if get_free_memory(device) > memory_required:
+                break
+        shift_model = current_loaded_models[i]
+        if shift_model.device == device:
+            if shift_model not in keep_loaded:
+                m = current_loaded_models.pop(i)
+                m.model_unload()
+                del m
+                unloaded_model = True
+
+    if unloaded_model:
+        soft_empty_cache()
+    else:
+        if vram_state != VRAMState.HIGH_VRAM:
+            mem_free_total, mem_free_torch = get_free_memory(device, torch_free_too=True)
+            if mem_free_torch > mem_free_total * 0.25:
+                soft_empty_cache()
+
+def load_models_gpu(models, memory_required=0):
+    global vram_state
+
+    inference_memory = minimum_inference_memory()
+    extra_mem = max(inference_memory, memory_required)
+
+    models_to_load = []
+    models_already_loaded = []
+    for x in models:
+        loaded_model = LoadedModel(x)
+
+        if loaded_model in current_loaded_models:
+            index = current_loaded_models.index(loaded_model)
+            current_loaded_models.insert(0, current_loaded_models.pop(index))
+            models_already_loaded.append(loaded_model)
+        else:
+            if hasattr(x, "model"):
+                print(f"Requested to load {x.model.__class__.__name__}")
+            models_to_load.append(loaded_model)
+
+    if len(models_to_load) == 0:
+        devs = set(map(lambda a: a.device, models_already_loaded))
+        for d in devs:
+            if d != torch.device("cpu"):
+                free_memory(extra_mem, d, models_already_loaded)
+        return
+
+    print(f"Loading {len(models_to_load)} new model{'s' if len(models_to_load) > 1 else ''}")
+
+    total_memory_required = {}
+    for loaded_model in models_to_load:
+        unload_model_clones(loaded_model.model)
+        total_memory_required[loaded_model.device] = total_memory_required.get(loaded_model.device, 0) + loaded_model.model_memory_required(loaded_model.device)
+
+    for device in total_memory_required:
+        if device != torch.device("cpu"):
+            free_memory(total_memory_required[device] * 1.3 + extra_mem, device, models_already_loaded)
+
+    for loaded_model in models_to_load:
+        model = loaded_model.model
+        torch_dev = model.load_device
+        if is_device_cpu(torch_dev):
+            vram_set_state = VRAMState.DISABLED
+        else:
+            vram_set_state = vram_state
+        lowvram_model_memory = 0
+        if lowvram_available and (vram_set_state == VRAMState.LOW_VRAM or vram_set_state == VRAMState.NORMAL_VRAM):
+            model_size = loaded_model.model_memory_required(torch_dev)
+            current_free_mem = get_free_memory(torch_dev)
+            lowvram_model_memory = int(max(256 * (1024 * 1024), (current_free_mem - 1024 * (1024 * 1024)) / 1.3 ))
+            if model_size > (current_free_mem - inference_memory): #only switch to lowvram if really necessary
+                vram_set_state = VRAMState.LOW_VRAM
+            else:
+                lowvram_model_memory = 0
+
+        if vram_set_state == VRAMState.NO_VRAM:
+            lowvram_model_memory = 256 * 1024 * 1024
+
+        cur_loaded_model = loaded_model.model_load(lowvram_model_memory)
+        current_loaded_models.insert(0, loaded_model)
+    return
+
+
+def load_model_gpu(model):
+    return load_models_gpu([model])
+
+def cleanup_models():
+    to_delete = []
+    for i in range(len(current_loaded_models)):
+        if sys.getrefcount(current_loaded_models[i].model) <= 2:
+            to_delete = [i] + to_delete
+
+    for i in to_delete:
+        x = current_loaded_models.pop(i)
+        x.model_unload()
+        del x
+
+def dtype_size(dtype):
+    dtype_size = 4
+    if dtype == torch.float16 or dtype == torch.bfloat16:
+        dtype_size = 2
+    elif dtype == torch.float32:
+        dtype_size = 4
+    else:
+        try:
+            dtype_size = dtype.itemsize
+        except: #Old pytorch doesn't have .itemsize
+            pass
+    return dtype_size
+
+def unet_offload_device():
+    if vram_state == VRAMState.HIGH_VRAM:
+        return get_torch_device()
+    else:
+        return torch.device("cpu")
+
+def unet_inital_load_device(parameters, dtype):
+    torch_dev = get_torch_device()
+    if vram_state == VRAMState.HIGH_VRAM:
+        return torch_dev
+
+    cpu_dev = torch.device("cpu")
+    if ALWAYS_VRAM_OFFLOAD:
+        return cpu_dev
+
+    model_size = dtype_size(dtype) * parameters
+
+    mem_dev = get_free_memory(torch_dev)
+    mem_cpu = get_free_memory(cpu_dev)
+    if mem_dev > mem_cpu and model_size < mem_dev:
+        return torch_dev
+    else:
+        return cpu_dev
+
+def unet_dtype(device=None, model_params=0):
+    if args.unet_in_bf16:
+        return torch.bfloat16
+    if args.unet_in_fp16:
+        return torch.float16
+    if args.unet_in_fp8_e4m3fn:
+        return torch.float8_e4m3fn
+    if args.unet_in_fp8_e5m2:
+        return torch.float8_e5m2
+    if should_use_fp16(device=device, model_params=model_params):
+        return torch.float16
+    return torch.float32
+
+# None means no manual cast
+def unet_manual_cast(weight_dtype, inference_device):
+    if weight_dtype == torch.float32:
+        return None
+
+    fp16_supported = ldm_patched.modules.model_management.should_use_fp16(inference_device, prioritize_performance=False)
+    if fp16_supported and weight_dtype == torch.float16:
+        return None
+
+    if fp16_supported:
+        return torch.float16
+    else:
+        return torch.float32
+
+def text_encoder_offload_device():
+    if args.always_gpu:
+        return get_torch_device()
+    else:
+        return torch.device("cpu")
+
+def text_encoder_device():
+    if args.always_gpu:
+        return get_torch_device()
+    elif vram_state == VRAMState.HIGH_VRAM or vram_state == VRAMState.NORMAL_VRAM:
+        if is_intel_xpu():
+            return torch.device("cpu")
+        if should_use_fp16(prioritize_performance=False):
+            return get_torch_device()
+        else:
+            return torch.device("cpu")
+    else:
+        return torch.device("cpu")
+
+def text_encoder_dtype(device=None):
+    if args.clip_in_fp8_e4m3fn:
+        return torch.float8_e4m3fn
+    elif args.clip_in_fp8_e5m2:
+        return torch.float8_e5m2
+    elif args.clip_in_fp16:
+        return torch.float16
+    elif args.clip_in_fp32:
+        return torch.float32
+
+    if is_device_cpu(device):
+        return torch.float16
+
+    if should_use_fp16(device, prioritize_performance=False):
+        return torch.float16
+    else:
+        return torch.float32
+
+def intermediate_device():
+    if args.always_gpu:
+        return get_torch_device()
+    else:
+        return torch.device("cpu")
+
+def vae_device():
+    return get_torch_device()
+
+def vae_offload_device():
+    if args.always_gpu:
+        return get_torch_device()
+    else:
+        return torch.device("cpu")
+
+def vae_dtype():
+    global VAE_DTYPE
+    return VAE_DTYPE
+
+def get_autocast_device(dev):
+    if hasattr(dev, 'type'):
+        return dev.type
+    return "cuda"
+
+def supports_dtype(device, dtype): #TODO
+    if dtype == torch.float32:
+        return True
+    if is_device_cpu(device):
+        return False
+    if dtype == torch.float16:
+        return True
+    if dtype == torch.bfloat16:
+        return True
+    return False
+
+def cast_to_device(tensor, device, dtype, copy=False):
+    device_supports_cast = False
+    if tensor.dtype == torch.float32 or tensor.dtype == torch.float16:
+        device_supports_cast = True
+    elif tensor.dtype == torch.bfloat16:
+        if hasattr(device, 'type') and device.type.startswith("cuda"):
+            device_supports_cast = True
+        elif is_intel_xpu():
+            device_supports_cast = True
+
+    non_blocking = True
+    if is_device_mps(device):
+        non_blocking = False #pytorch bug? mps doesn't support non blocking
+
+    if device_supports_cast:
+        if copy:
+            if tensor.device == device:
+                return tensor.to(dtype, copy=copy, non_blocking=non_blocking)
+            return tensor.to(device, copy=copy, non_blocking=non_blocking).to(dtype, non_blocking=non_blocking)
+        else:
+            return tensor.to(device, non_blocking=non_blocking).to(dtype, non_blocking=non_blocking)
+    else:
+        return tensor.to(device, dtype, copy=copy, non_blocking=non_blocking)
+
+def xformers_enabled():
+    global directml_enabled
+    global cpu_state
+    if cpu_state != CPUState.GPU:
+        return False
+    if is_intel_xpu():
+        return False
+    if directml_enabled:
+        return False
+    return XFORMERS_IS_AVAILABLE
+
+
+def xformers_enabled_vae():
+    enabled = xformers_enabled()
+    if not enabled:
+        return False
+
+    return XFORMERS_ENABLED_VAE
+
+def pytorch_attention_enabled():
+    global ENABLE_PYTORCH_ATTENTION
+    return ENABLE_PYTORCH_ATTENTION
+
+def pytorch_attention_flash_attention():
+    global ENABLE_PYTORCH_ATTENTION
+    if ENABLE_PYTORCH_ATTENTION:
+        #TODO: more reliable way of checking for flash attention?
+        if is_nvidia(): #pytorch flash attention only works on Nvidia
+            return True
+    return False
+
+def get_free_memory(dev=None, torch_free_too=False):
+    global directml_enabled
+    if dev is None:
+        dev = get_torch_device()
+
+    if hasattr(dev, 'type') and (dev.type == 'cpu' or dev.type == 'mps'):
+        mem_free_total = psutil.virtual_memory().available
+        mem_free_torch = mem_free_total
+    else:
+        if directml_enabled:
+            mem_free_total = 1024 * 1024 * 1024 #TODO
+            mem_free_torch = mem_free_total
+        elif is_intel_xpu():
+            stats = torch.xpu.memory_stats(dev)
+            mem_active = stats['active_bytes.all.current']
+            mem_allocated = stats['allocated_bytes.all.current']
+            mem_reserved = stats['reserved_bytes.all.current']
+            mem_free_torch = mem_reserved - mem_active
+            mem_free_total = torch.xpu.get_device_properties(dev).total_memory - mem_allocated
+        else:
+            stats = torch.cuda.memory_stats(dev)
+            mem_active = stats['active_bytes.all.current']
+            mem_reserved = stats['reserved_bytes.all.current']
+            mem_free_cuda, _ = torch.cuda.mem_get_info(dev)
+            mem_free_torch = mem_reserved - mem_active
+            mem_free_total = mem_free_cuda + mem_free_torch
+
+    if torch_free_too:
+        return (mem_free_total, mem_free_torch)
+    else:
+        return mem_free_total
+
+def cpu_mode():
+    global cpu_state
+    return cpu_state == CPUState.CPU
+
+def mps_mode():
+    global cpu_state
+    return cpu_state == CPUState.MPS
+
+def is_device_cpu(device):
+    if hasattr(device, 'type'):
+        if (device.type == 'cpu'):
+            return True
+    return False
+
+def is_device_mps(device):
+    if hasattr(device, 'type'):
+        if (device.type == 'mps'):
+            return True
+    return False
+
+def should_use_fp16(device=None, model_params=0, prioritize_performance=True):
+    global directml_enabled
+
+    if device is not None:
+        if is_device_cpu(device):
+            return False
+
+    if FORCE_FP16:
+        return True
+
+    if device is not None: #TODO
+        if is_device_mps(device):
+            return False
+
+    if FORCE_FP32:
+        return False
+
+    if directml_enabled:
+        return False
+
+    if cpu_mode() or mps_mode():
+        return False #TODO ?
+
+    if is_intel_xpu():
+        return True
+
+    if torch.cuda.is_bf16_supported():
+        return True
+
+    props = torch.cuda.get_device_properties("cuda")
+    if props.major < 6:
+        return False
+
+    fp16_works = False
+    #FP16 is confirmed working on a 1080 (GP104) but it's a bit slower than FP32 so it should only be enabled
+    #when the model doesn't actually fit on the card
+    #TODO: actually test if GP106 and others have the same type of behavior
+    nvidia_10_series = ["1080", "1070", "titan x", "p3000", "p3200", "p4000", "p4200", "p5000", "p5200", "p6000", "1060", "1050"]
+    for x in nvidia_10_series:
+        if x in props.name.lower():
+            fp16_works = True
+
+    if fp16_works:
+        free_model_memory = (get_free_memory() * 0.9 - minimum_inference_memory())
+        if (not prioritize_performance) or model_params * 4 > free_model_memory:
+            return True
+
+    if props.major < 7:
+        return False
+
+    #FP16 is just broken on these cards
+    nvidia_16_series = ["1660", "1650", "1630", "T500", "T550", "T600", "MX550", "MX450", "CMP 30HX", "T2000", "T1000", "T1200"]
+    for x in nvidia_16_series:
+        if x in props.name:
+            return False
+
+    return True
+
+def soft_empty_cache(force=False):
+    global cpu_state
+    if cpu_state == CPUState.MPS:
+        torch.mps.empty_cache()
+    elif is_intel_xpu():
+        torch.xpu.empty_cache()
+    elif torch.cuda.is_available():
+        if force or is_nvidia(): #This seems to make things worse on ROCm so I only do it for cuda
+            torch.cuda.empty_cache()
+            torch.cuda.ipc_collect()
+
+def resolve_lowvram_weight(weight, model, key):
+    if weight.device == torch.device("meta"): #lowvram NOTE: this depends on the inner working of the accelerate library so it might break.
+        key_split = key.split('.')              # I have no idea why they don't just leave the weight there instead of using the meta device.
+        op = ldm_patched.modules.utils.get_attr(model, '.'.join(key_split[:-1]))
+        weight = op._hf_hook.weights_map[key_split[-1]]
+    return weight
+
+#TODO: might be cleaner to put this somewhere else
+import threading
+
+class InterruptProcessingException(Exception):
+    pass
+
+interrupt_processing_mutex = threading.RLock()
+
+interrupt_processing = False
+def interrupt_current_processing(value=True):
+    global interrupt_processing
+    global interrupt_processing_mutex
+    with interrupt_processing_mutex:
+        interrupt_processing = value
+
+def processing_interrupted():
+    global interrupt_processing
+    global interrupt_processing_mutex
+    with interrupt_processing_mutex:
+        return interrupt_processing
+
+def throw_exception_if_processing_interrupted():
+    global interrupt_processing
+    global interrupt_processing_mutex
+    with interrupt_processing_mutex:
+        if interrupt_processing:
+            interrupt_processing = False
+            raise InterruptProcessingException()

