Announcement

Minor optimization of logs. Added lazy loading of images.

.gitignore

@@ -7,6 +7,7 @@ __pycache__
 *.patch
 *.backup
 *.corrupted
+sorted_styles.json
 /language/default.json
 lena.png
 lena_result.png

args_manager.py

@@ -15,6 +15,10 @@ fcbh_cli.parser.add_argument("--enable-smart-memory", action="store_true",
                              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)
+fcbh_cli.parser.add_argument("--disable-image-log", action='store_true',
+                             help="Prevent writing images and logs to hard drive.")
+
 fcbh_cli.parser.set_defaults(
     disable_cuda_malloc=True,
     auto_launch=True,
@@ -23,9 +27,7 @@ fcbh_cli.parser.set_defaults(
 
 fcbh_cli.args = fcbh_cli.parser.parse_args()
 
-# (beta, enabled by default. )
-# (Probably disable by default because of issues like https://github.com/lllyasviel/Fooocus/issues/724)
-if fcbh_cli.args.enable_smart_memory:
-    fcbh_cli.args.disable_smart_memory = False
+# (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 = fcbh_cli.args

backend/headless/fcbh/cli_args.py

@@ -62,6 +62,13 @@ fpvae_group.add_argument("--fp16-vae", action="store_true", help="Run the VAE in
 fpvae_group.add_argument("--fp32-vae", action="store_true", help="Run the VAE in full precision fp32.")
 fpvae_group.add_argument("--bf16-vae", action="store_true", help="Run the VAE in bf16.")
 
+fpte_group = parser.add_mutually_exclusive_group()
+fpte_group.add_argument("--fp8_e4m3fn-text-enc", action="store_true", help="Store text encoder weights in fp8 (e4m3fn variant).")
+fpte_group.add_argument("--fp8_e5m2-text-enc", action="store_true", help="Store text encoder weights in fp8 (e5m2 variant).")
+fpte_group.add_argument("--fp16-text-enc", action="store_true", help="Store text encoder weights in fp16.")
+fpte_group.add_argument("--fp32-text-enc", action="store_true", help="Store text encoder weights in fp32.")
+
+
 parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE", const=-1, help="Use torch-directml.")
 
 parser.add_argument("--disable-ipex-optimize", action="store_true", help="Disables ipex.optimize when loading models with Intel GPUs.")

backend/headless/fcbh/clip_vision.py

@@ -99,7 +99,7 @@ def load_clipvision_from_sd(sd, prefix="", convert_keys=False):
     clip = ClipVisionModel(json_config)
     m, u = clip.load_sd(sd)
     if len(m) > 0:
-        print("missing clip vision:", m)
+        print("extra keys clip vision:", m)
     u = set(u)
     keys = list(sd.keys())
     for k in keys:

backend/headless/fcbh/controlnet.py

@@ -33,7 +33,7 @@ class ControlBase:
         self.cond_hint_original = None
         self.cond_hint = None
         self.strength = 1.0
-        self.timestep_percent_range = (1.0, 0.0)
+        self.timestep_percent_range = (0.0, 1.0)
         self.timestep_range = None
 
         if device is None:
@@ -42,7 +42,7 @@ class ControlBase:
         self.previous_controlnet = None
         self.global_average_pooling = False
 
-    def set_cond_hint(self, cond_hint, strength=1.0, timestep_percent_range=(1.0, 0.0)):
+    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

backend/headless/fcbh/extra_samplers/uni_pc.py

@@ -858,7 +858,7 @@ def predict_eps_sigma(model, input, sigma_in, **kwargs):
     return  (input - model(input, sigma_in, **kwargs)) / sigma
 
 
-def sample_unipc(model, noise, image, sigmas, sampling_function, max_denoise, extra_args=None, callback=None, disable=False, noise_mask=None, variant='bh1'):
+def sample_unipc(model, noise, image, sigmas, max_denoise, extra_args=None, callback=None, disable=False, noise_mask=None, variant='bh1'):
         timesteps = sigmas.clone()
         if sigmas[-1] == 0:
             timesteps = sigmas[:]

backend/headless/fcbh/k_diffusion/sampling.py

@@ -750,3 +750,61 @@ def sample_lcm(model, x, sigmas, extra_args=None, callback=None, disable=None, n
         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

backend/headless/fcbh/ldm/modules/diffusionmodules/openaimodel.py

@@ -28,25 +28,6 @@ class TimestepBlock(nn.Module):
         Apply the module to `x` given `emb` timestep embeddings.
         """
 
-
-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, x, emb, context=None, transformer_options={}, output_shape=None):
-        for layer in self:
-            if isinstance(layer, TimestepBlock):
-                x = layer(x, emb)
-            elif isinstance(layer, SpatialTransformer):
-                x = layer(x, context, transformer_options)
-            elif isinstance(layer, Upsample):
-                x = layer(x, output_shape=output_shape)
-            else:
-                x = layer(x)
-        return x
-
 #This is needed because accelerate makes a copy of transformer_options which breaks "current_index"
 def forward_timestep_embed(ts, x, emb, context=None, transformer_options={}, output_shape=None):
     for layer in ts:
@@ -54,6 +35,7 @@ def forward_timestep_embed(ts, x, emb, context=None, transformer_options={}, out
             x = layer(x, emb)
         elif isinstance(layer, SpatialTransformer):
             x = layer(x, context, transformer_options)
+            if "current_index" in transformer_options:
                 transformer_options["current_index"] += 1
         elif isinstance(layer, Upsample):
             x = layer(x, output_shape=output_shape)
@@ -61,6 +43,15 @@ def forward_timestep_embed(ts, x, emb, context=None, transformer_options={}, out
             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.
@@ -255,7 +246,10 @@ 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):
@@ -624,7 +618,16 @@ class UNetModel(nn.Module):
             transformer_options["block"] = ("input", id)
             h = forward_timestep_embed(module, h, emb, context, transformer_options)
             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)

backend/headless/fcbh/model_base.py

@@ -121,6 +121,7 @@ class BaseModel(torch.nn.Module):
             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)
@@ -157,6 +158,16 @@ class BaseModel(torch.nn.Module):
     def set_inpaint(self):
         self.inpaint_model = True
 
+    def memory_required(self, input_shape):
+        area = input_shape[0] * input_shape[2] * input_shape[3]
+        if fcbh.model_management.xformers_enabled() or fcbh.model_management.pytorch_attention_flash_attention():
+            #TODO: this needs to be tweaked
+            return (area / (fcbh.model_management.dtype_size(self.get_dtype()) * 10)) * (1024 * 1024)
+        else:
+            #TODO: this formula might be too aggressive since I tweaked the sub-quad and split algorithms to use less memory.
+            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 = []

backend/headless/fcbh/model_detection.py

@@ -186,17 +186,24 @@ def convert_config(unet_config):
 
 def unet_config_from_diffusers_unet(state_dict, dtype):
     match = {}
-    attention_resolutions = []
+    transformer_depth = []
 
     attn_res = 1
-    for i in range(5):
-        k = "down_blocks.{}.attentions.1.transformer_blocks.0.attn2.to_k.weight".format(i)
-        if k in state_dict:
-            match["context_dim"] = state_dict[k].shape[1]
-            attention_resolutions.append(attn_res)
-        attn_res *= 2
+    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]
 
-    match["attention_resolutions"] = attention_resolutions
+        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]
@@ -208,50 +215,55 @@ def unet_config_from_diffusers_unet(state_dict, dtype):
 
     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, 'attention_resolutions': [2, 4], 'transformer_depth': [0, 2, 10], 'channel_mult': [1, 2, 4],
-            'transformer_depth_middle': 10, 'use_linear_in_transformer': True, 'context_dim': 2048, "num_head_channels": 64}
+            '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]}
 
     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, 'attention_resolutions': [2, 4], 'transformer_depth': [0, 4, 4, 0], 'channel_mult': [1, 2, 4, 4],
-                    'transformer_depth_middle': 4, 'use_linear_in_transformer': True, 'context_dim': 1280, "num_head_channels": 64}
+                    '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]}
 
     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,
-            'attention_resolutions': [1, 2, 4], 'transformer_depth': [1, 1, 1, 0], 'channel_mult': [1, 2, 4, 4],
-            'transformer_depth_middle': 1, 'use_linear_in_transformer': True, 'context_dim': 1024, "num_head_channels": 64}
+            '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]}
 
     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, 'attention_resolutions': [1, 2, 4], 'transformer_depth': [1, 1, 1, 0], 'channel_mult': [1, 2, 4, 4],
-                    'transformer_depth_middle': 1, 'use_linear_in_transformer': True, 'context_dim': 1024, "num_head_channels": 64}
+                    '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]}
 
     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, 'attention_resolutions': [1, 2, 4], 'transformer_depth': [1, 1, 1, 0], 'channel_mult': [1, 2, 4, 4],
-                    'transformer_depth_middle': 1, 'use_linear_in_transformer': True, 'context_dim': 1024}
+                    '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]}
 
-    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,
-            'attention_resolutions': [1, 2, 4], 'transformer_depth': [1, 1, 1, 0], 'channel_mult': [1, 2, 4, 4],
-            'transformer_depth_middle': 1, 'use_linear_in_transformer': False, 'context_dim': 768, "num_heads": 8}
+    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]}
 
     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, 'attention_resolutions': [4], 'transformer_depth': [0, 0, 1], 'channel_mult': [1, 2, 4],
-            'transformer_depth_middle': 1, 'use_linear_in_transformer': True, 'context_dim': 2048, "num_head_channels": 64}
+                     '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]}
 
     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, 'attention_resolutions': [], 'transformer_depth': [0, 0, 0], 'channel_mult': [1, 2, 4],
-            'transformer_depth_middle': 0, 'use_linear_in_transformer': True, "num_head_channels": 64, 'context_dim': 1}
+                       '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]}
 
     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, 'attention_resolutions': [2, 4], 'transformer_depth': [0, 2, 10], 'channel_mult': [1, 2, 4],
-            'transformer_depth_middle': 10, 'use_linear_in_transformer': True, 'context_dim': 2048, "num_head_channels": 64}
+                              '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]}
 
-    supported_models = [SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet, SDXL_diffusers_inpaint]
+    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}
+
+    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

backend/headless/fcbh/model_management.py

@@ -133,6 +133,10 @@ else:
         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)
@@ -478,6 +482,21 @@ def text_encoder_device():
     else:
         return torch.device("cpu")
 
+def text_encoder_dtype(device=None):
+    if args.fp8_e4m3fn_text_enc:
+        return torch.float8_e4m3fn
+    elif args.fp8_e5m2_text_enc:
+        return torch.float8_e5m2
+    elif args.fp16_text_enc:
+        return torch.float16
+    elif args.fp32_text_enc:
+        return torch.float32
+
+    if should_use_fp16(device, prioritize_performance=False):
+        return torch.float16
+    else:
+        return torch.float32
+
 def vae_device():
     return get_torch_device()
 
@@ -579,27 +598,6 @@ def get_free_memory(dev=None, torch_free_too=False):
     else:
         return mem_free_total
 
-def batch_area_memory(area):
-    if xformers_enabled() or pytorch_attention_flash_attention():
-        #TODO: these formulas are copied from maximum_batch_area below
-        return (area / 20) * (1024 * 1024)
-    else:
-        return (((area * 0.6) / 0.9) + 1024) * (1024 * 1024)
-
-def maximum_batch_area():
-    global vram_state
-    if vram_state == VRAMState.NO_VRAM:
-        return 0
-
-    memory_free = get_free_memory() / (1024 * 1024)
-    if xformers_enabled() or pytorch_attention_flash_attention():
-        #TODO: this needs to be tweaked
-        area = 20 * memory_free
-    else:
-        #TODO: this formula is because AMD sucks and has memory management issues which might be fixed in the future
-        area = ((memory_free - 1024) * 0.9) / (0.6)
-    return int(max(area, 0))
-
 def cpu_mode():
     global cpu_state
     return cpu_state == CPUState.CPU

backend/headless/fcbh/model_patcher.py

@@ -6,7 +6,7 @@ import fcbh.utils
 import fcbh.model_management
 
 class ModelPatcher:
-    def __init__(self, model, load_device, offload_device, size=0, current_device=None):
+    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 = {}
@@ -22,6 +22,8 @@ class ModelPatcher:
         else:
             self.current_device = current_device
 
+        self.weight_inplace_update = weight_inplace_update
+
     def model_size(self):
         if self.size > 0:
             return self.size
@@ -35,7 +37,7 @@ class ModelPatcher:
         return size
 
     def clone(self):
-        n = ModelPatcher(self.model, self.load_device, self.offload_device, self.size, self.current_device)
+        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][:]
@@ -50,6 +52,9 @@ class ModelPatcher:
             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
@@ -91,6 +96,12 @@ class ModelPatcher:
     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")
 
@@ -171,14 +182,19 @@ class ModelPatcher:
 
             weight = model_sd[key]
 
+            inplace_update = self.weight_inplace_update
+
             if key not in self.backup:
-                self.backup[key] = weight.to(self.offload_device)
+                self.backup[key] = weight.to(device=self.offload_device, copy=inplace_update)
 
             if device_to is not None:
                 temp_weight = fcbh.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:
+                fcbh.utils.copy_to_param(self.model, key, out_weight)
+            else:
                 fcbh.utils.set_attr(self.model, key, out_weight)
             del temp_weight
 
@@ -295,6 +311,10 @@ class ModelPatcher:
     def unpatch_model(self, device_to=None):
         keys = list(self.backup.keys())
 
+        if self.weight_inplace_update:
+            for k in keys:
+                fcbh.utils.copy_to_param(self.model, k, self.backup[k])
+        else:
             for k in keys:
                 fcbh.utils.set_attr(self.model, k, self.backup[k])
 

backend/headless/fcbh/model_sampling.py

@@ -24,7 +24,7 @@ class ModelSamplingDiscrete(torch.nn.Module):
         super().__init__()
         beta_schedule = "linear"
         if model_config is not None:
-            beta_schedule = model_config.beta_schedule
+            beta_schedule = model_config.sampling_settings.get("beta_schedule", beta_schedule)
         self._register_schedule(given_betas=None, beta_schedule=beta_schedule, timesteps=1000, linear_start=0.00085, linear_end=0.012, cosine_s=8e-3)
         self.sigma_data = 1.0
 
@@ -76,5 +76,10 @@ class ModelSamplingDiscrete(torch.nn.Module):
         return log_sigma.exp()
 
     def percent_to_sigma(self, percent):
-        return self.sigma(torch.tensor(percent * 999.0))
+        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()
 

backend/headless/fcbh/sample.py

@@ -83,7 +83,7 @@ def prepare_sampling(model, noise_shape, positive, negative, noise_mask):
 
     real_model = None
     models, inference_memory = get_additional_models(positive, negative, model.model_dtype())
-    fcbh.model_management.load_models_gpu([model] + models, fcbh.model_management.batch_area_memory(noise_shape[0] * noise_shape[2] * noise_shape[3]) + inference_memory)
+    fcbh.model_management.load_models_gpu([model] + models, model.memory_required(noise_shape) + inference_memory)
     real_model = model.model
 
     return real_model, positive, negative, noise_mask, models

backend/headless/fcbh/samplers.py

@@ -11,7 +11,7 @@ import fcbh.conds
 
 #The main sampling function shared by all the samplers
 #Returns denoised
-def sampling_function(model_function, x, timestep, uncond, cond, cond_scale, model_options={}, seed=None):
+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
@@ -134,7 +134,7 @@ def sampling_function(model_function, x, timestep, uncond, cond, cond_scale, mod
 
             return out
 
-        def calc_cond_uncond_batch(model_function, cond, uncond, x_in, timestep, max_total_area, model_options):
+        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
 
@@ -170,9 +170,11 @@ def sampling_function(model_function, x, timestep, uncond, cond, cond_scale, mod
                 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]
-                    if (len(batch_amount) * first_shape[0] * first_shape[2] * first_shape[3] < max_total_area):
+                    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
 
@@ -218,12 +220,14 @@ def sampling_function(model_function, x, timestep, uncond, cond, cond_scale, mod
                         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_function, {"input": input_x, "timestep": timestep_, "c": c, "cond_or_uncond": cond_or_uncond}).chunk(batch_chunks)
+                    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_function(input_x, timestep_, **c).chunk(batch_chunks)
+                    output = model.apply_model(input_x, timestep_, **c).chunk(batch_chunks)
                 del input_x
 
                 for o in range(batch_chunks):
@@ -242,11 +246,10 @@ def sampling_function(model_function, x, timestep, uncond, cond, cond_scale, mod
             return out_cond, out_uncond
 
 
-        max_total_area = model_management.maximum_batch_area()
         if math.isclose(cond_scale, 1.0):
             uncond = None
 
-        cond, uncond = calc_cond_uncond_batch(model_function, cond, uncond, x, timestep, max_total_area, model_options)
+        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)
@@ -258,7 +261,7 @@ class CFGNoisePredictor(torch.nn.Module):
         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.apply_model, x, timestep, uncond, cond, cond_scale, model_options=model_options, seed=seed)
+        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)
@@ -511,23 +514,27 @@ class Sampler:
 
 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, sampling_function=sampling_function, max_denoise=self.max_denoise(model_wrap, sigmas), extra_args=extra_args, noise_mask=denoise_mask, callback=callback, disable=disable_pbar)
+        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, sampling_function=sampling_function, max_denoise=self.max_denoise(model_wrap, sigmas), extra_args=extra_args, noise_mask=denoise_mask, callback=callback, variant='bh2', disable=disable_pbar)
+        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", "dpm_2", "dpm_2_ancestral",
+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"]
 
-def ksampler(sampler_name, extra_options={}, inpaint_options={}):
-    class KSAMPLER(Sampler):
+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 inpaint_options.get("random", False): #TODO: Should this be the default?
+        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:
@@ -543,20 +550,33 @@ def ksampler(sampler_name, extra_options={}, inpaint_options={}):
         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]
-
-            if latent_image is not None:
-                noise += latent_image
-            if sampler_name == "dpm_fast":
-                samples = k_diffusion_sampling.sample_dpm_fast(model_k, noise, sigma_min, sigmas[0], total_steps, extra_args=extra_args, callback=k_callback, disable=disable_pbar)
+            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":
-                samples = k_diffusion_sampling.sample_dpm_adaptive(model_k, noise, sigma_min, sigmas[0], extra_args=extra_args, callback=k_callback, disable=disable_pbar)
+        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:
-                samples = getattr(k_diffusion_sampling, "sample_{}".format(sampler_name))(model_k, noise, sigmas, extra_args=extra_args, callback=k_callback, disable=disable_pbar, **extra_options)
-            return samples
-    return KSAMPLER
+        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)
@@ -617,11 +637,11 @@ def calculate_sigmas_scheduler(model, scheduler_name, steps):
         print("error invalid scheduler", self.scheduler)
     return sigmas
 
-def sampler_class(name):
+def sampler_object(name):
     if name == "uni_pc":
-        sampler = UNIPC
+        sampler = UNIPC()
     elif name == "uni_pc_bh2":
-        sampler = UNIPCBH2
+        sampler = UNIPCBH2()
     elif name == "ddim":
         sampler = ksampler("euler", inpaint_options={"random": True})
     else:
@@ -686,6 +706,6 @@ class KSampler:
                 else:
                     return torch.zeros_like(noise)
 
-        sampler = sampler_class(self.sampler)
+        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)
+        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)

backend/headless/fcbh/sd.py

@@ -23,6 +23,7 @@ import fcbh.model_patcher
 import fcbh.lora
 import fcbh.t2i_adapter.adapter
 import fcbh.supported_models_base
+import fcbh.taesd.taesd
 
 def load_model_weights(model, sd):
     m, u = model.load_state_dict(sd, strict=False)
@@ -35,7 +36,7 @@ def load_model_weights(model, sd):
             w = sd.pop(x)
             del w
     if len(m) > 0:
-        print("missing", m)
+        print("extra keys", m)
     return model
 
 def load_clip_weights(model, sd):
@@ -95,10 +96,7 @@ class CLIP:
         load_device = model_management.text_encoder_device()
         offload_device = model_management.text_encoder_offload_device()
         params['device'] = offload_device
-        if model_management.should_use_fp16(load_device, prioritize_performance=False):
-            params['dtype'] = torch.float16
-        else:
-            params['dtype'] = torch.float32
+        params['dtype'] = model_management.text_encoder_dtype(load_device)
 
         self.cond_stage_model = clip(**(params))
 
@@ -157,7 +155,13 @@ class VAE:
         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 "taesd_decoder.1.weight" in sd:
+                self.first_stage_model = fcbh.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)
@@ -209,7 +213,7 @@ class VAE:
     def decode(self, samples_in):
         self.first_stage_model = self.first_stage_model.to(self.device)
         try:
-            memory_used = (2562 * samples_in.shape[2] * samples_in.shape[3] * 64) * 1.7
+            memory_used = self.memory_used_decode(samples_in.shape, self.vae_dtype)
             model_management.free_memory(memory_used, self.device)
             free_memory = model_management.get_free_memory(self.device)
             batch_number = int(free_memory / memory_used)
@@ -237,7 +241,7 @@ class VAE:
         self.first_stage_model = self.first_stage_model.to(self.device)
         pixel_samples = pixel_samples.movedim(-1,1)
         try:
-            memory_used = (2078 * pixel_samples.shape[2] * pixel_samples.shape[3]) * 1.7 #NOTE: this constant along with the one in the decode above are estimated from the mem usage for the VAE and could change.
+            memory_used = self.memory_used_encode(pixel_samples.shape, self.vae_dtype)
             model_management.free_memory(memory_used, self.device)
             free_memory = model_management.get_free_memory(self.device)
             batch_number = int(free_memory / memory_used)
@@ -444,6 +448,7 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o
 
     if output_vae:
         vae_sd = fcbh.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:

backend/headless/fcbh/sd1_clip.py

@@ -173,9 +173,9 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         if getattr(self.transformer, self.inner_name).final_layer_norm.weight.dtype != torch.float32:
             precision_scope = torch.autocast
         else:
-            precision_scope = lambda a, b: contextlib.nullcontext(a)
+            precision_scope = lambda a, dtype: contextlib.nullcontext(a)
 
-        with precision_scope(model_management.get_autocast_device(device), torch.float32):
+        with precision_scope(model_management.get_autocast_device(device), dtype=torch.float32):
             attention_mask = None
             if self.enable_attention_masks:
                 attention_mask = torch.zeros_like(tokens)

backend/headless/fcbh/supported_models_base.py

@@ -19,7 +19,7 @@ class BASE:
     clip_prefix = []
     clip_vision_prefix = None
     noise_aug_config = None
-    beta_schedule = "linear"
+    sampling_settings = {}
     latent_format = latent_formats.LatentFormat
 
     @classmethod
@@ -53,6 +53,12 @@ class BASE:
     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)

backend/headless/fcbh/taesd/taesd.py

@@ -46,15 +46,16 @@ class TAESD(nn.Module):
     latent_magnitude = 3
     latent_shift = 0.5
 
-    def __init__(self, encoder_path="taesd_encoder.pth", decoder_path="taesd_decoder.pth"):
+    def __init__(self, encoder_path=None, decoder_path=None):
         """Initialize pretrained TAESD on the given device from the given checkpoints."""
         super().__init__()
-        self.encoder = Encoder()
-        self.decoder = Decoder()
+        self.taesd_encoder = Encoder()
+        self.taesd_decoder = Decoder()
+        self.vae_scale = torch.nn.Parameter(torch.tensor(1.0))
         if encoder_path is not None:
-            self.encoder.load_state_dict(fcbh.utils.load_torch_file(encoder_path, safe_load=True))
+            self.taesd_encoder.load_state_dict(fcbh.utils.load_torch_file(encoder_path, safe_load=True))
         if decoder_path is not None:
-            self.decoder.load_state_dict(fcbh.utils.load_torch_file(decoder_path, safe_load=True))
+            self.taesd_decoder.load_state_dict(fcbh.utils.load_torch_file(decoder_path, safe_load=True))
 
     @staticmethod
     def scale_latents(x):
@@ -65,3 +66,11 @@ class TAESD(nn.Module):
     def unscale_latents(x):
         """[0, 1] -> raw latents"""
         return x.sub(TAESD.latent_shift).mul(2 * TAESD.latent_magnitude)
+
+    def decode(self, x):
+        x_sample = self.taesd_decoder(x * self.vae_scale)
+        x_sample = x_sample.sub(0.5).mul(2)
+        return x_sample
+
+    def encode(self, x):
+        return self.taesd_encoder(x * 0.5 + 0.5) / self.vae_scale

backend/headless/fcbh/utils.py

@@ -258,9 +258,17 @@ def set_attr(obj, attr, value):
     for name in attrs[:-1]:
         obj = getattr(obj, name)
     prev = getattr(obj, attrs[-1])
-    setattr(obj, attrs[-1], torch.nn.Parameter(value))
+    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:
@@ -299,23 +307,25 @@ def bislerp(samples, width, height):
         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):
-        coords_1 = torch.arange(length_old).reshape((1,1,1,-1)).to(torch.float32)
+    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).reshape((1,1,1,-1)).to(torch.float32) + 1
+        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)
+    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))
@@ -328,7 +338,7 @@ def bislerp(samples, width, height):
     result = result.reshape(n, h, w_new, c).movedim(-1, 1)
 
     #linear h
-    ratios, coords_1, coords_2 = generate_bilinear_data(h, h_new)
+    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))
@@ -339,7 +349,7 @@ def bislerp(samples, width, height):
 
     result = slerp(pass_1, pass_2, ratios)
     result = result.reshape(n, h_new, w_new, c).movedim(-1, 1)
-    return result
+    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]

backend/headless/fcbh_extras/nodes_custom_sampler.py

@@ -16,7 +16,7 @@ class BasicScheduler:
                       }
                }
     RETURN_TYPES = ("SIGMAS",)
-    CATEGORY = "sampling/custom_sampling"
+    CATEGORY = "sampling/custom_sampling/schedulers"
 
     FUNCTION = "get_sigmas"
 
@@ -36,7 +36,7 @@ class KarrasScheduler:
                     }
                }
     RETURN_TYPES = ("SIGMAS",)
-    CATEGORY = "sampling/custom_sampling"
+    CATEGORY = "sampling/custom_sampling/schedulers"
 
     FUNCTION = "get_sigmas"
 
@@ -54,7 +54,7 @@ class ExponentialScheduler:
                     }
                }
     RETURN_TYPES = ("SIGMAS",)
-    CATEGORY = "sampling/custom_sampling"
+    CATEGORY = "sampling/custom_sampling/schedulers"
 
     FUNCTION = "get_sigmas"
 
@@ -73,7 +73,7 @@ class PolyexponentialScheduler:
                     }
                }
     RETURN_TYPES = ("SIGMAS",)
-    CATEGORY = "sampling/custom_sampling"
+    CATEGORY = "sampling/custom_sampling/schedulers"
 
     FUNCTION = "get_sigmas"
 
@@ -92,7 +92,7 @@ class VPScheduler:
                     }
                }
     RETURN_TYPES = ("SIGMAS",)
-    CATEGORY = "sampling/custom_sampling"
+    CATEGORY = "sampling/custom_sampling/schedulers"
 
     FUNCTION = "get_sigmas"
 
@@ -109,7 +109,7 @@ class SplitSigmas:
                      }
                 }
     RETURN_TYPES = ("SIGMAS","SIGMAS")
-    CATEGORY = "sampling/custom_sampling"
+    CATEGORY = "sampling/custom_sampling/sigmas"
 