ldm_patched/modules/model_patcher.py

@@ -0,0 +1,349 @@
+import torch
+import copy
+import inspect
+
+import ldm_patched.modules.utils
+import ldm_patched.modules.model_management
+
+class ModelPatcher:
+    def __init__(self, model, load_device, offload_device, size=0, current_device=None, weight_inplace_update=False):
+        self.size = size
+        self.model = model
+        self.patches = {}
+        self.backup = {}
+        self.object_patches = {}
+        self.object_patches_backup = {}
+        self.model_options = {"transformer_options":{}}
+        self.model_size()
+        self.load_device = load_device
+        self.offload_device = offload_device
+        if current_device is None:
+            self.current_device = self.offload_device
+        else:
+            self.current_device = current_device
+
+        self.weight_inplace_update = weight_inplace_update
+
+    def model_size(self):
+        if self.size > 0:
+            return self.size
+        model_sd = self.model.state_dict()
+        size = 0
+        for k in model_sd:
+            t = model_sd[k]
+            size += t.nelement() * t.element_size()
+        self.size = size
+        self.model_keys = set(model_sd.keys())
+        return size
+
+    def clone(self):
+        n = ModelPatcher(self.model, self.load_device, self.offload_device, self.size, self.current_device, weight_inplace_update=self.weight_inplace_update)
+        n.patches = {}
+        for k in self.patches:
+            n.patches[k] = self.patches[k][:]
+
+        n.object_patches = self.object_patches.copy()
+        n.model_options = copy.deepcopy(self.model_options)
+        n.model_keys = self.model_keys
+        return n
+
+    def is_clone(self, other):
+        if hasattr(other, 'model') and self.model is other.model:
+            return True
+        return False
+
+    def memory_required(self, input_shape):
+        return self.model.memory_required(input_shape=input_shape)
+
+    def set_model_sampler_cfg_function(self, sampler_cfg_function):
+        if len(inspect.signature(sampler_cfg_function).parameters) == 3:
+            self.model_options["sampler_cfg_function"] = lambda args: sampler_cfg_function(args["cond"], args["uncond"], args["cond_scale"]) #Old way
+        else:
+            self.model_options["sampler_cfg_function"] = sampler_cfg_function
+
+    def set_model_unet_function_wrapper(self, unet_wrapper_function):
+        self.model_options["model_function_wrapper"] = unet_wrapper_function
+
+    def set_model_patch(self, patch, name):
+        to = self.model_options["transformer_options"]
+        if "patches" not in to:
+            to["patches"] = {}
+        to["patches"][name] = to["patches"].get(name, []) + [patch]
+
+    def set_model_patch_replace(self, patch, name, block_name, number):
+        to = self.model_options["transformer_options"]
+        if "patches_replace" not in to:
+            to["patches_replace"] = {}
+        if name not in to["patches_replace"]:
+            to["patches_replace"][name] = {}
+        to["patches_replace"][name][(block_name, number)] = patch
+
+    def set_model_attn1_patch(self, patch):
+        self.set_model_patch(patch, "attn1_patch")
+
+    def set_model_attn2_patch(self, patch):
+        self.set_model_patch(patch, "attn2_patch")
+
+    def set_model_attn1_replace(self, patch, block_name, number):
+        self.set_model_patch_replace(patch, "attn1", block_name, number)
+
+    def set_model_attn2_replace(self, patch, block_name, number):
+        self.set_model_patch_replace(patch, "attn2", block_name, number)
+
+    def set_model_attn1_output_patch(self, patch):
+        self.set_model_patch(patch, "attn1_output_patch")
+
+    def set_model_attn2_output_patch(self, patch):
+        self.set_model_patch(patch, "attn2_output_patch")
+
+    def set_model_input_block_patch(self, patch):
+        self.set_model_patch(patch, "input_block_patch")
+
+    def set_model_input_block_patch_after_skip(self, patch):
+        self.set_model_patch(patch, "input_block_patch_after_skip")
+
+    def set_model_output_block_patch(self, patch):
+        self.set_model_patch(patch, "output_block_patch")
+
+    def add_object_patch(self, name, obj):
+        self.object_patches[name] = obj
+
+    def model_patches_to(self, device):
+        to = self.model_options["transformer_options"]
+        if "patches" in to:
+            patches = to["patches"]
+            for name in patches:
+                patch_list = patches[name]
+                for i in range(len(patch_list)):
+                    if hasattr(patch_list[i], "to"):
+                        patch_list[i] = patch_list[i].to(device)
+        if "patches_replace" in to:
+            patches = to["patches_replace"]
+            for name in patches:
+                patch_list = patches[name]
+                for k in patch_list:
+                    if hasattr(patch_list[k], "to"):
+                        patch_list[k] = patch_list[k].to(device)
+        if "model_function_wrapper" in self.model_options:
+            wrap_func = self.model_options["model_function_wrapper"]
+            if hasattr(wrap_func, "to"):
+                self.model_options["model_function_wrapper"] = wrap_func.to(device)
+
+    def model_dtype(self):
+        if hasattr(self.model, "get_dtype"):
+            return self.model.get_dtype()
+
+    def add_patches(self, patches, strength_patch=1.0, strength_model=1.0):
+        p = set()
+        for k in patches:
+            if k in self.model_keys:
+                p.add(k)
+                current_patches = self.patches.get(k, [])
+                current_patches.append((strength_patch, patches[k], strength_model))
+                self.patches[k] = current_patches
+
+        return list(p)
+
+    def get_key_patches(self, filter_prefix=None):
+        ldm_patched.modules.model_management.unload_model_clones(self)
+        model_sd = self.model_state_dict()
+        p = {}
+        for k in model_sd:
+            if filter_prefix is not None:
+                if not k.startswith(filter_prefix):
+                    continue
+            if k in self.patches:
+                p[k] = [model_sd[k]] + self.patches[k]
+            else:
+                p[k] = (model_sd[k],)
+        return p
+
+    def model_state_dict(self, filter_prefix=None):
+        sd = self.model.state_dict()
+        keys = list(sd.keys())
+        if filter_prefix is not None:
+            for k in keys:
+                if not k.startswith(filter_prefix):
+                    sd.pop(k)
+        return sd
+
+    def patch_model(self, device_to=None):
+        for k in self.object_patches:
+            old = getattr(self.model, k)
+            if k not in self.object_patches_backup:
+                self.object_patches_backup[k] = old
+            setattr(self.model, k, self.object_patches[k])
+
+        model_sd = self.model_state_dict()
+        for key in self.patches:
+            if key not in model_sd:
+                print("could not patch. key doesn't exist in model:", key)
+                continue
+
+            weight = model_sd[key]
+
+            inplace_update = self.weight_inplace_update
+
+            if key not in self.backup:
+                self.backup[key] = weight.to(device=self.offload_device, copy=inplace_update)
+
+            if device_to is not None:
+                temp_weight = ldm_patched.modules.model_management.cast_to_device(weight, device_to, torch.float32, copy=True)
+            else:
+                temp_weight = weight.to(torch.float32, copy=True)
+            out_weight = self.calculate_weight(self.patches[key], temp_weight, key).to(weight.dtype)
+            if inplace_update:
+                ldm_patched.modules.utils.copy_to_param(self.model, key, out_weight)
+            else:
+                ldm_patched.modules.utils.set_attr(self.model, key, out_weight)
+            del temp_weight
+
+        if device_to is not None:
+            self.model.to(device_to)
+            self.current_device = device_to
+
+        return self.model
+
+    def calculate_weight(self, patches, weight, key):
+        for p in patches:
+            alpha = p[0]
+            v = p[1]
+            strength_model = p[2]
+
+            if strength_model != 1.0:
+                weight *= strength_model
+
+            if isinstance(v, list):
+                v = (self.calculate_weight(v[1:], v[0].clone(), key), )
+
+            if len(v) == 1:
+                patch_type = "diff"
+            elif len(v) == 2:
+                patch_type = v[0]
+                v = v[1]
+
+            if patch_type == "diff":
+                w1 = v[0]
+                if alpha != 0.0:
+                    if w1.shape != weight.shape:
+                        print("WARNING SHAPE MISMATCH {} WEIGHT NOT MERGED {} != {}".format(key, w1.shape, weight.shape))
+                    else:
+                        weight += alpha * ldm_patched.modules.model_management.cast_to_device(w1, weight.device, weight.dtype)
+            elif patch_type == "lora": #lora/locon
+                mat1 = ldm_patched.modules.model_management.cast_to_device(v[0], weight.device, torch.float32)
+                mat2 = ldm_patched.modules.model_management.cast_to_device(v[1], weight.device, torch.float32)
+                if v[2] is not None:
+                    alpha *= v[2] / mat2.shape[0]
+                if v[3] is not None:
+                    #locon mid weights, hopefully the math is fine because I didn't properly test it
+                    mat3 = ldm_patched.modules.model_management.cast_to_device(v[3], weight.device, torch.float32)
+                    final_shape = [mat2.shape[1], mat2.shape[0], mat3.shape[2], mat3.shape[3]]
+                    mat2 = torch.mm(mat2.transpose(0, 1).flatten(start_dim=1), mat3.transpose(0, 1).flatten(start_dim=1)).reshape(final_shape).transpose(0, 1)
+                try:
+                    weight += (alpha * torch.mm(mat1.flatten(start_dim=1), mat2.flatten(start_dim=1))).reshape(weight.shape).type(weight.dtype)
+                except Exception as e:
+                    print("ERROR", key, e)
+            elif patch_type == "lokr":
+                w1 = v[0]
+                w2 = v[1]
+                w1_a = v[3]
+                w1_b = v[4]
+                w2_a = v[5]
+                w2_b = v[6]
+                t2 = v[7]
+                dim = None
+
+                if w1 is None:
+                    dim = w1_b.shape[0]
+                    w1 = torch.mm(ldm_patched.modules.model_management.cast_to_device(w1_a, weight.device, torch.float32),
+                                  ldm_patched.modules.model_management.cast_to_device(w1_b, weight.device, torch.float32))
+                else:
+                    w1 = ldm_patched.modules.model_management.cast_to_device(w1, weight.device, torch.float32)
+
+                if w2 is None:
+                    dim = w2_b.shape[0]
+                    if t2 is None:
+                        w2 = torch.mm(ldm_patched.modules.model_management.cast_to_device(w2_a, weight.device, torch.float32),
+                                      ldm_patched.modules.model_management.cast_to_device(w2_b, weight.device, torch.float32))
+                    else:
+                        w2 = torch.einsum('i j k l, j r, i p -> p r k l',
+                                          ldm_patched.modules.model_management.cast_to_device(t2, weight.device, torch.float32),
+                                          ldm_patched.modules.model_management.cast_to_device(w2_b, weight.device, torch.float32),
+                                          ldm_patched.modules.model_management.cast_to_device(w2_a, weight.device, torch.float32))
+                else:
+                    w2 = ldm_patched.modules.model_management.cast_to_device(w2, weight.device, torch.float32)
+
+                if len(w2.shape) == 4:
+                    w1 = w1.unsqueeze(2).unsqueeze(2)
+                if v[2] is not None and dim is not None:
+                    alpha *= v[2] / dim
+
+                try:
+                    weight += alpha * torch.kron(w1, w2).reshape(weight.shape).type(weight.dtype)
+                except Exception as e:
+                    print("ERROR", key, e)
+            elif patch_type == "loha":
+                w1a = v[0]
+                w1b = v[1]
+                if v[2] is not None:
+                    alpha *= v[2] / w1b.shape[0]
+                w2a = v[3]
+                w2b = v[4]
+                if v[5] is not None: #cp decomposition
+                    t1 = v[5]
+                    t2 = v[6]
+                    m1 = torch.einsum('i j k l, j r, i p -> p r k l',
+                                      ldm_patched.modules.model_management.cast_to_device(t1, weight.device, torch.float32),
+                                      ldm_patched.modules.model_management.cast_to_device(w1b, weight.device, torch.float32),
+                                      ldm_patched.modules.model_management.cast_to_device(w1a, weight.device, torch.float32))
+
+                    m2 = torch.einsum('i j k l, j r, i p -> p r k l',
+                                      ldm_patched.modules.model_management.cast_to_device(t2, weight.device, torch.float32),
+                                      ldm_patched.modules.model_management.cast_to_device(w2b, weight.device, torch.float32),
+                                      ldm_patched.modules.model_management.cast_to_device(w2a, weight.device, torch.float32))
+                else:
+                    m1 = torch.mm(ldm_patched.modules.model_management.cast_to_device(w1a, weight.device, torch.float32),
+                                  ldm_patched.modules.model_management.cast_to_device(w1b, weight.device, torch.float32))
+                    m2 = torch.mm(ldm_patched.modules.model_management.cast_to_device(w2a, weight.device, torch.float32),
+                                  ldm_patched.modules.model_management.cast_to_device(w2b, weight.device, torch.float32))
+
+                try:
+                    weight += (alpha * m1 * m2).reshape(weight.shape).type(weight.dtype)
+                except Exception as e:
+                    print("ERROR", key, e)
+            elif patch_type == "glora":
+                if v[4] is not None:
+                    alpha *= v[4] / v[0].shape[0]
+
+                a1 = ldm_patched.modules.model_management.cast_to_device(v[0].flatten(start_dim=1), weight.device, torch.float32)
+                a2 = ldm_patched.modules.model_management.cast_to_device(v[1].flatten(start_dim=1), weight.device, torch.float32)
+                b1 = ldm_patched.modules.model_management.cast_to_device(v[2].flatten(start_dim=1), weight.device, torch.float32)
+                b2 = ldm_patched.modules.model_management.cast_to_device(v[3].flatten(start_dim=1), weight.device, torch.float32)
+
+                weight += ((torch.mm(b2, b1) + torch.mm(torch.mm(weight.flatten(start_dim=1), a2), a1)) * alpha).reshape(weight.shape).type(weight.dtype)
+            else:
+                print("patch type not recognized", patch_type, key)
+
+        return weight
+
+    def unpatch_model(self, device_to=None):
+        keys = list(self.backup.keys())
+
+        if self.weight_inplace_update:
+            for k in keys:
+                ldm_patched.modules.utils.copy_to_param(self.model, k, self.backup[k])
+        else:
+            for k in keys:
+                ldm_patched.modules.utils.set_attr(self.model, k, self.backup[k])
+
+        self.backup = {}
+
+        if device_to is not None:
+            self.model.to(device_to)
+            self.current_device = device_to
+
+        keys = list(self.object_patches_backup.keys())
+        for k in keys:
+            setattr(self.model, k, self.object_patches_backup[k])
+
+        self.object_patches_backup = {}

ldm_patched/modules/model_sampling.py

@@ -0,0 +1,136 @@
+import torch
+import numpy as np
+from ldm_patched.ldm.modules.diffusionmodules.util import make_beta_schedule
+import math
+
+class EPS:
+    def calculate_input(self, sigma, noise):
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (noise.ndim - 1))
+        return noise / (sigma ** 2 + self.sigma_data ** 2) ** 0.5
+
+    def calculate_denoised(self, sigma, model_output, model_input):
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (model_output.ndim - 1))
+        return model_input - model_output * sigma
+
+
+class V_PREDICTION(EPS):
+    def calculate_denoised(self, sigma, model_output, model_input):
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (model_output.ndim - 1))
+        return model_input * self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) - model_output * sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2) ** 0.5
+
+
+class ModelSamplingDiscrete(torch.nn.Module):
+    def __init__(self, model_config=None):
+        super().__init__()
+
+        if model_config is not None:
+            sampling_settings = model_config.sampling_settings
+        else:
+            sampling_settings = {}
+
+        beta_schedule = sampling_settings.get("beta_schedule", "linear")
+        linear_start = sampling_settings.get("linear_start", 0.00085)
+        linear_end = sampling_settings.get("linear_end", 0.012)
+
+        self._register_schedule(given_betas=None, beta_schedule=beta_schedule, timesteps=1000, linear_start=linear_start, linear_end=linear_end, cosine_s=8e-3)
+        self.sigma_data = 1.0
+
+    def _register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000,
+                          linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+        if given_betas is not None:
+            betas = given_betas
+        else:
+            betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s)
+        alphas = 1. - betas
+        alphas_cumprod = torch.tensor(np.cumprod(alphas, axis=0), dtype=torch.float32)
+        # alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
+
+        timesteps, = betas.shape
+        self.num_timesteps = int(timesteps)
+        self.linear_start = linear_start
+        self.linear_end = linear_end
+
+        # self.register_buffer('betas', torch.tensor(betas, dtype=torch.float32))
+        # self.register_buffer('alphas_cumprod', torch.tensor(alphas_cumprod, dtype=torch.float32))
+        # self.register_buffer('alphas_cumprod_prev', torch.tensor(alphas_cumprod_prev, dtype=torch.float32))
+
+        sigmas = ((1 - alphas_cumprod) / alphas_cumprod) ** 0.5
+        self.set_sigmas(sigmas)
+
+    def set_sigmas(self, sigmas):
+        self.register_buffer('sigmas', sigmas)
+        self.register_buffer('log_sigmas', sigmas.log())
+
+    @property
+    def sigma_min(self):
+        return self.sigmas[0]
+
+    @property
+    def sigma_max(self):
+        return self.sigmas[-1]
+
+    def timestep(self, sigma):
+        log_sigma = sigma.log()
+        dists = log_sigma.to(self.log_sigmas.device) - self.log_sigmas[:, None]
+        return dists.abs().argmin(dim=0).view(sigma.shape).to(sigma.device)
+
+    def sigma(self, timestep):
+        t = torch.clamp(timestep.float().to(self.log_sigmas.device), min=0, max=(len(self.sigmas) - 1))
+        low_idx = t.floor().long()
+        high_idx = t.ceil().long()
+        w = t.frac()
+        log_sigma = (1 - w) * self.log_sigmas[low_idx] + w * self.log_sigmas[high_idx]
+        return log_sigma.exp().to(timestep.device)
+
+    def percent_to_sigma(self, percent):
+        if percent <= 0.0:
+            return 999999999.9
+        if percent >= 1.0:
+            return 0.0
+        percent = 1.0 - percent
+        return self.sigma(torch.tensor(percent * 999.0)).item()
+
+
+class ModelSamplingContinuousEDM(torch.nn.Module):
+    def __init__(self, model_config=None):
+        super().__init__()
+        self.sigma_data = 1.0
+
+        if model_config is not None:
+            sampling_settings = model_config.sampling_settings
+        else:
+            sampling_settings = {}
+
+        sigma_min = sampling_settings.get("sigma_min", 0.002)
+        sigma_max = sampling_settings.get("sigma_max", 120.0)
+        self.set_sigma_range(sigma_min, sigma_max)
+
+    def set_sigma_range(self, sigma_min, sigma_max):
+        sigmas = torch.linspace(math.log(sigma_min), math.log(sigma_max), 1000).exp()
+
+        self.register_buffer('sigmas', sigmas) #for compatibility with some schedulers
+        self.register_buffer('log_sigmas', sigmas.log())
+
+    @property
+    def sigma_min(self):
+        return self.sigmas[0]
+
+    @property
+    def sigma_max(self):
+        return self.sigmas[-1]
+
+    def timestep(self, sigma):
+        return 0.25 * sigma.log()
+
+    def sigma(self, timestep):
+        return (timestep / 0.25).exp()
+
+    def percent_to_sigma(self, percent):
+        if percent <= 0.0:
+            return 999999999.9
+        if percent >= 1.0:
+            return 0.0
+        percent = 1.0 - percent
+
+        log_sigma_min = math.log(self.sigma_min)
+        return math.exp((math.log(self.sigma_max) - log_sigma_min) * percent + log_sigma_min)

ldm_patched/modules/ops.py

@@ -0,0 +1,65 @@
+import torch
+from contextlib import contextmanager
+
+class disable_weight_init:
+    class Linear(torch.nn.Linear):
+        def reset_parameters(self):
+            return None
+
+    class Conv2d(torch.nn.Conv2d):
+        def reset_parameters(self):
+            return None
+
+    class Conv3d(torch.nn.Conv3d):
+        def reset_parameters(self):
+            return None
+
+    class GroupNorm(torch.nn.GroupNorm):
+        def reset_parameters(self):
+            return None
+
+    class LayerNorm(torch.nn.LayerNorm):
+        def reset_parameters(self):
+            return None
+
+    @classmethod
+    def conv_nd(s, dims, *args, **kwargs):
+        if dims == 2:
+            return s.Conv2d(*args, **kwargs)
+        elif dims == 3:
+            return s.Conv3d(*args, **kwargs)
+        else:
+            raise ValueError(f"unsupported dimensions: {dims}")
+
+def cast_bias_weight(s, input):
+    bias = None
+    if s.bias is not None:
+        bias = s.bias.to(device=input.device, dtype=input.dtype)
+    weight = s.weight.to(device=input.device, dtype=input.dtype)
+    return weight, bias
+
+class manual_cast(disable_weight_init):
+    class Linear(disable_weight_init.Linear):
+        def forward(self, input):
+            weight, bias = cast_bias_weight(self, input)
+            return torch.nn.functional.linear(input, weight, bias)
+
+    class Conv2d(disable_weight_init.Conv2d):
+        def forward(self, input):
+            weight, bias = cast_bias_weight(self, input)
+            return self._conv_forward(input, weight, bias)
+
+    class Conv3d(disable_weight_init.Conv3d):
+        def forward(self, input):
+            weight, bias = cast_bias_weight(self, input)
+            return self._conv_forward(input, weight, bias)
+
+    class GroupNorm(disable_weight_init.GroupNorm):
+        def forward(self, input):
+            weight, bias = cast_bias_weight(self, input)
+            return torch.nn.functional.group_norm(input, self.num_groups, weight, bias, self.eps)
+
+    class LayerNorm(disable_weight_init.LayerNorm):
+        def forward(self, input):
+            weight, bias = cast_bias_weight(self, input)
+            return torch.nn.functional.layer_norm(input, self.normalized_shape, weight, bias, self.eps)

ldm_patched/modules/options.py

@@ -0,0 +1,6 @@
+
+args_parsing = False
+
+def enable_args_parsing(enable=True):
+    global args_parsing
+    args_parsing = enable

ldm_patched/modules/sample.py

@@ -0,0 +1,118 @@
+import torch
+import ldm_patched.modules.model_management
+import ldm_patched.modules.samplers
+import ldm_patched.modules.conds
+import ldm_patched.modules.utils
+import math
+import numpy as np
+
+def prepare_noise(latent_image, seed, noise_inds=None):
+    """
+    creates random noise given a latent image and a seed.
+    optional arg skip can be used to skip and discard x number of noise generations for a given seed
+    """
+    generator = torch.manual_seed(seed)
+    if noise_inds is None:
+        return torch.randn(latent_image.size(), dtype=latent_image.dtype, layout=latent_image.layout, generator=generator, device="cpu")
+    
+    unique_inds, inverse = np.unique(noise_inds, return_inverse=True)
+    noises = []
+    for i in range(unique_inds[-1]+1):
+        noise = torch.randn([1] + list(latent_image.size())[1:], dtype=latent_image.dtype, layout=latent_image.layout, generator=generator, device="cpu")
+        if i in unique_inds:
+            noises.append(noise)
+    noises = [noises[i] for i in inverse]
+    noises = torch.cat(noises, axis=0)
+    return noises
+
+def prepare_mask(noise_mask, shape, device):
+    """ensures noise mask is of proper dimensions"""
+    noise_mask = torch.nn.functional.interpolate(noise_mask.reshape((-1, 1, noise_mask.shape[-2], noise_mask.shape[-1])), size=(shape[2], shape[3]), mode="bilinear")
+    noise_mask = noise_mask.round()
+    noise_mask = torch.cat([noise_mask] * shape[1], dim=1)
+    noise_mask = ldm_patched.modules.utils.repeat_to_batch_size(noise_mask, shape[0])
+    noise_mask = noise_mask.to(device)
+    return noise_mask
+
+def get_models_from_cond(cond, model_type):
+    models = []
+    for c in cond:
+        if model_type in c:
+            models += [c[model_type]]
+    return models
+
+def convert_cond(cond):
+    out = []
+    for c in cond:
+        temp = c[1].copy()
+        model_conds = temp.get("model_conds", {})
+        if c[0] is not None:
+            model_conds["c_crossattn"] = ldm_patched.modules.conds.CONDCrossAttn(c[0])
+        temp["model_conds"] = model_conds
+        out.append(temp)
+    return out
+
+def get_additional_models(positive, negative, dtype):
+    """loads additional models in positive and negative conditioning"""
+    control_nets = set(get_models_from_cond(positive, "control") + get_models_from_cond(negative, "control"))
+
+    inference_memory = 0
+    control_models = []
+    for m in control_nets:
+        control_models += m.get_models()
+        inference_memory += m.inference_memory_requirements(dtype)
+
+    gligen = get_models_from_cond(positive, "gligen") + get_models_from_cond(negative, "gligen")
+    gligen = [x[1] for x in gligen]
+    models = control_models + gligen
+    return models, inference_memory
+
+def cleanup_additional_models(models):
+    """cleanup additional models that were loaded"""
+    for m in models:
+        if hasattr(m, 'cleanup'):
+            m.cleanup()
+
+def prepare_sampling(model, noise_shape, positive, negative, noise_mask):
+    device = model.load_device
+    positive = convert_cond(positive)
+    negative = convert_cond(negative)
+
+    if noise_mask is not None:
+        noise_mask = prepare_mask(noise_mask, noise_shape, device)
+
+    real_model = None
+    models, inference_memory = get_additional_models(positive, negative, model.model_dtype())
+    ldm_patched.modules.model_management.load_models_gpu([model] + models, model.memory_required([noise_shape[0] * 2] + list(noise_shape[1:])) + inference_memory)
+    real_model = model.model
+
+    return real_model, positive, negative, noise_mask, models
+
+
+def sample(model, noise, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, denoise=1.0, disable_noise=False, start_step=None, last_step=None, force_full_denoise=False, noise_mask=None, sigmas=None, callback=None, disable_pbar=False, seed=None):
+    real_model, positive_copy, negative_copy, noise_mask, models = prepare_sampling(model, noise.shape, positive, negative, noise_mask)
+
+    noise = noise.to(model.load_device)
+    latent_image = latent_image.to(model.load_device)
+
+    sampler = ldm_patched.modules.samplers.KSampler(real_model, steps=steps, device=model.load_device, sampler=sampler_name, scheduler=scheduler, denoise=denoise, model_options=model.model_options)
+
+    samples = sampler.sample(noise, positive_copy, negative_copy, cfg=cfg, latent_image=latent_image, start_step=start_step, last_step=last_step, force_full_denoise=force_full_denoise, denoise_mask=noise_mask, sigmas=sigmas, callback=callback, disable_pbar=disable_pbar, seed=seed)
+    samples = samples.to(ldm_patched.modules.model_management.intermediate_device())
+
+    cleanup_additional_models(models)
+    cleanup_additional_models(set(get_models_from_cond(positive_copy, "control") + get_models_from_cond(negative_copy, "control")))
+    return samples
+
+def sample_custom(model, noise, cfg, sampler, sigmas, positive, negative, latent_image, noise_mask=None, callback=None, disable_pbar=False, seed=None):
+    real_model, positive_copy, negative_copy, noise_mask, models = prepare_sampling(model, noise.shape, positive, negative, noise_mask)
+    noise = noise.to(model.load_device)
+    latent_image = latent_image.to(model.load_device)
+    sigmas = sigmas.to(model.load_device)
+
+    samples = ldm_patched.modules.samplers.sample(real_model, noise, positive_copy, negative_copy, cfg, model.load_device, sampler, sigmas, model_options=model.model_options, latent_image=latent_image, denoise_mask=noise_mask, callback=callback, disable_pbar=disable_pbar, seed=seed)
+    samples = samples.to(ldm_patched.modules.model_management.intermediate_device())
+    cleanup_additional_models(models)
+    cleanup_additional_models(set(get_models_from_cond(positive_copy, "control") + get_models_from_cond(negative_copy, "control")))
+    return samples
+

ldm_patched/modules/samplers.py

@@ -0,0 +1,708 @@
+from ldm_patched.k_diffusion import sampling as k_diffusion_sampling
+from ldm_patched.unipc import uni_pc
+import torch
+import enum
+from ldm_patched.modules import model_management
+import math
+from ldm_patched.modules import model_base
+import ldm_patched.modules.utils
+import ldm_patched.modules.conds
+
+
+#The main sampling function shared by all the samplers
+#Returns denoised
+def sampling_function(model, x, timestep, uncond, cond, cond_scale, model_options={}, seed=None):
+        def get_area_and_mult(conds, x_in, timestep_in):
+            area = (x_in.shape[2], x_in.shape[3], 0, 0)
+            strength = 1.0
+
+            if 'timestep_start' in conds:
+                timestep_start = conds['timestep_start']
+                if timestep_in[0] > timestep_start:
+                    return None
+            if 'timestep_end' in conds:
+                timestep_end = conds['timestep_end']
+                if timestep_in[0] < timestep_end:
+                    return None
+            if 'area' in conds:
+                area = conds['area']
+            if 'strength' in conds:
+                strength = conds['strength']
+
+            input_x = x_in[:,:,area[2]:area[0] + area[2],area[3]:area[1] + area[3]]
+            if 'mask' in conds:
+                # Scale the mask to the size of the input
+                # The mask should have been resized as we began the sampling process
+                mask_strength = 1.0
+                if "mask_strength" in conds:
+                    mask_strength = conds["mask_strength"]
+                mask = conds['mask']
+                assert(mask.shape[1] == x_in.shape[2])
+                assert(mask.shape[2] == x_in.shape[3])
+                mask = mask[:,area[2]:area[0] + area[2],area[3]:area[1] + area[3]] * mask_strength
+                mask = mask.unsqueeze(1).repeat(input_x.shape[0] // mask.shape[0], input_x.shape[1], 1, 1)
+            else:
+                mask = torch.ones_like(input_x)
+            mult = mask * strength
+
+            if 'mask' not in conds:
+                rr = 8
+                if area[2] != 0:
+                    for t in range(rr):
+                        mult[:,:,t:1+t,:] *= ((1.0/rr) * (t + 1))
+                if (area[0] + area[2]) < x_in.shape[2]:
+                    for t in range(rr):
+                        mult[:,:,area[0] - 1 - t:area[0] - t,:] *= ((1.0/rr) * (t + 1))
+                if area[3] != 0:
+                    for t in range(rr):
+                        mult[:,:,:,t:1+t] *= ((1.0/rr) * (t + 1))
+                if (area[1] + area[3]) < x_in.shape[3]:
+                    for t in range(rr):
+                        mult[:,:,:,area[1] - 1 - t:area[1] - t] *= ((1.0/rr) * (t + 1))
+
+            conditionning = {}
+            model_conds = conds["model_conds"]
+            for c in model_conds:
+                conditionning[c] = model_conds[c].process_cond(batch_size=x_in.shape[0], device=x_in.device, area=area)
+
+            control = None
+            if 'control' in conds:
+                control = conds['control']
+
+            patches = None
+            if 'gligen' in conds:
+                gligen = conds['gligen']
+                patches = {}
+                gligen_type = gligen[0]
+                gligen_model = gligen[1]
+                if gligen_type == "position":
+                    gligen_patch = gligen_model.model.set_position(input_x.shape, gligen[2], input_x.device)
+                else:
+                    gligen_patch = gligen_model.model.set_empty(input_x.shape, input_x.device)
+
+                patches['middle_patch'] = [gligen_patch]
+
+            return (input_x, mult, conditionning, area, control, patches)
+
+        def cond_equal_size(c1, c2):
+            if c1 is c2:
+                return True
+            if c1.keys() != c2.keys():
+                return False
+            for k in c1:
+                if not c1[k].can_concat(c2[k]):
+                    return False
+            return True
+
+        def can_concat_cond(c1, c2):
+            if c1[0].shape != c2[0].shape:
+                return False
+
+            #control
+            if (c1[4] is None) != (c2[4] is None):
+                return False
+            if c1[4] is not None:
+                if c1[4] is not c2[4]:
+                    return False
+
+            #patches
+            if (c1[5] is None) != (c2[5] is None):
+                return False
+            if (c1[5] is not None):
+                if c1[5] is not c2[5]:
+                    return False
+
+            return cond_equal_size(c1[2], c2[2])
+
+        def cond_cat(c_list):
+            c_crossattn = []
+            c_concat = []
+            c_adm = []
+            crossattn_max_len = 0
+
+            temp = {}
+            for x in c_list:
+                for k in x:
+                    cur = temp.get(k, [])
+                    cur.append(x[k])
+                    temp[k] = cur
+
+            out = {}
+            for k in temp:
+                conds = temp[k]
+                out[k] = conds[0].concat(conds[1:])
+
+            return out
+
+        def calc_cond_uncond_batch(model, cond, uncond, x_in, timestep, model_options):
+            out_cond = torch.zeros_like(x_in)
+            out_count = torch.ones_like(x_in) * 1e-37
+
+            out_uncond = torch.zeros_like(x_in)
+            out_uncond_count = torch.ones_like(x_in) * 1e-37
+
+            COND = 0
+            UNCOND = 1
+
+            to_run = []
+            for x in cond:
+                p = get_area_and_mult(x, x_in, timestep)
+                if p is None:
+                    continue
+
+                to_run += [(p, COND)]
+            if uncond is not None:
+                for x in uncond:
+                    p = get_area_and_mult(x, x_in, timestep)
+                    if p is None:
+                        continue
+
+                    to_run += [(p, UNCOND)]
+
+            while len(to_run) > 0:
+                first = to_run[0]
+                first_shape = first[0][0].shape
+                to_batch_temp = []
+                for x in range(len(to_run)):
+                    if can_concat_cond(to_run[x][0], first[0]):
+                        to_batch_temp += [x]
+
+                to_batch_temp.reverse()
+                to_batch = to_batch_temp[:1]
+
+                free_memory = model_management.get_free_memory(x_in.device)
+                for i in range(1, len(to_batch_temp) + 1):
+                    batch_amount = to_batch_temp[:len(to_batch_temp)//i]
+                    input_shape = [len(batch_amount) * first_shape[0]] + list(first_shape)[1:]
+                    if model.memory_required(input_shape) < free_memory:
+                        to_batch = batch_amount
+                        break
+
+                input_x = []
+                mult = []
+                c = []
+                cond_or_uncond = []
+                area = []
+                control = None
+                patches = None
+                for x in to_batch:
+                    o = to_run.pop(x)
+                    p = o[0]
+                    input_x += [p[0]]
+                    mult += [p[1]]
+                    c += [p[2]]
+                    area += [p[3]]
+                    cond_or_uncond += [o[1]]
+                    control = p[4]
+                    patches = p[5]
+
+                batch_chunks = len(cond_or_uncond)
+                input_x = torch.cat(input_x)
+                c = cond_cat(c)
+                timestep_ = torch.cat([timestep] * batch_chunks)
+
+                if control is not None:
+                    c['control'] = control.get_control(input_x, timestep_, c, len(cond_or_uncond))
+
+                transformer_options = {}
+                if 'transformer_options' in model_options:
+                    transformer_options = model_options['transformer_options'].copy()
+
+                if patches is not None:
+                    if "patches" in transformer_options:
+                        cur_patches = transformer_options["patches"].copy()
+                        for p in patches:
+                            if p in cur_patches:
+                                cur_patches[p] = cur_patches[p] + patches[p]
+                            else:
+                                cur_patches[p] = patches[p]
+                    else:
+                        transformer_options["patches"] = patches
+
+                transformer_options["cond_or_uncond"] = cond_or_uncond[:]
+                transformer_options["sigmas"] = timestep
+
+                c['transformer_options'] = transformer_options
+
+                if 'model_function_wrapper' in model_options:
+                    output = model_options['model_function_wrapper'](model.apply_model, {"input": input_x, "timestep": timestep_, "c": c, "cond_or_uncond": cond_or_uncond}).chunk(batch_chunks)
+                else:
+                    output = model.apply_model(input_x, timestep_, **c).chunk(batch_chunks)
+                del input_x
+
+                for o in range(batch_chunks):
+                    if cond_or_uncond[o] == COND:
+                        out_cond[:,:,area[o][2]:area[o][0] + area[o][2],area[o][3]:area[o][1] + area[o][3]] += output[o] * mult[o]
+                        out_count[:,:,area[o][2]:area[o][0] + area[o][2],area[o][3]:area[o][1] + area[o][3]] += mult[o]
+                    else:
+                        out_uncond[:,:,area[o][2]:area[o][0] + area[o][2],area[o][3]:area[o][1] + area[o][3]] += output[o] * mult[o]
+                        out_uncond_count[:,:,area[o][2]:area[o][0] + area[o][2],area[o][3]:area[o][1] + area[o][3]] += mult[o]
+                del mult
+
+            out_cond /= out_count
+            del out_count
+            out_uncond /= out_uncond_count
+            del out_uncond_count
+            return out_cond, out_uncond
+
+
+        if math.isclose(cond_scale, 1.0):
+            uncond = None
+
+        cond, uncond = calc_cond_uncond_batch(model, cond, uncond, x, timestep, model_options)
+        if "sampler_cfg_function" in model_options:
+            args = {"cond": x - cond, "uncond": x - uncond, "cond_scale": cond_scale, "timestep": timestep, "input": x, "sigma": timestep}
+            return x - model_options["sampler_cfg_function"](args)
+        else:
+            return uncond + (cond - uncond) * cond_scale
+
+class CFGNoisePredictor(torch.nn.Module):
+    def __init__(self, model):
+        super().__init__()
+        self.inner_model = model
+    def apply_model(self, x, timestep, cond, uncond, cond_scale, model_options={}, seed=None):
+        out = sampling_function(self.inner_model, x, timestep, uncond, cond, cond_scale, model_options=model_options, seed=seed)
+        return out
+    def forward(self, *args, **kwargs):
+        return self.apply_model(*args, **kwargs)
+
+class KSamplerX0Inpaint(torch.nn.Module):
+    def __init__(self, model):
+        super().__init__()
+        self.inner_model = model
+    def forward(self, x, sigma, uncond, cond, cond_scale, denoise_mask, model_options={}, seed=None):
+        if denoise_mask is not None:
+            latent_mask = 1. - denoise_mask
+            x = x * denoise_mask + (self.latent_image + self.noise * sigma.reshape([sigma.shape[0]] + [1] * (len(self.noise.shape) - 1))) * latent_mask
+        out = self.inner_model(x, sigma, cond=cond, uncond=uncond, cond_scale=cond_scale, model_options=model_options, seed=seed)
+        if denoise_mask is not None:
+            out = out * denoise_mask + self.latent_image * latent_mask
+        return out
+