     FUNCTION = "get_sigmas"
 
@@ -118,6 +118,24 @@ class SplitSigmas:
         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):
@@ -126,12 +144,12 @@ class KSamplerSelect:
                       }
                }
     RETURN_TYPES = ("SAMPLER",)
-    CATEGORY = "sampling/custom_sampling"
+    CATEGORY = "sampling/custom_sampling/samplers"
 
     FUNCTION = "get_sampler"
 
     def get_sampler(self, sampler_name):
-        sampler = fcbh.samplers.sampler_class(sampler_name)()
+        sampler = fcbh.samplers.sampler_object(sampler_name)
         return (sampler, )
 
 class SamplerDPMPP_2M_SDE:
@@ -145,7 +163,7 @@ class SamplerDPMPP_2M_SDE:
                       }
                }
     RETURN_TYPES = ("SAMPLER",)
-    CATEGORY = "sampling/custom_sampling"
+    CATEGORY = "sampling/custom_sampling/samplers"
 
     FUNCTION = "get_sampler"
 
@@ -154,7 +172,7 @@ class SamplerDPMPP_2M_SDE:
             sampler_name = "dpmpp_2m_sde"
         else:
             sampler_name = "dpmpp_2m_sde_gpu"
-        sampler = fcbh.samplers.ksampler(sampler_name, {"eta": eta, "s_noise": s_noise, "solver_type": solver_type})()
+        sampler = fcbh.samplers.ksampler(sampler_name, {"eta": eta, "s_noise": s_noise, "solver_type": solver_type})
         return (sampler, )
 
 
@@ -169,7 +187,7 @@ class SamplerDPMPP_SDE:
                       }
                }
     RETURN_TYPES = ("SAMPLER",)
-    CATEGORY = "sampling/custom_sampling"
+    CATEGORY = "sampling/custom_sampling/samplers"
 
     FUNCTION = "get_sampler"
 
@@ -178,7 +196,7 @@ class SamplerDPMPP_SDE:
             sampler_name = "dpmpp_sde"
         else:
             sampler_name = "dpmpp_sde_gpu"
-        sampler = fcbh.samplers.ksampler(sampler_name, {"eta": eta, "s_noise": s_noise, "r": r})()
+        sampler = fcbh.samplers.ksampler(sampler_name, {"eta": eta, "s_noise": s_noise, "r": r})
         return (sampler, )
 
 class SamplerCustom:
@@ -234,6 +252,7 @@ class SamplerCustom:
 
 NODE_CLASS_MAPPINGS = {
     "SamplerCustom": SamplerCustom,
+    "BasicScheduler": BasicScheduler,
     "KarrasScheduler": KarrasScheduler,
     "ExponentialScheduler": ExponentialScheduler,
     "PolyexponentialScheduler": PolyexponentialScheduler,
@@ -241,6 +260,6 @@ NODE_CLASS_MAPPINGS = {
     "KSamplerSelect": KSamplerSelect,
     "SamplerDPMPP_2M_SDE": SamplerDPMPP_2M_SDE,
     "SamplerDPMPP_SDE": SamplerDPMPP_SDE,
-    "BasicScheduler": BasicScheduler,
     "SplitSigmas": SplitSigmas,
+    "FlipSigmas": FlipSigmas,
 }

backend/headless/fcbh_extras/nodes_images.py

@@ -0,0 +1,120 @@
+import nodes
+import folder_paths
+from fcbh.cli_args import args
+
+from PIL import Image
+import numpy as np
+import json
+import os
+
+MAX_RESOLUTION = nodes.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 = folder_paths.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": "fcbh_backend"}),
+                     "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, "aoeu")
+        filename_prefix += self.prefix_append
+        full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.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_metadata:
+            metadata = pil_images[0].getexif()
+            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,) } }
+
+NODE_CLASS_MAPPINGS = {
+    "ImageCrop": ImageCrop,
+    "RepeatImageBatch": RepeatImageBatch,
+    "SaveAnimatedWEBP": SaveAnimatedWEBP,
+}

backend/headless/fcbh_extras/nodes_latent.py

@@ -1,4 +1,5 @@
 import fcbh.utils
+import torch
 
 def reshape_latent_to(target_shape, latent):
     if latent.shape[1:] != target_shape[1:]:
@@ -67,8 +68,43 @@ class LatentMultiply:
         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,
 }

backend/headless/fcbh_extras/nodes_model_advanced.py

@@ -66,7 +66,12 @@ class ModelSamplingDiscreteLCM(torch.nn.Module):
         return log_sigma.exp()
 
     def percent_to_sigma(self, percent):
-        return self.sigma(torch.tensor(percent * 999.0))
+        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):

backend/headless/fcbh_extras/nodes_model_downscale.py

@@ -0,0 +1,53 @@
+import torch
+import fcbh.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 = fcbh.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 = fcbh.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)",
+}

backend/headless/latent_preview.py

@@ -22,10 +22,7 @@ class TAESDPreviewerImpl(LatentPreviewer):
         self.taesd = taesd
 
     def decode_latent_to_preview(self, x0):
-        x_sample = self.taesd.decoder(x0[:1])[0].detach()
-        # x_sample = self.taesd.unscale_latents(x_sample).div(4).add(0.5)  # returns value in [-2, 2]
-        x_sample = x_sample.sub(0.5).mul(2)
-
+        x_sample = self.taesd.decode(x0[:1])[0].detach()
         x_sample = torch.clamp((x_sample + 1.0) / 2.0, min=0.0, max=1.0)
         x_sample = 255. * np.moveaxis(x_sample.cpu().numpy(), 0, 2)
         x_sample = x_sample.astype(np.uint8)

backend/headless/nodes.py

@@ -248,8 +248,8 @@ class ConditioningSetTimestepRange:
         c = []
         for t in conditioning:
             d = t[1].copy()
-            d['start_percent'] = 1.0 - start
-            d['end_percent'] = 1.0 - end
+            d['start_percent'] = start
+            d['end_percent'] = end
             n = [t[0], d]
             c.append(n)
         return (c, )
@@ -573,9 +573,55 @@ class LoraLoader:
         return (model_lora, clip_lora)
 
 class VAELoader:
+    @staticmethod
+    def vae_list():
+        vaes = folder_paths.get_filename_list("vae")
+        approx_vaes = folder_paths.get_filename_list("vae_approx")
+        sdxl_taesd_enc = False
+        sdxl_taesd_dec = False
+        sd1_taesd_enc = False
+        sd1_taesd_dec = False
+
+        for v in approx_vaes:
+            if v.startswith("taesd_decoder."):
+                sd1_taesd_dec = True
+            elif v.startswith("taesd_encoder."):
+                sd1_taesd_enc = True
+            elif v.startswith("taesdxl_decoder."):
+                sdxl_taesd_dec = True
+            elif v.startswith("taesdxl_encoder."):
+                sdxl_taesd_enc = True
+        if sd1_taesd_dec and sd1_taesd_enc:
+            vaes.append("taesd")
+        if sdxl_taesd_dec and sdxl_taesd_enc:
+            vaes.append("taesdxl")
+        return vaes
+
+    @staticmethod
+    def load_taesd(name):
+        sd = {}
+        approx_vaes = folder_paths.get_filename_list("vae_approx")
+
+        encoder = next(filter(lambda a: a.startswith("{}_encoder.".format(name)), approx_vaes))
+        decoder = next(filter(lambda a: a.startswith("{}_decoder.".format(name)), approx_vaes))
+
+        enc = fcbh.utils.load_torch_file(folder_paths.get_full_path("vae_approx", encoder))
+        for k in enc:
+            sd["taesd_encoder.{}".format(k)] = enc[k]
+
+        dec = fcbh.utils.load_torch_file(folder_paths.get_full_path("vae_approx", decoder))
+        for k in dec:
+            sd["taesd_decoder.{}".format(k)] = dec[k]
+
+        if name == "taesd":
+            sd["vae_scale"] = torch.tensor(0.18215)
+        elif name == "taesdxl":
+            sd["vae_scale"] = torch.tensor(0.13025)
+        return sd
+
     @classmethod
     def INPUT_TYPES(s):
-        return {"required": { "vae_name": (folder_paths.get_filename_list("vae"), )}}
+        return {"required": { "vae_name": (s.vae_list(), )}}
     RETURN_TYPES = ("VAE",)
     FUNCTION = "load_vae"
 
@@ -583,6 +629,9 @@ class VAELoader:
 
     #TODO: scale factor?
     def load_vae(self, vae_name):
+        if vae_name in ["taesd", "taesdxl"]:
+            sd = self.load_taesd(vae_name)
+        else:
             vae_path = folder_paths.get_full_path("vae", vae_name)
             sd = fcbh.utils.load_torch_file(vae_path)
         vae = fcbh.sd.VAE(sd=sd)
@@ -685,7 +734,7 @@ class ControlNetApplyAdvanced:
                 if prev_cnet in cnets:
                     c_net = cnets[prev_cnet]
                 else:
-                    c_net = control_net.copy().set_cond_hint(control_hint, strength, (1.0 - start_percent, 1.0 - end_percent))
+                    c_net = control_net.copy().set_cond_hint(control_hint, strength, (start_percent, end_percent))
                     c_net.set_previous_controlnet(prev_cnet)
                     cnets[prev_cnet] = c_net
 
@@ -1799,6 +1848,8 @@ def init_custom_nodes():
         "nodes_custom_sampler.py",
         "nodes_hypertile.py",
         "nodes_model_advanced.py",
+        "nodes_model_downscale.py",
+        "nodes_images.py",
     ]
 
     for node_file in extras_files:

css/style.css

@@ -29,8 +29,8 @@
 
 .canvas-tooltip-info {
   position: absolute;
-  top: 10px;
-  left: 10px;
+  top: 28px;
+  left: 2px;
   cursor: help;
   background-color: rgba(0, 0, 0, 0.3);
   width: 20px;
@@ -93,3 +93,102 @@
 .styler {
   overflow:inherit !important;
 }
+
+/* fullpage image viewer */
+
+#lightboxModal{
+    display: none;
+    position: fixed;
+    z-index: 1001;
+    left: 0;
+    top: 0;
+    width: 100%;
+    height: 100%;
+    overflow: auto;
+    background-color: rgba(20, 20, 20, 0.95);
+    user-select: none;
+    -webkit-user-select: none;
+    flex-direction: column;
+}
+
+.modalControls {
+    display: flex;
+    position: absolute;
+    right: 0px;
+    left: 0px;
+    gap: 1em;
+    padding: 1em;
+    background-color:rgba(0,0,0,0);
+    z-index: 1;
+    transition: 0.2s ease background-color;
+}
+.modalControls:hover {
+    background-color:rgba(0,0,0,0.9);
+}
+.modalClose {
+    margin-left: auto;
+}
+.modalControls span{
+    color: white;
+    text-shadow: 0px 0px 0.25em black;
+    font-size: 35px;
+    font-weight: bold;
+    cursor: pointer;
+    width: 1em;
+}
+
+.modalControls span:hover, .modalControls span:focus{
+    color: #999;
+    text-decoration: none;
+}
+
+#lightboxModal > img {
+    display: block;
+    margin: auto;
+    width: auto;
+}
+
+#lightboxModal > img.modalImageFullscreen{
+    object-fit: contain;
+    height: 100%;
+    width: 100%;
+    min-height: 0;
+}
+
+.modalPrev,
+.modalNext {
+  cursor: pointer;
+  position: absolute;
+  top: 50%;
+  width: auto;
+  padding: 16px;
+  margin-top: -50px;
+  color: white;
+  font-weight: bold;
+  font-size: 20px;
+  transition: 0.6s ease;
+  border-radius: 0 3px 3px 0;
+  user-select: none;
+  -webkit-user-select: none;
+}
+
+.modalNext {
+  right: 0;
+  border-radius: 3px 0 0 3px;
+}
+
+.modalPrev:hover,
+.modalNext:hover {
+  background-color: rgba(0, 0, 0, 0.8);
+}
+
+#imageARPreview {
+    position: absolute;
+    top: 0px;
+    left: 0px;
+    border: 2px solid red;
+    background: rgba(255, 0, 0, 0.3);
+    z-index: 900;
+    pointer-events: none;
+    display: none;
+}

face_experiments.py

@@ -0,0 +1,7 @@
+import cv2
+import fooocus_extras.face_crop as cropper
+
+
+img = cv2.imread('lena.png')
+result = cropper.crop_image(img)
+cv2.imwrite('lena_result.png', result)

fooocus_extras/face_crop.py

@@ -0,0 +1,50 @@
+import cv2
+import numpy as np
+import modules.config
+
+
+faceRestoreHelper = None
+
+
+def align_warp_face(self, landmark, border_mode='constant'):
+    affine_matrix = cv2.estimateAffinePartial2D(landmark, self.face_template, method=cv2.LMEDS)[0]
+    self.affine_matrices.append(affine_matrix)
+    if border_mode == 'constant':
+        border_mode = cv2.BORDER_CONSTANT
+    elif border_mode == 'reflect101':
+        border_mode = cv2.BORDER_REFLECT101
+    elif border_mode == 'reflect':
+        border_mode = cv2.BORDER_REFLECT
+    input_img = self.input_img
+    cropped_face = cv2.warpAffine(input_img, affine_matrix, self.face_size,
+                                  borderMode=border_mode, borderValue=(135, 133, 132))
+    return cropped_face
+
+
+def crop_image(img_rgb):
+    global faceRestoreHelper
+    
+    if faceRestoreHelper is None:
+        from fooocus_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
+        )
+
+    faceRestoreHelper.clean_all()
+    faceRestoreHelper.read_image(np.ascontiguousarray(img_rgb[:, :, ::-1].copy()))
+    faceRestoreHelper.get_face_landmarks_5()
+
+    landmarks = faceRestoreHelper.all_landmarks_5
+    # landmarks are already sorted with confidence.
+
+    if len(landmarks) == 0:
+        print('No face detected')
+        return img_rgb
+    else:
+        print(f'Detected {len(landmarks)} faces')
+
+    result = align_warp_face(faceRestoreHelper, landmarks[0])
+
+    return np.ascontiguousarray(result[:, :, ::-1].copy())

fooocus_extras/facexlib/detection/__init__.py

@@ -0,0 +1,31 @@
+import torch
+from copy import deepcopy
+
+from fooocus_extras.facexlib.utils import load_file_from_url
+from .retinaface import RetinaFace
+
+
+def init_detection_model(model_name, half=False, device='cuda', model_rootpath=None):
+    if model_name == 'retinaface_resnet50':
+        model = RetinaFace(network_name='resnet50', half=half, device=device)
+        model_url = 'https://github.com/xinntao/facexlib/releases/download/v0.1.0/detection_Resnet50_Final.pth'
+    elif model_name == 'retinaface_mobile0.25':
+        model = RetinaFace(network_name='mobile0.25', half=half, device=device)
+        model_url = 'https://github.com/xinntao/facexlib/releases/download/v0.1.0/detection_mobilenet0.25_Final.pth'
+    else:
+        raise NotImplementedError(f'{model_name} is not implemented.')
+
+    model_path = load_file_from_url(
+        url=model_url, model_dir='facexlib/weights', progress=True, file_name=None, save_dir=model_rootpath)
+
+    # TODO: clean pretrained model
+    load_net = torch.load(model_path, map_location=lambda storage, loc: storage)
+    # remove unnecessary 'module.'
+    for k, v in deepcopy(load_net).items():
+        if k.startswith('module.'):
+            load_net[k[7:]] = v
+            load_net.pop(k)
+    model.load_state_dict(load_net, strict=True)
+    model.eval()
+    model = model.to(device)
+    return model

fooocus_extras/facexlib/detection/align_trans.py

@@ -0,0 +1,219 @@
+import cv2
+import numpy as np
+
+from .matlab_cp2tform import get_similarity_transform_for_cv2
+
+# reference facial points, a list of coordinates (x,y)
+REFERENCE_FACIAL_POINTS = [[30.29459953, 51.69630051], [65.53179932, 51.50139999], [48.02519989, 71.73660278],
+                           [33.54930115, 92.3655014], [62.72990036, 92.20410156]]
+
+DEFAULT_CROP_SIZE = (96, 112)
+
+
+class FaceWarpException(Exception):
+
+    def __str__(self):
+        return 'In File {}:{}'.format(__file__, super.__str__(self))
+
+
+def get_reference_facial_points(output_size=None, inner_padding_factor=0.0, outer_padding=(0, 0), default_square=False):
+    """
+    Function:
+    ----------
+        get reference 5 key points according to crop settings:
+        0. Set default crop_size:
+            if default_square:
+                crop_size = (112, 112)
+            else:
+                crop_size = (96, 112)
+        1. Pad the crop_size by inner_padding_factor in each side;
+        2. Resize crop_size into (output_size - outer_padding*2),
+            pad into output_size with outer_padding;
+        3. Output reference_5point;
+    Parameters:
+    ----------
+        @output_size: (w, h) or None
+            size of aligned face image
+        @inner_padding_factor: (w_factor, h_factor)
+            padding factor for inner (w, h)
+        @outer_padding: (w_pad, h_pad)
+            each row is a pair of coordinates (x, y)
+        @default_square: True or False
+            if True:
+                default crop_size = (112, 112)
+            else:
+                default crop_size = (96, 112);
+        !!! make sure, if output_size is not None:
+                (output_size - outer_padding)
+                = some_scale * (default crop_size * (1.0 +
+                inner_padding_factor))
+    Returns:
+    ----------
+        @reference_5point: 5x2 np.array
+            each row is a pair of transformed coordinates (x, y)
+    """
+
+    tmp_5pts = np.array(REFERENCE_FACIAL_POINTS)
+    tmp_crop_size = np.array(DEFAULT_CROP_SIZE)
+
+    # 0) make the inner region a square
+    if default_square:
+        size_diff = max(tmp_crop_size) - tmp_crop_size
+        tmp_5pts += size_diff / 2
+        tmp_crop_size += size_diff
+
+    if (output_size and output_size[0] == tmp_crop_size[0] and output_size[1] == tmp_crop_size[1]):
+
+        return tmp_5pts
+
+    if (inner_padding_factor == 0 and outer_padding == (0, 0)):
+        if output_size is None:
+            return tmp_5pts
+        else:
+            raise FaceWarpException('No paddings to do, output_size must be None or {}'.format(tmp_crop_size))
+
+    # check output size
+    if not (0 <= inner_padding_factor <= 1.0):
+        raise FaceWarpException('Not (0 <= inner_padding_factor <= 1.0)')
+
+    if ((inner_padding_factor > 0 or outer_padding[0] > 0 or outer_padding[1] > 0) and output_size is None):
+        output_size = tmp_crop_size * \
+            (1 + inner_padding_factor * 2).astype(np.int32)
+        output_size += np.array(outer_padding)
+    if not (outer_padding[0] < output_size[0] and outer_padding[1] < output_size[1]):
+        raise FaceWarpException('Not (outer_padding[0] < output_size[0] and outer_padding[1] < output_size[1])')
+
+    # 1) pad the inner region according inner_padding_factor
+    if inner_padding_factor > 0:
+        size_diff = tmp_crop_size * inner_padding_factor * 2
+        tmp_5pts += size_diff / 2
+        tmp_crop_size += np.round(size_diff).astype(np.int32)
+
+    # 2) resize the padded inner region
+    size_bf_outer_pad = np.array(output_size) - np.array(outer_padding) * 2
+
+    if size_bf_outer_pad[0] * tmp_crop_size[1] != size_bf_outer_pad[1] * tmp_crop_size[0]:
+        raise FaceWarpException('Must have (output_size - outer_padding)'
+                                '= some_scale * (crop_size * (1.0 + inner_padding_factor)')
+
+    scale_factor = size_bf_outer_pad[0].astype(np.float32) / tmp_crop_size[0]
+    tmp_5pts = tmp_5pts * scale_factor
+    #    size_diff = tmp_crop_size * (scale_factor - min(scale_factor))
+    #    tmp_5pts = tmp_5pts + size_diff / 2
+    tmp_crop_size = size_bf_outer_pad
+
+    # 3) add outer_padding to make output_size
+    reference_5point = tmp_5pts + np.array(outer_padding)
+    tmp_crop_size = output_size
+
+    return reference_5point
+
+
+def get_affine_transform_matrix(src_pts, dst_pts):
+    """
+    Function:
+    ----------
+        get affine transform matrix 'tfm' from src_pts to dst_pts
+    Parameters:
+    ----------
+        @src_pts: Kx2 np.array
+            source points matrix, each row is a pair of coordinates (x, y)
+        @dst_pts: Kx2 np.array
+            destination points matrix, each row is a pair of coordinates (x, y)
+    Returns:
+    ----------
+        @tfm: 2x3 np.array
+            transform matrix from src_pts to dst_pts
+    """
+
+    tfm = np.float32([[1, 0, 0], [0, 1, 0]])
+    n_pts = src_pts.shape[0]
+    ones = np.ones((n_pts, 1), src_pts.dtype)
+    src_pts_ = np.hstack([src_pts, ones])
+    dst_pts_ = np.hstack([dst_pts, ones])
+
+    A, res, rank, s = np.linalg.lstsq(src_pts_, dst_pts_)
+
+    if rank == 3:
+        tfm = np.float32([[A[0, 0], A[1, 0], A[2, 0]], [A[0, 1], A[1, 1], A[2, 1]]])
+    elif rank == 2:
+        tfm = np.float32([[A[0, 0], A[1, 0], 0], [A[0, 1], A[1, 1], 0]])
+
+    return tfm
+
+
+def warp_and_crop_face(src_img, facial_pts, reference_pts=None, crop_size=(96, 112), align_type='smilarity'):
+    """
+    Function:
+    ----------
+        apply affine transform 'trans' to uv
+    Parameters:
+    ----------
+        @src_img: 3x3 np.array
+            input image
+        @facial_pts: could be
+            1)a list of K coordinates (x,y)
+        or
+            2) Kx2 or 2xK np.array
+            each row or col is a pair of coordinates (x, y)
+        @reference_pts: could be
+            1) a list of K coordinates (x,y)
+        or
+            2) Kx2 or 2xK np.array
+            each row or col is a pair of coordinates (x, y)
+        or
+            3) None
+            if None, use default reference facial points
+        @crop_size: (w, h)
+            output face image size
+        @align_type: transform type, could be one of
+            1) 'similarity': use similarity transform
+            2) 'cv2_affine': use the first 3 points to do affine transform,
+                    by calling cv2.getAffineTransform()
+            3) 'affine': use all points to do affine transform
+    Returns:
+    ----------
+        @face_img: output face image with size (w, h) = @crop_size
+    """
+
+    if reference_pts is None:
+        if crop_size[0] == 96 and crop_size[1] == 112:
+            reference_pts = REFERENCE_FACIAL_POINTS
+        else:
+            default_square = False
+            inner_padding_factor = 0
+            outer_padding = (0, 0)
+            output_size = crop_size
+
+            reference_pts = get_reference_facial_points(output_size, inner_padding_factor, outer_padding,
+                                                        default_square)
+
+    ref_pts = np.float32(reference_pts)
+    ref_pts_shp = ref_pts.shape
+    if max(ref_pts_shp) < 3 or min(ref_pts_shp) != 2:
+        raise FaceWarpException('reference_pts.shape must be (K,2) or (2,K) and K>2')
+
+    if ref_pts_shp[0] == 2:
+        ref_pts = ref_pts.T
+
+    src_pts = np.float32(facial_pts)
+    src_pts_shp = src_pts.shape
+    if max(src_pts_shp) < 3 or min(src_pts_shp) != 2:
+        raise FaceWarpException('facial_pts.shape must be (K,2) or (2,K) and K>2')
+
+    if src_pts_shp[0] == 2:
+        src_pts = src_pts.T
+
+    if src_pts.shape != ref_pts.shape:
+        raise FaceWarpException('facial_pts and reference_pts must have the same shape')
+
+    if align_type == 'cv2_affine':
+        tfm = cv2.getAffineTransform(src_pts[0:3], ref_pts[0:3])
+    elif align_type == 'affine':
+        tfm = get_affine_transform_matrix(src_pts, ref_pts)
+    else:
+        tfm = get_similarity_transform_for_cv2(src_pts, ref_pts)
+
+    face_img = cv2.warpAffine(src_img, tfm, (crop_size[0], crop_size[1]))
+
+    return face_img