+def simple_scheduler(model, steps):
+    s = model.model_sampling
+    sigs = []
+    ss = len(s.sigmas) / steps
+    for x in range(steps):
+        sigs += [float(s.sigmas[-(1 + int(x * ss))])]
+    sigs += [0.0]
+    return torch.FloatTensor(sigs)
+
+def ddim_scheduler(model, steps):
+    s = model.model_sampling
+    sigs = []
+    ss = len(s.sigmas) // steps
+    x = 1
+    while x < len(s.sigmas):
+        sigs += [float(s.sigmas[x])]
+        x += ss
+    sigs = sigs[::-1]
+    sigs += [0.0]
+    return torch.FloatTensor(sigs)
+
+def normal_scheduler(model, steps, sgm=False, floor=False):
+    s = model.model_sampling
+    start = s.timestep(s.sigma_max)
+    end = s.timestep(s.sigma_min)
+
+    if sgm:
+        timesteps = torch.linspace(start, end, steps + 1)[:-1]
+    else:
+        timesteps = torch.linspace(start, end, steps)
+
+    sigs = []
+    for x in range(len(timesteps)):
+        ts = timesteps[x]
+        sigs.append(s.sigma(ts))
+    sigs += [0.0]
+    return torch.FloatTensor(sigs)
+
+def get_mask_aabb(masks):
+    if masks.numel() == 0:
+        return torch.zeros((0, 4), device=masks.device, dtype=torch.int)
+
+    b = masks.shape[0]
+
+    bounding_boxes = torch.zeros((b, 4), device=masks.device, dtype=torch.int)
+    is_empty = torch.zeros((b), device=masks.device, dtype=torch.bool)
+    for i in range(b):
+        mask = masks[i]
+        if mask.numel() == 0:
+            continue
+        if torch.max(mask != 0) == False:
+            is_empty[i] = True
+            continue
+        y, x = torch.where(mask)
+        bounding_boxes[i, 0] = torch.min(x)
+        bounding_boxes[i, 1] = torch.min(y)
+        bounding_boxes[i, 2] = torch.max(x)
+        bounding_boxes[i, 3] = torch.max(y)
+
+    return bounding_boxes, is_empty
+
+def resolve_areas_and_cond_masks(conditions, h, w, device):
+    # We need to decide on an area outside the sampling loop in order to properly generate opposite areas of equal sizes.
+    # While we're doing this, we can also resolve the mask device and scaling for performance reasons
+    for i in range(len(conditions)):
+        c = conditions[i]
+        if 'area' in c:
+            area = c['area']
+            if area[0] == "percentage":
+                modified = c.copy()
+                area = (max(1, round(area[1] * h)), max(1, round(area[2] * w)), round(area[3] * h), round(area[4] * w))
+                modified['area'] = area
+                c = modified
+                conditions[i] = c
+
+        if 'mask' in c:
+            mask = c['mask']
+            mask = mask.to(device=device)
+            modified = c.copy()
+            if len(mask.shape) == 2:
+                mask = mask.unsqueeze(0)
+            if mask.shape[1] != h or mask.shape[2] != w:
+                mask = torch.nn.functional.interpolate(mask.unsqueeze(1), size=(h, w), mode='bilinear', align_corners=False).squeeze(1)
+
+            if modified.get("set_area_to_bounds", False):
+                bounds = torch.max(torch.abs(mask),dim=0).values.unsqueeze(0)
+                boxes, is_empty = get_mask_aabb(bounds)
+                if is_empty[0]:
+                    # Use the minimum possible size for efficiency reasons. (Since the mask is all-0, this becomes a noop anyway)
+                    modified['area'] = (8, 8, 0, 0)
+                else:
+                    box = boxes[0]
+                    H, W, Y, X = (box[3] - box[1] + 1, box[2] - box[0] + 1, box[1], box[0])
+                    H = max(8, H)
+                    W = max(8, W)
+                    area = (int(H), int(W), int(Y), int(X))
+                    modified['area'] = area
+
+            modified['mask'] = mask
+            conditions[i] = modified
+
+def create_cond_with_same_area_if_none(conds, c):
+    if 'area' not in c:
+        return
+
+    c_area = c['area']
+    smallest = None
+    for x in conds:
+        if 'area' in x:
+            a = x['area']
+            if c_area[2] >= a[2] and c_area[3] >= a[3]:
+                if a[0] + a[2] >= c_area[0] + c_area[2]:
+                    if a[1] + a[3] >= c_area[1] + c_area[3]:
+                        if smallest is None:
+                            smallest = x
+                        elif 'area' not in smallest:
+                            smallest = x
+                        else:
+                            if smallest['area'][0] * smallest['area'][1] > a[0] * a[1]:
+                                smallest = x
+        else:
+            if smallest is None:
+                smallest = x
+    if smallest is None:
+        return
+    if 'area' in smallest:
+        if smallest['area'] == c_area:
+            return
+
+    out = c.copy()
+    out['model_conds'] = smallest['model_conds'].copy() #TODO: which fields should be copied?
+    conds += [out]
+
+def calculate_start_end_timesteps(model, conds):
+    s = model.model_sampling
+    for t in range(len(conds)):
+        x = conds[t]
+
+        timestep_start = None
+        timestep_end = None
+        if 'start_percent' in x:
+            timestep_start = s.percent_to_sigma(x['start_percent'])
+        if 'end_percent' in x:
+            timestep_end = s.percent_to_sigma(x['end_percent'])
+
+        if (timestep_start is not None) or (timestep_end is not None):
+            n = x.copy()
+            if (timestep_start is not None):
+                n['timestep_start'] = timestep_start
+            if (timestep_end is not None):
+                n['timestep_end'] = timestep_end
+            conds[t] = n
+
+def pre_run_control(model, conds):
+    s = model.model_sampling
+    for t in range(len(conds)):
+        x = conds[t]
+
+        timestep_start = None
+        timestep_end = None
+        percent_to_timestep_function = lambda a: s.percent_to_sigma(a)
+        if 'control' in x:
+            x['control'].pre_run(model, percent_to_timestep_function)
+
+def apply_empty_x_to_equal_area(conds, uncond, name, uncond_fill_func):
+    cond_cnets = []
+    cond_other = []
+    uncond_cnets = []
+    uncond_other = []
+    for t in range(len(conds)):
+        x = conds[t]
+        if 'area' not in x:
+            if name in x and x[name] is not None:
+                cond_cnets.append(x[name])
+            else:
+                cond_other.append((x, t))
+    for t in range(len(uncond)):
+        x = uncond[t]
+        if 'area' not in x:
+            if name in x and x[name] is not None:
+                uncond_cnets.append(x[name])
+            else:
+                uncond_other.append((x, t))
+
+    if len(uncond_cnets) > 0:
+        return
+
+    for x in range(len(cond_cnets)):
+        temp = uncond_other[x % len(uncond_other)]
+        o = temp[0]
+        if name in o and o[name] is not None:
+            n = o.copy()
+            n[name] = uncond_fill_func(cond_cnets, x)
+            uncond += [n]
+        else:
+            n = o.copy()
+            n[name] = uncond_fill_func(cond_cnets, x)
+            uncond[temp[1]] = n
+
+def encode_model_conds(model_function, conds, noise, device, prompt_type, **kwargs):
+    for t in range(len(conds)):
+        x = conds[t]
+        params = x.copy()
+        params["device"] = device
+        params["noise"] = noise
+        params["width"] = params.get("width", noise.shape[3] * 8)
+        params["height"] = params.get("height", noise.shape[2] * 8)
+        params["prompt_type"] = params.get("prompt_type", prompt_type)
+        for k in kwargs:
+            if k not in params:
+                params[k] = kwargs[k]
+
+        out = model_function(**params)
+        x = x.copy()
+        model_conds = x['model_conds'].copy()
+        for k in out:
+            model_conds[k] = out[k]
+        x['model_conds'] = model_conds
+        conds[t] = x
+    return conds
+
+class Sampler:
+    def sample(self):
+        pass
+
+    def max_denoise(self, model_wrap, sigmas):
+        max_sigma = float(model_wrap.inner_model.model_sampling.sigma_max)
+        sigma = float(sigmas[0])
+        return math.isclose(max_sigma, sigma, rel_tol=1e-05) or sigma > max_sigma
+
+class UNIPC(Sampler):
+    def sample(self, model_wrap, sigmas, extra_args, callback, noise, latent_image=None, denoise_mask=None, disable_pbar=False):
+        return uni_pc.sample_unipc(model_wrap, noise, latent_image, sigmas, max_denoise=self.max_denoise(model_wrap, sigmas), extra_args=extra_args, noise_mask=denoise_mask, callback=callback, disable=disable_pbar)
+
+class UNIPCBH2(Sampler):
+    def sample(self, model_wrap, sigmas, extra_args, callback, noise, latent_image=None, denoise_mask=None, disable_pbar=False):
+        return uni_pc.sample_unipc(model_wrap, noise, latent_image, sigmas, max_denoise=self.max_denoise(model_wrap, sigmas), extra_args=extra_args, noise_mask=denoise_mask, callback=callback, variant='bh2', disable=disable_pbar)
+
+KSAMPLER_NAMES = ["euler", "euler_ancestral", "heun", "heunpp2","dpm_2", "dpm_2_ancestral",
+                  "lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_sde", "dpmpp_sde_gpu",
+                  "dpmpp_2m", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm"]
+
+class KSAMPLER(Sampler):
+    def __init__(self, sampler_function, extra_options={}, inpaint_options={}):
+        self.sampler_function = sampler_function
+        self.extra_options = extra_options
+        self.inpaint_options = inpaint_options
+
+    def sample(self, model_wrap, sigmas, extra_args, callback, noise, latent_image=None, denoise_mask=None, disable_pbar=False):
+        extra_args["denoise_mask"] = denoise_mask
+        model_k = KSamplerX0Inpaint(model_wrap)
+        model_k.latent_image = latent_image
+        if self.inpaint_options.get("random", False): #TODO: Should this be the default?
+            generator = torch.manual_seed(extra_args.get("seed", 41) + 1)
+            model_k.noise = torch.randn(noise.shape, generator=generator, device="cpu").to(noise.dtype).to(noise.device)
+        else:
+            model_k.noise = noise
+
+        if self.max_denoise(model_wrap, sigmas):
+            noise = noise * torch.sqrt(1.0 + sigmas[0] ** 2.0)
+        else:
+            noise = noise * sigmas[0]
+
+        k_callback = None
+        total_steps = len(sigmas) - 1
+        if callback is not None:
+            k_callback = lambda x: callback(x["i"], x["denoised"], x["x"], total_steps)
+
+        if latent_image is not None:
+            noise += latent_image
+
+        samples = self.sampler_function(model_k, noise, sigmas, extra_args=extra_args, callback=k_callback, disable=disable_pbar, **self.extra_options)
+        return samples
+
+
+def ksampler(sampler_name, extra_options={}, inpaint_options={}):
+    if sampler_name == "dpm_fast":
+        def dpm_fast_function(model, noise, sigmas, extra_args, callback, disable):
+            sigma_min = sigmas[-1]
+            if sigma_min == 0:
+                sigma_min = sigmas[-2]
+            total_steps = len(sigmas) - 1
+            return k_diffusion_sampling.sample_dpm_fast(model, noise, sigma_min, sigmas[0], total_steps, extra_args=extra_args, callback=callback, disable=disable)
+        sampler_function = dpm_fast_function
+    elif sampler_name == "dpm_adaptive":
+        def dpm_adaptive_function(model, noise, sigmas, extra_args, callback, disable):
+            sigma_min = sigmas[-1]
+            if sigma_min == 0:
+                sigma_min = sigmas[-2]
+            return k_diffusion_sampling.sample_dpm_adaptive(model, noise, sigma_min, sigmas[0], extra_args=extra_args, callback=callback, disable=disable)
+        sampler_function = dpm_adaptive_function
+    else:
+        sampler_function = getattr(k_diffusion_sampling, "sample_{}".format(sampler_name))
+
+    return KSAMPLER(sampler_function, extra_options, inpaint_options)
+
+def wrap_model(model):
+    model_denoise = CFGNoisePredictor(model)
+    return model_denoise
+
+def sample(model, noise, positive, negative, cfg, device, sampler, sigmas, model_options={}, latent_image=None, denoise_mask=None, callback=None, disable_pbar=False, seed=None):
+    positive = positive[:]
+    negative = negative[:]
+
+    resolve_areas_and_cond_masks(positive, noise.shape[2], noise.shape[3], device)
+    resolve_areas_and_cond_masks(negative, noise.shape[2], noise.shape[3], device)
+
+    model_wrap = wrap_model(model)
+
+    calculate_start_end_timesteps(model, negative)
+    calculate_start_end_timesteps(model, positive)
+
+    #make sure each cond area has an opposite one with the same area
+    for c in positive:
+        create_cond_with_same_area_if_none(negative, c)
+    for c in negative:
+        create_cond_with_same_area_if_none(positive, c)
+
+    pre_run_control(model, negative + positive)
+
+    apply_empty_x_to_equal_area(list(filter(lambda c: c.get('control_apply_to_uncond', False) == True, positive)), negative, 'control', lambda cond_cnets, x: cond_cnets[x])
+    apply_empty_x_to_equal_area(positive, negative, 'gligen', lambda cond_cnets, x: cond_cnets[x])
+
+    if latent_image is not None:
+        latent_image = model.process_latent_in(latent_image)
+
+    if hasattr(model, 'extra_conds'):
+        positive = encode_model_conds(model.extra_conds, positive, noise, device, "positive", latent_image=latent_image, denoise_mask=denoise_mask)
+        negative = encode_model_conds(model.extra_conds, negative, noise, device, "negative", latent_image=latent_image, denoise_mask=denoise_mask)
+
+    extra_args = {"cond":positive, "uncond":negative, "cond_scale": cfg, "model_options": model_options, "seed":seed}
+
+    samples = sampler.sample(model_wrap, sigmas, extra_args, callback, noise, latent_image, denoise_mask, disable_pbar)
+    return model.process_latent_out(samples.to(torch.float32))
+
+SCHEDULER_NAMES = ["normal", "karras", "exponential", "sgm_uniform", "simple", "ddim_uniform"]
+SAMPLER_NAMES = KSAMPLER_NAMES + ["ddim", "uni_pc", "uni_pc_bh2"]
+
+def calculate_sigmas_scheduler(model, scheduler_name, steps):
+    if scheduler_name == "karras":
+        sigmas = k_diffusion_sampling.get_sigmas_karras(n=steps, sigma_min=float(model.model_sampling.sigma_min), sigma_max=float(model.model_sampling.sigma_max))
+    elif scheduler_name == "exponential":
+        sigmas = k_diffusion_sampling.get_sigmas_exponential(n=steps, sigma_min=float(model.model_sampling.sigma_min), sigma_max=float(model.model_sampling.sigma_max))
+    elif scheduler_name == "normal":
+        sigmas = normal_scheduler(model, steps)
+    elif scheduler_name == "simple":
+        sigmas = simple_scheduler(model, steps)
+    elif scheduler_name == "ddim_uniform":
+        sigmas = ddim_scheduler(model, steps)
+    elif scheduler_name == "sgm_uniform":
+        sigmas = normal_scheduler(model, steps, sgm=True)
+    else:
+        print("error invalid scheduler", self.scheduler)
+    return sigmas
+
+def sampler_object(name):
+    if name == "uni_pc":
+        sampler = UNIPC()
+    elif name == "uni_pc_bh2":
+        sampler = UNIPCBH2()
+    elif name == "ddim":
+        sampler = ksampler("euler", inpaint_options={"random": True})
+    else:
+        sampler = ksampler(name)
+    return sampler
+
+class KSampler:
+    SCHEDULERS = SCHEDULER_NAMES
+    SAMPLERS = SAMPLER_NAMES
+
+    def __init__(self, model, steps, device, sampler=None, scheduler=None, denoise=None, model_options={}):
+        self.model = model
+        self.device = device
+        if scheduler not in self.SCHEDULERS:
+            scheduler = self.SCHEDULERS[0]
+        if sampler not in self.SAMPLERS:
+            sampler = self.SAMPLERS[0]
+        self.scheduler = scheduler
+        self.sampler = sampler
+        self.set_steps(steps, denoise)
+        self.denoise = denoise
+        self.model_options = model_options
+
+    def calculate_sigmas(self, steps):
+        sigmas = None
+
+        discard_penultimate_sigma = False
+        if self.sampler in ['dpm_2', 'dpm_2_ancestral', 'uni_pc', 'uni_pc_bh2']:
+            steps += 1
+            discard_penultimate_sigma = True
+
+        sigmas = calculate_sigmas_scheduler(self.model, self.scheduler, steps)
+
+        if discard_penultimate_sigma:
+            sigmas = torch.cat([sigmas[:-2], sigmas[-1:]])
+        return sigmas
+
+    def set_steps(self, steps, denoise=None):
+        self.steps = steps
+        if denoise is None or denoise > 0.9999:
+            self.sigmas = self.calculate_sigmas(steps).to(self.device)
+        else:
+            new_steps = int(steps/denoise)
+            sigmas = self.calculate_sigmas(new_steps).to(self.device)
+            self.sigmas = sigmas[-(steps + 1):]
+
+    def sample(self, noise, positive, negative, cfg, latent_image=None, start_step=None, last_step=None, force_full_denoise=False, denoise_mask=None, sigmas=None, callback=None, disable_pbar=False, seed=None):
+        if sigmas is None:
+            sigmas = self.sigmas
+
+        if last_step is not None and last_step < (len(sigmas) - 1):
+            sigmas = sigmas[:last_step + 1]
+            if force_full_denoise:
+                sigmas[-1] = 0
+
+        if start_step is not None:
+            if start_step < (len(sigmas) - 1):
+                sigmas = sigmas[start_step:]
+            else:
+                if latent_image is not None:
+                    return latent_image
+                else:
+                    return torch.zeros_like(noise)
+
+        sampler = sampler_object(self.sampler)
+
+        return sample(self.model, noise, positive, negative, cfg, self.device, sampler, sigmas, self.model_options, latent_image=latent_image, denoise_mask=denoise_mask, callback=callback, disable_pbar=disable_pbar, seed=seed)

ldm_patched/modules/sd.py

@@ -0,0 +1,533 @@
+import torch
+import contextlib
+import math
+
+from ldm_patched.modules import model_management
+from ldm_patched.ldm.util import instantiate_from_config
+from ldm_patched.ldm.models.autoencoder import AutoencoderKL, AutoencodingEngine
+import yaml
+
+import ldm_patched.modules.utils
+
+from . import clip_vision
+from . import gligen
+from . import diffusers_convert
+from . import model_base
+from . import model_detection
+
+from . import sd1_clip
+from . import sd2_clip
+from . import sdxl_clip
+
+import ldm_patched.modules.model_patcher
+import ldm_patched.modules.lora
+import ldm_patched.t2ia.adapter
+import ldm_patched.modules.supported_models_base
+import ldm_patched.taesd.taesd
+
+def load_model_weights(model, sd):
+    m, u = model.load_state_dict(sd, strict=False)
+    m = set(m)
+    unexpected_keys = set(u)
+
+    k = list(sd.keys())
+    for x in k:
+        if x not in unexpected_keys:
+            w = sd.pop(x)
+            del w
+    if len(m) > 0:
+        print("extra", m)
+    return model
+
+def load_clip_weights(model, sd):
+    k = list(sd.keys())
+    for x in k:
+        if x.startswith("cond_stage_model.transformer.") and not x.startswith("cond_stage_model.transformer.text_model."):
+            y = x.replace("cond_stage_model.transformer.", "cond_stage_model.transformer.text_model.")
+            sd[y] = sd.pop(x)
+
+    if 'cond_stage_model.transformer.text_model.embeddings.position_ids' in sd:
+        ids = sd['cond_stage_model.transformer.text_model.embeddings.position_ids']
+        if ids.dtype == torch.float32:
+            sd['cond_stage_model.transformer.text_model.embeddings.position_ids'] = ids.round()
+
+    sd = ldm_patched.modules.utils.transformers_convert(sd, "cond_stage_model.model.", "cond_stage_model.transformer.text_model.", 24)
+    return load_model_weights(model, sd)
+
+
+def load_lora_for_models(model, clip, lora, strength_model, strength_clip):
+    key_map = {}
+    if model is not None:
+        key_map = ldm_patched.modules.lora.model_lora_keys_unet(model.model, key_map)
+    if clip is not None:
+        key_map = ldm_patched.modules.lora.model_lora_keys_clip(clip.cond_stage_model, key_map)
+
+    loaded = ldm_patched.modules.lora.load_lora(lora, key_map)
+    if model is not None:
+        new_modelpatcher = model.clone()
+        k = new_modelpatcher.add_patches(loaded, strength_model)
+    else:
+        k = ()
+        new_modelpatcher = None
+
+    if clip is not None:
+        new_clip = clip.clone()
+        k1 = new_clip.add_patches(loaded, strength_clip)
+    else:
+        k1 = ()
+        new_clip = None
+    k = set(k)
+    k1 = set(k1)
+    for x in loaded:
+        if (x not in k) and (x not in k1):
+            print("NOT LOADED", x)
+
+    return (new_modelpatcher, new_clip)
+
+
+class CLIP:
+    def __init__(self, target=None, embedding_directory=None, no_init=False):
+        if no_init:
+            return
+        params = target.params.copy()
+        clip = target.clip
+        tokenizer = target.tokenizer
+
+        load_device = model_management.text_encoder_device()
+        offload_device = model_management.text_encoder_offload_device()
+        params['device'] = offload_device
+        params['dtype'] = model_management.text_encoder_dtype(load_device)
+
+        self.cond_stage_model = clip(**(params))
+
+        self.tokenizer = tokenizer(embedding_directory=embedding_directory)
+        self.patcher = ldm_patched.modules.model_patcher.ModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device)
+        self.layer_idx = None
+
+    def clone(self):
+        n = CLIP(no_init=True)
+        n.patcher = self.patcher.clone()
+        n.cond_stage_model = self.cond_stage_model
+        n.tokenizer = self.tokenizer
+        n.layer_idx = self.layer_idx
+        return n
+
+    def add_patches(self, patches, strength_patch=1.0, strength_model=1.0):
+        return self.patcher.add_patches(patches, strength_patch, strength_model)
+
+    def clip_layer(self, layer_idx):
+        self.layer_idx = layer_idx
+
+    def tokenize(self, text, return_word_ids=False):
+        return self.tokenizer.tokenize_with_weights(text, return_word_ids)
+
+    def encode_from_tokens(self, tokens, return_pooled=False):
+        if self.layer_idx is not None:
+            self.cond_stage_model.clip_layer(self.layer_idx)
+        else:
+            self.cond_stage_model.reset_clip_layer()
+
+        self.load_model()
+        cond, pooled = self.cond_stage_model.encode_token_weights(tokens)
+        if return_pooled:
+            return cond, pooled
+        return cond
+
+    def encode(self, text):
+        tokens = self.tokenize(text)
+        return self.encode_from_tokens(tokens)
+
+    def load_sd(self, sd):
+        return self.cond_stage_model.load_sd(sd)
+
+    def get_sd(self):
+        return self.cond_stage_model.state_dict()
+
+    def load_model(self):
+        model_management.load_model_gpu(self.patcher)
+        return self.patcher
+
+    def get_key_patches(self):
+        return self.patcher.get_key_patches()
+
+class VAE:
+    def __init__(self, sd=None, device=None, config=None, dtype=None):
+        if 'decoder.up_blocks.0.resnets.0.norm1.weight' in sd.keys(): #diffusers format
+            sd = diffusers_convert.convert_vae_state_dict(sd)
+
+        self.memory_used_encode = lambda shape, dtype: (1767 * shape[2] * shape[3]) * model_management.dtype_size(dtype) #These are for AutoencoderKL and need tweaking (should be lower)
+        self.memory_used_decode = lambda shape, dtype: (2178 * shape[2] * shape[3] * 64) * model_management.dtype_size(dtype)
+
+        if config is None:
+            if "decoder.mid.block_1.mix_factor" in sd:
+                encoder_config = {'double_z': True, 'z_channels': 4, 'resolution': 256, 'in_channels': 3, 'out_ch': 3, 'ch': 128, 'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2, 'attn_resolutions': [], 'dropout': 0.0}
+                decoder_config = encoder_config.copy()
+                decoder_config["video_kernel_size"] = [3, 1, 1]
+                decoder_config["alpha"] = 0.0
+                self.first_stage_model = AutoencodingEngine(regularizer_config={'target': "ldm_patched.ldm.models.autoencoder.DiagonalGaussianRegularizer"},
+                                                            encoder_config={'target': "ldm_patched.ldm.modules.diffusionmodules.model.Encoder", 'params': encoder_config},
+                                                            decoder_config={'target': "ldm_patched.ldm.modules.temporal_ae.VideoDecoder", 'params': decoder_config})
+            elif "taesd_decoder.1.weight" in sd:
+                self.first_stage_model = ldm_patched.taesd.taesd.TAESD()
+            else:
+                #default SD1.x/SD2.x VAE parameters
+                ddconfig = {'double_z': True, 'z_channels': 4, 'resolution': 256, 'in_channels': 3, 'out_ch': 3, 'ch': 128, 'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2, 'attn_resolutions': [], 'dropout': 0.0}
+                self.first_stage_model = AutoencoderKL(ddconfig=ddconfig, embed_dim=4)
+        else:
+            self.first_stage_model = AutoencoderKL(**(config['params']))
+        self.first_stage_model = self.first_stage_model.eval()
+
+        m, u = self.first_stage_model.load_state_dict(sd, strict=False)
+        if len(m) > 0:
+            print("Missing VAE keys", m)
+
+        if len(u) > 0:
+            print("Leftover VAE keys", u)
+
+        if device is None:
+            device = model_management.vae_device()
+        self.device = device
+        offload_device = model_management.vae_offload_device()
+        if dtype is None:
+            dtype = model_management.vae_dtype()
+        self.vae_dtype = dtype
+        self.first_stage_model.to(self.vae_dtype)
+        self.output_device = model_management.intermediate_device()
+
+        self.patcher = ldm_patched.modules.model_patcher.ModelPatcher(self.first_stage_model, load_device=self.device, offload_device=offload_device)
+
+    def decode_tiled_(self, samples, tile_x=64, tile_y=64, overlap = 16):
+        steps = samples.shape[0] * ldm_patched.modules.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x, tile_y, overlap)
+        steps += samples.shape[0] * ldm_patched.modules.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x // 2, tile_y * 2, overlap)
+        steps += samples.shape[0] * ldm_patched.modules.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x * 2, tile_y // 2, overlap)
+        pbar = ldm_patched.modules.utils.ProgressBar(steps)
+
+        decode_fn = lambda a: (self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)) + 1.0).float()
+        output = torch.clamp((
+            (ldm_patched.modules.utils.tiled_scale(samples, decode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = 8, output_device=self.output_device, pbar = pbar) +
+            ldm_patched.modules.utils.tiled_scale(samples, decode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = 8, output_device=self.output_device, pbar = pbar) +
+             ldm_patched.modules.utils.tiled_scale(samples, decode_fn, tile_x, tile_y, overlap, upscale_amount = 8, output_device=self.output_device, pbar = pbar))
+            / 3.0) / 2.0, min=0.0, max=1.0)
+        return output
+
+    def encode_tiled_(self, pixel_samples, tile_x=512, tile_y=512, overlap = 64):
+        steps = pixel_samples.shape[0] * ldm_patched.modules.utils.get_tiled_scale_steps(pixel_samples.shape[3], pixel_samples.shape[2], tile_x, tile_y, overlap)
+        steps += pixel_samples.shape[0] * ldm_patched.modules.utils.get_tiled_scale_steps(pixel_samples.shape[3], pixel_samples.shape[2], tile_x // 2, tile_y * 2, overlap)
+        steps += pixel_samples.shape[0] * ldm_patched.modules.utils.get_tiled_scale_steps(pixel_samples.shape[3], pixel_samples.shape[2], tile_x * 2, tile_y // 2, overlap)
+        pbar = ldm_patched.modules.utils.ProgressBar(steps)
+
+        encode_fn = lambda a: self.first_stage_model.encode((2. * a - 1.).to(self.vae_dtype).to(self.device)).float()
+        samples = ldm_patched.modules.utils.tiled_scale(pixel_samples, encode_fn, tile_x, tile_y, overlap, upscale_amount = (1/8), out_channels=4, output_device=self.output_device, pbar=pbar)
+        samples += ldm_patched.modules.utils.tiled_scale(pixel_samples, encode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = (1/8), out_channels=4, output_device=self.output_device, pbar=pbar)
+        samples += ldm_patched.modules.utils.tiled_scale(pixel_samples, encode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = (1/8), out_channels=4, output_device=self.output_device, pbar=pbar)
+        samples /= 3.0
+        return samples
+
+    def decode(self, samples_in):
+        try:
+            memory_used = self.memory_used_decode(samples_in.shape, self.vae_dtype)
+            model_management.load_models_gpu([self.patcher], memory_required=memory_used)
+            free_memory = model_management.get_free_memory(self.device)
+            batch_number = int(free_memory / memory_used)
+            batch_number = max(1, batch_number)
+
+            pixel_samples = torch.empty((samples_in.shape[0], 3, round(samples_in.shape[2] * 8), round(samples_in.shape[3] * 8)), device=self.output_device)
+            for x in range(0, samples_in.shape[0], batch_number):
+                samples = samples_in[x:x+batch_number].to(self.vae_dtype).to(self.device)
+                pixel_samples[x:x+batch_number] = torch.clamp((self.first_stage_model.decode(samples).to(self.output_device).float() + 1.0) / 2.0, min=0.0, max=1.0)
+        except model_management.OOM_EXCEPTION as e:
+            print("Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.")
+            pixel_samples = self.decode_tiled_(samples_in)
+
+        pixel_samples = pixel_samples.to(self.output_device).movedim(1,-1)
+        return pixel_samples
+
+    def decode_tiled(self, samples, tile_x=64, tile_y=64, overlap = 16):
+        model_management.load_model_gpu(self.patcher)
+        output = self.decode_tiled_(samples, tile_x, tile_y, overlap)
+        return output.movedim(1,-1)
+
+    def encode(self, pixel_samples):
+        pixel_samples = pixel_samples.movedim(-1,1)
+        try:
+            memory_used = self.memory_used_encode(pixel_samples.shape, self.vae_dtype)
+            model_management.load_models_gpu([self.patcher], memory_required=memory_used)
+            free_memory = model_management.get_free_memory(self.device)
+            batch_number = int(free_memory / memory_used)
+            batch_number = max(1, batch_number)
+            samples = torch.empty((pixel_samples.shape[0], 4, round(pixel_samples.shape[2] // 8), round(pixel_samples.shape[3] // 8)), device=self.output_device)
+            for x in range(0, pixel_samples.shape[0], batch_number):
+                pixels_in = (2. * pixel_samples[x:x+batch_number] - 1.).to(self.vae_dtype).to(self.device)
+                samples[x:x+batch_number] = self.first_stage_model.encode(pixels_in).to(self.output_device).float()
+
+        except model_management.OOM_EXCEPTION as e:
+            print("Warning: Ran out of memory when regular VAE encoding, retrying with tiled VAE encoding.")
+            samples = self.encode_tiled_(pixel_samples)
+
+        return samples
+
+    def encode_tiled(self, pixel_samples, tile_x=512, tile_y=512, overlap = 64):
+        model_management.load_model_gpu(self.patcher)
+        pixel_samples = pixel_samples.movedim(-1,1)
+        samples = self.encode_tiled_(pixel_samples, tile_x=tile_x, tile_y=tile_y, overlap=overlap)
+        return samples
+
+    def get_sd(self):
+        return self.first_stage_model.state_dict()
+
+class StyleModel:
+    def __init__(self, model, device="cpu"):
+        self.model = model
+
+    def get_cond(self, input):
+        return self.model(input.last_hidden_state)
+
+
+def load_style_model(ckpt_path):
+    model_data = ldm_patched.modules.utils.load_torch_file(ckpt_path, safe_load=True)
+    keys = model_data.keys()
+    if "style_embedding" in keys:
+        model = ldm_patched.t2ia.adapter.StyleAdapter(width=1024, context_dim=768, num_head=8, n_layes=3, num_token=8)
+    else:
+        raise Exception("invalid style model {}".format(ckpt_path))
+    model.load_state_dict(model_data)
+    return StyleModel(model)
+
+
+def load_clip(ckpt_paths, embedding_directory=None):
+    clip_data = []
+    for p in ckpt_paths:
+        clip_data.append(ldm_patched.modules.utils.load_torch_file(p, safe_load=True))
+
+    class EmptyClass:
+        pass
+
+    for i in range(len(clip_data)):
+        if "transformer.resblocks.0.ln_1.weight" in clip_data[i]:
+            clip_data[i] = ldm_patched.modules.utils.transformers_convert(clip_data[i], "", "text_model.", 32)
+
+    clip_target = EmptyClass()
+    clip_target.params = {}
+    if len(clip_data) == 1:
+        if "text_model.encoder.layers.30.mlp.fc1.weight" in clip_data[0]:
+            clip_target.clip = sdxl_clip.SDXLRefinerClipModel
+            clip_target.tokenizer = sdxl_clip.SDXLTokenizer
+        elif "text_model.encoder.layers.22.mlp.fc1.weight" in clip_data[0]:
+            clip_target.clip = sd2_clip.SD2ClipModel
+            clip_target.tokenizer = sd2_clip.SD2Tokenizer
+        else:
+            clip_target.clip = sd1_clip.SD1ClipModel
+            clip_target.tokenizer = sd1_clip.SD1Tokenizer
+    else:
+        clip_target.clip = sdxl_clip.SDXLClipModel
+        clip_target.tokenizer = sdxl_clip.SDXLTokenizer
+
+    clip = CLIP(clip_target, embedding_directory=embedding_directory)
+    for c in clip_data:
+        m, u = clip.load_sd(c)
+        if len(m) > 0:
+            print("clip missing:", m)
+
+        if len(u) > 0:
+            print("clip unexpected:", u)
+    return clip
+
+def load_gligen(ckpt_path):
+    data = ldm_patched.modules.utils.load_torch_file(ckpt_path, safe_load=True)
+    model = gligen.load_gligen(data)
+    if model_management.should_use_fp16():
+        model = model.half()
+    return ldm_patched.modules.model_patcher.ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=model_management.unet_offload_device())
+
+def load_checkpoint(config_path=None, ckpt_path=None, output_vae=True, output_clip=True, embedding_directory=None, state_dict=None, config=None):
+    #TODO: this function is a mess and should be removed eventually
+    if config is None:
+        with open(config_path, 'r') as stream:
+            config = yaml.safe_load(stream)
+    model_config_params = config['model']['params']
+    clip_config = model_config_params['cond_stage_config']
+    scale_factor = model_config_params['scale_factor']
+    vae_config = model_config_params['first_stage_config']
+
+    fp16 = False
+    if "unet_config" in model_config_params:
+        if "params" in model_config_params["unet_config"]:
+            unet_config = model_config_params["unet_config"]["params"]
+            if "use_fp16" in unet_config:
+                fp16 = unet_config.pop("use_fp16")
+                if fp16:
+                    unet_config["dtype"] = torch.float16
+
+    noise_aug_config = None
+    if "noise_aug_config" in model_config_params:
+        noise_aug_config = model_config_params["noise_aug_config"]
+
+    model_type = model_base.ModelType.EPS
+
+    if "parameterization" in model_config_params:
+        if model_config_params["parameterization"] == "v":
+            model_type = model_base.ModelType.V_PREDICTION
+
+    clip = None
+    vae = None
+
+    class WeightsLoader(torch.nn.Module):
+        pass
+
+    if state_dict is None:
+        state_dict = ldm_patched.modules.utils.load_torch_file(ckpt_path)
+
+    class EmptyClass:
+        pass
+
+    model_config = ldm_patched.modules.supported_models_base.BASE({})
+
+    from . import latent_formats
+    model_config.latent_format = latent_formats.SD15(scale_factor=scale_factor)
+    model_config.unet_config = model_detection.convert_config(unet_config)
+
+    if config['model']["target"].endswith("ImageEmbeddingConditionedLatentDiffusion"):
+        model = model_base.SD21UNCLIP(model_config, noise_aug_config["params"], model_type=model_type)
+    else:
+        model = model_base.BaseModel(model_config, model_type=model_type)
+
+    if config['model']["target"].endswith("LatentInpaintDiffusion"):
+        model.set_inpaint()
+
+    if fp16:
+        model = model.half()
+
+    offload_device = model_management.unet_offload_device()
+    model = model.to(offload_device)
+    model.load_model_weights(state_dict, "model.diffusion_model.")
+
+    if output_vae:
+        vae_sd = ldm_patched.modules.utils.state_dict_prefix_replace(state_dict, {"first_stage_model.": ""}, filter_keys=True)
+        vae = VAE(sd=vae_sd, config=vae_config)
+
+    if output_clip:
+        w = WeightsLoader()
+        clip_target = EmptyClass()
+        clip_target.params = clip_config.get("params", {})
+        if clip_config["target"].endswith("FrozenOpenCLIPEmbedder"):
+            clip_target.clip = sd2_clip.SD2ClipModel
+            clip_target.tokenizer = sd2_clip.SD2Tokenizer
+            clip = CLIP(clip_target, embedding_directory=embedding_directory)
+            w.cond_stage_model = clip.cond_stage_model.clip_h
+        elif clip_config["target"].endswith("FrozenCLIPEmbedder"):
+            clip_target.clip = sd1_clip.SD1ClipModel
+            clip_target.tokenizer = sd1_clip.SD1Tokenizer
+            clip = CLIP(clip_target, embedding_directory=embedding_directory)
+            w.cond_stage_model = clip.cond_stage_model.clip_l
+        load_clip_weights(w, state_dict)
+
+    return (ldm_patched.modules.model_patcher.ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=offload_device), clip, vae)
+
+def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True):
+    sd = ldm_patched.modules.utils.load_torch_file(ckpt_path)
+    sd_keys = sd.keys()
+    clip = None
+    clipvision = None
+    vae = None
+    model = None
+    model_patcher = None
+    clip_target = None
+
+    parameters = ldm_patched.modules.utils.calculate_parameters(sd, "model.diffusion_model.")
+    unet_dtype = model_management.unet_dtype(model_params=parameters)
+    load_device = model_management.get_torch_device()
+    manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device)
+
+    class WeightsLoader(torch.nn.Module):
+        pass
+
+    model_config = model_detection.model_config_from_unet(sd, "model.diffusion_model.", unet_dtype)
+    model_config.set_manual_cast(manual_cast_dtype)
+
+    if model_config is None:
+        raise RuntimeError("ERROR: Could not detect model type of: {}".format(ckpt_path))
+
+    if model_config.clip_vision_prefix is not None:
+        if output_clipvision:
+            clipvision = clip_vision.load_clipvision_from_sd(sd, model_config.clip_vision_prefix, True)
+
+    if output_model:
+        inital_load_device = model_management.unet_inital_load_device(parameters, unet_dtype)
+        offload_device = model_management.unet_offload_device()
+        model = model_config.get_model(sd, "model.diffusion_model.", device=inital_load_device)
+        model.load_model_weights(sd, "model.diffusion_model.")
+
+    if output_vae:
+        vae_sd = ldm_patched.modules.utils.state_dict_prefix_replace(sd, {"first_stage_model.": ""}, filter_keys=True)
+        vae_sd = model_config.process_vae_state_dict(vae_sd)
+        vae = VAE(sd=vae_sd)
+
+    if output_clip:
+        w = WeightsLoader()
+        clip_target = model_config.clip_target()
+        if clip_target is not None:
+            clip = CLIP(clip_target, embedding_directory=embedding_directory)
+            w.cond_stage_model = clip.cond_stage_model
+            sd = model_config.process_clip_state_dict(sd)
+            load_model_weights(w, sd)
+
+    left_over = sd.keys()
+    if len(left_over) > 0:
+        print("left over keys:", left_over)
+
+    if output_model:
+        model_patcher = ldm_patched.modules.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=model_management.unet_offload_device(), current_device=inital_load_device)
+        if inital_load_device != torch.device("cpu"):
+            print("loaded straight to GPU")
+            model_management.load_model_gpu(model_patcher)
+
+    return (model_patcher, clip, vae, clipvision)
+
+
+def load_unet_state_dict(sd): #load unet in diffusers format
+    parameters = ldm_patched.modules.utils.calculate_parameters(sd)
+    unet_dtype = model_management.unet_dtype(model_params=parameters)
+    load_device = model_management.get_torch_device()
+    manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device)
+
+    if "input_blocks.0.0.weight" in sd: #ldm
+        model_config = model_detection.model_config_from_unet(sd, "", unet_dtype)
+        if model_config is None:
+            return None
+        new_sd = sd
+
+    else: #diffusers
+        model_config = model_detection.model_config_from_diffusers_unet(sd, unet_dtype)
+        if model_config is None:
+            return None
+
+        diffusers_keys = ldm_patched.modules.utils.unet_to_diffusers(model_config.unet_config)
+
+        new_sd = {}
+        for k in diffusers_keys:
+            if k in sd:
+                new_sd[diffusers_keys[k]] = sd.pop(k)
+            else:
+                print(diffusers_keys[k], k)
+    offload_device = model_management.unet_offload_device()
+    model_config.set_manual_cast(manual_cast_dtype)
+    model = model_config.get_model(new_sd, "")
+    model = model.to(offload_device)
+    model.load_model_weights(new_sd, "")
+    left_over = sd.keys()
+    if len(left_over) > 0:
+        print("left over keys in unet:", left_over)
+    return ldm_patched.modules.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=offload_device)
+
+def load_unet(unet_path):
+    sd = ldm_patched.modules.utils.load_torch_file(unet_path)
+    model = load_unet_state_dict(sd)
+    if model is None:
+        print("ERROR UNSUPPORTED UNET", unet_path)
+        raise RuntimeError("ERROR: Could not detect model type of: {}".format(unet_path))
+    return model
+
+def save_checkpoint(output_path, model, clip, vae, metadata=None):
+    model_management.load_models_gpu([model, clip.load_model()])
+    sd = model.model.state_dict_for_saving(clip.get_sd(), vae.get_sd())
+    ldm_patched.modules.utils.save_torch_file(sd, output_path, metadata=metadata)