fooocus_extras/facexlib/detection/matlab_cp2tform.py

@@ -0,0 +1,317 @@
+import numpy as np
+from numpy.linalg import inv, lstsq
+from numpy.linalg import matrix_rank as rank
+from numpy.linalg import norm
+
+
+class MatlabCp2tormException(Exception):
+
+    def __str__(self):
+        return 'In File {}:{}'.format(__file__, super.__str__(self))
+
+
+def tformfwd(trans, uv):
+    """
+    Function:
+    ----------
+        apply affine transform 'trans' to uv
+
+    Parameters:
+    ----------
+        @trans: 3x3 np.array
+            transform matrix
+        @uv: Kx2 np.array
+            each row is a pair of coordinates (x, y)
+
+    Returns:
+    ----------
+        @xy: Kx2 np.array
+            each row is a pair of transformed coordinates (x, y)
+    """
+    uv = np.hstack((uv, np.ones((uv.shape[0], 1))))
+    xy = np.dot(uv, trans)
+    xy = xy[:, 0:-1]
+    return xy
+
+
+def tforminv(trans, uv):
+    """
+    Function:
+    ----------
+        apply the inverse of affine transform 'trans' to uv
+
+    Parameters:
+    ----------
+        @trans: 3x3 np.array
+            transform matrix
+        @uv: Kx2 np.array
+            each row is a pair of coordinates (x, y)
+
+    Returns:
+    ----------
+        @xy: Kx2 np.array
+            each row is a pair of inverse-transformed coordinates (x, y)
+    """
+    Tinv = inv(trans)
+    xy = tformfwd(Tinv, uv)
+    return xy
+
+
+def findNonreflectiveSimilarity(uv, xy, options=None):
+    options = {'K': 2}
+
+    K = options['K']
+    M = xy.shape[0]
+    x = xy[:, 0].reshape((-1, 1))  # use reshape to keep a column vector
+    y = xy[:, 1].reshape((-1, 1))  # use reshape to keep a column vector
+
+    tmp1 = np.hstack((x, y, np.ones((M, 1)), np.zeros((M, 1))))
+    tmp2 = np.hstack((y, -x, np.zeros((M, 1)), np.ones((M, 1))))
+    X = np.vstack((tmp1, tmp2))
+
+    u = uv[:, 0].reshape((-1, 1))  # use reshape to keep a column vector
+    v = uv[:, 1].reshape((-1, 1))  # use reshape to keep a column vector
+    U = np.vstack((u, v))
+
+    # We know that X * r = U
+    if rank(X) >= 2 * K:
+        r, _, _, _ = lstsq(X, U, rcond=-1)
+        r = np.squeeze(r)
+    else:
+        raise Exception('cp2tform:twoUniquePointsReq')
+    sc = r[0]
+    ss = r[1]
+    tx = r[2]
+    ty = r[3]
+
+    Tinv = np.array([[sc, -ss, 0], [ss, sc, 0], [tx, ty, 1]])
+    T = inv(Tinv)
+    T[:, 2] = np.array([0, 0, 1])
+
+    return T, Tinv
+
+
+def findSimilarity(uv, xy, options=None):
+    options = {'K': 2}
+
+    #    uv = np.array(uv)
+    #    xy = np.array(xy)
+
+    # Solve for trans1
+    trans1, trans1_inv = findNonreflectiveSimilarity(uv, xy, options)
+
+    # Solve for trans2
+
+    # manually reflect the xy data across the Y-axis
+    xyR = xy
+    xyR[:, 0] = -1 * xyR[:, 0]
+
+    trans2r, trans2r_inv = findNonreflectiveSimilarity(uv, xyR, options)
+
+    # manually reflect the tform to undo the reflection done on xyR
+    TreflectY = np.array([[-1, 0, 0], [0, 1, 0], [0, 0, 1]])
+
+    trans2 = np.dot(trans2r, TreflectY)
+
+    # Figure out if trans1 or trans2 is better
+    xy1 = tformfwd(trans1, uv)
+    norm1 = norm(xy1 - xy)
+
+    xy2 = tformfwd(trans2, uv)
+    norm2 = norm(xy2 - xy)
+
+    if norm1 <= norm2:
+        return trans1, trans1_inv
+    else:
+        trans2_inv = inv(trans2)
+        return trans2, trans2_inv
+
+
+def get_similarity_transform(src_pts, dst_pts, reflective=True):
+    """
+    Function:
+    ----------
+        Find Similarity Transform Matrix 'trans':
+            u = src_pts[:, 0]
+            v = src_pts[:, 1]
+            x = dst_pts[:, 0]
+            y = dst_pts[:, 1]
+            [x, y, 1] = [u, v, 1] * trans
+
+    Parameters:
+    ----------
+        @src_pts: Kx2 np.array
+            source points, each row is a pair of coordinates (x, y)
+        @dst_pts: Kx2 np.array
+            destination points, each row is a pair of transformed
+            coordinates (x, y)
+        @reflective: True or False
+            if True:
+                use reflective similarity transform
+            else:
+                use non-reflective similarity transform
+
+    Returns:
+    ----------
+       @trans: 3x3 np.array
+            transform matrix from uv to xy
+        trans_inv: 3x3 np.array
+            inverse of trans, transform matrix from xy to uv
+    """
+
+    if reflective:
+        trans, trans_inv = findSimilarity(src_pts, dst_pts)
+    else:
+        trans, trans_inv = findNonreflectiveSimilarity(src_pts, dst_pts)
+
+    return trans, trans_inv
+
+
+def cvt_tform_mat_for_cv2(trans):
+    """
+    Function:
+    ----------
+        Convert Transform Matrix 'trans' into 'cv2_trans' which could be
+        directly used by cv2.warpAffine():
+            u = src_pts[:, 0]
+            v = src_pts[:, 1]
+            x = dst_pts[:, 0]
+            y = dst_pts[:, 1]
+            [x, y].T = cv_trans * [u, v, 1].T
+
+    Parameters:
+    ----------
+        @trans: 3x3 np.array
+            transform matrix from uv to xy
+
+    Returns:
+    ----------
+        @cv2_trans: 2x3 np.array
+            transform matrix from src_pts to dst_pts, could be directly used
+            for cv2.warpAffine()
+    """
+    cv2_trans = trans[:, 0:2].T
+
+    return cv2_trans
+
+
+def get_similarity_transform_for_cv2(src_pts, dst_pts, reflective=True):
+    """
+    Function:
+    ----------
+        Find Similarity Transform Matrix 'cv2_trans' which could be
+        directly used by cv2.warpAffine():
+            u = src_pts[:, 0]
+            v = src_pts[:, 1]
+            x = dst_pts[:, 0]
+            y = dst_pts[:, 1]
+            [x, y].T = cv_trans * [u, v, 1].T
+
+    Parameters:
+    ----------
+        @src_pts: Kx2 np.array
+            source points, each row is a pair of coordinates (x, y)
+        @dst_pts: Kx2 np.array
+            destination points, each row is a pair of transformed
+            coordinates (x, y)
+        reflective: True or False
+            if True:
+                use reflective similarity transform
+            else:
+                use non-reflective similarity transform
+
+    Returns:
+    ----------
+        @cv2_trans: 2x3 np.array
+            transform matrix from src_pts to dst_pts, could be directly used
+            for cv2.warpAffine()
+    """
+    trans, trans_inv = get_similarity_transform(src_pts, dst_pts, reflective)
+    cv2_trans = cvt_tform_mat_for_cv2(trans)
+
+    return cv2_trans
+
+
+if __name__ == '__main__':
+    """
+    u = [0, 6, -2]
+    v = [0, 3, 5]
+    x = [-1, 0, 4]
+    y = [-1, -10, 4]
+
+    # In Matlab, run:
+    #
+    #   uv = [u'; v'];
+    #   xy = [x'; y'];
+    #   tform_sim=cp2tform(uv,xy,'similarity');
+    #
+    #   trans = tform_sim.tdata.T
+    #   ans =
+    #       -0.0764   -1.6190         0
+    #        1.6190   -0.0764         0
+    #       -3.2156    0.0290    1.0000
+    #   trans_inv = tform_sim.tdata.Tinv
+    #    ans =
+    #
+    #       -0.0291    0.6163         0
+    #       -0.6163   -0.0291         0
+    #       -0.0756    1.9826    1.0000
+    #    xy_m=tformfwd(tform_sim, u,v)
+    #
+    #    xy_m =
+    #
+    #       -3.2156    0.0290
+    #        1.1833   -9.9143
+    #        5.0323    2.8853
+    #    uv_m=tforminv(tform_sim, x,y)
+    #
+    #    uv_m =
+    #
+    #        0.5698    1.3953
+    #        6.0872    2.2733
+    #       -2.6570    4.3314
+    """
+    u = [0, 6, -2]
+    v = [0, 3, 5]
+    x = [-1, 0, 4]
+    y = [-1, -10, 4]
+
+    uv = np.array((u, v)).T
+    xy = np.array((x, y)).T
+
+    print('\n--->uv:')
+    print(uv)
+    print('\n--->xy:')
+    print(xy)
+
+    trans, trans_inv = get_similarity_transform(uv, xy)
+
+    print('\n--->trans matrix:')
+    print(trans)
+
+    print('\n--->trans_inv matrix:')
+    print(trans_inv)
+
+    print('\n---> apply transform to uv')
+    print('\nxy_m = uv_augmented * trans')
+    uv_aug = np.hstack((uv, np.ones((uv.shape[0], 1))))
+    xy_m = np.dot(uv_aug, trans)
+    print(xy_m)
+
+    print('\nxy_m = tformfwd(trans, uv)')
+    xy_m = tformfwd(trans, uv)
+    print(xy_m)
+
+    print('\n---> apply inverse transform to xy')
+    print('\nuv_m = xy_augmented * trans_inv')
+    xy_aug = np.hstack((xy, np.ones((xy.shape[0], 1))))
+    uv_m = np.dot(xy_aug, trans_inv)
+    print(uv_m)
+
+    print('\nuv_m = tformfwd(trans_inv, xy)')
+    uv_m = tformfwd(trans_inv, xy)
+    print(uv_m)
+
+    uv_m = tforminv(trans, xy)
+    print('\nuv_m = tforminv(trans, xy)')
+    print(uv_m)

fooocus_extras/facexlib/detection/retinaface.py

@@ -0,0 +1,366 @@
+import cv2
+import numpy as np
+import torch
+import torch.nn as nn
+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 fooocus_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,
+                                                 py_cpu_nms)
+
+
+def generate_config(network_name):
+
+    cfg_mnet = {
+        'name': 'mobilenet0.25',
+        'min_sizes': [[16, 32], [64, 128], [256, 512]],
+        'steps': [8, 16, 32],
+        'variance': [0.1, 0.2],
+        'clip': False,
+        'loc_weight': 2.0,
+        'gpu_train': True,
+        'batch_size': 32,
+        'ngpu': 1,
+        'epoch': 250,
+        'decay1': 190,
+        'decay2': 220,
+        'image_size': 640,
+        'return_layers': {
+            'stage1': 1,
+            'stage2': 2,
+            'stage3': 3
+        },
+        'in_channel': 32,
+        'out_channel': 64
+    }
+
+    cfg_re50 = {
+        'name': 'Resnet50',
+        'min_sizes': [[16, 32], [64, 128], [256, 512]],
+        'steps': [8, 16, 32],
+        'variance': [0.1, 0.2],
+        'clip': False,
+        'loc_weight': 2.0,
+        'gpu_train': True,
+        'batch_size': 24,
+        'ngpu': 4,
+        'epoch': 100,
+        'decay1': 70,
+        'decay2': 90,
+        'image_size': 840,
+        'return_layers': {
+            'layer2': 1,
+            'layer3': 2,
+            'layer4': 3
+        },
+        'in_channel': 256,
+        'out_channel': 256
+    }
+
+    if network_name == 'mobile0.25':
+        return cfg_mnet
+    elif network_name == 'resnet50':
+        return cfg_re50
+    else:
+        raise NotImplementedError(f'network_name={network_name}')
+
+
+class RetinaFace(nn.Module):
+
+    def __init__(self, network_name='resnet50', half=False, phase='test', device=None):
+        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') if device is None else device
+
+        super(RetinaFace, self).__init__()
+        self.half_inference = half
+        cfg = generate_config(network_name)
+        self.backbone = cfg['name']
+
+        self.model_name = f'retinaface_{network_name}'
+        self.cfg = cfg
+        self.phase = phase
+        self.target_size, self.max_size = 1600, 2150
+        self.resize, self.scale, self.scale1 = 1., None, None
+        self.mean_tensor = torch.tensor([[[[104.]], [[117.]], [[123.]]]], device=self.device)
+        self.reference = get_reference_facial_points(default_square=True)
+        # Build network.
+        backbone = None
+        if cfg['name'] == 'mobilenet0.25':
+            backbone = MobileNetV1()
+            self.body = IntermediateLayerGetter(backbone, cfg['return_layers'])
+        elif cfg['name'] == 'Resnet50':
+            import torchvision.models as models
+            backbone = models.resnet50(weights=None)
+            self.body = IntermediateLayerGetter(backbone, cfg['return_layers'])
+
+        in_channels_stage2 = cfg['in_channel']
+        in_channels_list = [
+            in_channels_stage2 * 2,
+            in_channels_stage2 * 4,
+            in_channels_stage2 * 8,
+        ]
+
+        out_channels = cfg['out_channel']
+        self.fpn = FPN(in_channels_list, out_channels)
+        self.ssh1 = SSH(out_channels, out_channels)
+        self.ssh2 = SSH(out_channels, out_channels)
+        self.ssh3 = SSH(out_channels, out_channels)
+
+        self.ClassHead = make_class_head(fpn_num=3, inchannels=cfg['out_channel'])
+        self.BboxHead = make_bbox_head(fpn_num=3, inchannels=cfg['out_channel'])
+        self.LandmarkHead = make_landmark_head(fpn_num=3, inchannels=cfg['out_channel'])
+
+        self.to(self.device)
+        self.eval()
+        if self.half_inference:
+            self.half()
+
+    def forward(self, inputs):
+        out = self.body(inputs)
+
+        if self.backbone == 'mobilenet0.25' or self.backbone == 'Resnet50':
+            out = list(out.values())
+        # FPN
+        fpn = self.fpn(out)
+
+        # SSH
+        feature1 = self.ssh1(fpn[0])
+        feature2 = self.ssh2(fpn[1])
+        feature3 = self.ssh3(fpn[2])
+        features = [feature1, feature2, feature3]
+
+        bbox_regressions = torch.cat([self.BboxHead[i](feature) for i, feature in enumerate(features)], dim=1)
+        classifications = torch.cat([self.ClassHead[i](feature) for i, feature in enumerate(features)], dim=1)
+        tmp = [self.LandmarkHead[i](feature) for i, feature in enumerate(features)]
+        ldm_regressions = (torch.cat(tmp, dim=1))
+
+        if self.phase == 'train':
+            output = (bbox_regressions, classifications, ldm_regressions)
+        else:
+            output = (bbox_regressions, F.softmax(classifications, dim=-1), ldm_regressions)
+        return output
+
+    def __detect_faces(self, inputs):
+        # get scale
+        height, width = inputs.shape[2:]
+        self.scale = torch.tensor([width, height, width, height], dtype=torch.float32, device=self.device)
+        tmp = [width, height, width, height, width, height, width, height, width, height]
+        self.scale1 = torch.tensor(tmp, dtype=torch.float32, device=self.device)
+
+        # forawrd
+        inputs = inputs.to(self.device)
+        if self.half_inference:
+            inputs = inputs.half()
+        loc, conf, landmarks = self(inputs)
+
+        # get priorbox
+        priorbox = PriorBox(self.cfg, image_size=inputs.shape[2:])
+        priors = priorbox.forward().to(self.device)
+
+        return loc, conf, landmarks, priors
+
+    # single image detection
+    def transform(self, image, use_origin_size):
+        # convert to opencv format
+        if isinstance(image, Image.Image):
+            image = cv2.cvtColor(np.asarray(image), cv2.COLOR_RGB2BGR)
+        image = image.astype(np.float32)
+
+        # testing scale
+        im_size_min = np.min(image.shape[0:2])
+        im_size_max = np.max(image.shape[0:2])
+        resize = float(self.target_size) / float(im_size_min)
+
+        # prevent bigger axis from being more than max_size
+        if np.round(resize * im_size_max) > self.max_size:
+            resize = float(self.max_size) / float(im_size_max)
+        resize = 1 if use_origin_size else resize
+
+        # resize
+        if resize != 1:
+            image = cv2.resize(image, None, None, fx=resize, fy=resize, interpolation=cv2.INTER_LINEAR)
+
+        # convert to torch.tensor format
+        # image -= (104, 117, 123)
+        image = image.transpose(2, 0, 1)
+        image = torch.from_numpy(image).unsqueeze(0)
+
+        return image, resize
+
+    def detect_faces(
+        self,
+        image,
+        conf_threshold=0.8,
+        nms_threshold=0.4,
+        use_origin_size=True,
+    ):
+        image, self.resize = self.transform(image, use_origin_size)
+        image = image.to(self.device)
+        if self.half_inference:
+            image = image.half()
+        image = image - self.mean_tensor
+
+        loc, conf, landmarks, priors = self.__detect_faces(image)
+
+        boxes = decode(loc.data.squeeze(0), priors.data, self.cfg['variance'])
+        boxes = boxes * self.scale / self.resize
+        boxes = boxes.cpu().numpy()
+
+        scores = conf.squeeze(0).data.cpu().numpy()[:, 1]
+
+        landmarks = decode_landm(landmarks.squeeze(0), priors, self.cfg['variance'])
+        landmarks = landmarks * self.scale1 / self.resize
+        landmarks = landmarks.cpu().numpy()
+
+        # ignore low scores
+        inds = np.where(scores > conf_threshold)[0]
+        boxes, landmarks, scores = boxes[inds], landmarks[inds], scores[inds]
+
+        # sort
+        order = scores.argsort()[::-1]
+        boxes, landmarks, scores = boxes[order], landmarks[order], scores[order]
+
+        # do NMS
+        bounding_boxes = np.hstack((boxes, scores[:, np.newaxis])).astype(np.float32, copy=False)
+        keep = py_cpu_nms(bounding_boxes, nms_threshold)
+        bounding_boxes, landmarks = bounding_boxes[keep, :], landmarks[keep]
+        # self.t['forward_pass'].toc()
+        # print(self.t['forward_pass'].average_time)
+        # import sys
+        # sys.stdout.flush()
+        return np.concatenate((bounding_boxes, landmarks), axis=1)
+
+    def __align_multi(self, image, boxes, landmarks, limit=None):
+
+        if len(boxes) < 1:
+            return [], []
+
+        if limit:
+            boxes = boxes[:limit]
+            landmarks = landmarks[:limit]
+
+        faces = []
+        for landmark in landmarks:
+            facial5points = [[landmark[2 * j], landmark[2 * j + 1]] for j in range(5)]
+
+            warped_face = warp_and_crop_face(np.array(image), facial5points, self.reference, crop_size=(112, 112))
+            faces.append(warped_face)
+
+        return np.concatenate((boxes, landmarks), axis=1), faces
+
+    def align_multi(self, img, conf_threshold=0.8, limit=None):
+
+        rlt = self.detect_faces(img, conf_threshold=conf_threshold)
+        boxes, landmarks = rlt[:, 0:5], rlt[:, 5:]
+
+        return self.__align_multi(img, boxes, landmarks, limit)
+
+    # batched detection
+    def batched_transform(self, frames, use_origin_size):
+        """
+        Arguments:
+            frames: a list of PIL.Image, or torch.Tensor(shape=[n, h, w, c],
+                type=np.float32, BGR format).
+            use_origin_size: whether to use origin size.
+        """
+        from_PIL = True if isinstance(frames[0], Image.Image) else False
+
+        # convert to opencv format
+        if from_PIL:
+            frames = [cv2.cvtColor(np.asarray(frame), cv2.COLOR_RGB2BGR) for frame in frames]
+            frames = np.asarray(frames, dtype=np.float32)
+
+        # testing scale
+        im_size_min = np.min(frames[0].shape[0:2])
+        im_size_max = np.max(frames[0].shape[0:2])
+        resize = float(self.target_size) / float(im_size_min)
+
+        # prevent bigger axis from being more than max_size
+        if np.round(resize * im_size_max) > self.max_size:
+            resize = float(self.max_size) / float(im_size_max)
+        resize = 1 if use_origin_size else resize
+
+        # resize
+        if resize != 1:
+            if not from_PIL:
+                frames = F.interpolate(frames, scale_factor=resize)
+            else:
+                frames = [
+                    cv2.resize(frame, None, None, fx=resize, fy=resize, interpolation=cv2.INTER_LINEAR)
+                    for frame in frames
+                ]
+
+        # convert to torch.tensor format
+        if not from_PIL:
+            frames = frames.transpose(1, 2).transpose(1, 3).contiguous()
+        else:
+            frames = frames.transpose((0, 3, 1, 2))
+            frames = torch.from_numpy(frames)
+
+        return frames, resize
+
+    def batched_detect_faces(self, frames, conf_threshold=0.8, nms_threshold=0.4, use_origin_size=True):
+        """
+        Arguments:
+            frames: a list of PIL.Image, or np.array(shape=[n, h, w, c],
+                type=np.uint8, BGR format).
+            conf_threshold: confidence threshold.
+            nms_threshold: nms threshold.
+            use_origin_size: whether to use origin size.
+        Returns:
+            final_bounding_boxes: list of np.array ([n_boxes, 5],
+                type=np.float32).
+            final_landmarks: list of np.array ([n_boxes, 10], type=np.float32).
+        """
+        # self.t['forward_pass'].tic()
+        frames, self.resize = self.batched_transform(frames, use_origin_size)
+        frames = frames.to(self.device)
+        frames = frames - self.mean_tensor
+
+        b_loc, b_conf, b_landmarks, priors = self.__detect_faces(frames)
+
+        final_bounding_boxes, final_landmarks = [], []
+
+        # decode
+        priors = priors.unsqueeze(0)
+        b_loc = batched_decode(b_loc, priors, self.cfg['variance']) * self.scale / self.resize
+        b_landmarks = batched_decode_landm(b_landmarks, priors, self.cfg['variance']) * self.scale1 / self.resize
+        b_conf = b_conf[:, :, 1]
+
+        # index for selection
+        b_indice = b_conf > conf_threshold
+
+        # concat
+        b_loc_and_conf = torch.cat((b_loc, b_conf.unsqueeze(-1)), dim=2).float()
+
+        for pred, landm, inds in zip(b_loc_and_conf, b_landmarks, b_indice):
+
+            # ignore low scores
+            pred, landm = pred[inds, :], landm[inds, :]
+            if pred.shape[0] == 0:
+                final_bounding_boxes.append(np.array([], dtype=np.float32))
+                final_landmarks.append(np.array([], dtype=np.float32))
+                continue
+
+            # sort
+            # order = score.argsort(descending=True)
+            # box, landm, score = box[order], landm[order], score[order]
+
+            # to CPU
+            bounding_boxes, landm = pred.cpu().numpy(), landm.cpu().numpy()
+
+            # NMS
+            keep = py_cpu_nms(bounding_boxes, nms_threshold)
+            bounding_boxes, landmarks = bounding_boxes[keep, :], landm[keep]
+
+            # append
+            final_bounding_boxes.append(bounding_boxes)
+            final_landmarks.append(landmarks)
+        # self.t['forward_pass'].toc(average=True)
+        # self.batch_time += self.t['forward_pass'].diff
+        # self.total_frame += len(frames)
+        # print(self.batch_time / self.total_frame)
+
+        return final_bounding_boxes, final_landmarks

fooocus_extras/facexlib/detection/retinaface_net.py

@@ -0,0 +1,196 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+def conv_bn(inp, oup, stride=1, leaky=0):
+    return nn.Sequential(
+        nn.Conv2d(inp, oup, 3, stride, 1, bias=False), nn.BatchNorm2d(oup),
+        nn.LeakyReLU(negative_slope=leaky, inplace=True))
+
+
+def conv_bn_no_relu(inp, oup, stride):
+    return nn.Sequential(
+        nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
+        nn.BatchNorm2d(oup),
+    )
+
+
+def conv_bn1X1(inp, oup, stride, leaky=0):
+    return nn.Sequential(
+        nn.Conv2d(inp, oup, 1, stride, padding=0, bias=False), nn.BatchNorm2d(oup),
+        nn.LeakyReLU(negative_slope=leaky, inplace=True))
+
+
+def conv_dw(inp, oup, stride, leaky=0.1):
+    return nn.Sequential(
+        nn.Conv2d(inp, inp, 3, stride, 1, groups=inp, bias=False),
+        nn.BatchNorm2d(inp),
+        nn.LeakyReLU(negative_slope=leaky, inplace=True),
+        nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
+        nn.BatchNorm2d(oup),
+        nn.LeakyReLU(negative_slope=leaky, inplace=True),
+    )
+
+
+class SSH(nn.Module):
+
+    def __init__(self, in_channel, out_channel):
+        super(SSH, self).__init__()
+        assert out_channel % 4 == 0
+        leaky = 0
+        if (out_channel <= 64):
+            leaky = 0.1
+        self.conv3X3 = conv_bn_no_relu(in_channel, out_channel // 2, stride=1)
+
+        self.conv5X5_1 = conv_bn(in_channel, out_channel // 4, stride=1, leaky=leaky)
+        self.conv5X5_2 = conv_bn_no_relu(out_channel // 4, out_channel // 4, stride=1)
+
+        self.conv7X7_2 = conv_bn(out_channel // 4, out_channel // 4, stride=1, leaky=leaky)
+        self.conv7x7_3 = conv_bn_no_relu(out_channel // 4, out_channel // 4, stride=1)
+
+    def forward(self, input):
+        conv3X3 = self.conv3X3(input)
+
+        conv5X5_1 = self.conv5X5_1(input)
+        conv5X5 = self.conv5X5_2(conv5X5_1)
+
+        conv7X7_2 = self.conv7X7_2(conv5X5_1)
+        conv7X7 = self.conv7x7_3(conv7X7_2)
+
+        out = torch.cat([conv3X3, conv5X5, conv7X7], dim=1)
+        out = F.relu(out)
+        return out
+
+
+class FPN(nn.Module):
+
+    def __init__(self, in_channels_list, out_channels):
+        super(FPN, self).__init__()
+        leaky = 0
+        if (out_channels <= 64):
+            leaky = 0.1
+        self.output1 = conv_bn1X1(in_channels_list[0], out_channels, stride=1, leaky=leaky)
+        self.output2 = conv_bn1X1(in_channels_list[1], out_channels, stride=1, leaky=leaky)
+        self.output3 = conv_bn1X1(in_channels_list[2], out_channels, stride=1, leaky=leaky)
+
+        self.merge1 = conv_bn(out_channels, out_channels, leaky=leaky)
+        self.merge2 = conv_bn(out_channels, out_channels, leaky=leaky)
+
+    def forward(self, input):
+        # names = list(input.keys())
+        # input = list(input.values())
+
+        output1 = self.output1(input[0])
+        output2 = self.output2(input[1])
+        output3 = self.output3(input[2])
+
+        up3 = F.interpolate(output3, size=[output2.size(2), output2.size(3)], mode='nearest')
+        output2 = output2 + up3
+        output2 = self.merge2(output2)
+
+        up2 = F.interpolate(output2, size=[output1.size(2), output1.size(3)], mode='nearest')
+        output1 = output1 + up2
+        output1 = self.merge1(output1)
+
+        out = [output1, output2, output3]
+        return out
+
+
+class MobileNetV1(nn.Module):
+
+    def __init__(self):
+        super(MobileNetV1, self).__init__()
+        self.stage1 = nn.Sequential(
+            conv_bn(3, 8, 2, leaky=0.1),  # 3
+            conv_dw(8, 16, 1),  # 7
+            conv_dw(16, 32, 2),  # 11
+            conv_dw(32, 32, 1),  # 19
+            conv_dw(32, 64, 2),  # 27
+            conv_dw(64, 64, 1),  # 43
+        )
+        self.stage2 = nn.Sequential(
+            conv_dw(64, 128, 2),  # 43 + 16 = 59
+            conv_dw(128, 128, 1),  # 59 + 32 = 91
+            conv_dw(128, 128, 1),  # 91 + 32 = 123
+            conv_dw(128, 128, 1),  # 123 + 32 = 155
+            conv_dw(128, 128, 1),  # 155 + 32 = 187
+            conv_dw(128, 128, 1),  # 187 + 32 = 219
+        )
+        self.stage3 = nn.Sequential(
+            conv_dw(128, 256, 2),  # 219 +3 2 = 241
+            conv_dw(256, 256, 1),  # 241 + 64 = 301
+        )
+        self.avg = nn.AdaptiveAvgPool2d((1, 1))
+        self.fc = nn.Linear(256, 1000)
+
+    def forward(self, x):
+        x = self.stage1(x)
+        x = self.stage2(x)
+        x = self.stage3(x)
+        x = self.avg(x)
+        # x = self.model(x)
+        x = x.view(-1, 256)
+        x = self.fc(x)
+        return x
+
+
+class ClassHead(nn.Module):
+
+    def __init__(self, inchannels=512, num_anchors=3):
+        super(ClassHead, self).__init__()
+        self.num_anchors = num_anchors
+        self.conv1x1 = nn.Conv2d(inchannels, self.num_anchors * 2, kernel_size=(1, 1), stride=1, padding=0)
+
+    def forward(self, x):
+        out = self.conv1x1(x)
+        out = out.permute(0, 2, 3, 1).contiguous()
+
+        return out.view(out.shape[0], -1, 2)
+
+
+class BboxHead(nn.Module):
+
+    def __init__(self, inchannels=512, num_anchors=3):
+        super(BboxHead, self).__init__()
+        self.conv1x1 = nn.Conv2d(inchannels, num_anchors * 4, kernel_size=(1, 1), stride=1, padding=0)
+
+    def forward(self, x):
+        out = self.conv1x1(x)
+        out = out.permute(0, 2, 3, 1).contiguous()
+
+        return out.view(out.shape[0], -1, 4)
+
+
+class LandmarkHead(nn.Module):
+
+    def __init__(self, inchannels=512, num_anchors=3):
+        super(LandmarkHead, self).__init__()
+        self.conv1x1 = nn.Conv2d(inchannels, num_anchors * 10, kernel_size=(1, 1), stride=1, padding=0)
+
+    def forward(self, x):
+        out = self.conv1x1(x)
+        out = out.permute(0, 2, 3, 1).contiguous()
+
+        return out.view(out.shape[0], -1, 10)
+
+
+def make_class_head(fpn_num=3, inchannels=64, anchor_num=2):
+    classhead = nn.ModuleList()
+    for i in range(fpn_num):
+        classhead.append(ClassHead(inchannels, anchor_num))
+    return classhead
+
+
+def make_bbox_head(fpn_num=3, inchannels=64, anchor_num=2):
+    bboxhead = nn.ModuleList()
+    for i in range(fpn_num):
+        bboxhead.append(BboxHead(inchannels, anchor_num))
+    return bboxhead
+
+
+def make_landmark_head(fpn_num=3, inchannels=64, anchor_num=2):
+    landmarkhead = nn.ModuleList()
+    for i in range(fpn_num):
+        landmarkhead.append(LandmarkHead(inchannels, anchor_num))
+    return landmarkhead

fooocus_extras/facexlib/detection/retinaface_utils.py

@@ -0,0 +1,421 @@
+import numpy as np
+import torch
+import torchvision
+from itertools import product as product
+from math import ceil
+
+
+class PriorBox(object):
+
+    def __init__(self, cfg, image_size=None, phase='train'):
+        super(PriorBox, self).__init__()
+        self.min_sizes = cfg['min_sizes']
+        self.steps = cfg['steps']
+        self.clip = cfg['clip']
+        self.image_size = image_size
+        self.feature_maps = [[ceil(self.image_size[0] / step), ceil(self.image_size[1] / step)] for step in self.steps]
+        self.name = 's'
+
+    def forward(self):
+        anchors = []
+        for k, f in enumerate(self.feature_maps):
+            min_sizes = self.min_sizes[k]
+            for i, j in product(range(f[0]), range(f[1])):
+                for min_size in min_sizes:
+                    s_kx = min_size / self.image_size[1]
+                    s_ky = min_size / self.image_size[0]
+                    dense_cx = [x * self.steps[k] / self.image_size[1] for x in [j + 0.5]]
+                    dense_cy = [y * self.steps[k] / self.image_size[0] for y in [i + 0.5]]
+                    for cy, cx in product(dense_cy, dense_cx):
+                        anchors += [cx, cy, s_kx, s_ky]
+
+        # back to torch land
+        output = torch.Tensor(anchors).view(-1, 4)
+        if self.clip:
+            output.clamp_(max=1, min=0)
+        return output
+
+
+def py_cpu_nms(dets, thresh):
+    """Pure Python NMS baseline."""
+    keep = torchvision.ops.nms(
+        boxes=torch.Tensor(dets[:, :4]),
+        scores=torch.Tensor(dets[:, 4]),
+        iou_threshold=thresh,
+    )
+
+    return list(keep)
+
+
+def point_form(boxes):
+    """ Convert prior_boxes to (xmin, ymin, xmax, ymax)
+    representation for comparison to point form ground truth data.
+    Args:
+        boxes: (tensor) center-size default boxes from priorbox layers.
+    Return:
+        boxes: (tensor) Converted xmin, ymin, xmax, ymax form of boxes.
+    """
+    return torch.cat(
+        (
+            boxes[:, :2] - boxes[:, 2:] / 2,  # xmin, ymin
+            boxes[:, :2] + boxes[:, 2:] / 2),
+        1)  # xmax, ymax
+
+
+def center_size(boxes):
+    """ Convert prior_boxes to (cx, cy, w, h)
+    representation for comparison to center-size form ground truth data.
+    Args:
+        boxes: (tensor) point_form boxes
+    Return:
+        boxes: (tensor) Converted xmin, ymin, xmax, ymax form of boxes.
+    """
+    return torch.cat(
+        (boxes[:, 2:] + boxes[:, :2]) / 2,  # cx, cy
+        boxes[:, 2:] - boxes[:, :2],
+        1)  # w, h
+
+
+def intersect(box_a, box_b):
+    """ We resize both tensors to [A,B,2] without new malloc:
+    [A,2] -> [A,1,2] -> [A,B,2]
+    [B,2] -> [1,B,2] -> [A,B,2]
+    Then we compute the area of intersect between box_a and box_b.
+    Args:
+      box_a: (tensor) bounding boxes, Shape: [A,4].
+      box_b: (tensor) bounding boxes, Shape: [B,4].
+    Return:
+      (tensor) intersection area, Shape: [A,B].
+    """
+    A = box_a.size(0)
+    B = box_b.size(0)
+    max_xy = torch.min(box_a[:, 2:].unsqueeze(1).expand(A, B, 2), box_b[:, 2:].unsqueeze(0).expand(A, B, 2))
+    min_xy = torch.max(box_a[:, :2].unsqueeze(1).expand(A, B, 2), box_b[:, :2].unsqueeze(0).expand(A, B, 2))
+    inter = torch.clamp((max_xy - min_xy), min=0)
+    return inter[:, :, 0] * inter[:, :, 1]
+
+
+def jaccard(box_a, box_b):
+    """Compute the jaccard overlap of two sets of boxes.  The jaccard overlap
+    is simply the intersection over union of two boxes.  Here we operate on
+    ground truth boxes and default boxes.
+    E.g.:
+        A ∩ B / A ∪ B = A ∩ B / (area(A) + area(B) - A ∩ B)
+    Args:
+        box_a: (tensor) Ground truth bounding boxes, Shape: [num_objects,4]
+        box_b: (tensor) Prior boxes from priorbox layers, Shape: [num_priors,4]
+    Return:
+        jaccard overlap: (tensor) Shape: [box_a.size(0), box_b.size(0)]
+    """
+    inter = intersect(box_a, box_b)
+    area_a = ((box_a[:, 2] - box_a[:, 0]) * (box_a[:, 3] - box_a[:, 1])).unsqueeze(1).expand_as(inter)  # [A,B]
+    area_b = ((box_b[:, 2] - box_b[:, 0]) * (box_b[:, 3] - box_b[:, 1])).unsqueeze(0).expand_as(inter)  # [A,B]
+    union = area_a + area_b - inter
+    return inter / union  # [A,B]
+
+
+def matrix_iou(a, b):
+    """
+    return iou of a and b, numpy version for data augenmentation
+    """
+    lt = np.maximum(a[:, np.newaxis, :2], b[:, :2])
+    rb = np.minimum(a[:, np.newaxis, 2:], b[:, 2:])
+
+    area_i = np.prod(rb - lt, axis=2) * (lt < rb).all(axis=2)
+    area_a = np.prod(a[:, 2:] - a[:, :2], axis=1)
+    area_b = np.prod(b[:, 2:] - b[:, :2], axis=1)
+    return area_i / (area_a[:, np.newaxis] + area_b - area_i)
+
+
+def matrix_iof(a, b):
+    """
+    return iof of a and b, numpy version for data augenmentation
+    """
+    lt = np.maximum(a[:, np.newaxis, :2], b[:, :2])
+    rb = np.minimum(a[:, np.newaxis, 2:], b[:, 2:])
+
+    area_i = np.prod(rb - lt, axis=2) * (lt < rb).all(axis=2)
+    area_a = np.prod(a[:, 2:] - a[:, :2], axis=1)
+    return area_i / np.maximum(area_a[:, np.newaxis], 1)
+
+
+def match(threshold, truths, priors, variances, labels, landms, loc_t, conf_t, landm_t, idx):
+    """Match each prior box with the ground truth box of the highest jaccard
+    overlap, encode the bounding boxes, then return the matched indices
+    corresponding to both confidence and location preds.
+    Args:
+        threshold: (float) The overlap threshold used when matching boxes.
+        truths: (tensor) Ground truth boxes, Shape: [num_obj, 4].
+        priors: (tensor) Prior boxes from priorbox layers, Shape: [n_priors,4].
+        variances: (tensor) Variances corresponding to each prior coord,
+            Shape: [num_priors, 4].
+        labels: (tensor) All the class labels for the image, Shape: [num_obj].
+        landms: (tensor) Ground truth landms, Shape [num_obj, 10].
+        loc_t: (tensor) Tensor to be filled w/ encoded location targets.
+        conf_t: (tensor) Tensor to be filled w/ matched indices for conf preds.
+        landm_t: (tensor) Tensor to be filled w/ encoded landm targets.
+        idx: (int) current batch index
+    Return:
+        The matched indices corresponding to 1)location 2)confidence
+        3)landm preds.
+    """
+    # jaccard index
+    overlaps = jaccard(truths, point_form(priors))
+    # (Bipartite Matching)
+    # [1,num_objects] best prior for each ground truth
+    best_prior_overlap, best_prior_idx = overlaps.max(1, keepdim=True)
+
+    # ignore hard gt
+    valid_gt_idx = best_prior_overlap[:, 0] >= 0.2
+    best_prior_idx_filter = best_prior_idx[valid_gt_idx, :]
+    if best_prior_idx_filter.shape[0] <= 0:
+        loc_t[idx] = 0
+        conf_t[idx] = 0
+        return
+
+    # [1,num_priors] best ground truth for each prior
+    best_truth_overlap, best_truth_idx = overlaps.max(0, keepdim=True)
+    best_truth_idx.squeeze_(0)
+    best_truth_overlap.squeeze_(0)
+    best_prior_idx.squeeze_(1)
+    best_prior_idx_filter.squeeze_(1)
+    best_prior_overlap.squeeze_(1)
+    best_truth_overlap.index_fill_(0, best_prior_idx_filter, 2)  # ensure best prior
+    # TODO refactor: index  best_prior_idx with long tensor
+    # ensure every gt matches with its prior of max overlap
+    for j in range(best_prior_idx.size(0)):  # 判别此anchor是预测哪一个boxes
+        best_truth_idx[best_prior_idx[j]] = j
+    matches = truths[best_truth_idx]  # Shape: [num_priors,4] 此处为每一个anchor对应的bbox取出来
+    conf = labels[best_truth_idx]  # Shape: [num_priors]      此处为每一个anchor对应的label取出来
+    conf[best_truth_overlap < threshold] = 0  # label as background   overlap<0.35的全部作为负样本
+    loc = encode(matches, priors, variances)
+
+    matches_landm = landms[best_truth_idx]
+    landm = encode_landm(matches_landm, priors, variances)
+    loc_t[idx] = loc  # [num_priors,4] encoded offsets to learn
+    conf_t[idx] = conf  # [num_priors] top class label for each prior
+    landm_t[idx] = landm
+
+
+def encode(matched, priors, variances):
+    """Encode the variances from the priorbox layers into the ground truth boxes
+    we have matched (based on jaccard overlap) with the prior boxes.
+    Args:
+        matched: (tensor) Coords of ground truth for each prior in point-form
+            Shape: [num_priors, 4].
+        priors: (tensor) Prior boxes in center-offset form
+            Shape: [num_priors,4].
+        variances: (list[float]) Variances of priorboxes
+    Return:
+        encoded boxes (tensor), Shape: [num_priors, 4]
+    """
+
+    # dist b/t match center and prior's center
+    g_cxcy = (matched[:, :2] + matched[:, 2:]) / 2 - priors[:, :2]
+    # encode variance
+    g_cxcy /= (variances[0] * priors[:, 2:])
+    # match wh / prior wh
+    g_wh = (matched[:, 2:] - matched[:, :2]) / priors[:, 2:]
+    g_wh = torch.log(g_wh) / variances[1]
+    # return target for smooth_l1_loss
+    return torch.cat([g_cxcy, g_wh], 1)  # [num_priors,4]
+
+
+def encode_landm(matched, priors, variances):
+    """Encode the variances from the priorbox layers into the ground truth boxes
+    we have matched (based on jaccard overlap) with the prior boxes.
+    Args:
+        matched: (tensor) Coords of ground truth for each prior in point-form
+            Shape: [num_priors, 10].
+        priors: (tensor) Prior boxes in center-offset form
+            Shape: [num_priors,4].
+        variances: (list[float]) Variances of priorboxes
+    Return:
+        encoded landm (tensor), Shape: [num_priors, 10]
+    """
+
+    # dist b/t match center and prior's center
+    matched = torch.reshape(matched, (matched.size(0), 5, 2))
+    priors_cx = priors[:, 0].unsqueeze(1).expand(matched.size(0), 5).unsqueeze(2)
+    priors_cy = priors[:, 1].unsqueeze(1).expand(matched.size(0), 5).unsqueeze(2)
+    priors_w = priors[:, 2].unsqueeze(1).expand(matched.size(0), 5).unsqueeze(2)
+    priors_h = priors[:, 3].unsqueeze(1).expand(matched.size(0), 5).unsqueeze(2)
+    priors = torch.cat([priors_cx, priors_cy, priors_w, priors_h], dim=2)
+    g_cxcy = matched[:, :, :2] - priors[:, :, :2]
+    # encode variance
+    g_cxcy /= (variances[0] * priors[:, :, 2:])
+    # g_cxcy /= priors[:, :, 2:]
+    g_cxcy = g_cxcy.reshape(g_cxcy.size(0), -1)
+    # return target for smooth_l1_loss
+    return g_cxcy
+
+
+# Adapted from https://github.com/Hakuyume/chainer-ssd
+def decode(loc, priors, variances):
+    """Decode locations from predictions using priors to undo
+    the encoding we did for offset regression at train time.
+    Args:
+        loc (tensor): location predictions for loc layers,
+            Shape: [num_priors,4]
+        priors (tensor): Prior boxes in center-offset form.
+            Shape: [num_priors,4].
+        variances: (list[float]) Variances of priorboxes
+    Return:
+        decoded bounding box predictions
+    """
+
+    boxes = torch.cat((priors[:, :2] + loc[:, :2] * variances[0] * priors[:, 2:],
+                       priors[:, 2:] * torch.exp(loc[:, 2:] * variances[1])), 1)
+    boxes[:, :2] -= boxes[:, 2:] / 2
+    boxes[:, 2:] += boxes[:, :2]
+    return boxes
+
+
+def decode_landm(pre, priors, variances):
+    """Decode landm from predictions using priors to undo
+    the encoding we did for offset regression at train time.
+    Args:
+        pre (tensor): landm predictions for loc layers,
+            Shape: [num_priors,10]
+        priors (tensor): Prior boxes in center-offset form.
+            Shape: [num_priors,4].
+        variances: (list[float]) Variances of priorboxes
+    Return:
+        decoded landm predictions
+    """
+    tmp = (
+        priors[:, :2] + pre[:, :2] * variances[0] * priors[:, 2:],
+        priors[:, :2] + pre[:, 2:4] * variances[0] * priors[:, 2:],
+        priors[:, :2] + pre[:, 4:6] * variances[0] * priors[:, 2:],
+        priors[:, :2] + pre[:, 6:8] * variances[0] * priors[:, 2:],
+        priors[:, :2] + pre[:, 8:10] * variances[0] * priors[:, 2:],
+    )
+    landms = torch.cat(tmp, dim=1)
+    return landms
+
+
+def batched_decode(b_loc, priors, variances):
+    """Decode locations from predictions using priors to undo
+    the encoding we did for offset regression at train time.
+    Args:
+        b_loc (tensor): location predictions for loc layers,
+            Shape: [num_batches,num_priors,4]
+        priors (tensor): Prior boxes in center-offset form.
+            Shape: [1,num_priors,4].
+        variances: (list[float]) Variances of priorboxes
+    Return:
+        decoded bounding box predictions
+    """
+    boxes = (
+        priors[:, :, :2] + b_loc[:, :, :2] * variances[0] * priors[:, :, 2:],
+        priors[:, :, 2:] * torch.exp(b_loc[:, :, 2:] * variances[1]),
+    )
+    boxes = torch.cat(boxes, dim=2)
+
+    boxes[:, :, :2] -= boxes[:, :, 2:] / 2
+    boxes[:, :, 2:] += boxes[:, :, :2]
+    return boxes
+
+
+def batched_decode_landm(pre, priors, variances):
+    """Decode landm from predictions using priors to undo
+    the encoding we did for offset regression at train time.
+    Args:
+        pre (tensor): landm predictions for loc layers,
+            Shape: [num_batches,num_priors,10]
+        priors (tensor): Prior boxes in center-offset form.
+            Shape: [1,num_priors,4].
+        variances: (list[float]) Variances of priorboxes
+    Return:
+        decoded landm predictions
+    """
+    landms = (
+        priors[:, :, :2] + pre[:, :, :2] * variances[0] * priors[:, :, 2:],
+        priors[:, :, :2] + pre[:, :, 2:4] * variances[0] * priors[:, :, 2:],
+        priors[:, :, :2] + pre[:, :, 4:6] * variances[0] * priors[:, :, 2:],
+        priors[:, :, :2] + pre[:, :, 6:8] * variances[0] * priors[:, :, 2:],
+        priors[:, :, :2] + pre[:, :, 8:10] * variances[0] * priors[:, :, 2:],
+    )
+    landms = torch.cat(landms, dim=2)
+    return landms
+
+
+def log_sum_exp(x):
+    """Utility function for computing log_sum_exp while determining
+    This will be used to determine unaveraged confidence loss across
+    all examples in a batch.
+    Args:
+        x (Variable(tensor)): conf_preds from conf layers
+    """
+    x_max = x.data.max()
+    return torch.log(torch.sum(torch.exp(x - x_max), 1, keepdim=True)) + x_max
+
+
+# Original author: Francisco Massa:
+# https://github.com/fmassa/object-detection.torch
+# Ported to PyTorch by Max deGroot (02/01/2017)
+def nms(boxes, scores, overlap=0.5, top_k=200):
+    """Apply non-maximum suppression at test time to avoid detecting too many
+    overlapping bounding boxes for a given object.
+    Args:
+        boxes: (tensor) The location preds for the img, Shape: [num_priors,4].
+        scores: (tensor) The class predscores for the img, Shape:[num_priors].
+        overlap: (float) The overlap thresh for suppressing unnecessary boxes.
+        top_k: (int) The Maximum number of box preds to consider.
+    Return:
+        The indices of the kept boxes with respect to num_priors.
+    """
+
+    keep = torch.Tensor(scores.size(0)).fill_(0).long()
+    if boxes.numel() == 0:
+        return keep
+    x1 = boxes[:, 0]
+    y1 = boxes[:, 1]
+    x2 = boxes[:, 2]
+    y2 = boxes[:, 3]
+    area = torch.mul(x2 - x1, y2 - y1)
+    v, idx = scores.sort(0)  # sort in ascending order
+    # I = I[v >= 0.01]
+    idx = idx[-top_k:]  # indices of the top-k largest vals
+    xx1 = boxes.new()
+    yy1 = boxes.new()
+    xx2 = boxes.new()
+    yy2 = boxes.new()
+    w = boxes.new()
+    h = boxes.new()
+
+    # keep = torch.Tensor()
+    count = 0
+    while idx.numel() > 0:
+        i = idx[-1]  # index of current largest val
+        # keep.append(i)
+        keep[count] = i
+        count += 1
+        if idx.size(0) == 1:
+            break
+        idx = idx[:-1]  # remove kept element from view
+        # load bboxes of next highest vals
+        torch.index_select(x1, 0, idx, out=xx1)
+        torch.index_select(y1, 0, idx, out=yy1)
+        torch.index_select(x2, 0, idx, out=xx2)
+        torch.index_select(y2, 0, idx, out=yy2)
+        # store element-wise max with next highest score
+        xx1 = torch.clamp(xx1, min=x1[i])
+        yy1 = torch.clamp(yy1, min=y1[i])
+        xx2 = torch.clamp(xx2, max=x2[i])
+        yy2 = torch.clamp(yy2, max=y2[i])
+        w.resize_as_(xx2)
+        h.resize_as_(yy2)
+        w = xx2 - xx1
+        h = yy2 - yy1
+        # check sizes of xx1 and xx2.. after each iteration
+        w = torch.clamp(w, min=0.0)
+        h = torch.clamp(h, min=0.0)
+        inter = w * h
+        # IoU = i / (area(a) + area(b) - i)
+        rem_areas = torch.index_select(area, 0, idx)  # load remaining areas)
+        union = (rem_areas - inter) + area[i]
+        IoU = inter / union  # store result in iou
+        # keep only elements with an IoU <= overlap
+        idx = idx[IoU.le(overlap)]
+    return keep, count

fooocus_extras/facexlib/parsing/__init__.py

@@ -0,0 +1,24 @@
+import torch
+
+from fooocus_extras.facexlib.utils import load_file_from_url
+from .bisenet import BiSeNet
+from .parsenet import ParseNet
+
+
+def init_parsing_model(model_name='bisenet', half=False, device='cuda', model_rootpath=None):
+    if model_name == 'bisenet':
+        model = BiSeNet(num_class=19)
+        model_url = 'https://github.com/xinntao/facexlib/releases/download/v0.2.0/parsing_bisenet.pth'
+    elif model_name == 'parsenet':
+        model = ParseNet(in_size=512, out_size=512, parsing_ch=19)
+        model_url = 'https://github.com/xinntao/facexlib/releases/download/v0.2.2/parsing_parsenet.pth'
+    else:
+        raise NotImplementedError(f'{model_name} is not implemented.')
+
+    model_path = load_file_from_url(
+        url=model_url, model_dir='facexlib/weights', progress=True, file_name=None, save_dir=model_rootpath)
+    load_net = torch.load(model_path, map_location=lambda storage, loc: storage)
+    model.load_state_dict(load_net, strict=True)
+    model.eval()
+    model = model.to(device)
+    return model

fooocus_extras/facexlib/parsing/bisenet.py

@@ -0,0 +1,140 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from .resnet import ResNet18
+
+
+class ConvBNReLU(nn.Module):
+
+    def __init__(self, in_chan, out_chan, ks=3, stride=1, padding=1):
+        super(ConvBNReLU, self).__init__()
+        self.conv = nn.Conv2d(in_chan, out_chan, kernel_size=ks, stride=stride, padding=padding, bias=False)
+        self.bn = nn.BatchNorm2d(out_chan)
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = F.relu(self.bn(x))
+        return x
+
+
+class BiSeNetOutput(nn.Module):
+
+    def __init__(self, in_chan, mid_chan, num_class):
+        super(BiSeNetOutput, self).__init__()
+        self.conv = ConvBNReLU(in_chan, mid_chan, ks=3, stride=1, padding=1)
+        self.conv_out = nn.Conv2d(mid_chan, num_class, kernel_size=1, bias=False)
+
+    def forward(self, x):
+        feat = self.conv(x)
+        out = self.conv_out(feat)
+        return out, feat
+
+
+class AttentionRefinementModule(nn.Module):
+
+    def __init__(self, in_chan, out_chan):
+        super(AttentionRefinementModule, self).__init__()
+        self.conv = ConvBNReLU(in_chan, out_chan, ks=3, stride=1, padding=1)
+        self.conv_atten = nn.Conv2d(out_chan, out_chan, kernel_size=1, bias=False)
+        self.bn_atten = nn.BatchNorm2d(out_chan)
+        self.sigmoid_atten = nn.Sigmoid()
+
+    def forward(self, x):
+        feat = self.conv(x)
+        atten = F.avg_pool2d(feat, feat.size()[2:])
+        atten = self.conv_atten(atten)
+        atten = self.bn_atten(atten)
+        atten = self.sigmoid_atten(atten)
+        out = torch.mul(feat, atten)
+        return out
+
+
+class ContextPath(nn.Module):
+
+    def __init__(self):
+        super(ContextPath, self).__init__()
+        self.resnet = ResNet18()
+        self.arm16 = AttentionRefinementModule(256, 128)
+        self.arm32 = AttentionRefinementModule(512, 128)
+        self.conv_head32 = ConvBNReLU(128, 128, ks=3, stride=1, padding=1)
+        self.conv_head16 = ConvBNReLU(128, 128, ks=3, stride=1, padding=1)
+        self.conv_avg = ConvBNReLU(512, 128, ks=1, stride=1, padding=0)
+
+    def forward(self, x):
+        feat8, feat16, feat32 = self.resnet(x)
+        h8, w8 = feat8.size()[2:]
+        h16, w16 = feat16.size()[2:]
+        h32, w32 = feat32.size()[2:]
+
+        avg = F.avg_pool2d(feat32, feat32.size()[2:])
+        avg = self.conv_avg(avg)
+        avg_up = F.interpolate(avg, (h32, w32), mode='nearest')
+
+        feat32_arm = self.arm32(feat32)
+        feat32_sum = feat32_arm + avg_up
+        feat32_up = F.interpolate(feat32_sum, (h16, w16), mode='nearest')
+        feat32_up = self.conv_head32(feat32_up)
+
+        feat16_arm = self.arm16(feat16)
+        feat16_sum = feat16_arm + feat32_up
+        feat16_up = F.interpolate(feat16_sum, (h8, w8), mode='nearest')
+        feat16_up = self.conv_head16(feat16_up)
+
+        return feat8, feat16_up, feat32_up  # x8, x8, x16
+
+
+class FeatureFusionModule(nn.Module):
+
+    def __init__(self, in_chan, out_chan):
+        super(FeatureFusionModule, self).__init__()
+        self.convblk = ConvBNReLU(in_chan, out_chan, ks=1, stride=1, padding=0)
+        self.conv1 = nn.Conv2d(out_chan, out_chan // 4, kernel_size=1, stride=1, padding=0, bias=False)
+        self.conv2 = nn.Conv2d(out_chan // 4, out_chan, kernel_size=1, stride=1, padding=0, bias=False)
+        self.relu = nn.ReLU(inplace=True)
+        self.sigmoid = nn.Sigmoid()
+
+    def forward(self, fsp, fcp):
+        fcat = torch.cat([fsp, fcp], dim=1)
+        feat = self.convblk(fcat)
+        atten = F.avg_pool2d(feat, feat.size()[2:])
+        atten = self.conv1(atten)
+        atten = self.relu(atten)
+        atten = self.conv2(atten)
+        atten = self.sigmoid(atten)
+        feat_atten = torch.mul(feat, atten)
+        feat_out = feat_atten + feat
+        return feat_out
+
+
+class BiSeNet(nn.Module):
+
+    def __init__(self, num_class):
+        super(BiSeNet, self).__init__()
+        self.cp = ContextPath()
+        self.ffm = FeatureFusionModule(256, 256)
+        self.conv_out = BiSeNetOutput(256, 256, num_class)
+        self.conv_out16 = BiSeNetOutput(128, 64, num_class)
+        self.conv_out32 = BiSeNetOutput(128, 64, num_class)
+
+    def forward(self, x, return_feat=False):
+        h, w = x.size()[2:]
+        feat_res8, feat_cp8, feat_cp16 = self.cp(x)  # return res3b1 feature
+        feat_sp = feat_res8  # replace spatial path feature with res3b1 feature
+        feat_fuse = self.ffm(feat_sp, feat_cp8)
+
+        out, feat = self.conv_out(feat_fuse)
+        out16, feat16 = self.conv_out16(feat_cp8)
+        out32, feat32 = self.conv_out32(feat_cp16)
+
+        out = F.interpolate(out, (h, w), mode='bilinear', align_corners=True)
+        out16 = F.interpolate(out16, (h, w), mode='bilinear', align_corners=True)
+        out32 = F.interpolate(out32, (h, w), mode='bilinear', align_corners=True)
+
+        if return_feat:
+            feat = F.interpolate(feat, (h, w), mode='bilinear', align_corners=True)
+            feat16 = F.interpolate(feat16, (h, w), mode='bilinear', align_corners=True)
+            feat32 = F.interpolate(feat32, (h, w), mode='bilinear', align_corners=True)
+            return out, out16, out32, feat, feat16, feat32
+        else:
+            return out, out16, out32