ldm_patched/modules/sd1_clip.py

@@ -0,0 +1,519 @@
+import os
+
+from transformers import CLIPTokenizer
+import ldm_patched.modules.ops
+import torch
+import traceback
+import zipfile
+from . import model_management
+import contextlib
+import ldm_patched.modules.clip_model
+import json
+
+def gen_empty_tokens(special_tokens, length):
+    start_token = special_tokens.get("start", None)
+    end_token = special_tokens.get("end", None)
+    pad_token = special_tokens.get("pad")
+    output = []
+    if start_token is not None:
+        output.append(start_token)
+    if end_token is not None:
+        output.append(end_token)
+    output += [pad_token] * (length - len(output))
+    return output
+
+class ClipTokenWeightEncoder:
+    def encode_token_weights(self, token_weight_pairs):
+        to_encode = list()
+        max_token_len = 0
+        has_weights = False
+        for x in token_weight_pairs:
+            tokens = list(map(lambda a: a[0], x))
+            max_token_len = max(len(tokens), max_token_len)
+            has_weights = has_weights or not all(map(lambda a: a[1] == 1.0, x))
+            to_encode.append(tokens)
+
+        sections = len(to_encode)
+        if has_weights or sections == 0:
+            to_encode.append(gen_empty_tokens(self.special_tokens, max_token_len))
+
+        out, pooled = self.encode(to_encode)
+        if pooled is not None:
+            first_pooled = pooled[0:1].to(model_management.intermediate_device())
+        else:
+            first_pooled = pooled
+
+        output = []
+        for k in range(0, sections):
+            z = out[k:k+1]
+            if has_weights:
+                z_empty = out[-1]
+                for i in range(len(z)):
+                    for j in range(len(z[i])):
+                        weight = token_weight_pairs[k][j][1]
+                        if weight != 1.0:
+                            z[i][j] = (z[i][j] - z_empty[j]) * weight + z_empty[j]
+            output.append(z)
+
+        if (len(output) == 0):
+            return out[-1:].to(model_management.intermediate_device()), first_pooled
+        return torch.cat(output, dim=-2).to(model_management.intermediate_device()), first_pooled
+
+class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
+    """Uses the CLIP transformer encoder for text (from huggingface)"""
+    LAYERS = [
+        "last",
+        "pooled",
+        "hidden"
+    ]
+    def __init__(self, version="openai/clip-vit-large-patch14", device="cpu", max_length=77,
+                 freeze=True, layer="last", layer_idx=None, textmodel_json_config=None, dtype=None, model_class=ldm_patched.modules.clip_model.CLIPTextModel,
+                 special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=True):  # clip-vit-base-patch32
+        super().__init__()
+        assert layer in self.LAYERS
+
+        if textmodel_json_config is None:
+            textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_clip_config.json")
+
+        with open(textmodel_json_config) as f:
+            config = json.load(f)
+
+        self.transformer = model_class(config, dtype, device, ldm_patched.modules.ops.manual_cast)
+        self.num_layers = self.transformer.num_layers
+
+        self.max_length = max_length
+        if freeze:
+            self.freeze()
+        self.layer = layer
+        self.layer_idx = None
+        self.special_tokens = special_tokens
+        self.text_projection = torch.nn.Parameter(torch.eye(self.transformer.get_input_embeddings().weight.shape[1]))
+        self.logit_scale = torch.nn.Parameter(torch.tensor(4.6055))
+        self.enable_attention_masks = False
+
+        self.layer_norm_hidden_state = layer_norm_hidden_state
+        if layer == "hidden":
+            assert layer_idx is not None
+            assert abs(layer_idx) < self.num_layers
+            self.clip_layer(layer_idx)
+        self.layer_default = (self.layer, self.layer_idx)
+
+    def freeze(self):
+        self.transformer = self.transformer.eval()
+        #self.train = disabled_train
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def clip_layer(self, layer_idx):
+        if abs(layer_idx) > self.num_layers:
+            self.layer = "last"
+        else:
+            self.layer = "hidden"
+            self.layer_idx = layer_idx
+
+    def reset_clip_layer(self):
+        self.layer = self.layer_default[0]
+        self.layer_idx = self.layer_default[1]
+
+    def set_up_textual_embeddings(self, tokens, current_embeds):
+        out_tokens = []
+        next_new_token = token_dict_size = current_embeds.weight.shape[0] - 1
+        embedding_weights = []
+
+        for x in tokens:
+            tokens_temp = []
+            for y in x:
+                if isinstance(y, int):
+                    if y == token_dict_size: #EOS token
+                        y = -1
+                    tokens_temp += [y]
+                else:
+                    if y.shape[0] == current_embeds.weight.shape[1]:
+                        embedding_weights += [y]
+                        tokens_temp += [next_new_token]
+                        next_new_token += 1
+                    else:
+                        print("WARNING: shape mismatch when trying to apply embedding, embedding will be ignored", y.shape[0], current_embeds.weight.shape[1])
+            while len(tokens_temp) < len(x):
+                tokens_temp += [self.special_tokens["pad"]]
+            out_tokens += [tokens_temp]
+
+        n = token_dict_size
+        if len(embedding_weights) > 0:
+            new_embedding = torch.nn.Embedding(next_new_token + 1, current_embeds.weight.shape[1], device=current_embeds.weight.device, dtype=current_embeds.weight.dtype)
+            new_embedding.weight[:token_dict_size] = current_embeds.weight[:-1]
+            for x in embedding_weights:
+                new_embedding.weight[n] = x
+                n += 1
+            new_embedding.weight[n] = current_embeds.weight[-1] #EOS embedding
+            self.transformer.set_input_embeddings(new_embedding)
+
+        processed_tokens = []
+        for x in out_tokens:
+            processed_tokens += [list(map(lambda a: n if a == -1 else a, x))] #The EOS token should always be the largest one
+
+        return processed_tokens
+
+    def forward(self, tokens):
+        backup_embeds = self.transformer.get_input_embeddings()
+        device = backup_embeds.weight.device
+        tokens = self.set_up_textual_embeddings(tokens, backup_embeds)
+        tokens = torch.LongTensor(tokens).to(device)
+
+        attention_mask = None
+        if self.enable_attention_masks:
+            attention_mask = torch.zeros_like(tokens)
+            max_token = self.transformer.get_input_embeddings().weight.shape[0] - 1
+            for x in range(attention_mask.shape[0]):
+                for y in range(attention_mask.shape[1]):
+                    attention_mask[x, y] = 1
+                    if tokens[x, y] == max_token:
+                        break
+
+        outputs = self.transformer(tokens, attention_mask, intermediate_output=self.layer_idx, final_layer_norm_intermediate=self.layer_norm_hidden_state)
+        self.transformer.set_input_embeddings(backup_embeds)
+
+        if self.layer == "last":
+            z = outputs[0]
+        else:
+            z = outputs[1]
+
+        if outputs[2] is not None:
+            pooled_output = outputs[2].float()
+        else:
+            pooled_output = None
+
+        if self.text_projection is not None and pooled_output is not None:
+            pooled_output = pooled_output.float().to(self.text_projection.device) @ self.text_projection.float()
+        return z.float(), pooled_output
+
+    def encode(self, tokens):
+        return self(tokens)
+
+    def load_sd(self, sd):
+        if "text_projection" in sd:
+            self.text_projection[:] = sd.pop("text_projection")
+        if "text_projection.weight" in sd:
+            self.text_projection[:] = sd.pop("text_projection.weight").transpose(0, 1)
+        return self.transformer.load_state_dict(sd, strict=False)
+
+def parse_parentheses(string):
+    result = []
+    current_item = ""
+    nesting_level = 0
+    for char in string:
+        if char == "(":
+            if nesting_level == 0:
+                if current_item:
+                    result.append(current_item)
+                    current_item = "("
+                else:
+                    current_item = "("
+            else:
+                current_item += char
+            nesting_level += 1
+        elif char == ")":
+            nesting_level -= 1
+            if nesting_level == 0:
+                result.append(current_item + ")")
+                current_item = ""
+            else:
+                current_item += char
+        else:
+            current_item += char
+    if current_item:
+        result.append(current_item)
+    return result
+
+def token_weights(string, current_weight):
+    a = parse_parentheses(string)
+    out = []
+    for x in a:
+        weight = current_weight
+        if len(x) >= 2 and x[-1] == ')' and x[0] == '(':
+            x = x[1:-1]
+            xx = x.rfind(":")
+            weight *= 1.1
+            if xx > 0:
+                try:
+                    weight = float(x[xx+1:])
+                    x = x[:xx]
+                except:
+                    pass
+            out += token_weights(x, weight)
+        else:
+            out += [(x, current_weight)]
+    return out
+
+def escape_important(text):
+    text = text.replace("\\)", "\0\1")
+    text = text.replace("\\(", "\0\2")
+    return text
+
+def unescape_important(text):
+    text = text.replace("\0\1", ")")
+    text = text.replace("\0\2", "(")
+    return text
+
+def safe_load_embed_zip(embed_path):
+    with zipfile.ZipFile(embed_path) as myzip:
+        names = list(filter(lambda a: "data/" in a, myzip.namelist()))
+        names.reverse()
+        for n in names:
+            with myzip.open(n) as myfile:
+                data = myfile.read()
+                number = len(data) // 4
+                length_embed = 1024 #sd2.x
+                if number < 768:
+                    continue
+                if number % 768 == 0:
+                    length_embed = 768 #sd1.x
+                num_embeds = number // length_embed
+                embed = torch.frombuffer(data, dtype=torch.float)
+                out = embed.reshape((num_embeds, length_embed)).clone()
+                del embed
+                return out
+
+def expand_directory_list(directories):
+    dirs = set()
+    for x in directories:
+        dirs.add(x)
+        for root, subdir, file in os.walk(x, followlinks=True):
+            dirs.add(root)
+    return list(dirs)
+
+def load_embed(embedding_name, embedding_directory, embedding_size, embed_key=None):
+    if isinstance(embedding_directory, str):
+        embedding_directory = [embedding_directory]
+
+    embedding_directory = expand_directory_list(embedding_directory)
+
+    valid_file = None
+    for embed_dir in embedding_directory:
+        embed_path = os.path.abspath(os.path.join(embed_dir, embedding_name))
+        embed_dir = os.path.abspath(embed_dir)
+        try:
+            if os.path.commonpath((embed_dir, embed_path)) != embed_dir:
+                continue
+        except:
+            continue
+        if not os.path.isfile(embed_path):
+            extensions = ['.safetensors', '.pt', '.bin']
+            for x in extensions:
+                t = embed_path + x
+                if os.path.isfile(t):
+                    valid_file = t
+                    break
+        else:
+            valid_file = embed_path
+        if valid_file is not None:
+            break
+
+    if valid_file is None:
+        return None
+
+    embed_path = valid_file
+
+    embed_out = None
+
+    try:
+        if embed_path.lower().endswith(".safetensors"):
+            import safetensors.torch
+            embed = safetensors.torch.load_file(embed_path, device="cpu")
+        else:
+            if 'weights_only' in torch.load.__code__.co_varnames:
+                try:
+                    embed = torch.load(embed_path, weights_only=True, map_location="cpu")
+                except:
+                    embed_out = safe_load_embed_zip(embed_path)
+            else:
+                embed = torch.load(embed_path, map_location="cpu")
+    except Exception as e:
+        print(traceback.format_exc())
+        print()
+        print("error loading embedding, skipping loading:", embedding_name)
+        return None
+
+    if embed_out is None:
+        if 'string_to_param' in embed:
+            values = embed['string_to_param'].values()
+            embed_out = next(iter(values))
+        elif isinstance(embed, list):
+            out_list = []
+            for x in range(len(embed)):
+                for k in embed[x]:
+                    t = embed[x][k]
+                    if t.shape[-1] != embedding_size:
+                        continue
+                    out_list.append(t.reshape(-1, t.shape[-1]))
+            embed_out = torch.cat(out_list, dim=0)
+        elif embed_key is not None and embed_key in embed:
+            embed_out = embed[embed_key]
+        else:
+            values = embed.values()
+            embed_out = next(iter(values))
+    return embed_out
+
+class SDTokenizer:
+    def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, pad_to_max_length=True):
+        if tokenizer_path is None:
+            tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_tokenizer")
+        self.tokenizer = tokenizer_class.from_pretrained(tokenizer_path)
+        self.max_length = max_length
+
+        empty = self.tokenizer('')["input_ids"]
+        if has_start_token:
+            self.tokens_start = 1
+            self.start_token = empty[0]
+            self.end_token = empty[1]
+        else:
+            self.tokens_start = 0
+            self.start_token = None
+            self.end_token = empty[0]
+        self.pad_with_end = pad_with_end
+        self.pad_to_max_length = pad_to_max_length
+
+        vocab = self.tokenizer.get_vocab()
+        self.inv_vocab = {v: k for k, v in vocab.items()}
+        self.embedding_directory = embedding_directory
+        self.max_word_length = 8
+        self.embedding_identifier = "embedding:"
+        self.embedding_size = embedding_size
+        self.embedding_key = embedding_key
+
+    def _try_get_embedding(self, embedding_name:str):
+        '''
+        Takes a potential embedding name and tries to retrieve it.
+        Returns a Tuple consisting of the embedding and any leftover string, embedding can be None.
+        '''
+        embed = load_embed(embedding_name, self.embedding_directory, self.embedding_size, self.embedding_key)
+        if embed is None:
+            stripped = embedding_name.strip(',')
+            if len(stripped) < len(embedding_name):
+                embed = load_embed(stripped, self.embedding_directory, self.embedding_size, self.embedding_key)
+                return (embed, embedding_name[len(stripped):])
+        return (embed, "")
+
+
+    def tokenize_with_weights(self, text:str, return_word_ids=False):
+        '''
+        Takes a prompt and converts it to a list of (token, weight, word id) elements.
+        Tokens can both be integer tokens and pre computed CLIP tensors.
+        Word id values are unique per word and embedding, where the id 0 is reserved for non word tokens.
+        Returned list has the dimensions NxM where M is the input size of CLIP
+        '''
+        if self.pad_with_end:
+            pad_token = self.end_token
+        else:
+            pad_token = 0
+
+        text = escape_important(text)
+        parsed_weights = token_weights(text, 1.0)
+
+        #tokenize words
+        tokens = []
+        for weighted_segment, weight in parsed_weights:
+            to_tokenize = unescape_important(weighted_segment).replace("\n", " ").split(' ')
+            to_tokenize = [x for x in to_tokenize if x != ""]
+            for word in to_tokenize:
+                #if we find an embedding, deal with the embedding
+                if word.startswith(self.embedding_identifier) and self.embedding_directory is not None:
+                    embedding_name = word[len(self.embedding_identifier):].strip('\n')
+                    embed, leftover = self._try_get_embedding(embedding_name)
+                    if embed is None:
+                        print(f"warning, embedding:{embedding_name} does not exist, ignoring")
+                    else:
+                        if len(embed.shape) == 1:
+                            tokens.append([(embed, weight)])
+                        else:
+                            tokens.append([(embed[x], weight) for x in range(embed.shape[0])])
+                    #if we accidentally have leftover text, continue parsing using leftover, else move on to next word
+                    if leftover != "":
+                        word = leftover
+                    else:
+                        continue
+                #parse word
+                tokens.append([(t, weight) for t in self.tokenizer(word)["input_ids"][self.tokens_start:-1]])
+
+        #reshape token array to CLIP input size
+        batched_tokens = []
+        batch = []
+        if self.start_token is not None:
+            batch.append((self.start_token, 1.0, 0))
+        batched_tokens.append(batch)
+        for i, t_group in enumerate(tokens):
+            #determine if we're going to try and keep the tokens in a single batch
+            is_large = len(t_group) >= self.max_word_length
+
+            while len(t_group) > 0:
+                if len(t_group) + len(batch) > self.max_length - 1:
+                    remaining_length = self.max_length - len(batch) - 1
+                    #break word in two and add end token
+                    if is_large:
+                        batch.extend([(t,w,i+1) for t,w in t_group[:remaining_length]])
+                        batch.append((self.end_token, 1.0, 0))
+                        t_group = t_group[remaining_length:]
+                    #add end token and pad
+                    else:
+                        batch.append((self.end_token, 1.0, 0))
+                        if self.pad_to_max_length:
+                            batch.extend([(pad_token, 1.0, 0)] * (remaining_length))
+                    #start new batch
+                    batch = []
+                    if self.start_token is not None:
+                        batch.append((self.start_token, 1.0, 0))
+                    batched_tokens.append(batch)
+                else:
+                    batch.extend([(t,w,i+1) for t,w in t_group])
+                    t_group = []
+
+        #fill last batch
+        batch.append((self.end_token, 1.0, 0))
+        if self.pad_to_max_length:
+            batch.extend([(pad_token, 1.0, 0)] * (self.max_length - len(batch)))
+
+        if not return_word_ids:
+            batched_tokens = [[(t, w) for t, w,_ in x] for x in batched_tokens]
+
+        return batched_tokens
+
+
+    def untokenize(self, token_weight_pair):
+        return list(map(lambda a: (a, self.inv_vocab[a[0]]), token_weight_pair))
+
+
+class SD1Tokenizer:
+    def __init__(self, embedding_directory=None, clip_name="l", tokenizer=SDTokenizer):
+        self.clip_name = clip_name
+        self.clip = "clip_{}".format(self.clip_name)
+        setattr(self, self.clip, tokenizer(embedding_directory=embedding_directory))
+
+    def tokenize_with_weights(self, text:str, return_word_ids=False):
+        out = {}
+        out[self.clip_name] = getattr(self, self.clip).tokenize_with_weights(text, return_word_ids)
+        return out
+
+    def untokenize(self, token_weight_pair):
+        return getattr(self, self.clip).untokenize(token_weight_pair)
+
+
+class SD1ClipModel(torch.nn.Module):
+    def __init__(self, device="cpu", dtype=None, clip_name="l", clip_model=SDClipModel, **kwargs):
+        super().__init__()
+        self.clip_name = clip_name
+        self.clip = "clip_{}".format(self.clip_name)
+        setattr(self, self.clip, clip_model(device=device, dtype=dtype, **kwargs))
+
+    def clip_layer(self, layer_idx):
+        getattr(self, self.clip).clip_layer(layer_idx)
+
+    def reset_clip_layer(self):
+        getattr(self, self.clip).reset_clip_layer()
+
+    def encode_token_weights(self, token_weight_pairs):
+        token_weight_pairs = token_weight_pairs[self.clip_name]
+        out, pooled = getattr(self, self.clip).encode_token_weights(token_weight_pairs)
+        return out, pooled
+
+    def load_sd(self, sd):
+        return getattr(self, self.clip).load_sd(sd)

ldm_patched/modules/sd1_clip_config.json

@@ -0,0 +1,25 @@
+{
+  "_name_or_path": "openai/clip-vit-large-patch14",
+  "architectures": [
+    "CLIPTextModel"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "dropout": 0.0,
+  "eos_token_id": 2,
+  "hidden_act": "quick_gelu",
+  "hidden_size": 768,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 77,
+  "model_type": "clip_text_model",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 1,
+  "projection_dim": 768,
+  "torch_dtype": "float32",
+  "transformers_version": "4.24.0",
+  "vocab_size": 49408
+}

ldm_patched/modules/sd1_tokenizer/special_tokens_map.json

@@ -0,0 +1,24 @@
+{
+  "bos_token": {
+    "content": "<|startoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<|endoftext|>",
+  "unk_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

ldm_patched/modules/sd1_tokenizer/tokenizer_config.json

@@ -0,0 +1,34 @@
+{
+  "add_prefix_space": false,
+  "bos_token": {
+    "__type": "AddedToken",
+    "content": "<|startoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "do_lower_case": true,
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "errors": "replace",
+  "model_max_length": 77,
+  "name_or_path": "openai/clip-vit-large-patch14",
+  "pad_token": "<|endoftext|>",
+  "special_tokens_map_file": "./special_tokens_map.json",
+  "tokenizer_class": "CLIPTokenizer",
+  "unk_token": {
+    "__type": "AddedToken",
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