fooocus_extras/facexlib/parsing/parsenet.py

@@ -0,0 +1,194 @@
+"""Modified from https://github.com/chaofengc/PSFRGAN
+"""
+import numpy as np
+import torch.nn as nn
+from torch.nn import functional as F
+
+
+class NormLayer(nn.Module):
+    """Normalization Layers.
+
+    Args:
+        channels: input channels, for batch norm and instance norm.
+        input_size: input shape without batch size, for layer norm.
+    """
+
+    def __init__(self, channels, normalize_shape=None, norm_type='bn'):
+        super(NormLayer, self).__init__()
+        norm_type = norm_type.lower()
+        self.norm_type = norm_type
+        if norm_type == 'bn':
+            self.norm = nn.BatchNorm2d(channels, affine=True)
+        elif norm_type == 'in':
+            self.norm = nn.InstanceNorm2d(channels, affine=False)
+        elif norm_type == 'gn':
+            self.norm = nn.GroupNorm(32, channels, affine=True)
+        elif norm_type == 'pixel':
+            self.norm = lambda x: F.normalize(x, p=2, dim=1)
+        elif norm_type == 'layer':
+            self.norm = nn.LayerNorm(normalize_shape)
+        elif norm_type == 'none':
+            self.norm = lambda x: x * 1.0
+        else:
+            assert 1 == 0, f'Norm type {norm_type} not support.'
+
+    def forward(self, x, ref=None):
+        if self.norm_type == 'spade':
+            return self.norm(x, ref)
+        else:
+            return self.norm(x)
+
+
+class ReluLayer(nn.Module):
+    """Relu Layer.
+
+    Args:
+        relu type: type of relu layer, candidates are
+            - ReLU
+            - LeakyReLU: default relu slope 0.2
+            - PRelu
+            - SELU
+            - none: direct pass
+    """
+
+    def __init__(self, channels, relu_type='relu'):
+        super(ReluLayer, self).__init__()
+        relu_type = relu_type.lower()
+        if relu_type == 'relu':
+            self.func = nn.ReLU(True)
+        elif relu_type == 'leakyrelu':
+            self.func = nn.LeakyReLU(0.2, inplace=True)
+        elif relu_type == 'prelu':
+            self.func = nn.PReLU(channels)
+        elif relu_type == 'selu':
+            self.func = nn.SELU(True)
+        elif relu_type == 'none':
+            self.func = lambda x: x * 1.0
+        else:
+            assert 1 == 0, f'Relu type {relu_type} not support.'
+
+    def forward(self, x):
+        return self.func(x)
+
+
+class ConvLayer(nn.Module):
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size=3,
+                 scale='none',
+                 norm_type='none',
+                 relu_type='none',
+                 use_pad=True,
+                 bias=True):
+        super(ConvLayer, self).__init__()
+        self.use_pad = use_pad
+        self.norm_type = norm_type
+        if norm_type in ['bn']:
+            bias = False
+
+        stride = 2 if scale == 'down' else 1
+
+        self.scale_func = lambda x: x
+        if scale == 'up':
+            self.scale_func = lambda x: nn.functional.interpolate(x, scale_factor=2, mode='nearest')
+
+        self.reflection_pad = nn.ReflectionPad2d(int(np.ceil((kernel_size - 1.) / 2)))
+        self.conv2d = nn.Conv2d(in_channels, out_channels, kernel_size, stride, bias=bias)
+
+        self.relu = ReluLayer(out_channels, relu_type)
+        self.norm = NormLayer(out_channels, norm_type=norm_type)
+
+    def forward(self, x):
+        out = self.scale_func(x)
+        if self.use_pad:
+            out = self.reflection_pad(out)
+        out = self.conv2d(out)
+        out = self.norm(out)
+        out = self.relu(out)
+        return out
+
+
+class ResidualBlock(nn.Module):
+    """
+    Residual block recommended in: http://torch.ch/blog/2016/02/04/resnets.html
+    """
+
+    def __init__(self, c_in, c_out, relu_type='prelu', norm_type='bn', scale='none'):
+        super(ResidualBlock, self).__init__()
+
+        if scale == 'none' and c_in == c_out:
+            self.shortcut_func = lambda x: x
+        else:
+            self.shortcut_func = ConvLayer(c_in, c_out, 3, scale)
+
+        scale_config_dict = {'down': ['none', 'down'], 'up': ['up', 'none'], 'none': ['none', 'none']}
+        scale_conf = scale_config_dict[scale]
+
+        self.conv1 = ConvLayer(c_in, c_out, 3, scale_conf[0], norm_type=norm_type, relu_type=relu_type)
+        self.conv2 = ConvLayer(c_out, c_out, 3, scale_conf[1], norm_type=norm_type, relu_type='none')
+
+    def forward(self, x):
+        identity = self.shortcut_func(x)
+
+        res = self.conv1(x)
+        res = self.conv2(res)
+        return identity + res
+
+
+class ParseNet(nn.Module):
+
+    def __init__(self,
+                 in_size=128,
+                 out_size=128,
+                 min_feat_size=32,
+                 base_ch=64,
+                 parsing_ch=19,
+                 res_depth=10,
+                 relu_type='LeakyReLU',
+                 norm_type='bn',
+                 ch_range=[32, 256]):
+        super().__init__()
+        self.res_depth = res_depth
+        act_args = {'norm_type': norm_type, 'relu_type': relu_type}
+        min_ch, max_ch = ch_range
+
+        ch_clip = lambda x: max(min_ch, min(x, max_ch))  # noqa: E731
+        min_feat_size = min(in_size, min_feat_size)
+
+        down_steps = int(np.log2(in_size // min_feat_size))
+        up_steps = int(np.log2(out_size // min_feat_size))
+
+        # =============== define encoder-body-decoder ====================
+        self.encoder = []
+        self.encoder.append(ConvLayer(3, base_ch, 3, 1))
+        head_ch = base_ch
+        for i in range(down_steps):
+            cin, cout = ch_clip(head_ch), ch_clip(head_ch * 2)
+            self.encoder.append(ResidualBlock(cin, cout, scale='down', **act_args))
+            head_ch = head_ch * 2
+
+        self.body = []
+        for i in range(res_depth):
+            self.body.append(ResidualBlock(ch_clip(head_ch), ch_clip(head_ch), **act_args))
+
+        self.decoder = []
+        for i in range(up_steps):
+            cin, cout = ch_clip(head_ch), ch_clip(head_ch // 2)
+            self.decoder.append(ResidualBlock(cin, cout, scale='up', **act_args))
+            head_ch = head_ch // 2
+
+        self.encoder = nn.Sequential(*self.encoder)
+        self.body = nn.Sequential(*self.body)
+        self.decoder = nn.Sequential(*self.decoder)
+        self.out_img_conv = ConvLayer(ch_clip(head_ch), 3)
+        self.out_mask_conv = ConvLayer(ch_clip(head_ch), parsing_ch)
+
+    def forward(self, x):
+        feat = self.encoder(x)
+        x = feat + self.body(feat)
+        x = self.decoder(x)
+        out_img = self.out_img_conv(x)
+        out_mask = self.out_mask_conv(x)
+        return out_mask, out_img

fooocus_extras/facexlib/parsing/resnet.py

@@ -0,0 +1,69 @@
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+def conv3x3(in_planes, out_planes, stride=1):
+    """3x3 convolution with padding"""
+    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False)
+
+
+class BasicBlock(nn.Module):
+
+    def __init__(self, in_chan, out_chan, stride=1):
+        super(BasicBlock, self).__init__()
+        self.conv1 = conv3x3(in_chan, out_chan, stride)
+        self.bn1 = nn.BatchNorm2d(out_chan)
+        self.conv2 = conv3x3(out_chan, out_chan)
+        self.bn2 = nn.BatchNorm2d(out_chan)
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = None
+        if in_chan != out_chan or stride != 1:
+            self.downsample = nn.Sequential(
+                nn.Conv2d(in_chan, out_chan, kernel_size=1, stride=stride, bias=False),
+                nn.BatchNorm2d(out_chan),
+            )
+
+    def forward(self, x):
+        residual = self.conv1(x)
+        residual = F.relu(self.bn1(residual))
+        residual = self.conv2(residual)
+        residual = self.bn2(residual)
+
+        shortcut = x
+        if self.downsample is not None:
+            shortcut = self.downsample(x)
+
+        out = shortcut + residual
+        out = self.relu(out)
+        return out
+
+
+def create_layer_basic(in_chan, out_chan, bnum, stride=1):
+    layers = [BasicBlock(in_chan, out_chan, stride=stride)]
+    for i in range(bnum - 1):
+        layers.append(BasicBlock(out_chan, out_chan, stride=1))
+    return nn.Sequential(*layers)
+
+
+class ResNet18(nn.Module):
+
+    def __init__(self):
+        super(ResNet18, self).__init__()
+        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False)
+        self.bn1 = nn.BatchNorm2d(64)
+        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        self.layer1 = create_layer_basic(64, 64, bnum=2, stride=1)
+        self.layer2 = create_layer_basic(64, 128, bnum=2, stride=2)
+        self.layer3 = create_layer_basic(128, 256, bnum=2, stride=2)
+        self.layer4 = create_layer_basic(256, 512, bnum=2, stride=2)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = F.relu(self.bn1(x))
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        feat8 = self.layer2(x)  # 1/8
+        feat16 = self.layer3(feat8)  # 1/16
+        feat32 = self.layer4(feat16)  # 1/32
+        return feat8, feat16, feat32

fooocus_extras/facexlib/utils/__init__.py

@@ -0,0 +1,7 @@
+from .face_utils import align_crop_face_landmarks, compute_increased_bbox, get_valid_bboxes, paste_face_back
+from .misc import img2tensor, load_file_from_url, scandir
+
+__all__ = [
+    'align_crop_face_landmarks', 'compute_increased_bbox', 'get_valid_bboxes', 'load_file_from_url', 'paste_face_back',
+    'img2tensor', 'scandir'
+]

fooocus_extras/facexlib/utils/face_restoration_helper.py

@@ -0,0 +1,374 @@
+import cv2
+import numpy as np
+import os
+import torch
+from torchvision.transforms.functional import normalize
+
+from fooocus_extras.facexlib.detection import init_detection_model
+from fooocus_extras.facexlib.parsing import init_parsing_model
+from fooocus_extras.facexlib.utils.misc import img2tensor, imwrite
+
+
+def get_largest_face(det_faces, h, w):
+
+    def get_location(val, length):
+        if val < 0:
+            return 0
+        elif val > length:
+            return length
+        else:
+            return val
+
+    face_areas = []
+    for det_face in det_faces:
+        left = get_location(det_face[0], w)
+        right = get_location(det_face[2], w)
+        top = get_location(det_face[1], h)
+        bottom = get_location(det_face[3], h)
+        face_area = (right - left) * (bottom - top)
+        face_areas.append(face_area)
+    largest_idx = face_areas.index(max(face_areas))
+    return det_faces[largest_idx], largest_idx
+
+
+def get_center_face(det_faces, h=0, w=0, center=None):
+    if center is not None:
+        center = np.array(center)
+    else:
+        center = np.array([w / 2, h / 2])
+    center_dist = []
+    for det_face in det_faces:
+        face_center = np.array([(det_face[0] + det_face[2]) / 2, (det_face[1] + det_face[3]) / 2])
+        dist = np.linalg.norm(face_center - center)
+        center_dist.append(dist)
+    center_idx = center_dist.index(min(center_dist))
+    return det_faces[center_idx], center_idx
+
+
+class FaceRestoreHelper(object):
+    """Helper for the face restoration pipeline (base class)."""
+
+    def __init__(self,
+                 upscale_factor,
+                 face_size=512,
+                 crop_ratio=(1, 1),
+                 det_model='retinaface_resnet50',
+                 save_ext='png',
+                 template_3points=False,
+                 pad_blur=False,
+                 use_parse=False,
+                 device=None,
+                 model_rootpath=None):
+        self.template_3points = template_3points  # improve robustness
+        self.upscale_factor = upscale_factor
+        # the cropped face ratio based on the square face
+        self.crop_ratio = crop_ratio  # (h, w)
+        assert (self.crop_ratio[0] >= 1 and self.crop_ratio[1] >= 1), 'crop ration only supports >=1'
+        self.face_size = (int(face_size * self.crop_ratio[1]), int(face_size * self.crop_ratio[0]))
+
+        if self.template_3points:
+            self.face_template = np.array([[192, 240], [319, 240], [257, 371]])
+        else:
+            # standard 5 landmarks for FFHQ faces with 512 x 512
+            self.face_template = np.array([[192.98138, 239.94708], [318.90277, 240.1936], [256.63416, 314.01935],
+                                           [201.26117, 371.41043], [313.08905, 371.15118]])
+        self.face_template = self.face_template * (face_size / 512.0)
+        if self.crop_ratio[0] > 1:
+            self.face_template[:, 1] += face_size * (self.crop_ratio[0] - 1) / 2
+        if self.crop_ratio[1] > 1:
+            self.face_template[:, 0] += face_size * (self.crop_ratio[1] - 1) / 2
+        self.save_ext = save_ext
+        self.pad_blur = pad_blur
+        if self.pad_blur is True:
+            self.template_3points = False
+
+        self.all_landmarks_5 = []
+        self.det_faces = []
+        self.affine_matrices = []
+        self.inverse_affine_matrices = []
+        self.cropped_faces = []
+        self.restored_faces = []
+        self.pad_input_imgs = []
+
+        if device is None:
+            self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+        else:
+            self.device = device
+
+        # init face detection model
+        self.face_det = init_detection_model(det_model, half=False, device=self.device, model_rootpath=model_rootpath)
+
+        # init face parsing model
+        self.use_parse = use_parse
+        self.face_parse = init_parsing_model(model_name='parsenet', device=self.device, model_rootpath=model_rootpath)
+
+    def set_upscale_factor(self, upscale_factor):
+        self.upscale_factor = upscale_factor
+
+    def read_image(self, img):
+        """img can be image path or cv2 loaded image."""
+        # self.input_img is Numpy array, (h, w, c), BGR, uint8, [0, 255]
+        if isinstance(img, str):
+            img = cv2.imread(img)
+
+        if np.max(img) > 256:  # 16-bit image
+            img = img / 65535 * 255
+        if len(img.shape) == 2:  # gray image
+            img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
+        elif img.shape[2] == 4:  # RGBA image with alpha channel
+            img = img[:, :, 0:3]
+
+        self.input_img = img
+
+    def get_face_landmarks_5(self,
+                             only_keep_largest=False,
+                             only_center_face=False,
+                             resize=None,
+                             blur_ratio=0.01,
+                             eye_dist_threshold=None):
+        if resize is None:
+            scale = 1
+            input_img = self.input_img
+        else:
+            h, w = self.input_img.shape[0:2]
+            scale = min(h, w) / resize
+            h, w = int(h / scale), int(w / scale)
+            input_img = cv2.resize(self.input_img, (w, h), interpolation=cv2.INTER_LANCZOS4)
+
+        with torch.no_grad():
+            bboxes = self.face_det.detect_faces(input_img, 0.97) * scale
+        for bbox in bboxes:
+            # remove faces with too small eye distance: side faces or too small faces
+            eye_dist = np.linalg.norm([bbox[5] - bbox[7], bbox[6] - bbox[8]])
+            if eye_dist_threshold is not None and (eye_dist < eye_dist_threshold):
+                continue
+
+            if self.template_3points:
+                landmark = np.array([[bbox[i], bbox[i + 1]] for i in range(5, 11, 2)])
+            else:
+                landmark = np.array([[bbox[i], bbox[i + 1]] for i in range(5, 15, 2)])
+            self.all_landmarks_5.append(landmark)
+            self.det_faces.append(bbox[0:5])
+        if len(self.det_faces) == 0:
+            return 0
+        if only_keep_largest:
+            h, w, _ = self.input_img.shape
+            self.det_faces, largest_idx = get_largest_face(self.det_faces, h, w)
+            self.all_landmarks_5 = [self.all_landmarks_5[largest_idx]]
+        elif only_center_face:
+            h, w, _ = self.input_img.shape
+            self.det_faces, center_idx = get_center_face(self.det_faces, h, w)
+            self.all_landmarks_5 = [self.all_landmarks_5[center_idx]]
+
+        # pad blurry images
+        if self.pad_blur:
+            self.pad_input_imgs = []
+            for landmarks in self.all_landmarks_5:
+                # get landmarks
+                eye_left = landmarks[0, :]
+                eye_right = landmarks[1, :]
+                eye_avg = (eye_left + eye_right) * 0.5
+                mouth_avg = (landmarks[3, :] + landmarks[4, :]) * 0.5
+                eye_to_eye = eye_right - eye_left
+                eye_to_mouth = mouth_avg - eye_avg
+
+                # Get the oriented crop rectangle
+                # x: half width of the oriented crop rectangle
+                x = eye_to_eye - np.flipud(eye_to_mouth) * [-1, 1]
+                #  - np.flipud(eye_to_mouth) * [-1, 1]: rotate 90 clockwise
+                # norm with the hypotenuse: get the direction
+                x /= np.hypot(*x)  # get the hypotenuse of a right triangle
+                rect_scale = 1.5
+                x *= max(np.hypot(*eye_to_eye) * 2.0 * rect_scale, np.hypot(*eye_to_mouth) * 1.8 * rect_scale)
+                # y: half height of the oriented crop rectangle
+                y = np.flipud(x) * [-1, 1]
+
+                # c: center
+                c = eye_avg + eye_to_mouth * 0.1
+                # quad: (left_top, left_bottom, right_bottom, right_top)
+                quad = np.stack([c - x - y, c - x + y, c + x + y, c + x - y])
+                # qsize: side length of the square
+                qsize = np.hypot(*x) * 2
+                border = max(int(np.rint(qsize * 0.1)), 3)
+
+                # get pad
+                # pad: (width_left, height_top, width_right, height_bottom)
+                pad = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))),
+                       int(np.ceil(max(quad[:, 1]))))
+                pad = [
+                    max(-pad[0] + border, 1),
+                    max(-pad[1] + border, 1),
+                    max(pad[2] - self.input_img.shape[0] + border, 1),
+                    max(pad[3] - self.input_img.shape[1] + border, 1)
+                ]
+
+                if max(pad) > 1:
+                    # pad image
+                    pad_img = np.pad(self.input_img, ((pad[1], pad[3]), (pad[0], pad[2]), (0, 0)), 'reflect')
+                    # modify landmark coords
+                    landmarks[:, 0] += pad[0]
+                    landmarks[:, 1] += pad[1]
+                    # blur pad images
+                    h, w, _ = pad_img.shape
+                    y, x, _ = np.ogrid[:h, :w, :1]
+                    mask = np.maximum(1.0 - np.minimum(np.float32(x) / pad[0],
+                                                       np.float32(w - 1 - x) / pad[2]),
+                                      1.0 - np.minimum(np.float32(y) / pad[1],
+                                                       np.float32(h - 1 - y) / pad[3]))
+                    blur = int(qsize * blur_ratio)
+                    if blur % 2 == 0:
+                        blur += 1
+                    blur_img = cv2.boxFilter(pad_img, 0, ksize=(blur, blur))
+                    # blur_img = cv2.GaussianBlur(pad_img, (blur, blur), 0)
+
+                    pad_img = pad_img.astype('float32')
+                    pad_img += (blur_img - pad_img) * np.clip(mask * 3.0 + 1.0, 0.0, 1.0)
+                    pad_img += (np.median(pad_img, axis=(0, 1)) - pad_img) * np.clip(mask, 0.0, 1.0)
+                    pad_img = np.clip(pad_img, 0, 255)  # float32, [0, 255]
+                    self.pad_input_imgs.append(pad_img)
+                else:
+                    self.pad_input_imgs.append(np.copy(self.input_img))
+
+        return len(self.all_landmarks_5)
+
+    def align_warp_face(self, save_cropped_path=None, border_mode='constant'):
+        """Align and warp faces with face template.
+        """
+        if self.pad_blur:
+            assert len(self.pad_input_imgs) == len(
+                self.all_landmarks_5), f'Mismatched samples: {len(self.pad_input_imgs)} and {len(self.all_landmarks_5)}'
+        for idx, landmark in enumerate(self.all_landmarks_5):
+            # use 5 landmarks to get affine matrix
+            # use cv2.LMEDS method for the equivalence to skimage transform
+            # ref: https://blog.csdn.net/yichxi/article/details/115827338
+            affine_matrix = cv2.estimateAffinePartial2D(landmark, self.face_template, method=cv2.LMEDS)[0]
+            self.affine_matrices.append(affine_matrix)
+            # warp and crop faces
+            if border_mode == 'constant':
+                border_mode = cv2.BORDER_CONSTANT
+            elif border_mode == 'reflect101':
+                border_mode = cv2.BORDER_REFLECT101
+            elif border_mode == 'reflect':
+                border_mode = cv2.BORDER_REFLECT
+            if self.pad_blur:
+                input_img = self.pad_input_imgs[idx]
+            else:
+                input_img = self.input_img
+            cropped_face = cv2.warpAffine(
+                input_img, affine_matrix, self.face_size, borderMode=border_mode, borderValue=(135, 133, 132))  # gray
+            self.cropped_faces.append(cropped_face)
+            # save the cropped face
+            if save_cropped_path is not None:
+                path = os.path.splitext(save_cropped_path)[0]
+                save_path = f'{path}_{idx:02d}.{self.save_ext}'
+                imwrite(cropped_face, save_path)
+
+    def get_inverse_affine(self, save_inverse_affine_path=None):
+        """Get inverse affine matrix."""
+        for idx, affine_matrix in enumerate(self.affine_matrices):
+            inverse_affine = cv2.invertAffineTransform(affine_matrix)
+            inverse_affine *= self.upscale_factor
+            self.inverse_affine_matrices.append(inverse_affine)
+            # save inverse affine matrices
+            if save_inverse_affine_path is not None:
+                path, _ = os.path.splitext(save_inverse_affine_path)
+                save_path = f'{path}_{idx:02d}.pth'
+                torch.save(inverse_affine, save_path)
+
+    def add_restored_face(self, face):
+        self.restored_faces.append(face)
+
+    def paste_faces_to_input_image(self, save_path=None, upsample_img=None):
+        h, w, _ = self.input_img.shape
+        h_up, w_up = int(h * self.upscale_factor), int(w * self.upscale_factor)
+
+        if upsample_img is None:
+            # simply resize the background
+            upsample_img = cv2.resize(self.input_img, (w_up, h_up), interpolation=cv2.INTER_LANCZOS4)
+        else:
+            upsample_img = cv2.resize(upsample_img, (w_up, h_up), interpolation=cv2.INTER_LANCZOS4)
+
+        assert len(self.restored_faces) == len(
+            self.inverse_affine_matrices), ('length of restored_faces and affine_matrices are different.')
+        for restored_face, inverse_affine in zip(self.restored_faces, self.inverse_affine_matrices):
+            # Add an offset to inverse affine matrix, for more precise back alignment
+            if self.upscale_factor > 1:
+                extra_offset = 0.5 * self.upscale_factor
+            else:
+                extra_offset = 0
+            inverse_affine[:, 2] += extra_offset
+            inv_restored = cv2.warpAffine(restored_face, inverse_affine, (w_up, h_up))
+
+            if self.use_parse:
+                # inference
+                face_input = cv2.resize(restored_face, (512, 512), interpolation=cv2.INTER_LINEAR)
+                face_input = img2tensor(face_input.astype('float32') / 255., bgr2rgb=True, float32=True)
+                normalize(face_input, (0.5, 0.5, 0.5), (0.5, 0.5, 0.5), inplace=True)
+                face_input = torch.unsqueeze(face_input, 0).to(self.device)
+                with torch.no_grad():
+                    out = self.face_parse(face_input)[0]
+                out = out.argmax(dim=1).squeeze().cpu().numpy()
+
+                mask = np.zeros(out.shape)
+                MASK_COLORMAP = [0, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 0, 255, 0, 0, 0]
+                for idx, color in enumerate(MASK_COLORMAP):
+                    mask[out == idx] = color
+                #  blur the mask
+                mask = cv2.GaussianBlur(mask, (101, 101), 11)
+                mask = cv2.GaussianBlur(mask, (101, 101), 11)
+                # remove the black borders
+                thres = 10
+                mask[:thres, :] = 0
+                mask[-thres:, :] = 0
+                mask[:, :thres] = 0
+                mask[:, -thres:] = 0
+                mask = mask / 255.
+
+                mask = cv2.resize(mask, restored_face.shape[:2])
+                mask = cv2.warpAffine(mask, inverse_affine, (w_up, h_up), flags=3)
+                inv_soft_mask = mask[:, :, None]
+                pasted_face = inv_restored
+
+            else:  # use square parse maps
+                mask = np.ones(self.face_size, dtype=np.float32)
+                inv_mask = cv2.warpAffine(mask, inverse_affine, (w_up, h_up))
+                # remove the black borders
+                inv_mask_erosion = cv2.erode(
+                    inv_mask, np.ones((int(2 * self.upscale_factor), int(2 * self.upscale_factor)), np.uint8))
+                pasted_face = inv_mask_erosion[:, :, None] * inv_restored
+                total_face_area = np.sum(inv_mask_erosion)  # // 3
+                # compute the fusion edge based on the area of face
+                w_edge = int(total_face_area**0.5) // 20
+                erosion_radius = w_edge * 2
+                inv_mask_center = cv2.erode(inv_mask_erosion, np.ones((erosion_radius, erosion_radius), np.uint8))
+                blur_size = w_edge * 2
+                inv_soft_mask = cv2.GaussianBlur(inv_mask_center, (blur_size + 1, blur_size + 1), 0)
+                if len(upsample_img.shape) == 2:  # upsample_img is gray image
+                    upsample_img = upsample_img[:, :, None]
+                inv_soft_mask = inv_soft_mask[:, :, None]
+
+            if len(upsample_img.shape) == 3 and upsample_img.shape[2] == 4:  # alpha channel
+                alpha = upsample_img[:, :, 3:]
+                upsample_img = inv_soft_mask * pasted_face + (1 - inv_soft_mask) * upsample_img[:, :, 0:3]
+                upsample_img = np.concatenate((upsample_img, alpha), axis=2)
+            else:
+                upsample_img = inv_soft_mask * pasted_face + (1 - inv_soft_mask) * upsample_img
+
+        if np.max(upsample_img) > 256:  # 16-bit image
+            upsample_img = upsample_img.astype(np.uint16)
+        else:
+            upsample_img = upsample_img.astype(np.uint8)
+        if save_path is not None:
+            path = os.path.splitext(save_path)[0]
+            save_path = f'{path}.{self.save_ext}'
+            imwrite(upsample_img, save_path)
+        return upsample_img
+
+    def clean_all(self):
+        self.all_landmarks_5 = []
+        self.restored_faces = []
+        self.affine_matrices = []
+        self.cropped_faces = []
+        self.inverse_affine_matrices = []
+        self.det_faces = []
+        self.pad_input_imgs = []

fooocus_extras/facexlib/utils/face_utils.py

@@ -0,0 +1,250 @@
+import cv2
+import numpy as np
+import torch
+
+
+def compute_increased_bbox(bbox, increase_area, preserve_aspect=True):
+    left, top, right, bot = bbox
+    width = right - left
+    height = bot - top
+
+    if preserve_aspect:
+        width_increase = max(increase_area, ((1 + 2 * increase_area) * height - width) / (2 * width))
+        height_increase = max(increase_area, ((1 + 2 * increase_area) * width - height) / (2 * height))
+    else:
+        width_increase = height_increase = increase_area
+    left = int(left - width_increase * width)
+    top = int(top - height_increase * height)
+    right = int(right + width_increase * width)
+    bot = int(bot + height_increase * height)
+    return (left, top, right, bot)
+
+
+def get_valid_bboxes(bboxes, h, w):
+    left = max(bboxes[0], 0)
+    top = max(bboxes[1], 0)
+    right = min(bboxes[2], w)
+    bottom = min(bboxes[3], h)
+    return (left, top, right, bottom)
+
+
+def align_crop_face_landmarks(img,
+                              landmarks,
+                              output_size,
+                              transform_size=None,
+                              enable_padding=True,
+                              return_inverse_affine=False,
+                              shrink_ratio=(1, 1)):
+    """Align and crop face with landmarks.
+
+    The output_size and transform_size are based on width. The height is
+    adjusted based on shrink_ratio_h/shring_ration_w.
+
+    Modified from:
+    https://github.com/NVlabs/ffhq-dataset/blob/master/download_ffhq.py
+
+    Args:
+        img (Numpy array): Input image.
+        landmarks (Numpy array): 5 or 68 or 98 landmarks.
+        output_size (int): Output face size.
+        transform_size (ing): Transform size. Usually the four time of
+            output_size.
+        enable_padding (float): Default: True.
+        shrink_ratio (float | tuple[float] | list[float]): Shring the whole
+            face for height and width (crop larger area). Default: (1, 1).
+
+    Returns:
+        (Numpy array): Cropped face.
+    """
+    lm_type = 'retinaface_5'  # Options: dlib_5, retinaface_5
+
+    if isinstance(shrink_ratio, (float, int)):
+        shrink_ratio = (shrink_ratio, shrink_ratio)
+    if transform_size is None:
+        transform_size = output_size * 4
+
+    # Parse landmarks
+    lm = np.array(landmarks)
+    if lm.shape[0] == 5 and lm_type == 'retinaface_5':
+        eye_left = lm[0]
+        eye_right = lm[1]
+        mouth_avg = (lm[3] + lm[4]) * 0.5
+    elif lm.shape[0] == 5 and lm_type == 'dlib_5':
+        lm_eye_left = lm[2:4]
+        lm_eye_right = lm[0:2]
+        eye_left = np.mean(lm_eye_left, axis=0)
+        eye_right = np.mean(lm_eye_right, axis=0)
+        mouth_avg = lm[4]
+    elif lm.shape[0] == 68:
+        lm_eye_left = lm[36:42]
+        lm_eye_right = lm[42:48]
+        eye_left = np.mean(lm_eye_left, axis=0)
+        eye_right = np.mean(lm_eye_right, axis=0)
+        mouth_avg = (lm[48] + lm[54]) * 0.5
+    elif lm.shape[0] == 98:
+        lm_eye_left = lm[60:68]
+        lm_eye_right = lm[68:76]
+        eye_left = np.mean(lm_eye_left, axis=0)
+        eye_right = np.mean(lm_eye_right, axis=0)
+        mouth_avg = (lm[76] + lm[82]) * 0.5
+
+    eye_avg = (eye_left + eye_right) * 0.5
+    eye_to_eye = eye_right - eye_left
+    eye_to_mouth = mouth_avg - eye_avg
+
+    # Get the oriented crop rectangle
+    # x: half width of the oriented crop rectangle
+    x = eye_to_eye - np.flipud(eye_to_mouth) * [-1, 1]
+    #  - np.flipud(eye_to_mouth) * [-1, 1]: rotate 90 clockwise
+    # norm with the hypotenuse: get the direction
+    x /= np.hypot(*x)  # get the hypotenuse of a right triangle
+    rect_scale = 1  # TODO: you can edit it to get larger rect
+    x *= max(np.hypot(*eye_to_eye) * 2.0 * rect_scale, np.hypot(*eye_to_mouth) * 1.8 * rect_scale)
+    # y: half height of the oriented crop rectangle
+    y = np.flipud(x) * [-1, 1]
+
+    x *= shrink_ratio[1]  # width
+    y *= shrink_ratio[0]  # height
+
+    # c: center
+    c = eye_avg + eye_to_mouth * 0.1
+    # quad: (left_top, left_bottom, right_bottom, right_top)
+    quad = np.stack([c - x - y, c - x + y, c + x + y, c + x - y])
+    # qsize: side length of the square
+    qsize = np.hypot(*x) * 2
+
+    quad_ori = np.copy(quad)
+    # Shrink, for large face
+    # TODO: do we really need shrink
+    shrink = int(np.floor(qsize / output_size * 0.5))
+    if shrink > 1:
+        h, w = img.shape[0:2]
+        rsize = (int(np.rint(float(w) / shrink)), int(np.rint(float(h) / shrink)))
+        img = cv2.resize(img, rsize, interpolation=cv2.INTER_AREA)
+        quad /= shrink
+        qsize /= shrink
+
+    # Crop
+    h, w = img.shape[0:2]
+    border = max(int(np.rint(qsize * 0.1)), 3)
+    crop = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))),
+            int(np.ceil(max(quad[:, 1]))))
+    crop = (max(crop[0] - border, 0), max(crop[1] - border, 0), min(crop[2] + border, w), min(crop[3] + border, h))
+    if crop[2] - crop[0] < w or crop[3] - crop[1] < h:
+        img = img[crop[1]:crop[3], crop[0]:crop[2], :]
+        quad -= crop[0:2]
+
+    # Pad
+    # pad: (width_left, height_top, width_right, height_bottom)
+    h, w = img.shape[0:2]
+    pad = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))),
+           int(np.ceil(max(quad[:, 1]))))
+    pad = (max(-pad[0] + border, 0), max(-pad[1] + border, 0), max(pad[2] - w + border, 0), max(pad[3] - h + border, 0))
+    if enable_padding and max(pad) > border - 4:
+        pad = np.maximum(pad, int(np.rint(qsize * 0.3)))
+        img = np.pad(img, ((pad[1], pad[3]), (pad[0], pad[2]), (0, 0)), 'reflect')
+        h, w = img.shape[0:2]
+        y, x, _ = np.ogrid[:h, :w, :1]
+        mask = np.maximum(1.0 - np.minimum(np.float32(x) / pad[0],
+                                           np.float32(w - 1 - x) / pad[2]),
+                          1.0 - np.minimum(np.float32(y) / pad[1],
+                                           np.float32(h - 1 - y) / pad[3]))
+        blur = int(qsize * 0.02)
+        if blur % 2 == 0:
+            blur += 1
+        blur_img = cv2.boxFilter(img, 0, ksize=(blur, blur))
+
+        img = img.astype('float32')
+        img += (blur_img - img) * np.clip(mask * 3.0 + 1.0, 0.0, 1.0)
+        img += (np.median(img, axis=(0, 1)) - img) * np.clip(mask, 0.0, 1.0)
+        img = np.clip(img, 0, 255)  # float32, [0, 255]
+        quad += pad[:2]
+
+    # Transform use cv2
+    h_ratio = shrink_ratio[0] / shrink_ratio[1]
+    dst_h, dst_w = int(transform_size * h_ratio), transform_size
+    template = np.array([[0, 0], [0, dst_h], [dst_w, dst_h], [dst_w, 0]])
+    # use cv2.LMEDS method for the equivalence to skimage transform
+    # ref: https://blog.csdn.net/yichxi/article/details/115827338
+    affine_matrix = cv2.estimateAffinePartial2D(quad, template, method=cv2.LMEDS)[0]
+    cropped_face = cv2.warpAffine(
+        img, affine_matrix, (dst_w, dst_h), borderMode=cv2.BORDER_CONSTANT, borderValue=(135, 133, 132))  # gray
+
+    if output_size < transform_size:
+        cropped_face = cv2.resize(
+            cropped_face, (output_size, int(output_size * h_ratio)), interpolation=cv2.INTER_LINEAR)
+
+    if return_inverse_affine:
+        dst_h, dst_w = int(output_size * h_ratio), output_size
+        template = np.array([[0, 0], [0, dst_h], [dst_w, dst_h], [dst_w, 0]])
+        # use cv2.LMEDS method for the equivalence to skimage transform
+        # ref: https://blog.csdn.net/yichxi/article/details/115827338
+        affine_matrix = cv2.estimateAffinePartial2D(
+            quad_ori, np.array([[0, 0], [0, output_size], [dst_w, dst_h], [dst_w, 0]]), method=cv2.LMEDS)[0]
+        inverse_affine = cv2.invertAffineTransform(affine_matrix)
+    else:
+        inverse_affine = None
+    return cropped_face, inverse_affine
+
+
+def paste_face_back(img, face, inverse_affine):
+    h, w = img.shape[0:2]
+    face_h, face_w = face.shape[0:2]
+    inv_restored = cv2.warpAffine(face, inverse_affine, (w, h))
+    mask = np.ones((face_h, face_w, 3), dtype=np.float32)
+    inv_mask = cv2.warpAffine(mask, inverse_affine, (w, h))
+    # remove the black borders
+    inv_mask_erosion = cv2.erode(inv_mask, np.ones((2, 2), np.uint8))
+    inv_restored_remove_border = inv_mask_erosion * inv_restored
+    total_face_area = np.sum(inv_mask_erosion) // 3
+    # compute the fusion edge based on the area of face
+    w_edge = int(total_face_area**0.5) // 20
+    erosion_radius = w_edge * 2
+    inv_mask_center = cv2.erode(inv_mask_erosion, np.ones((erosion_radius, erosion_radius), np.uint8))
+    blur_size = w_edge * 2
+    inv_soft_mask = cv2.GaussianBlur(inv_mask_center, (blur_size + 1, blur_size + 1), 0)
+    img = inv_soft_mask * inv_restored_remove_border + (1 - inv_soft_mask) * img
+    # float32, [0, 255]
+    return img
+
+
+if __name__ == '__main__':
+    import os
+
+    from fooocus_extras.facexlib.detection import init_detection_model
+    from fooocus_extras.facexlib.utils.face_restoration_helper import get_largest_face
+    from fooocus_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]
+
+    # initialize model
+    det_net = init_detection_model('retinaface_resnet50', half=False)
+    img_ori = cv2.imread(img_path)
+    h, w = img_ori.shape[0:2]
+    # if larger than 800, scale it
+    scale = max(h / 800, w / 800)
+    if scale > 1:
+        img = cv2.resize(img_ori, (int(w / scale), int(h / scale)), interpolation=cv2.INTER_LINEAR)
+
+    with torch.no_grad():
+        bboxes = det_net.detect_faces(img, 0.97)
+    if scale > 1:
+        bboxes *= scale  # the score is incorrect
+    bboxes = get_largest_face(bboxes, h, w)[0]
+    visualize_detection(img_ori, [bboxes], f'tmp/{img_name}_det.png')
+
+    landmarks = np.array([[bboxes[i], bboxes[i + 1]] for i in range(5, 15, 2)])
+
+    cropped_face, inverse_affine = align_crop_face_landmarks(
+        img_ori,
+        landmarks,
+        output_size=512,
+        transform_size=None,
+        enable_padding=True,
+        return_inverse_affine=True,
+        shrink_ratio=(1, 1))
+
+    cv2.imwrite(f'tmp/{img_name}_cropeed_face.png', cropped_face)
+    img = paste_face_back(img_ori, cropped_face, inverse_affine)
+    cv2.imwrite(f'tmp/{img_name}_back.png', img)

fooocus_extras/facexlib/utils/misc.py

@@ -0,0 +1,118 @@
+import cv2
+import os
+import os.path as osp
+import torch
+from torch.hub import download_url_to_file, get_dir
+from urllib.parse import urlparse
+
+ROOT_DIR = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+
+def imwrite(img, file_path, params=None, auto_mkdir=True):
+    """Write image to file.
+
+    Args:
+        img (ndarray): Image array to be written.
+        file_path (str): Image file path.
+        params (None or list): Same as opencv's :func:`imwrite` interface.
+        auto_mkdir (bool): If the parent folder of `file_path` does not exist,
+            whether to create it automatically.
+
+    Returns:
+        bool: Successful or not.
+    """
+    if auto_mkdir:
+        dir_name = os.path.abspath(os.path.dirname(file_path))
+        os.makedirs(dir_name, exist_ok=True)
+    return cv2.imwrite(file_path, img, params)
+
+
+def img2tensor(imgs, bgr2rgb=True, float32=True):
+    """Numpy array to tensor.
+
+    Args:
+        imgs (list[ndarray] | ndarray): Input images.
+        bgr2rgb (bool): Whether to change bgr to rgb.
+        float32 (bool): Whether to change to float32.
+
+    Returns:
+        list[tensor] | tensor: Tensor images. If returned results only have
+            one element, just return tensor.
+    """
+
+    def _totensor(img, bgr2rgb, float32):
+        if img.shape[2] == 3 and bgr2rgb:
+            if img.dtype == 'float64':
+                img = img.astype('float32')
+            img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+        img = torch.from_numpy(img.transpose(2, 0, 1))
+        if float32:
+            img = img.float()
+        return img
+
+    if isinstance(imgs, list):
+        return [_totensor(img, bgr2rgb, float32) for img in imgs]
+    else:
+        return _totensor(imgs, bgr2rgb, float32)
+
+
+def load_file_from_url(url, model_dir=None, progress=True, file_name=None, save_dir=None):
+    """Ref:https://github.com/1adrianb/face-alignment/blob/master/face_alignment/utils.py
+    """
+    if model_dir is None:
+        hub_dir = get_dir()
+        model_dir = os.path.join(hub_dir, 'checkpoints')
+
+    if save_dir is None:
+        save_dir = os.path.join(ROOT_DIR, model_dir)
+    os.makedirs(save_dir, exist_ok=True)
+
+    parts = urlparse(url)
+    filename = os.path.basename(parts.path)
+    if file_name is not None:
+        filename = file_name
+    cached_file = os.path.abspath(os.path.join(save_dir, filename))
+    if not os.path.exists(cached_file):
+        print(f'Downloading: "{url}" to {cached_file}\n')
+        download_url_to_file(url, cached_file, hash_prefix=None, progress=progress)
+    return cached_file
+
+
+def scandir(dir_path, suffix=None, recursive=False, full_path=False):
+    """Scan a directory to find the interested files.
+    Args:
+        dir_path (str): Path of the directory.
+        suffix (str | tuple(str), optional): File suffix that we are
+            interested in. Default: None.
+        recursive (bool, optional): If set to True, recursively scan the
+            directory. Default: False.
+        full_path (bool, optional): If set to True, include the dir_path.
+            Default: False.
+    Returns:
+        A generator for all the interested files with relative paths.
+    """
+
+    if (suffix is not None) and not isinstance(suffix, (str, tuple)):
+        raise TypeError('"suffix" must be a string or tuple of strings')
+
+    root = dir_path
+
+    def _scandir(dir_path, suffix, recursive):
+        for entry in os.scandir(dir_path):
+            if not entry.name.startswith('.') and entry.is_file():
+                if full_path:
+                    return_path = entry.path
+                else:
+                    return_path = osp.relpath(entry.path, root)
+
+                if suffix is None:
+                    yield return_path
+                elif return_path.endswith(suffix):
+                    yield return_path
+            else:
+                if recursive:
+                    yield from _scandir(entry.path, suffix=suffix, recursive=recursive)
+                else:
+                    continue
+
+    return _scandir(dir_path, suffix=suffix, recursive=recursive)

fooocus_extras/ip_adapter.py

@@ -84,27 +84,21 @@ class IPAdapterModel(torch.nn.Module):
 
 clip_vision: fcbh.clip_vision.ClipVisionModel = None
 ip_negative: torch.Tensor = None
-image_proj_model: ModelPatcher = None
-ip_layers: ModelPatcher = None
-ip_adapter: IPAdapterModel = None
-ip_unconds = None
+ip_adapters: dict = {}
 
 
 def load_ip_adapter(clip_vision_path, ip_negative_path, ip_adapter_path):
-    global clip_vision, image_proj_model, ip_layers, ip_negative, ip_adapter, ip_unconds
+    global clip_vision, ip_negative, ip_adapters
 
-    if clip_vision_path is None:
-        return
-    if ip_negative_path is None:
-        return
-    if ip_adapter_path is None:
-        return
-    if clip_vision is not None and image_proj_model is not None and ip_layers is not None and ip_negative is not None:
-        return
-
-    ip_negative = sf.load_file(ip_negative_path)['data']
+    if clip_vision is None and isinstance(clip_vision_path, str):
         clip_vision = fcbh.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']
+
+    if not isinstance(ip_adapter_path, str) or ip_adapter_path in ip_adapters:
+        return
+
     load_device = model_management.get_torch_device()
     offload_device = torch.device('cpu')
 
@@ -141,7 +135,13 @@ def load_ip_adapter(clip_vision_path, ip_negative_path, ip_adapter_path):
     ip_layers = ModelPatcher(model=ip_adapter.ip_layers, load_device=load_device,
                              offload_device=offload_device)
 
-    ip_unconds = None
+    ip_adapters[ip_adapter_path] = dict(
+        ip_adapter=ip_adapter,
+        image_proj_model=image_proj_model,
+        ip_layers=ip_layers,
+        ip_unconds=None
+    )
+
     return
 
 
@@ -161,8 +161,9 @@ def clip_preprocess(image):
 
 @torch.no_grad()
 @torch.inference_mode()
-def preprocess(img):
-    global ip_unconds
+def preprocess(img, ip_adapter_path):
+    global ip_adapters
+    entry = ip_adapters[ip_adapter_path]
 
     fcbh.model_management.load_model_gpu(clip_vision.patcher)
     pixel_values = clip_preprocess(numpy_to_pytorch(img).to(clip_vision.load_device))
@@ -175,6 +176,11 @@ def preprocess(img):
     with precision_scope(fcbh.model_management.get_autocast_device(clip_vision.load_device), torch.float32):
         outputs = clip_vision.model(pixel_values=pixel_values, output_hidden_states=True)
 
+    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.hidden_states[-2]
     else:
@@ -190,9 +196,11 @@ def preprocess(img):
     if ip_unconds is None:
         uncond = ip_negative.to(device=ip_adapter.load_device, dtype=ip_adapter.dtype)
         ip_unconds = [m(uncond).cpu() for m in ip_layers.model.to_kvs]
+        entry['ip_unconds'] = ip_unconds
 
     ip_conds = [m(cond).cpu() for m in ip_layers.model.to_kvs]
-    return ip_conds
+
+    return ip_conds, ip_unconds
 
 
 @torch.no_grad()
@@ -206,18 +214,17 @@ def patch_model(model, tasks):
             current_step = float(model.model.diffusion_model.current_step.detach().cpu().numpy()[0])
             cond_or_uncond = extra_options['cond_or_uncond']
 
-            with torch.autocast("cuda", dtype=ip_adapter.dtype):
             q = n
             k = [context_attn2]
             v = [value_attn2]
             b, _, _ = q.shape
 
-                for ip_conds, cn_stop, cn_weight in tasks:
+            for (cs, ucs), cn_stop, cn_weight in tasks:
                 if current_step < cn_stop:
-                        ip_k_c = ip_conds[ip_index * 2].to(q)
-                        ip_v_c = ip_conds[ip_index * 2 + 1].to(q)
-                        ip_k_uc = ip_unconds[ip_index * 2].to(q)
-                        ip_v_uc = ip_unconds[ip_index * 2 + 1].to(q)
+                    ip_k_c = cs[ip_index * 2].to(q)
+                    ip_v_c = cs[ip_index * 2 + 1].to(q)
+                    ip_k_uc = ucs[ip_index * 2].to(q)
+                    ip_v_uc = ucs[ip_index * 2 + 1].to(q)
 
                     ip_k = torch.cat([(ip_k_c, ip_k_uc)[i] for i in cond_or_uncond], dim=0)
                     ip_v = torch.cat([(ip_v_c, ip_v_uc)[i] for i in cond_or_uncond], dim=0)
@@ -247,6 +254,7 @@ def patch_model(model, tasks):
             v = torch.cat(v, dim=1)
             out = sdp(q, k, v, extra_options)
 
+
             return out.to(dtype=org_dtype)
         return patcher
 
@@ -260,25 +268,19 @@ def patch_model(model, tasks):
             to["patches_replace"]["attn2"][key] = make_attn_patcher(number)
 
     number = 0
-    if not ip_adapter.sdxl:
-        for id in [1, 2, 4, 5, 7, 8]:  # id of input_blocks that have cross attention
-            set_model_patch_replace(new_model, number, ("input", id))
-            number += 1
-        for id in [3, 4, 5, 6, 7, 8, 9, 10, 11]:  # id of output_blocks that have cross attention
-            set_model_patch_replace(new_model, number, ("output", id))
-            number += 1
-        set_model_patch_replace(new_model, number, ("middle", 0))
-    else:
-        for id in [4, 5, 7, 8]:  # id of input_blocks that have cross attention
-            block_indices = range(2) if id in [4, 5] else range(10)  # transformer_depth
+
+    for id in [4, 5, 7, 8]:
+        block_indices = range(2) if id in [4, 5] else range(10)
         for index in block_indices:
             set_model_patch_replace(new_model, number, ("input", id, index))
             number += 1
-        for id in range(6):  # id of output_blocks that have cross attention
-            block_indices = range(2) if id in [3, 4, 5] else range(10)  # transformer_depth
+
+    for id in range(6):
+        block_indices = range(2) if id in [3, 4, 5] else range(10)
         for index in block_indices:
             set_model_patch_replace(new_model, number, ("output", id, index))
             number += 1
+
     for index in range(10):
         set_model_patch_replace(new_model, number, ("middle", 0, index))
         number += 1

fooocus_version.py

@@ -1 +1 @@
-version = '2.1.786'
+version = '2.1.824'

javascript/imageviewer.js

@@ -0,0 +1,260 @@
+// From A1111
+
+function closeModal() {
+    gradioApp().getElementById("lightboxModal").style.display = "none";
+}
+
+function showModal(event) {
+    const source = event.target || event.srcElement;
+    const modalImage = gradioApp().getElementById("modalImage");
+    const lb = gradioApp().getElementById("lightboxModal");
+    modalImage.src = source.src;
+    if (modalImage.style.display === 'none') {
+        lb.style.setProperty('background-image', 'url(' + source.src + ')');
+    }
+    lb.style.display = "flex";
+    lb.focus();
+
+    event.stopPropagation();
+}
+
+function negmod(n, m) {
+    return ((n % m) + m) % m;
+}
+
+function updateOnBackgroundChange() {
+    const modalImage = gradioApp().getElementById("modalImage");
+    if (modalImage && modalImage.offsetParent) {
+        let currentButton = selected_gallery_button();
+
+        if (currentButton?.children?.length > 0 && modalImage.src != currentButton.children[0].src) {
+            modalImage.src = currentButton.children[0].src;
+            if (modalImage.style.display === 'none') {
+                const modal = gradioApp().getElementById("lightboxModal");
+                modal.style.setProperty('background-image', `url(${modalImage.src})`);
+            }
+        }
+    }
+}
+
+function all_gallery_buttons() {
+    var allGalleryButtons = gradioApp().querySelectorAll('.image_gallery .thumbnails > .thumbnail-item.thumbnail-small');
+    var visibleGalleryButtons = [];
+    allGalleryButtons.forEach(function(elem) {
+        if (elem.parentElement.offsetParent) {
+            visibleGalleryButtons.push(elem);
+        }
+    });
+    return visibleGalleryButtons;
+}
+
+function selected_gallery_button() {
+    return all_gallery_buttons().find(elem => elem.classList.contains('selected')) ?? null;
+}
+
+function selected_gallery_index() {
+    return all_gallery_buttons().findIndex(elem => elem.classList.contains('selected'));
+}
+
+function modalImageSwitch(offset) {
+    var galleryButtons = all_gallery_buttons();
+
+    if (galleryButtons.length > 1) {
+        var currentButton = selected_gallery_button();
+
+        var result = -1;
+        galleryButtons.forEach(function(v, i) {
+            if (v == currentButton) {
+                result = i;
+            }
+        });
+
+        if (result != -1) {
+            var nextButton = galleryButtons[negmod((result + offset), galleryButtons.length)];
+            nextButton.click();
+            const modalImage = gradioApp().getElementById("modalImage");
+            const modal = gradioApp().getElementById("lightboxModal");
+            modalImage.src = nextButton.children[0].src;
+            if (modalImage.style.display === 'none') {
+                modal.style.setProperty('background-image', `url(${modalImage.src})`);
+            }
+            setTimeout(function() {
+                modal.focus();
+            }, 10);
+        }
+    }
+}
+
+function saveImage() {
+
+}
+
+function modalSaveImage(event) {
+    event.stopPropagation();
+}
+
+function modalNextImage(event) {
+    modalImageSwitch(1);
+    event.stopPropagation();
+}
+
+function modalPrevImage(event) {
+    modalImageSwitch(-1);
+    event.stopPropagation();
+}
+
+function modalKeyHandler(event) {
+    switch (event.key) {
+    case "s":
+        saveImage();
+        break;
+    case "ArrowLeft":
+        modalPrevImage(event);
+        break;
+    case "ArrowRight":
+        modalNextImage(event);
+        break;
+    case "Escape":
+        closeModal();
+        break;
+    }
+}
+
+function setupImageForLightbox(e) {
+    if (e.dataset.modded) {
+        return;
+    }
+
+    e.dataset.modded = true;
+    e.style.cursor = 'pointer';
+    e.style.userSelect = 'none';
+
+    var isFirefox = navigator.userAgent.toLowerCase().indexOf('firefox') > -1;
+
+    // For Firefox, listening on click first switched to next image then shows the lightbox.
+    // If you know how to fix this without switching to mousedown event, please.
+    // For other browsers the event is click to make it possiblr to drag picture.
+    var event = isFirefox ? 'mousedown' : 'click';
+
+    e.addEventListener(event, function(evt) {
+        if (evt.button == 1) {
+            open(evt.target.src);
+            evt.preventDefault();
+            return;
+        }
+        if (evt.button != 0) return;
+
+        modalZoomSet(gradioApp().getElementById('modalImage'), true);
+        evt.preventDefault();
+        showModal(evt);
+    }, true);
+
+}
+
+function modalZoomSet(modalImage, enable) {
+    if (modalImage) modalImage.classList.toggle('modalImageFullscreen', !!enable);
+}
+
+function modalZoomToggle(event) {
+    var modalImage = gradioApp().getElementById("modalImage");
+    modalZoomSet(modalImage, !modalImage.classList.contains('modalImageFullscreen'));
+    event.stopPropagation();
+}
+
+function modalTileImageToggle(event) {
+    const modalImage = gradioApp().getElementById("modalImage");
+    const modal = gradioApp().getElementById("lightboxModal");
+    const isTiling = modalImage.style.display === 'none';
+    if (isTiling) {
+        modalImage.style.display = 'block';
+        modal.style.setProperty('background-image', 'none');
+    } else {
+        modalImage.style.display = 'none';
+        modal.style.setProperty('background-image', `url(${modalImage.src})`);
+    }
+
+    event.stopPropagation();
+}
+
+onAfterUiUpdate(function() {
+    var fullImg_preview = gradioApp().querySelectorAll('.image_gallery > div > img');
+    if (fullImg_preview != null) {
+        fullImg_preview.forEach(setupImageForLightbox);
+    }
+    updateOnBackgroundChange();
+});
+
+document.addEventListener("DOMContentLoaded", function() {
+    //const modalFragment = document.createDocumentFragment();
+    const modal = document.createElement('div');
+    modal.onclick = closeModal;
+    modal.id = "lightboxModal";
+    modal.tabIndex = 0;
+    modal.addEventListener('keydown', modalKeyHandler, true);
+
+    const modalControls = document.createElement('div');
+    modalControls.className = 'modalControls gradio-container';
+    modal.append(modalControls);
+
+    const modalZoom = document.createElement('span');
+    modalZoom.className = 'modalZoom cursor';
+    modalZoom.innerHTML = '⤡';
+    modalZoom.addEventListener('click', modalZoomToggle, true);
+    modalZoom.title = "Toggle zoomed view";
+    modalControls.appendChild(modalZoom);
+
+    // const modalTileImage = document.createElement('span');
+    // modalTileImage.className = 'modalTileImage cursor';
+    // modalTileImage.innerHTML = '⊞';
+    // modalTileImage.addEventListener('click', modalTileImageToggle, true);
+    // modalTileImage.title = "Preview tiling";
+    // modalControls.appendChild(modalTileImage);
+    //
+    // const modalSave = document.createElement("span");
+    // modalSave.className = "modalSave cursor";
+    // modalSave.id = "modal_save";
+    // modalSave.innerHTML = "🖫";
+    // modalSave.addEventListener("click", modalSaveImage, true);
+    // modalSave.title = "Save Image(s)";
+    // modalControls.appendChild(modalSave);
+
+    const modalClose = document.createElement('span');
+    modalClose.className = 'modalClose cursor';
+    modalClose.innerHTML = '×';
+    modalClose.onclick = closeModal;
+    modalClose.title = "Close image viewer";
+    modalControls.appendChild(modalClose);
+
+    const modalImage = document.createElement('img');
+    modalImage.id = 'modalImage';
+    modalImage.onclick = closeModal;
+    modalImage.tabIndex = 0;
+    modalImage.addEventListener('keydown', modalKeyHandler, true);
+    modal.appendChild(modalImage);
+
+    const modalPrev = document.createElement('a');
+    modalPrev.className = 'modalPrev';
+    modalPrev.innerHTML = '❮';
+    modalPrev.tabIndex = 0;
+    modalPrev.addEventListener('click', modalPrevImage, true);
+    modalPrev.addEventListener('keydown', modalKeyHandler, true);
+    modal.appendChild(modalPrev);
+
+    const modalNext = document.createElement('a');
+    modalNext.className = 'modalNext';
+    modalNext.innerHTML = '❯';
+    modalNext.tabIndex = 0;
+    modalNext.addEventListener('click', modalNextImage, true);
+    modalNext.addEventListener('keydown', modalKeyHandler, true);
+
+    modal.appendChild(modalNext);
+
+    try {
+        gradioApp().appendChild(modal);
+    } catch (e) {
+        gradioApp().body.appendChild(modal);
+    }
+
+    document.body.appendChild(modal);
+
+});