ldm_patched/modules/sd2_clip.py

@@ -0,0 +1,24 @@
+from ldm_patched.modules import sd1_clip
+import torch
+import os
+
+class SD2ClipHModel(sd1_clip.SDClipModel):
+    def __init__(self, arch="ViT-H-14", device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None):
+        if layer == "penultimate":
+            layer="hidden"
+            layer_idx=-2
+
+        textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd2_clip_config.json")
+        super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 49406, "end": 49407, "pad": 0})
+
+class SD2ClipHTokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, tokenizer_path=None, embedding_directory=None):
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=1024)
+
+class SD2Tokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None):
+        super().__init__(embedding_directory=embedding_directory, clip_name="h", tokenizer=SD2ClipHTokenizer)
+
+class SD2ClipModel(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, **kwargs):
+        super().__init__(device=device, dtype=dtype, clip_name="h", clip_model=SD2ClipHModel, **kwargs)

ldm_patched/modules/sd2_clip_config.json

@@ -0,0 +1,23 @@
+{
+  "architectures": [
+    "CLIPTextModel"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "dropout": 0.0,
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_size": 1024,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 77,
+  "model_type": "clip_text_model",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 24,
+  "pad_token_id": 1,
+  "projection_dim": 1024,
+  "torch_dtype": "float32",
+  "vocab_size": 49408
+}

ldm_patched/modules/sdxl_clip.py

@@ -0,0 +1,66 @@
+from ldm_patched.modules import sd1_clip
+import torch
+import os
+
+class SDXLClipG(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None):
+        if layer == "penultimate":
+            layer="hidden"
+            layer_idx=-2
+
+        textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_config_bigg.json")
+        super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype,
+                         special_tokens={"start": 49406, "end": 49407, "pad": 0}, layer_norm_hidden_state=False)
+
+    def load_sd(self, sd):
+        return super().load_sd(sd)
+
+class SDXLClipGTokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, tokenizer_path=None, embedding_directory=None):
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=1280, embedding_key='clip_g')
+
+
+class SDXLTokenizer:
+    def __init__(self, embedding_directory=None):
+        self.clip_l = sd1_clip.SDTokenizer(embedding_directory=embedding_directory)
+        self.clip_g = SDXLClipGTokenizer(embedding_directory=embedding_directory)
+
+    def tokenize_with_weights(self, text:str, return_word_ids=False):
+        out = {}
+        out["g"] = self.clip_g.tokenize_with_weights(text, return_word_ids)
+        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids)
+        return out
+
+    def untokenize(self, token_weight_pair):
+        return self.clip_g.untokenize(token_weight_pair)
+
+class SDXLClipModel(torch.nn.Module):
+    def __init__(self, device="cpu", dtype=None):
+        super().__init__()
+        self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False)
+        self.clip_g = SDXLClipG(device=device, dtype=dtype)
+
+    def clip_layer(self, layer_idx):
+        self.clip_l.clip_layer(layer_idx)
+        self.clip_g.clip_layer(layer_idx)
+
+    def reset_clip_layer(self):
+        self.clip_g.reset_clip_layer()
+        self.clip_l.reset_clip_layer()
+
+    def encode_token_weights(self, token_weight_pairs):
+        token_weight_pairs_g = token_weight_pairs["g"]
+        token_weight_pairs_l = token_weight_pairs["l"]
+        g_out, g_pooled = self.clip_g.encode_token_weights(token_weight_pairs_g)
+        l_out, l_pooled = self.clip_l.encode_token_weights(token_weight_pairs_l)
+        return torch.cat([l_out, g_out], dim=-1), g_pooled
+
+    def load_sd(self, sd):
+        if "text_model.encoder.layers.30.mlp.fc1.weight" in sd:
+            return self.clip_g.load_sd(sd)
+        else:
+            return self.clip_l.load_sd(sd)
+
+class SDXLRefinerClipModel(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None):
+        super().__init__(device=device, dtype=dtype, clip_name="g", clip_model=SDXLClipG)

ldm_patched/modules/supported_models.py

@@ -0,0 +1,246 @@
+import torch
+from . import model_base
+from . import utils
+
+from . import sd1_clip
+from . import sd2_clip
+from . import sdxl_clip
+
+from . import supported_models_base
+from . import latent_formats
+
+from . import diffusers_convert
+
+class SD15(supported_models_base.BASE):
+    unet_config = {
+        "context_dim": 768,
+        "model_channels": 320,
+        "use_linear_in_transformer": False,
+        "adm_in_channels": None,
+        "use_temporal_attention": False,
+    }
+
+    unet_extra_config = {
+        "num_heads": 8,
+        "num_head_channels": -1,
+    }
+
+    latent_format = latent_formats.SD15
+
+    def process_clip_state_dict(self, state_dict):
+        k = list(state_dict.keys())
+        for x in k:
+            if x.startswith("cond_stage_model.transformer.") and not x.startswith("cond_stage_model.transformer.text_model."):
+                y = x.replace("cond_stage_model.transformer.", "cond_stage_model.transformer.text_model.")
+                state_dict[y] = state_dict.pop(x)
+
+        if 'cond_stage_model.transformer.text_model.embeddings.position_ids' in state_dict:
+            ids = state_dict['cond_stage_model.transformer.text_model.embeddings.position_ids']
+            if ids.dtype == torch.float32:
+                state_dict['cond_stage_model.transformer.text_model.embeddings.position_ids'] = ids.round()
+
+        replace_prefix = {}
+        replace_prefix["cond_stage_model."] = "cond_stage_model.clip_l."
+        state_dict = utils.state_dict_prefix_replace(state_dict, replace_prefix)
+        return state_dict
+
+    def process_clip_state_dict_for_saving(self, state_dict):
+        replace_prefix = {"clip_l.": "cond_stage_model."}
+        return utils.state_dict_prefix_replace(state_dict, replace_prefix)
+
+    def clip_target(self):
+        return supported_models_base.ClipTarget(sd1_clip.SD1Tokenizer, sd1_clip.SD1ClipModel)
+
+class SD20(supported_models_base.BASE):
+    unet_config = {
+        "context_dim": 1024,
+        "model_channels": 320,
+        "use_linear_in_transformer": True,
+        "adm_in_channels": None,
+        "use_temporal_attention": False,
+    }
+
+    latent_format = latent_formats.SD15
+
+    def model_type(self, state_dict, prefix=""):
+        if self.unet_config["in_channels"] == 4: #SD2.0 inpainting models are not v prediction
+            k = "{}output_blocks.11.1.transformer_blocks.0.norm1.bias".format(prefix)
+            out = state_dict[k]
+            if torch.std(out, unbiased=False) > 0.09: # not sure how well this will actually work. I guess we will find out.
+                return model_base.ModelType.V_PREDICTION
+        return model_base.ModelType.EPS
+
+    def process_clip_state_dict(self, state_dict):
+        replace_prefix = {}
+        replace_prefix["conditioner.embedders.0.model."] = "cond_stage_model.model." #SD2 in sgm format
+        state_dict = utils.state_dict_prefix_replace(state_dict, replace_prefix)
+
+        state_dict = utils.transformers_convert(state_dict, "cond_stage_model.model.", "cond_stage_model.clip_h.transformer.text_model.", 24)
+        return state_dict
+
+    def process_clip_state_dict_for_saving(self, state_dict):
+        replace_prefix = {}
+        replace_prefix["clip_h"] = "cond_stage_model.model"
+        state_dict = utils.state_dict_prefix_replace(state_dict, replace_prefix)
+        state_dict = diffusers_convert.convert_text_enc_state_dict_v20(state_dict)
+        return state_dict
+
+    def clip_target(self):
+        return supported_models_base.ClipTarget(sd2_clip.SD2Tokenizer, sd2_clip.SD2ClipModel)
+
+class SD21UnclipL(SD20):
+    unet_config = {
+        "context_dim": 1024,
+        "model_channels": 320,
+        "use_linear_in_transformer": True,
+        "adm_in_channels": 1536,
+        "use_temporal_attention": False,
+    }
+
+    clip_vision_prefix = "embedder.model.visual."
+    noise_aug_config = {"noise_schedule_config": {"timesteps": 1000, "beta_schedule": "squaredcos_cap_v2"}, "timestep_dim": 768}
+
+
+class SD21UnclipH(SD20):
+    unet_config = {
+        "context_dim": 1024,
+        "model_channels": 320,
+        "use_linear_in_transformer": True,
+        "adm_in_channels": 2048,
+        "use_temporal_attention": False,
+    }
+
+    clip_vision_prefix = "embedder.model.visual."
+    noise_aug_config = {"noise_schedule_config": {"timesteps": 1000, "beta_schedule": "squaredcos_cap_v2"}, "timestep_dim": 1024}
+
+class SDXLRefiner(supported_models_base.BASE):
+    unet_config = {
+        "model_channels": 384,
+        "use_linear_in_transformer": True,
+        "context_dim": 1280,
+        "adm_in_channels": 2560,
+        "transformer_depth": [0, 0, 4, 4, 4, 4, 0, 0],
+        "use_temporal_attention": False,
+    }
+
+    latent_format = latent_formats.SDXL
+
+    def get_model(self, state_dict, prefix="", device=None):
+        return model_base.SDXLRefiner(self, device=device)
+
+    def process_clip_state_dict(self, state_dict):
+        keys_to_replace = {}
+        replace_prefix = {}
+
+        state_dict = utils.transformers_convert(state_dict, "conditioner.embedders.0.model.", "cond_stage_model.clip_g.transformer.text_model.", 32)
+        keys_to_replace["conditioner.embedders.0.model.text_projection"] = "cond_stage_model.clip_g.text_projection"
+        keys_to_replace["conditioner.embedders.0.model.logit_scale"] = "cond_stage_model.clip_g.logit_scale"
+
+        state_dict = utils.state_dict_key_replace(state_dict, keys_to_replace)
+        return state_dict
+
+    def process_clip_state_dict_for_saving(self, state_dict):
+        replace_prefix = {}
+        state_dict_g = diffusers_convert.convert_text_enc_state_dict_v20(state_dict, "clip_g")
+        if "clip_g.transformer.text_model.embeddings.position_ids" in state_dict_g:
+            state_dict_g.pop("clip_g.transformer.text_model.embeddings.position_ids")
+        replace_prefix["clip_g"] = "conditioner.embedders.0.model"
+        state_dict_g = utils.state_dict_prefix_replace(state_dict_g, replace_prefix)
+        return state_dict_g
+
+    def clip_target(self):
+        return supported_models_base.ClipTarget(sdxl_clip.SDXLTokenizer, sdxl_clip.SDXLRefinerClipModel)
+
+class SDXL(supported_models_base.BASE):
+    unet_config = {
+        "model_channels": 320,
+        "use_linear_in_transformer": True,
+        "transformer_depth": [0, 0, 2, 2, 10, 10],
+        "context_dim": 2048,
+        "adm_in_channels": 2816,
+        "use_temporal_attention": False,
+    }
+
+    latent_format = latent_formats.SDXL
+
+    def model_type(self, state_dict, prefix=""):
+        if "v_pred" in state_dict:
+            return model_base.ModelType.V_PREDICTION
+        else:
+            return model_base.ModelType.EPS
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.SDXL(self, model_type=self.model_type(state_dict, prefix), device=device)
+        if self.inpaint_model():
+            out.set_inpaint()
+        return out
+
+    def process_clip_state_dict(self, state_dict):
+        keys_to_replace = {}
+        replace_prefix = {}
+
+        replace_prefix["conditioner.embedders.0.transformer.text_model"] = "cond_stage_model.clip_l.transformer.text_model"
+        state_dict = utils.transformers_convert(state_dict, "conditioner.embedders.1.model.", "cond_stage_model.clip_g.transformer.text_model.", 32)
+        keys_to_replace["conditioner.embedders.1.model.text_projection"] = "cond_stage_model.clip_g.text_projection"
+        keys_to_replace["conditioner.embedders.1.model.text_projection.weight"] = "cond_stage_model.clip_g.text_projection"
+        keys_to_replace["conditioner.embedders.1.model.logit_scale"] = "cond_stage_model.clip_g.logit_scale"
+
+        state_dict = utils.state_dict_prefix_replace(state_dict, replace_prefix)
+        state_dict = utils.state_dict_key_replace(state_dict, keys_to_replace)
+        return state_dict
+
+    def process_clip_state_dict_for_saving(self, state_dict):
+        replace_prefix = {}
+        keys_to_replace = {}
+        state_dict_g = diffusers_convert.convert_text_enc_state_dict_v20(state_dict, "clip_g")
+        if "clip_g.transformer.text_model.embeddings.position_ids" in state_dict_g:
+            state_dict_g.pop("clip_g.transformer.text_model.embeddings.position_ids")
+        for k in state_dict:
+            if k.startswith("clip_l"):
+                state_dict_g[k] = state_dict[k]
+
+        replace_prefix["clip_g"] = "conditioner.embedders.1.model"
+        replace_prefix["clip_l"] = "conditioner.embedders.0"
+        state_dict_g = utils.state_dict_prefix_replace(state_dict_g, replace_prefix)
+        return state_dict_g
+
+    def clip_target(self):
+        return supported_models_base.ClipTarget(sdxl_clip.SDXLTokenizer, sdxl_clip.SDXLClipModel)
+
+class SSD1B(SDXL):
+    unet_config = {
+        "model_channels": 320,
+        "use_linear_in_transformer": True,
+        "transformer_depth": [0, 0, 2, 2, 4, 4],
+        "context_dim": 2048,
+        "adm_in_channels": 2816,
+        "use_temporal_attention": False,
+    }
+
+class SVD_img2vid(supported_models_base.BASE):
+    unet_config = {
+        "model_channels": 320,
+        "in_channels": 8,
+        "use_linear_in_transformer": True,
+        "transformer_depth": [1, 1, 1, 1, 1, 1, 0, 0],
+        "context_dim": 1024,
+        "adm_in_channels": 768,
+        "use_temporal_attention": True,
+        "use_temporal_resblock": True
+    }
+
+    clip_vision_prefix = "conditioner.embedders.0.open_clip.model.visual."
+
+    latent_format = latent_formats.SD15
+
+    sampling_settings = {"sigma_max": 700.0, "sigma_min": 0.002}
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.SVD_img2vid(self, device=device)
+        return out
+
+    def clip_target(self):
+        return None
+
+models = [SD15, SD20, SD21UnclipL, SD21UnclipH, SDXLRefiner, SDXL, SSD1B]
+models += [SVD_img2vid]

ldm_patched/modules/supported_models_base.py

@@ -0,0 +1,77 @@
+import torch
+from . import model_base
+from . import utils
+from . import latent_formats
+
+class ClipTarget:
+    def __init__(self, tokenizer, clip):
+        self.clip = clip
+        self.tokenizer = tokenizer
+        self.params = {}
+
+class BASE:
+    unet_config = {}
+    unet_extra_config = {
+        "num_heads": -1,
+        "num_head_channels": 64,
+    }
+
+    clip_prefix = []
+    clip_vision_prefix = None
+    noise_aug_config = None
+    sampling_settings = {}
+    latent_format = latent_formats.LatentFormat
+
+    manual_cast_dtype = None
+
+    @classmethod
+    def matches(s, unet_config):
+        for k in s.unet_config:
+            if s.unet_config[k] != unet_config[k]:
+                return False
+        return True
+
+    def model_type(self, state_dict, prefix=""):
+        return model_base.ModelType.EPS
+
+    def inpaint_model(self):
+        return self.unet_config["in_channels"] > 4
+
+    def __init__(self, unet_config):
+        self.unet_config = unet_config
+        self.latent_format = self.latent_format()
+        for x in self.unet_extra_config:
+            self.unet_config[x] = self.unet_extra_config[x]
+
+    def get_model(self, state_dict, prefix="", device=None):
+        if self.noise_aug_config is not None:
+            out = model_base.SD21UNCLIP(self, self.noise_aug_config, model_type=self.model_type(state_dict, prefix), device=device)
+        else:
+            out = model_base.BaseModel(self, model_type=self.model_type(state_dict, prefix), device=device)
+        if self.inpaint_model():
+            out.set_inpaint()
+        return out
+
+    def process_clip_state_dict(self, state_dict):
+        return state_dict
+
+    def process_unet_state_dict(self, state_dict):
+        return state_dict
+
+    def process_vae_state_dict(self, state_dict):
+        return state_dict
+
+    def process_clip_state_dict_for_saving(self, state_dict):
+        replace_prefix = {"": "cond_stage_model."}
+        return utils.state_dict_prefix_replace(state_dict, replace_prefix)
+
+    def process_unet_state_dict_for_saving(self, state_dict):
+        replace_prefix = {"": "model.diffusion_model."}
+        return utils.state_dict_prefix_replace(state_dict, replace_prefix)
+
+    def process_vae_state_dict_for_saving(self, state_dict):
+        replace_prefix = {"": "first_stage_model."}
+        return utils.state_dict_prefix_replace(state_dict, replace_prefix)
+
+    def set_manual_cast(self, manual_cast_dtype):
+        self.manual_cast_dtype = manual_cast_dtype