javascript/localization.js

@@ -73,6 +73,10 @@ function processNode(node) {
     });
 }
 
+function refresh_style_localization() {
+    processNode(document.querySelector('.style_selections'));
+}
+
 function localizeWholePage() {
     processNode(gradioApp());
 

javascript/script.js

@@ -166,3 +166,10 @@ function uiElementInSight(el) {
 function playNotification() {
     gradioApp().querySelector('#audio_notification audio')?.play();
 }
+
+function set_theme(theme) {
+    var gradioURL = window.location.href;
+    if (!gradioURL.includes('?__theme=')) {
+        window.location.replace(gradioURL + '?__theme=' + theme);
+    }
+}

javascript/viewer.js

@@ -0,0 +1,88 @@
+window.main_viewer_height = 512;
+
+function refresh_grid() {
+    let gridContainer = document.querySelector('#final_gallery .grid-container');
+    let final_gallery = document.getElementById('final_gallery');
+
+    if (gridContainer) if (final_gallery) {
+        let rect = final_gallery.getBoundingClientRect();
+        let cols = Math.ceil((rect.width - 16.0) / rect.height);
+        if (cols < 2) cols = 2;
+        gridContainer.style.setProperty('--grid-cols', cols);
+    }
+}
+
+function refresh_grid_delayed() {
+    refresh_grid();
+    setTimeout(refresh_grid, 100);
+    setTimeout(refresh_grid, 500);
+    setTimeout(refresh_grid, 1000);
+}
+
+function resized() {
+    let windowHeight = window.innerHeight - 260;
+    let elements = document.getElementsByClassName('main_view');
+
+    if (windowHeight > 745) windowHeight = 745;
+
+    for (let i = 0; i < elements.length; i++) {
+        elements[i].style.height = windowHeight + 'px';
+    }
+
+    window.main_viewer_height = windowHeight;
+
+    refresh_grid();
+}
+
+function viewer_to_top(delay = 100) {
+    setTimeout(() => window.scrollTo({top: 0, behavior: 'smooth'}), delay);
+}
+
+function viewer_to_bottom(delay = 100) {
+    let element = document.getElementById('positive_prompt');
+    let yPos = window.main_viewer_height;
+
+    if (element) {
+        yPos = element.getBoundingClientRect().top + window.scrollY;
+    }
+
+    setTimeout(() => window.scrollTo({top: yPos - 8, behavior: 'smooth'}), delay);
+}
+
+window.addEventListener('resize', (e) => {
+    resized();
+});
+
+onUiLoaded(async () => {
+    resized();
+});
+
+function on_style_selection_blur() {
+    let target = document.querySelector("#gradio_receiver_style_selections textarea");
+    target.value = "on_style_selection_blur " + Math.random();
+    let e = new Event("input", {bubbles: true})
+    Object.defineProperty(e, "target", {value: target})
+    target.dispatchEvent(e);
+}
+
+onUiLoaded(async () => {
+    let spans = document.querySelectorAll('.aspect_ratios span');
+
+    spans.forEach(function (span) {
+        span.innerHTML = span.innerHTML.replace(/&lt;/g, '<').replace(/&gt;/g, '>');
+    });
+
+    document.querySelector('.style_selections').addEventListener('focusout', function (event) {
+        setTimeout(() => {
+            if (!this.contains(document.activeElement)) {
+                on_style_selection_blur();
+            }
+        }, 200);
+    });
+
+    let inputs = document.querySelectorAll('.lora_weight input[type="range"]');
+
+    inputs.forEach(function (input) {
+        input.style.marginTop = '12px';
+    });
+});

language/en.json

@@ -42,32 +42,7 @@
     "Speed": "Speed",
     "Quality": "Quality",
     "Aspect Ratios": "Aspect Ratios",
-    "896\u00d71152": "896\u00d71152",
     "width \u00d7 height": "width \u00d7 height",
-    "704\u00d71408": "704\u00d71408",
-    "704\u00d71344": "704\u00d71344",
-    "768\u00d71344": "768\u00d71344",
-    "768\u00d71280": "768\u00d71280",
-    "832\u00d71216": "832\u00d71216",
-    "832\u00d71152": "832\u00d71152",
-    "896\u00d71088": "896\u00d71088",
-    "960\u00d71088": "960\u00d71088",
-    "960\u00d71024": "960\u00d71024",
-    "1024\u00d71024": "1024\u00d71024",
-    "1024\u00d7960": "1024\u00d7960",
-    "1088\u00d7960": "1088\u00d7960",
-    "1088\u00d7896": "1088\u00d7896",
-    "1152\u00d7832": "1152\u00d7832",
-    "1216\u00d7832": "1216\u00d7832",
-    "1280\u00d7768": "1280\u00d7768",
-    "1344\u00d7768": "1344\u00d7768",
-    "1344\u00d7704": "1344\u00d7704",
-    "1408\u00d7704": "1408\u00d7704",
-    "1472\u00d7704": "1472\u00d7704",
-    "1536\u00d7640": "1536\u00d7640",
-    "1600\u00d7640": "1600\u00d7640",
-    "1664\u00d7576": "1664\u00d7576",
-    "1728\u00d7576": "1728\u00d7576",
     "Image Number": "Image Number",
     "Negative Prompt": "Negative Prompt",
     "Describing what you do not want to see.": "Describing what you do not want to see.",
@@ -385,5 +360,13 @@
     "B1": "B1",
     "B2": "B2",
     "S1": "S1",
-    "S2": "S2"
+    "S2": "S2",
+    "Extreme Speed": "Extreme Speed",
+    "\uD83D\uDD0E Type here to search styles ...": "\uD83D\uDD0E Type here to search styles ...",
+    "Type prompt here.": "Type prompt here.",
+    "Outpaint Expansion Direction:": "Outpaint Expansion Direction:",
+    "* Powered by Fooocus Inpaint Engine (beta)": "* Powered by Fooocus Inpaint Engine (beta)",
+    "Fooocus Enhance": "Fooocus Enhance",
+    "Fooocus Cinematic": "Fooocus Cinematic",
+    "Fooocus Sharp": "Fooocus Sharp"
 }

launch.py

@@ -10,6 +10,7 @@ sys.path += [root, backend_path]
 
 os.chdir(root)
 os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
+os.environ["GRADIO_SERVER_PORT"] = "7865"
 
 
 import platform

modules/advanced_parameters.py

@@ -1,27 +1,30 @@
-adm_scaler_positive, adm_scaler_negative, adm_scaler_end, adaptive_cfg, sampler_name,  \
+disable_preview, adm_scaler_positive, adm_scaler_negative, adm_scaler_end, adaptive_cfg, sampler_name,  \
     scheduler_name, generate_image_grid, overwrite_step, overwrite_switch, overwrite_width, overwrite_height, \
     overwrite_vary_strength, overwrite_upscale_strength, \
     mixing_image_prompt_and_vary_upscale, mixing_image_prompt_and_inpaint, \
-    debugging_cn_preprocessor, controlnet_softness, canny_low_threshold, canny_high_threshold, inpaint_engine, \
+    debugging_cn_preprocessor, skipping_cn_preprocessor, controlnet_softness, canny_low_threshold, canny_high_threshold, \
     refiner_swap_method, \
-    freeu_enabled, freeu_b1, freeu_b2, freeu_s1, freeu_s2 = [None] * 26
+    freeu_enabled, freeu_b1, freeu_b2, freeu_s1, freeu_s2, \
+    debugging_inpaint_preprocessor, inpaint_disable_initial_latent, inpaint_engine, inpaint_strength, inpaint_respective_field = [None] * 32
 
 
 def set_all_advanced_parameters(*args):
-    global adm_scaler_positive, adm_scaler_negative, adm_scaler_end, adaptive_cfg, sampler_name, \
+    global disable_preview, adm_scaler_positive, adm_scaler_negative, adm_scaler_end, adaptive_cfg, sampler_name, \
         scheduler_name, generate_image_grid, overwrite_step, overwrite_switch, overwrite_width, overwrite_height, \
         overwrite_vary_strength, overwrite_upscale_strength, \
         mixing_image_prompt_and_vary_upscale, mixing_image_prompt_and_inpaint, \
-        debugging_cn_preprocessor, controlnet_softness, canny_low_threshold, canny_high_threshold, inpaint_engine, \
+        debugging_cn_preprocessor, skipping_cn_preprocessor, controlnet_softness, canny_low_threshold, canny_high_threshold, \
         refiner_swap_method, \
-        freeu_enabled, freeu_b1, freeu_b2, freeu_s1, freeu_s2
+        freeu_enabled, freeu_b1, freeu_b2, freeu_s1, freeu_s2, \
+        debugging_inpaint_preprocessor, inpaint_disable_initial_latent, inpaint_engine, inpaint_strength, inpaint_respective_field
 
-    adm_scaler_positive, adm_scaler_negative, adm_scaler_end, adaptive_cfg, sampler_name, \
+    disable_preview, adm_scaler_positive, adm_scaler_negative, adm_scaler_end, adaptive_cfg, sampler_name, \
         scheduler_name, generate_image_grid, overwrite_step, overwrite_switch, overwrite_width, overwrite_height, \
         overwrite_vary_strength, overwrite_upscale_strength, \
         mixing_image_prompt_and_vary_upscale, mixing_image_prompt_and_inpaint, \
-        debugging_cn_preprocessor, controlnet_softness, canny_low_threshold, canny_high_threshold, inpaint_engine, \
+        debugging_cn_preprocessor, skipping_cn_preprocessor, controlnet_softness, canny_low_threshold, canny_high_threshold, \
         refiner_swap_method, \
-        freeu_enabled, freeu_b1, freeu_b2, freeu_s1, freeu_s2 = args
+        freeu_enabled, freeu_b1, freeu_b2, freeu_s1, freeu_s2, \
+        debugging_inpaint_preprocessor, inpaint_disable_initial_latent, inpaint_engine, inpaint_strength, inpaint_respective_field = args
 
     return

modules/async_worker.py

@@ -1,13 +1,18 @@
 import threading
 
 
-buffer = []
-outputs = []
-global_results = []
+class AsyncTask:
+    def __init__(self, args):
+        self.args = args
+        self.yields = []
+        self.results = []
+
+
+async_tasks = []
 
 
 def worker():
-    global buffer, outputs, global_results
+    global async_tasks
 
     import traceback
     import math
@@ -28,11 +33,12 @@ def worker():
     import modules.constants as constants
     import modules.advanced_parameters as advanced_parameters
     import fooocus_extras.ip_adapter as ip_adapter
+    import fooocus_extras.face_crop
 
     from modules.sdxl_styles import apply_style, apply_wildcards, fooocus_expansion
     from modules.private_logger import log
     from modules.expansion import safe_str
-    from modules.util import join_prompts, remove_empty_str, HWC3, resize_image, \
+    from modules.util import remove_empty_str, HWC3, resize_image, \
         get_image_shape_ceil, set_image_shape_ceil, get_shape_ceil, resample_image
     from modules.upscaler import perform_upscale
 
@@ -45,42 +51,40 @@ def worker():
     except Exception as e:
         print(e)
 
-    def progressbar(number, text):
+    def progressbar(async_task, number, text):
         print(f'[Fooocus] {text}')
-        outputs.append(['preview', (number, text, None)])
-
-    def yield_result(imgs, do_not_show_finished_images=False):
-        global global_results
+        async_task.yields.append(['preview', (number, text, None)])
 
+    def yield_result(async_task, imgs, do_not_show_finished_images=False):
         if not isinstance(imgs, list):
             imgs = [imgs]
 
-        global_results = global_results + imgs
+        async_task.results = async_task.results + imgs
 
         if do_not_show_finished_images:
             return
 
-        outputs.append(['results', global_results])
+        async_task.yields.append(['results', async_task.results])
         return
 
-    def build_image_wall():
+    def build_image_wall(async_task):
         if not advanced_parameters.generate_image_grid:
             return
 
-        global global_results
+        results = async_task.results
 
-        if len(global_results) < 2:
+        if len(results) < 2:
             return
 
-        for img in global_results:
+        for img in results:
             if not isinstance(img, np.ndarray):
                 return
             if img.ndim != 3:
                 return
 
-        H, W, C = global_results[0].shape
+        H, W, C = results[0].shape
 
-        for img in global_results:
+        for img in results:
             Hn, Wn, Cn = img.shape
             if H != Hn:
                 return
@@ -89,28 +93,29 @@ def worker():
             if C != Cn:
                 return
 
-        cols = float(len(global_results)) ** 0.5
+        cols = float(len(results)) ** 0.5
         cols = int(math.ceil(cols))
-        rows = float(len(global_results)) / float(cols)
+        rows = float(len(results)) / float(cols)
         rows = int(math.ceil(rows))
 
         wall = np.zeros(shape=(H * rows, W * cols, C), dtype=np.uint8)
 
         for y in range(rows):
             for x in range(cols):
-                if y * cols + x < len(global_results):
-                    img = global_results[y * cols + x]
+                if y * cols + x < len(results):
+                    img = results[y * cols + x]
                     wall[y * H:y * H + H, x * W:x * W + W, :] = img
 
         # must use deep copy otherwise gradio is super laggy. Do not use list.append() .
-        global_results = global_results + [wall]
+        async_task.results = async_task.results + [wall]
         return
 
     @torch.no_grad()
     @torch.inference_mode()
-    def handler(args):
+    def handler(async_task):
         execution_start_time = time.perf_counter()
 
+        args = async_task.args
         args.reverse()
 
         prompt = args.pop()
@@ -125,15 +130,16 @@ def worker():
         base_model_name = args.pop()
         refiner_model_name = args.pop()
         refiner_switch = args.pop()
-        loras = [(args.pop(), args.pop()) for _ in range(5)]
+        loras = [[str(args.pop()), float(args.pop())] for _ in range(5)]
         input_image_checkbox = args.pop()
         current_tab = args.pop()
         uov_method = args.pop()
         uov_input_image = args.pop()
         outpaint_selections = args.pop()
         inpaint_input_image = args.pop()
+        inpaint_additional_prompt = args.pop()
 
-        cn_tasks = {flags.cn_ip: [], flags.cn_canny: [], flags.cn_cpds: []}
+        cn_tasks = {x: [] for x in flags.ip_list}
         for _ in range(4):
             cn_img = args.pop()
             cn_stop = args.pop()
@@ -159,6 +165,36 @@ def worker():
             print(f'Refiner disabled because base model and refiner are same.')
             refiner_model_name = 'None'
 
+        assert performance_selection in ['Speed', 'Quality', 'Extreme Speed']
+
+        steps = 30
+
+        if performance_selection == 'Speed':
+            steps = 30
+
+        if performance_selection == 'Quality':
+            steps = 60
+
+        if performance_selection == 'Extreme Speed':
+            print('Enter LCM mode.')
+            progressbar(async_task, 1, 'Downloading LCM components ...')
+            loras += [(modules.config.downloading_sdxl_lcm_lora(), 1.0)]
+
+            if refiner_model_name != 'None':
+                print(f'Refiner disabled in LCM mode.')
+
+            refiner_model_name = 'None'
+            sampler_name = advanced_parameters.sampler_name = 'lcm'
+            scheduler_name = advanced_parameters.scheduler_name = 'lcm'
+            modules.patch.sharpness = sharpness = 0.0
+            cfg_scale = guidance_scale = 1.0
+            modules.patch.adaptive_cfg = advanced_parameters.adaptive_cfg = 1.0
+            refiner_switch = 1.0
+            modules.patch.positive_adm_scale = advanced_parameters.adm_scaler_positive = 1.0
+            modules.patch.negative_adm_scale = advanced_parameters.adm_scaler_negative = 1.0
+            modules.patch.adm_scaler_end = advanced_parameters.adm_scaler_end = 0.0
+            steps = 8
+
         modules.patch.adaptive_cfg = advanced_parameters.adaptive_cfg
         print(f'[Parameters] Adaptive CFG = {modules.patch.adaptive_cfg}')
 
@@ -168,7 +204,10 @@ def worker():
         modules.patch.positive_adm_scale = advanced_parameters.adm_scaler_positive
         modules.patch.negative_adm_scale = advanced_parameters.adm_scaler_negative
         modules.patch.adm_scaler_end = advanced_parameters.adm_scaler_end
-        print(f'[Parameters] ADM Scale = {modules.patch.positive_adm_scale} : {modules.patch.negative_adm_scale} : {modules.patch.adm_scaler_end}')
+        print(f'[Parameters] ADM Scale = '
+              f'{modules.patch.positive_adm_scale} : '
+              f'{modules.patch.negative_adm_scale} : '
+              f'{modules.patch.adm_scaler_end}')
 
         cfg_scale = float(guidance_scale)
         print(f'[Parameters] CFG = {cfg_scale}')
@@ -176,29 +215,28 @@ def worker():
         initial_latent = None
         denoising_strength = 1.0
         tiled = False
-        inpaint_worker.current_task = None
 
-        width, height = aspect_ratios_selection.split('×')
+        width, height = aspect_ratios_selection.replace('×', ' ').split(' ')[:2]
         width, height = int(width), int(height)
 
         skip_prompt_processing = False
         refiner_swap_method = advanced_parameters.refiner_swap_method
 
+        inpaint_worker.current_task = None
+        inpaint_parameterized = advanced_parameters.inpaint_engine != 'None'
         inpaint_image = None
         inpaint_mask = None
         inpaint_head_model_path = None
+
+        use_synthetic_refiner = False
+
         controlnet_canny_path = None
         controlnet_cpds_path = None
-        clip_vision_path, ip_negative_path, ip_adapter_path = None, None, None
+        clip_vision_path, ip_negative_path, ip_adapter_path, ip_adapter_face_path = None, None, None, None
 
         seed = int(image_seed)
         print(f'[Parameters] Seed = {seed}')
 
-        if performance_selection == 'Speed':
-            steps = 30
-        else:
-            steps = 60
-
         sampler_name = advanced_parameters.sampler_name
         scheduler_name = advanced_parameters.scheduler_name
 
@@ -206,7 +244,8 @@ def worker():
         tasks = []
 
         if input_image_checkbox:
-            if (current_tab == 'uov' or (current_tab == 'ip' and advanced_parameters.mixing_image_prompt_and_vary_upscale)) \
+            if (current_tab == 'uov' or (
+                    current_tab == 'ip' and advanced_parameters.mixing_image_prompt_and_vary_upscale)) \
                     and uov_method != flags.disabled and uov_input_image is not None:
                 uov_input_image = HWC3(uov_input_image)
                 if 'vary' in uov_method:
@@ -216,40 +255,66 @@ def worker():
                     if 'fast' in uov_method:
                         skip_prompt_processing = True
                     else:
+                        steps = 18
+
                         if performance_selection == 'Speed':
                             steps = 18
-                        else:
+
+                        if performance_selection == 'Quality':
                             steps = 36
-                    progressbar(1, 'Downloading upscale models ...')
+
+                        if performance_selection == 'Extreme Speed':
+                            steps = 8
+
+                    progressbar(async_task, 1, 'Downloading upscale models ...')
                     modules.config.downloading_upscale_model()
-            if (current_tab == 'inpaint' or (current_tab == 'ip' and advanced_parameters.mixing_image_prompt_and_inpaint))\
+            if (current_tab == 'inpaint' or (
+                    current_tab == 'ip' and advanced_parameters.mixing_image_prompt_and_inpaint)) \
                     and isinstance(inpaint_input_image, dict):
                 inpaint_image = inpaint_input_image['image']
                 inpaint_mask = inpaint_input_image['mask'][:, :, 0]
                 inpaint_image = HWC3(inpaint_image)
                 if isinstance(inpaint_image, np.ndarray) and isinstance(inpaint_mask, np.ndarray) \
                         and (np.any(inpaint_mask > 127) or len(outpaint_selections) > 0):
-                    progressbar(1, 'Downloading inpainter ...')
-                    inpaint_head_model_path, inpaint_patch_model_path = modules.config.downloading_inpaint_models(advanced_parameters.inpaint_engine)
+                    if inpaint_parameterized:
+                        progressbar(async_task, 1, 'Downloading inpainter ...')
+                        modules.config.downloading_upscale_model()
+                        inpaint_head_model_path, inpaint_patch_model_path = modules.config.downloading_inpaint_models(
+                            advanced_parameters.inpaint_engine)
                         base_model_additional_loras += [(inpaint_patch_model_path, 1.0)]
                         print(f'[Inpaint] Current inpaint model is {inpaint_patch_model_path}')
+                        if refiner_model_name == 'None':
+                            use_synthetic_refiner = True
+                            refiner_switch = 0.5
+                    else:
+                        inpaint_head_model_path, inpaint_patch_model_path = None, None
+                        print(f'[Inpaint] Parameterized inpaint is disabled.')
+                    if inpaint_additional_prompt != '':
+                        if prompt == '':
+                            prompt = inpaint_additional_prompt
+                        else:
+                            prompt = inpaint_additional_prompt + '\n' + prompt
                     goals.append('inpaint')
             if current_tab == 'ip' or \
                     advanced_parameters.mixing_image_prompt_and_inpaint or \
                     advanced_parameters.mixing_image_prompt_and_vary_upscale:
                 goals.append('cn')
-                progressbar(1, 'Downloading control models ...')
+                progressbar(async_task, 1, 'Downloading control models ...')
                 if len(cn_tasks[flags.cn_canny]) > 0:
                     controlnet_canny_path = modules.config.downloading_controlnet_canny()
                 if len(cn_tasks[flags.cn_cpds]) > 0:
                     controlnet_cpds_path = modules.config.downloading_controlnet_cpds()
                 if len(cn_tasks[flags.cn_ip]) > 0:
-                    clip_vision_path, ip_negative_path, ip_adapter_path = modules.config.downloading_ip_adapters()
-                progressbar(1, 'Loading control models ...')
+                    clip_vision_path, ip_negative_path, ip_adapter_path = modules.config.downloading_ip_adapters('ip')
+                if len(cn_tasks[flags.cn_ip_face]) > 0:
+                    clip_vision_path, ip_negative_path, ip_adapter_face_path = modules.config.downloading_ip_adapters(
+                        'face')
+                progressbar(async_task, 1, 'Loading control models ...')
 