ldm_patched/modules/utils.py

@@ -0,0 +1,461 @@
+import torch
+import math
+import struct
+import ldm_patched.modules.checkpoint_pickle
+import safetensors.torch
+import numpy as np
+from PIL import Image
+
+def load_torch_file(ckpt, safe_load=False, device=None):
+    if device is None:
+        device = torch.device("cpu")
+    if ckpt.lower().endswith(".safetensors"):
+        sd = safetensors.torch.load_file(ckpt, device=device.type)
+    else:
+        if safe_load:
+            if not 'weights_only' in torch.load.__code__.co_varnames:
+                print("Warning torch.load doesn't support weights_only on this pytorch version, loading unsafely.")
+                safe_load = False
+        if safe_load:
+            pl_sd = torch.load(ckpt, map_location=device, weights_only=True)
+        else:
+            pl_sd = torch.load(ckpt, map_location=device, pickle_module=ldm_patched.modules.checkpoint_pickle)
+        if "global_step" in pl_sd:
+            print(f"Global Step: {pl_sd['global_step']}")
+        if "state_dict" in pl_sd:
+            sd = pl_sd["state_dict"]
+        else:
+            sd = pl_sd
+    return sd
+
+def save_torch_file(sd, ckpt, metadata=None):
+    if metadata is not None:
+        safetensors.torch.save_file(sd, ckpt, metadata=metadata)
+    else:
+        safetensors.torch.save_file(sd, ckpt)
+
+def calculate_parameters(sd, prefix=""):
+    params = 0
+    for k in sd.keys():
+        if k.startswith(prefix):
+            params += sd[k].nelement()
+    return params
+
+def state_dict_key_replace(state_dict, keys_to_replace):
+    for x in keys_to_replace:
+        if x in state_dict:
+            state_dict[keys_to_replace[x]] = state_dict.pop(x)
+    return state_dict
+
+def state_dict_prefix_replace(state_dict, replace_prefix, filter_keys=False):
+    if filter_keys:
+        out = {}
+    else:
+        out = state_dict
+    for rp in replace_prefix:
+        replace = list(map(lambda a: (a, "{}{}".format(replace_prefix[rp], a[len(rp):])), filter(lambda a: a.startswith(rp), state_dict.keys())))
+        for x in replace:
+            w = state_dict.pop(x[0])
+            out[x[1]] = w
+    return out
+
+
+def transformers_convert(sd, prefix_from, prefix_to, number):
+    keys_to_replace = {
+        "{}positional_embedding": "{}embeddings.position_embedding.weight",
+        "{}token_embedding.weight": "{}embeddings.token_embedding.weight",
+        "{}ln_final.weight": "{}final_layer_norm.weight",
+        "{}ln_final.bias": "{}final_layer_norm.bias",
+    }
+
+    for k in keys_to_replace:
+        x = k.format(prefix_from)
+        if x in sd:
+            sd[keys_to_replace[k].format(prefix_to)] = sd.pop(x)
+
+    resblock_to_replace = {
+        "ln_1": "layer_norm1",
+        "ln_2": "layer_norm2",
+        "mlp.c_fc": "mlp.fc1",
+        "mlp.c_proj": "mlp.fc2",
+        "attn.out_proj": "self_attn.out_proj",
+    }
+
+    for resblock in range(number):
+        for x in resblock_to_replace:
+            for y in ["weight", "bias"]:
+                k = "{}transformer.resblocks.{}.{}.{}".format(prefix_from, resblock, x, y)
+                k_to = "{}encoder.layers.{}.{}.{}".format(prefix_to, resblock, resblock_to_replace[x], y)
+                if k in sd:
+                    sd[k_to] = sd.pop(k)
+
+        for y in ["weight", "bias"]:
+            k_from = "{}transformer.resblocks.{}.attn.in_proj_{}".format(prefix_from, resblock, y)
+            if k_from in sd:
+                weights = sd.pop(k_from)
+                shape_from = weights.shape[0] // 3
+                for x in range(3):
+                    p = ["self_attn.q_proj", "self_attn.k_proj", "self_attn.v_proj"]
+                    k_to = "{}encoder.layers.{}.{}.{}".format(prefix_to, resblock, p[x], y)
+                    sd[k_to] = weights[shape_from*x:shape_from*(x + 1)]
+    return sd
+
+UNET_MAP_ATTENTIONS = {
+    "proj_in.weight",
+    "proj_in.bias",
+    "proj_out.weight",
+    "proj_out.bias",
+    "norm.weight",
+    "norm.bias",
+}
+
+TRANSFORMER_BLOCKS = {
+    "norm1.weight",
+    "norm1.bias",
+    "norm2.weight",
+    "norm2.bias",
+    "norm3.weight",
+    "norm3.bias",
+    "attn1.to_q.weight",
+    "attn1.to_k.weight",
+    "attn1.to_v.weight",
+    "attn1.to_out.0.weight",
+    "attn1.to_out.0.bias",
+    "attn2.to_q.weight",
+    "attn2.to_k.weight",
+    "attn2.to_v.weight",
+    "attn2.to_out.0.weight",
+    "attn2.to_out.0.bias",
+    "ff.net.0.proj.weight",
+    "ff.net.0.proj.bias",
+    "ff.net.2.weight",
+    "ff.net.2.bias",
+}
+
+UNET_MAP_RESNET = {
+    "in_layers.2.weight": "conv1.weight",
+    "in_layers.2.bias": "conv1.bias",
+    "emb_layers.1.weight": "time_emb_proj.weight",
+    "emb_layers.1.bias": "time_emb_proj.bias",
+    "out_layers.3.weight": "conv2.weight",
+    "out_layers.3.bias": "conv2.bias",
+    "skip_connection.weight": "conv_shortcut.weight",
+    "skip_connection.bias": "conv_shortcut.bias",
+    "in_layers.0.weight": "norm1.weight",
+    "in_layers.0.bias": "norm1.bias",
+    "out_layers.0.weight": "norm2.weight",
+    "out_layers.0.bias": "norm2.bias",
+}
+
+UNET_MAP_BASIC = {
+    ("label_emb.0.0.weight", "class_embedding.linear_1.weight"),
+    ("label_emb.0.0.bias", "class_embedding.linear_1.bias"),
+    ("label_emb.0.2.weight", "class_embedding.linear_2.weight"),
+    ("label_emb.0.2.bias", "class_embedding.linear_2.bias"),
+    ("label_emb.0.0.weight", "add_embedding.linear_1.weight"),
+    ("label_emb.0.0.bias", "add_embedding.linear_1.bias"),
+    ("label_emb.0.2.weight", "add_embedding.linear_2.weight"),
+    ("label_emb.0.2.bias", "add_embedding.linear_2.bias"),
+    ("input_blocks.0.0.weight", "conv_in.weight"),
+    ("input_blocks.0.0.bias", "conv_in.bias"),
+    ("out.0.weight", "conv_norm_out.weight"),
+    ("out.0.bias", "conv_norm_out.bias"),
+    ("out.2.weight", "conv_out.weight"),
+    ("out.2.bias", "conv_out.bias"),
+    ("time_embed.0.weight", "time_embedding.linear_1.weight"),
+    ("time_embed.0.bias", "time_embedding.linear_1.bias"),
+    ("time_embed.2.weight", "time_embedding.linear_2.weight"),
+    ("time_embed.2.bias", "time_embedding.linear_2.bias")
+}
+
+def unet_to_diffusers(unet_config):
+    num_res_blocks = unet_config["num_res_blocks"]
+    channel_mult = unet_config["channel_mult"]
+    transformer_depth = unet_config["transformer_depth"][:]
+    transformer_depth_output = unet_config["transformer_depth_output"][:]
+    num_blocks = len(channel_mult)
+
+    transformers_mid = unet_config.get("transformer_depth_middle", None)
+
+    diffusers_unet_map = {}
+    for x in range(num_blocks):
+        n = 1 + (num_res_blocks[x] + 1) * x
+        for i in range(num_res_blocks[x]):
+            for b in UNET_MAP_RESNET:
+                diffusers_unet_map["down_blocks.{}.resnets.{}.{}".format(x, i, UNET_MAP_RESNET[b])] = "input_blocks.{}.0.{}".format(n, b)
+            num_transformers = transformer_depth.pop(0)
+            if num_transformers > 0:
+                for b in UNET_MAP_ATTENTIONS:
+                    diffusers_unet_map["down_blocks.{}.attentions.{}.{}".format(x, i, b)] = "input_blocks.{}.1.{}".format(n, b)
+                for t in range(num_transformers):
+                    for b in TRANSFORMER_BLOCKS:
+                        diffusers_unet_map["down_blocks.{}.attentions.{}.transformer_blocks.{}.{}".format(x, i, t, b)] = "input_blocks.{}.1.transformer_blocks.{}.{}".format(n, t, b)
+            n += 1
+        for k in ["weight", "bias"]:
+            diffusers_unet_map["down_blocks.{}.downsamplers.0.conv.{}".format(x, k)] = "input_blocks.{}.0.op.{}".format(n, k)
+
+    i = 0
+    for b in UNET_MAP_ATTENTIONS:
+        diffusers_unet_map["mid_block.attentions.{}.{}".format(i, b)] = "middle_block.1.{}".format(b)
+    for t in range(transformers_mid):
+        for b in TRANSFORMER_BLOCKS:
+            diffusers_unet_map["mid_block.attentions.{}.transformer_blocks.{}.{}".format(i, t, b)] = "middle_block.1.transformer_blocks.{}.{}".format(t, b)
+
+    for i, n in enumerate([0, 2]):
+        for b in UNET_MAP_RESNET:
+            diffusers_unet_map["mid_block.resnets.{}.{}".format(i, UNET_MAP_RESNET[b])] = "middle_block.{}.{}".format(n, b)
+
+    num_res_blocks = list(reversed(num_res_blocks))
+    for x in range(num_blocks):
+        n = (num_res_blocks[x] + 1) * x
+        l = num_res_blocks[x] + 1
+        for i in range(l):
+            c = 0
+            for b in UNET_MAP_RESNET:
+                diffusers_unet_map["up_blocks.{}.resnets.{}.{}".format(x, i, UNET_MAP_RESNET[b])] = "output_blocks.{}.0.{}".format(n, b)
+            c += 1
+            num_transformers = transformer_depth_output.pop()
+            if num_transformers > 0:
+                c += 1
+                for b in UNET_MAP_ATTENTIONS:
+                    diffusers_unet_map["up_blocks.{}.attentions.{}.{}".format(x, i, b)] = "output_blocks.{}.1.{}".format(n, b)
+                for t in range(num_transformers):
+                    for b in TRANSFORMER_BLOCKS:
+                        diffusers_unet_map["up_blocks.{}.attentions.{}.transformer_blocks.{}.{}".format(x, i, t, b)] = "output_blocks.{}.1.transformer_blocks.{}.{}".format(n, t, b)
+            if i == l - 1:
+                for k in ["weight", "bias"]:
+                    diffusers_unet_map["up_blocks.{}.upsamplers.0.conv.{}".format(x, k)] = "output_blocks.{}.{}.conv.{}".format(n, c, k)
+            n += 1
+
+    for k in UNET_MAP_BASIC:
+        diffusers_unet_map[k[1]] = k[0]
+
+    return diffusers_unet_map
+
+def repeat_to_batch_size(tensor, batch_size):
+    if tensor.shape[0] > batch_size:
+        return tensor[:batch_size]
+    elif tensor.shape[0] < batch_size:
+        return tensor.repeat([math.ceil(batch_size / tensor.shape[0])] + [1] * (len(tensor.shape) - 1))[:batch_size]
+    return tensor
+
+def resize_to_batch_size(tensor, batch_size):
+    in_batch_size = tensor.shape[0]
+    if in_batch_size == batch_size:
+        return tensor
+
+    if batch_size <= 1:
+        return tensor[:batch_size]
+
+    output = torch.empty([batch_size] + list(tensor.shape)[1:], dtype=tensor.dtype, device=tensor.device)
+    if batch_size < in_batch_size:
+        scale = (in_batch_size - 1) / (batch_size - 1)
+        for i in range(batch_size):
+            output[i] = tensor[min(round(i * scale), in_batch_size - 1)]
+    else:
+        scale = in_batch_size / batch_size
+        for i in range(batch_size):
+            output[i] = tensor[min(math.floor((i + 0.5) * scale), in_batch_size - 1)]
+
+    return output
+
+def convert_sd_to(state_dict, dtype):
+    keys = list(state_dict.keys())
+    for k in keys:
+        state_dict[k] = state_dict[k].to(dtype)
+    return state_dict
+
+def safetensors_header(safetensors_path, max_size=100*1024*1024):
+    with open(safetensors_path, "rb") as f:
+        header = f.read(8)
+        length_of_header = struct.unpack('<Q', header)[0]
+        if length_of_header > max_size:
+            return None
+        return f.read(length_of_header)
+
+def set_attr(obj, attr, value):
+    attrs = attr.split(".")
+    for name in attrs[:-1]:
+        obj = getattr(obj, name)
+    prev = getattr(obj, attrs[-1])
+    setattr(obj, attrs[-1], torch.nn.Parameter(value, requires_grad=False))
+    del prev
+
+def copy_to_param(obj, attr, value):
+    # inplace update tensor instead of replacing it
+    attrs = attr.split(".")
+    for name in attrs[:-1]:
+        obj = getattr(obj, name)
+    prev = getattr(obj, attrs[-1])
+    prev.data.copy_(value)
+
+def get_attr(obj, attr):
+    attrs = attr.split(".")
+    for name in attrs:
+        obj = getattr(obj, name)
+    return obj
+
+def bislerp(samples, width, height):
+    def slerp(b1, b2, r):
+        '''slerps batches b1, b2 according to ratio r, batches should be flat e.g. NxC'''
+        
+        c = b1.shape[-1]
+
+        #norms
+        b1_norms = torch.norm(b1, dim=-1, keepdim=True)
+        b2_norms = torch.norm(b2, dim=-1, keepdim=True)
+
+        #normalize
+        b1_normalized = b1 / b1_norms
+        b2_normalized = b2 / b2_norms
+
+        #zero when norms are zero
+        b1_normalized[b1_norms.expand(-1,c) == 0.0] = 0.0
+        b2_normalized[b2_norms.expand(-1,c) == 0.0] = 0.0
+
+        #slerp
+        dot = (b1_normalized*b2_normalized).sum(1)
+        omega = torch.acos(dot)
+        so = torch.sin(omega)
+
+        #technically not mathematically correct, but more pleasing?
+        res = (torch.sin((1.0-r.squeeze(1))*omega)/so).unsqueeze(1)*b1_normalized + (torch.sin(r.squeeze(1)*omega)/so).unsqueeze(1) * b2_normalized
+        res *= (b1_norms * (1.0-r) + b2_norms * r).expand(-1,c)
+
+        #edge cases for same or polar opposites
+        res[dot > 1 - 1e-5] = b1[dot > 1 - 1e-5] 
+        res[dot < 1e-5 - 1] = (b1 * (1.0-r) + b2 * r)[dot < 1e-5 - 1]
+        return res
+    
+    def generate_bilinear_data(length_old, length_new, device):
+        coords_1 = torch.arange(length_old, dtype=torch.float32, device=device).reshape((1,1,1,-1))
+        coords_1 = torch.nn.functional.interpolate(coords_1, size=(1, length_new), mode="bilinear")
+        ratios = coords_1 - coords_1.floor()
+        coords_1 = coords_1.to(torch.int64)
+        
+        coords_2 = torch.arange(length_old, dtype=torch.float32, device=device).reshape((1,1,1,-1)) + 1
+        coords_2[:,:,:,-1] -= 1
+        coords_2 = torch.nn.functional.interpolate(coords_2, size=(1, length_new), mode="bilinear")
+        coords_2 = coords_2.to(torch.int64)
+        return ratios, coords_1, coords_2
+
+    orig_dtype = samples.dtype
+    samples = samples.float()
+    n,c,h,w = samples.shape
+    h_new, w_new = (height, width)
+    
+    #linear w
+    ratios, coords_1, coords_2 = generate_bilinear_data(w, w_new, samples.device)
+    coords_1 = coords_1.expand((n, c, h, -1))
+    coords_2 = coords_2.expand((n, c, h, -1))
+    ratios = ratios.expand((n, 1, h, -1))
+
+    pass_1 = samples.gather(-1,coords_1).movedim(1, -1).reshape((-1,c))
+    pass_2 = samples.gather(-1,coords_2).movedim(1, -1).reshape((-1,c))
+    ratios = ratios.movedim(1, -1).reshape((-1,1))
+
+    result = slerp(pass_1, pass_2, ratios)
+    result = result.reshape(n, h, w_new, c).movedim(-1, 1)
+
+    #linear h
+    ratios, coords_1, coords_2 = generate_bilinear_data(h, h_new, samples.device)
+    coords_1 = coords_1.reshape((1,1,-1,1)).expand((n, c, -1, w_new))
+    coords_2 = coords_2.reshape((1,1,-1,1)).expand((n, c, -1, w_new))
+    ratios = ratios.reshape((1,1,-1,1)).expand((n, 1, -1, w_new))
+
+    pass_1 = result.gather(-2,coords_1).movedim(1, -1).reshape((-1,c))
+    pass_2 = result.gather(-2,coords_2).movedim(1, -1).reshape((-1,c))
+    ratios = ratios.movedim(1, -1).reshape((-1,1))
+
+    result = slerp(pass_1, pass_2, ratios)
+    result = result.reshape(n, h_new, w_new, c).movedim(-1, 1)
+    return result.to(orig_dtype)
+
+def lanczos(samples, width, height):
+    images = [Image.fromarray(np.clip(255. * image.movedim(0, -1).cpu().numpy(), 0, 255).astype(np.uint8)) for image in samples]
+    images = [image.resize((width, height), resample=Image.Resampling.LANCZOS) for image in images]
+    images = [torch.from_numpy(np.array(image).astype(np.float32) / 255.0).movedim(-1, 0) for image in images]
+    result = torch.stack(images)
+    return result.to(samples.device, samples.dtype)
+
+def common_upscale(samples, width, height, upscale_method, crop):
+        if crop == "center":
+            old_width = samples.shape[3]
+            old_height = samples.shape[2]
+            old_aspect = old_width / old_height
+            new_aspect = width / height
+            x = 0
+            y = 0
+            if old_aspect > new_aspect:
+                x = round((old_width - old_width * (new_aspect / old_aspect)) / 2)
+            elif old_aspect < new_aspect:
+                y = round((old_height - old_height * (old_aspect / new_aspect)) / 2)
+            s = samples[:,:,y:old_height-y,x:old_width-x]
+        else:
+            s = samples
+
+        if upscale_method == "bislerp":
+            return bislerp(s, width, height)
+        elif upscale_method == "lanczos":
+            return lanczos(s, width, height)
+        else:
+            return torch.nn.functional.interpolate(s, size=(height, width), mode=upscale_method)
+
+def get_tiled_scale_steps(width, height, tile_x, tile_y, overlap):
+    return math.ceil((height / (tile_y - overlap))) * math.ceil((width / (tile_x - overlap)))
+
+@torch.inference_mode()
+def tiled_scale(samples, function, tile_x=64, tile_y=64, overlap = 8, upscale_amount = 4, out_channels = 3, output_device="cpu", pbar = None):
+    output = torch.empty((samples.shape[0], out_channels, round(samples.shape[2] * upscale_amount), round(samples.shape[3] * upscale_amount)), device=output_device)
+    for b in range(samples.shape[0]):
+        s = samples[b:b+1]
+        out = torch.zeros((s.shape[0], out_channels, round(s.shape[2] * upscale_amount), round(s.shape[3] * upscale_amount)), device=output_device)
+        out_div = torch.zeros((s.shape[0], out_channels, round(s.shape[2] * upscale_amount), round(s.shape[3] * upscale_amount)), device=output_device)
+        for y in range(0, s.shape[2], tile_y - overlap):
+            for x in range(0, s.shape[3], tile_x - overlap):
+                s_in = s[:,:,y:y+tile_y,x:x+tile_x]
+
+                ps = function(s_in).to(output_device)
+                mask = torch.ones_like(ps)
+                feather = round(overlap * upscale_amount)
+                for t in range(feather):
+                        mask[:,:,t:1+t,:] *= ((1.0/feather) * (t + 1))
+                        mask[:,:,mask.shape[2] -1 -t: mask.shape[2]-t,:] *= ((1.0/feather) * (t + 1))
+                        mask[:,:,:,t:1+t] *= ((1.0/feather) * (t + 1))
+                        mask[:,:,:,mask.shape[3]- 1 - t: mask.shape[3]- t] *= ((1.0/feather) * (t + 1))
+                out[:,:,round(y*upscale_amount):round((y+tile_y)*upscale_amount),round(x*upscale_amount):round((x+tile_x)*upscale_amount)] += ps * mask
+                out_div[:,:,round(y*upscale_amount):round((y+tile_y)*upscale_amount),round(x*upscale_amount):round((x+tile_x)*upscale_amount)] += mask
+                if pbar is not None:
+                    pbar.update(1)
+
+        output[b:b+1] = out/out_div
+    return output
+
+PROGRESS_BAR_ENABLED = True
+def set_progress_bar_enabled(enabled):
+    global PROGRESS_BAR_ENABLED
+    PROGRESS_BAR_ENABLED = enabled
+
+PROGRESS_BAR_HOOK = None
+def set_progress_bar_global_hook(function):
+    global PROGRESS_BAR_HOOK
+    PROGRESS_BAR_HOOK = function
+
+class ProgressBar:
+    def __init__(self, total):
+        global PROGRESS_BAR_HOOK
+        self.total = total
+        self.current = 0
+        self.hook = PROGRESS_BAR_HOOK
+
+    def update_absolute(self, value, total=None, preview=None):
+        if total is not None:
+            self.total = total
+        if value > self.total:
+            value = self.total
+        self.current = value
+        if self.hook is not None:
+            self.hook(self.current, self.total, preview)
+
+    def update(self, value):
+        self.update_absolute(self.current + value)