         # Load or unload CNs
         pipeline.refresh_controlnets([controlnet_canny_path, controlnet_cpds_path])
         ip_adapter.load_ip_adapter(clip_vision_path, ip_negative_path, ip_adapter_path)
+        ip_adapter.load_ip_adapter(clip_vision_path, ip_negative_path, ip_adapter_face_path)
 
         switch = int(round(steps * refiner_switch))
 
@@ -268,7 +333,7 @@ def worker():
         print(f'[Parameters] Sampler = {sampler_name} - {scheduler_name}')
         print(f'[Parameters] Steps = {steps} - {switch}')
 
-        progressbar(1, 'Initializing ...')
+        progressbar(async_task, 1, 'Initializing ...')
 
         if not skip_prompt_processing:
 
@@ -285,11 +350,12 @@ def worker():
             extra_positive_prompts = prompts[1:] if len(prompts) > 1 else []
             extra_negative_prompts = negative_prompts[1:] if len(negative_prompts) > 1 else []
 
-            progressbar(3, 'Loading models ...')
+            progressbar(async_task, 3, 'Loading models ...')
             pipeline.refresh_everything(refiner_model_name=refiner_model_name, base_model_name=base_model_name,
-                                        loras=loras, base_model_additional_loras=base_model_additional_loras)
+                                        loras=loras, base_model_additional_loras=base_model_additional_loras,
+                                        use_synthetic_refiner=use_synthetic_refiner)
 
-            progressbar(3, 'Processing prompts ...')
+            progressbar(async_task, 3, 'Processing prompts ...')
             tasks = []
             for i in range(image_number):
                 task_seed = (seed + i) % (constants.MAX_SEED + 1)  # randint is inclusive, % is not
@@ -330,28 +396,31 @@ def worker():
                     uc=None,
                     positive_top_k=len(positive_basic_workloads),
                     negative_top_k=len(negative_basic_workloads),
-                    log_positive_prompt='\n'.join([task_prompt] + task_extra_positive_prompts),
-                    log_negative_prompt='\n'.join([task_negative_prompt] + task_extra_negative_prompts),
+                    log_positive_prompt='; '.join([task_prompt] + task_extra_positive_prompts),
+                    log_negative_prompt='; '.join([task_negative_prompt] + task_extra_negative_prompts),
                 ))
 
             if use_expansion:
                 for i, t in enumerate(tasks):
-                    progressbar(5, f'Preparing Fooocus text #{i + 1} ...')
+                    progressbar(async_task, 5, f'Preparing Fooocus text #{i + 1} ...')
                     expansion = pipeline.final_expansion(t['task_prompt'], t['task_seed'])
                     print(f'[Prompt Expansion] {expansion}')
                     t['expansion'] = expansion
                     t['positive'] = copy.deepcopy(t['positive']) + [expansion]  # Deep copy.
 
             for i, t in enumerate(tasks):
-                progressbar(7, f'Encoding positive #{i + 1} ...')
+                progressbar(async_task, 7, f'Encoding positive #{i + 1} ...')
                 t['c'] = pipeline.clip_encode(texts=t['positive'], pool_top_k=t['positive_top_k'])
 
             for i, t in enumerate(tasks):
-                progressbar(10, f'Encoding negative #{i + 1} ...')
+                if abs(float(cfg_scale) - 1.0) < 1e-4:
+                    t['uc'] = pipeline.clone_cond(t['c'])
+                else:
+                    progressbar(async_task, 10, f'Encoding negative #{i + 1} ...')
                     t['uc'] = pipeline.clip_encode(texts=t['negative'], pool_top_k=t['negative_top_k'])
 
         if len(goals) > 0:
-            progressbar(13, 'Image processing ...')
+            progressbar(async_task, 13, 'Image processing ...')
 
         if 'vary' in goals:
             if 'subtle' in uov_method:
@@ -372,8 +441,16 @@ def worker():
             uov_input_image = set_image_shape_ceil(uov_input_image, shape_ceil)
 
             initial_pixels = core.numpy_to_pytorch(uov_input_image)
-            progressbar(13, 'VAE encoding ...')
-            initial_latent = core.encode_vae(vae=pipeline.final_vae, pixels=initial_pixels)
+            progressbar(async_task, 13, 'VAE encoding ...')
+
+            candidate_vae, _ = pipeline.get_candidate_vae(
+                steps=steps,
+                switch=switch,
+                denoise=denoising_strength,
+                refiner_swap_method=refiner_swap_method
+            )
+
+            initial_latent = core.encode_vae(vae=candidate_vae, pixels=initial_pixels)
             B, C, H, W = initial_latent['samples'].shape
             width = W * 8
             height = H * 8
@@ -381,11 +458,8 @@ def worker():
 
         if 'upscale' in goals:
             H, W, C = uov_input_image.shape
-            progressbar(13, f'Upscaling image from {str((H, W))} ...')
-
-            uov_input_image = core.numpy_to_pytorch(uov_input_image)
+            progressbar(async_task, 13, f'Upscaling image from {str((H, W))} ...')
             uov_input_image = perform_upscale(uov_input_image)
-            uov_input_image = core.pytorch_to_numpy(uov_input_image)[0]
             print(f'Image upscaled.')
 
             if '1.5x' in uov_method:
@@ -419,7 +493,7 @@ def worker():
             if direct_return:
                 d = [('Upscale (Fast)', '2x')]
                 log(uov_input_image, d, single_line_number=1)
-                yield_result(uov_input_image, do_not_show_finished_images=True)
+                yield_result(async_task, uov_input_image, do_not_show_finished_images=True)
                 return
 
             tiled = True
@@ -429,16 +503,22 @@ def worker():
                 denoising_strength = advanced_parameters.overwrite_upscale_strength
 
             initial_pixels = core.numpy_to_pytorch(uov_input_image)
-            progressbar(13, 'VAE encoding ...')
+            progressbar(async_task, 13, 'VAE encoding ...')
+
+            candidate_vae, _ = pipeline.get_candidate_vae(
+                steps=steps,
+                switch=switch,
+                denoise=denoising_strength,
+                refiner_swap_method=refiner_swap_method
+            )
 
             initial_latent = core.encode_vae(
-                vae=pipeline.final_vae if pipeline.final_refiner_vae is None else pipeline.final_refiner_vae,
+                vae=candidate_vae,
                 pixels=initial_pixels, tiled=True)
             B, C, H, W = initial_latent['samples'].shape
             width = W * 8
             height = H * 8
             print(f'Final resolution is {str((height, width))}.')
-            refiner_swap_method = 'upscale'
 
         if 'inpaint' in goals:
             if len(outpaint_selections) > 0:
@@ -464,69 +544,96 @@ def worker():
 
                 inpaint_image = np.ascontiguousarray(inpaint_image.copy())
                 inpaint_mask = np.ascontiguousarray(inpaint_mask.copy())
+                advanced_parameters.inpaint_strength = 1.0
+                advanced_parameters.inpaint_respective_field = 1.0
 
-            inpaint_worker.current_task = inpaint_worker.InpaintWorker(image=inpaint_image, mask=inpaint_mask,
-                                                                       is_outpaint=len(outpaint_selections) > 0)
+            denoising_strength = advanced_parameters.inpaint_strength
 
-            pipeline.final_unet.model.diffusion_model.in_inpaint = True
+            inpaint_worker.current_task = inpaint_worker.InpaintWorker(
+                image=inpaint_image,
+                mask=inpaint_mask,
+                use_fill=denoising_strength > 0.99,
+                k=advanced_parameters.inpaint_respective_field
+            )
 
-            if advanced_parameters.debugging_cn_preprocessor:
-                yield_result(inpaint_worker.current_task.visualize_mask_processing(), do_not_show_finished_images=True)
+            if advanced_parameters.debugging_inpaint_preprocessor:
+                yield_result(async_task, inpaint_worker.current_task.visualize_mask_processing(),
+                             do_not_show_finished_images=True)
                 return
 
-            progressbar(13, 'VAE Inpaint encoding ...')
+            progressbar(async_task, 13, 'VAE Inpaint encoding ...')
 
             inpaint_pixel_fill = core.numpy_to_pytorch(inpaint_worker.current_task.interested_fill)
             inpaint_pixel_image = core.numpy_to_pytorch(inpaint_worker.current_task.interested_image)
             inpaint_pixel_mask = core.numpy_to_pytorch(inpaint_worker.current_task.interested_mask)
 
+            candidate_vae, candidate_vae_swap = pipeline.get_candidate_vae(
+                steps=steps,
+                switch=switch,
+                denoise=denoising_strength,
+                refiner_swap_method=refiner_swap_method
+            )
+
             latent_inpaint, latent_mask = core.encode_vae_inpaint(
                 mask=inpaint_pixel_mask,
-                vae=pipeline.final_vae,
+                vae=candidate_vae,
                 pixels=inpaint_pixel_image)
 
             latent_swap = None
-            if pipeline.final_refiner_vae is not None:
-                progressbar(13, 'VAE Inpaint SD15 encoding ...')
+            if candidate_vae_swap is not None:
+                progressbar(async_task, 13, 'VAE SD15 encoding ...')
                 latent_swap = core.encode_vae(
-                    vae=pipeline.final_refiner_vae,
+                    vae=candidate_vae_swap,
                     pixels=inpaint_pixel_fill)['samples']
 
-            progressbar(13, 'VAE encoding ...')
+            progressbar(async_task, 13, 'VAE encoding ...')
             latent_fill = core.encode_vae(
-                vae=pipeline.final_vae,
+                vae=candidate_vae,
                 pixels=inpaint_pixel_fill)['samples']
 
-            inpaint_worker.current_task.load_latent(latent_fill=latent_fill,
-                                                    latent_inpaint=latent_inpaint,
-                                                    latent_mask=latent_mask,
-                                                    latent_swap=latent_swap,
-                                                    inpaint_head_model_path=inpaint_head_model_path)
+            inpaint_worker.current_task.load_latent(
+                latent_fill=latent_fill, latent_mask=latent_mask, latent_swap=latent_swap)
+
+            if inpaint_parameterized:
+                pipeline.final_unet = inpaint_worker.current_task.patch(
+                    inpaint_head_model_path=inpaint_head_model_path,
+                    inpaint_latent=latent_inpaint,
+                    inpaint_latent_mask=latent_mask,
+                    model=pipeline.final_unet
+                )
+
+            if not advanced_parameters.inpaint_disable_initial_latent:
+                initial_latent = {'samples': latent_fill}
 
             B, C, H, W = latent_fill.shape
             height, width = H * 8, W * 8
             final_height, final_width = inpaint_worker.current_task.image.shape[:2]
-            initial_latent = {'samples': latent_fill}
             print(f'Final resolution is {str((final_height, final_width))}, latent is {str((height, width))}.')
 
         if 'cn' in goals:
             for task in cn_tasks[flags.cn_canny]:
                 cn_img, cn_stop, cn_weight = task
                 cn_img = resize_image(HWC3(cn_img), width=width, height=height)
+
+                if not advanced_parameters.skipping_cn_preprocessor:
                     cn_img = preprocessors.canny_pyramid(cn_img)
+
                 cn_img = HWC3(cn_img)
                 task[0] = core.numpy_to_pytorch(cn_img)
                 if advanced_parameters.debugging_cn_preprocessor:
-                    yield_result(cn_img, do_not_show_finished_images=True)
+                    yield_result(async_task, cn_img, do_not_show_finished_images=True)
                     return
             for task in cn_tasks[flags.cn_cpds]:
                 cn_img, cn_stop, cn_weight = task
                 cn_img = resize_image(HWC3(cn_img), width=width, height=height)
+
+                if not advanced_parameters.skipping_cn_preprocessor:
                     cn_img = preprocessors.cpds(cn_img)
+
                 cn_img = HWC3(cn_img)
                 task[0] = core.numpy_to_pytorch(cn_img)
                 if advanced_parameters.debugging_cn_preprocessor:
-                    yield_result(cn_img, do_not_show_finished_images=True)
+                    yield_result(async_task, cn_img, do_not_show_finished_images=True)
                     return
             for task in cn_tasks[flags.cn_ip]:
                 cn_img, cn_stop, cn_weight = task
@@ -535,13 +642,29 @@ def worker():
                 # https://github.com/tencent-ailab/IP-Adapter/blob/d580c50a291566bbf9fc7ac0f760506607297e6d/README.md?plain=1#L75
                 cn_img = resize_image(cn_img, width=224, height=224, resize_mode=0)
 
-                task[0] = ip_adapter.preprocess(cn_img)
+                task[0] = ip_adapter.preprocess(cn_img, ip_adapter_path=ip_adapter_path)
                 if advanced_parameters.debugging_cn_preprocessor:
-                    yield_result(cn_img, do_not_show_finished_images=True)
+                    yield_result(async_task, cn_img, do_not_show_finished_images=True)
+                    return
+            for task in cn_tasks[flags.cn_ip_face]:
+                cn_img, cn_stop, cn_weight = task
+                cn_img = HWC3(cn_img)
+
+                if not advanced_parameters.skipping_cn_preprocessor:
+                    cn_img = fooocus_extras.face_crop.crop_image(cn_img)
+
+                # https://github.com/tencent-ailab/IP-Adapter/blob/d580c50a291566bbf9fc7ac0f760506607297e6d/README.md?plain=1#L75
+                cn_img = resize_image(cn_img, width=224, height=224, resize_mode=0)
+
+                task[0] = ip_adapter.preprocess(cn_img, ip_adapter_path=ip_adapter_face_path)
+                if advanced_parameters.debugging_cn_preprocessor:
+                    yield_result(async_task, cn_img, do_not_show_finished_images=True)
                     return
 
-            if len(cn_tasks[flags.cn_ip]) > 0:
-                pipeline.final_unet = ip_adapter.patch_model(pipeline.final_unet, cn_tasks[flags.cn_ip])
+            all_ip_tasks = cn_tasks[flags.cn_ip] + cn_tasks[flags.cn_ip_face]
+
+            if len(all_ip_tasks) > 0:
+                pipeline.final_unet = ip_adapter.patch_model(pipeline.final_unet, all_ip_tasks)
 
         if advanced_parameters.freeu_enabled:
             print(f'FreeU is enabled!')
@@ -555,14 +678,40 @@ def worker():
 
         all_steps = steps * image_number
 
+        print(f'[Parameters] Denoising Strength = {denoising_strength}')
+
+        if isinstance(initial_latent, dict) and 'samples' in initial_latent:
+            log_shape = initial_latent['samples'].shape
+        else:
+            log_shape = f'Image Space {(height, width)}'
+
+        print(f'[Parameters] Initial Latent shape: {log_shape}')
+
         preparation_time = time.perf_counter() - execution_start_time
         print(f'Preparation time: {preparation_time:.2f} seconds')
 
-        outputs.append(['preview', (13, 'Moving model to GPU ...', None)])
+        final_sampler_name = sampler_name
+        final_scheduler_name = scheduler_name
+
+        if scheduler_name == 'lcm':
+            final_scheduler_name = 'sgm_uniform'
+            if pipeline.final_unet is not None:
+                pipeline.final_unet = core.opModelSamplingDiscrete.patch(
+                    pipeline.final_unet,
+                    sampling='lcm',
+                    zsnr=False)[0]
+            if pipeline.final_refiner_unet is not None:
+                pipeline.final_refiner_unet = core.opModelSamplingDiscrete.patch(
+                    pipeline.final_refiner_unet,
+                    sampling='lcm',
+                    zsnr=False)[0]
+            print('Using lcm scheduler.')
+
+        async_task.yields.append(['preview', (13, 'Moving model to GPU ...', None)])
 
         def callback(step, x0, x, total_steps, y):
             done_steps = current_task_id * steps + step
-            outputs.append(['preview', (
+            async_task.yields.append(['preview', (
                 int(15.0 + 85.0 * float(done_steps) / float(all_steps)),
                 f'Step {step}/{total_steps} in the {current_task_id + 1}-th Sampling',
                 y)])
@@ -592,8 +741,8 @@ def worker():
                     height=height,
                     image_seed=task['task_seed'],
                     callback=callback,
-                    sampler_name=sampler_name,
-                    scheduler_name=scheduler_name,
+                    sampler_name=final_sampler_name,
+                    scheduler_name=final_scheduler_name,
                     latent=initial_latent,
                     denoise=denoising_strength,
                     tiled=tiled,
@@ -616,7 +765,10 @@ def worker():
                         ('Resolution', str((width, height))),
                         ('Sharpness', sharpness),
                         ('Guidance Scale', guidance_scale),
-                        ('ADM Guidance', str((modules.patch.positive_adm_scale, modules.patch.negative_adm_scale))),
+                        ('ADM Guidance', str((
+                            modules.patch.positive_adm_scale,
+                            modules.patch.negative_adm_scale,
+                            modules.patch.adm_scaler_end))),
                         ('Base Model', base_model_name),
                         ('Refiner Model', refiner_model_name),
                         ('Refiner Switch', refiner_switch),
@@ -629,7 +781,7 @@ def worker():
                             d.append((f'LoRA [{n}] weight', w))
                     log(x, d, single_line_number=3)
 
-                yield_result(imgs, do_not_show_finished_images=len(tasks) == 1)
+                yield_result(async_task, imgs, do_not_show_finished_images=len(tasks) == 1)
             except fcbh.model_management.InterruptProcessingException as e:
                 if shared.last_stop == 'skip':
                     print('User skipped')
@@ -645,16 +797,15 @@ def worker():
 
     while True:
         time.sleep(0.01)
-        if len(buffer) > 0:
-            task = buffer.pop(0)
+        if len(async_tasks) > 0:
+            task = async_tasks.pop(0)
             try:
                 handler(task)
             except:
                 traceback.print_exc()
-            if len(buffer) == 0:
-                build_image_wall()
-                outputs.append(['finish', global_results])
-                global_results = []
+            finally:
+                build_image_wall(task)
+                task.yields.append(['finish', task.results])
                 pipeline.prepare_text_encoder(async_call=True)
     pass
 

modules/config.py

@@ -1,5 +1,7 @@
 import os
 import json
+import math
+import numbers
 import args_manager
 import modules.flags
 import modules.sdxl_styles
@@ -20,8 +22,12 @@ try:
             config_dict = json.load(json_file)
             always_save_keys = list(config_dict.keys())
 except Exception as e:
-    print('Load path config failed')
-    print(e)
+    print(f'Failed to load config file "{config_path}" . The reason is: {str(e)}')
+    print('Please make sure that:')
+    print(f'1. The file "{config_path}" is a valid text file, and you have access to read it.')
+    print('2. Use "\\\\" instead of "\\" when describing paths.')
+    print('3. There is no "," before the last "}".')
+    print('4. All key/value formats are correct.')
 
 
 def try_load_deprecated_user_path_config():
@@ -76,20 +82,18 @@ try_load_deprecated_user_path_config()
 preset = args_manager.args.preset
 
 if isinstance(preset, str):
-    preset = os.path.abspath(f'./presets/{preset}.json')
+    preset_path = os.path.abspath(f'./presets/{preset}.json')
     try:
-        if os.path.exists(preset):
-            with open(preset, "r", encoding="utf-8") as json_file:
-                preset = json.load(json_file)
+        if os.path.exists(preset_path):
+            with open(preset_path, "r", encoding="utf-8") as json_file:
+                config_dict.update(json.load(json_file))
+                print(f'Loaded preset: {preset_path}')
+        else:
+            raise FileNotFoundError
     except Exception as e:
-        print('Load preset config failed')
+        print(f'Load preset [{preset_path}] failed')
         print(e)
 
-preset = preset if isinstance(preset, dict) else None
-
-if preset is not None:
-    config_dict.update(preset)
-
 
 def get_dir_or_set_default(key, default_value):
     global config_dict, visited_keys, always_save_keys
@@ -104,6 +108,8 @@ def get_dir_or_set_default(key, default_value):
     if isinstance(v, str) and os.path.exists(v) and os.path.isdir(v):
         return v
     else:
+        if v is not None:
+            print(f'Failed to load config key: {json.dumps({key:v})} is invalid or does not exist; will use {json.dumps({key:default_value})} instead.')
         dp = os.path.abspath(os.path.join(os.path.dirname(__file__), default_value))
         os.makedirs(dp, exist_ok=True)
         config_dict[key] = dp
@@ -139,6 +145,8 @@ def get_config_item_or_set_default(key, default_value, validator, disable_empty_
     if validator(v):
         return v
     else:
+        if v is not None:
+            print(f'Failed to load config key: {json.dumps({key:v})} is invalid; will use {json.dumps({key:default_value})} instead.')
         config_dict[key] = default_value
         return default_value
 
@@ -156,27 +164,43 @@ default_refiner_model_name = get_config_item_or_set_default(
 default_refiner_switch = get_config_item_or_set_default(
     key='default_refiner_switch',
     default_value=0.5,
-    validator=lambda x: isinstance(x, float)
+    validator=lambda x: isinstance(x, numbers.Number) and 0 <= x <= 1
 )
-default_lora_name = get_config_item_or_set_default(
-    key='default_lora',
-    default_value='sd_xl_offset_example-lora_1.0.safetensors',
-    validator=lambda x: isinstance(x, str)
-)
-default_lora_weight = get_config_item_or_set_default(
-    key='default_lora_weight',
-    default_value=0.1,
-    validator=lambda x: isinstance(x, float)
+default_loras = get_config_item_or_set_default(
+    key='default_loras',
+    default_value=[
+        [
+            "sd_xl_offset_example-lora_1.0.safetensors",
+            0.1
+        ],
+        [
+            "None",
+            1.0
+        ],
+        [
+            "None",
+            1.0
+        ],
+        [
+            "None",
+            1.0
+        ],
+        [
+            "None",
+            1.0
+        ]
+    ],
+    validator=lambda x: isinstance(x, list) and all(len(y) == 2 and isinstance(y[0], str) and isinstance(y[1], numbers.Number) for y in x)
 )
 default_cfg_scale = get_config_item_or_set_default(
     key='default_cfg_scale',
     default_value=4.0,
-    validator=lambda x: isinstance(x, float)
+    validator=lambda x: isinstance(x, numbers.Number)
 )
 default_sample_sharpness = get_config_item_or_set_default(
     key='default_sample_sharpness',
-    default_value=2,
-    validator=lambda x: isinstance(x, float)
+    default_value=2.0,
+    validator=lambda x: isinstance(x, numbers.Number)
 )
 default_sampler = get_config_item_or_set_default(
     key='default_sampler',
@@ -190,7 +214,11 @@ default_scheduler = get_config_item_or_set_default(
 )
 default_styles = get_config_item_or_set_default(
     key='default_styles',
-    default_value=['Fooocus V2', 'Fooocus Enhance', 'Fooocus Sharp'],
+    default_value=[
+        "Fooocus V2",
+        "Fooocus Enhance",
+        "Fooocus Sharp"
+    ],
     validator=lambda x: isinstance(x, list) and all(y in modules.sdxl_styles.legal_style_names for y in x)
 )
 default_prompt_negative = get_config_item_or_set_default(
@@ -205,6 +233,11 @@ default_prompt = get_config_item_or_set_default(
     validator=lambda x: isinstance(x, str),
     disable_empty_as_none=True
 )
+default_performance = get_config_item_or_set_default(
+    key='default_performance',
+    default_value='Speed',
+    validator=lambda x: x in modules.flags.performance_selections
+)
 default_advanced_checkbox = get_config_item_or_set_default(
     key='default_advanced_checkbox',
     default_value=False,
@@ -213,21 +246,19 @@ default_advanced_checkbox = get_config_item_or_set_default(
 default_image_number = get_config_item_or_set_default(
     key='default_image_number',
     default_value=2,
-    validator=lambda x: isinstance(x, int) and x >= 1 and x <= 32
+    validator=lambda x: isinstance(x, int) and 1 <= x <= 32
 )
 checkpoint_downloads = get_config_item_or_set_default(
     key='checkpoint_downloads',
     default_value={
-        'juggernautXL_version6Rundiffusion.safetensors':
-            'https://huggingface.co/lllyasviel/fav_models/resolve/main/fav/juggernautXL_version6Rundiffusion.safetensors'
+        "juggernautXL_version6Rundiffusion.safetensors": "https://huggingface.co/lllyasviel/fav_models/resolve/main/fav/juggernautXL_version6Rundiffusion.safetensors"
     },
     validator=lambda x: isinstance(x, dict) and all(isinstance(k, str) and isinstance(v, str) for k, v in x.items())
 )
 lora_downloads = get_config_item_or_set_default(
     key='lora_downloads',
     default_value={
-        'sd_xl_offset_example-lora_1.0.safetensors':
-            'https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_offset_example-lora_1.0.safetensors'
+        "sd_xl_offset_example-lora_1.0.safetensors": "https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_offset_example-lora_1.0.safetensors"
     },
     validator=lambda x: isinstance(x, dict) and all(isinstance(k, str) and isinstance(v, str) for k, v in x.items())
 )
@@ -238,7 +269,13 @@ embeddings_downloads = get_config_item_or_set_default(
 )
 available_aspect_ratios = get_config_item_or_set_default(
     key='available_aspect_ratios',
-    default_value=['704*1408', '704*1344', '768*1344', '768*1280', '832*1216', '832*1152', '896*1152', '896*1088', '960*1088', '960*1024', '1024*1024', '1024*960', '1088*960', '1088*896', '1152*896', '1152*832', '1216*832', '1280*768', '1344*768', '1344*704', '1408*704', '1472*704', '1536*640', '1600*640', '1664*576', '1728*576'],
+    default_value=[
+        '704*1408', '704*1344', '768*1344', '768*1280', '832*1216', '832*1152',
+        '896*1152', '896*1088', '960*1088', '960*1024', '1024*1024', '1024*960',
+        '1088*960', '1088*896', '1152*896', '1152*832', '1216*832', '1280*768',
+        '1344*768', '1344*704', '1408*704', '1472*704', '1536*640', '1600*640',
+        '1664*576', '1728*576'
+    ],
     validator=lambda x: isinstance(x, list) and all('*' in v for v in x) and len(x) > 1
 )
 default_aspect_ratio = get_config_item_or_set_default(
@@ -246,16 +283,93 @@ default_aspect_ratio = get_config_item_or_set_default(
     default_value='1152*896' if '1152*896' in available_aspect_ratios else available_aspect_ratios[0],
     validator=lambda x: x in available_aspect_ratios
 )
+default_inpaint_engine_version = get_config_item_or_set_default(
+    key='default_inpaint_engine_version',
+    default_value='v2.6',
+    validator=lambda x: x in modules.flags.inpaint_engine_versions
+)
+default_cfg_tsnr = get_config_item_or_set_default(
+    key='default_cfg_tsnr',
+    default_value=7.0,
+    validator=lambda x: isinstance(x, numbers.Number)
+)
+default_overwrite_step = get_config_item_or_set_default(
+    key='default_overwrite_step',
+    default_value=-1,
+    validator=lambda x: isinstance(x, int)
+)
+default_overwrite_switch = get_config_item_or_set_default(
+    key='default_overwrite_switch',
+    default_value=-1,
+    validator=lambda x: isinstance(x, int)
+)
+example_inpaint_prompts = get_config_item_or_set_default(
+    key='example_inpaint_prompts',
+    default_value=[
+        'highly detailed face', 'detailed girl face', 'detailed man face', 'detailed hand', 'beautiful eyes'
+    ],
+    validator=lambda x: isinstance(x, list) and all(isinstance(v, str) for v in x)
+)
 
-with open(config_path, "w", encoding="utf-8") as json_file:
+example_inpaint_prompts = [[x] for x in example_inpaint_prompts]
+
+config_dict["default_loras"] = default_loras = default_loras[:5] + [['None', 1.0] for _ in range(5 - len(default_loras))]
+
+possible_preset_keys = [
+    "default_model",
+    "default_refiner",
+    "default_refiner_switch",
+    "default_loras",
+    "default_cfg_scale",
+    "default_sample_sharpness",
+    "default_sampler",
+    "default_scheduler",
+    "default_performance",
+    "default_prompt",
+    "default_prompt_negative",
+    "default_styles",
+    "default_aspect_ratio",
+    "checkpoint_downloads",
+    "embeddings_downloads",
+    "lora_downloads",
+]
+
+
+REWRITE_PRESET = False
+
+if REWRITE_PRESET and isinstance(args_manager.args.preset, str):
+    save_path = 'presets/' + args_manager.args.preset + '.json'
+    with open(save_path, "w", encoding="utf-8") as json_file:
+        json.dump({k: config_dict[k] for k in possible_preset_keys}, json_file, indent=4)
+    print(f'Preset saved to {save_path}. Exiting ...')
+    exit(0)
+
+
+def add_ratio(x):
+    a, b = x.replace('*', ' ').split(' ')[:2]
+    a, b = int(a), int(b)
+    g = math.gcd(a, b)
+    return f'{a}×{b} <span style="color: grey;"> \U00002223 {a // g}:{b // g}</span>'
+
+
+default_aspect_ratio = add_ratio(default_aspect_ratio)
+available_aspect_ratios = [add_ratio(x) for x in available_aspect_ratios]
+
+
+# Only write config in the first launch.
+if not os.path.exists(config_path):
+    with open(config_path, "w", encoding="utf-8") as json_file:
         json.dump({k: config_dict[k] for k in always_save_keys}, json_file, indent=4)
 
+
+# Always write tutorials.
 with open(config_example_path, "w", encoding="utf-8") as json_file:
     cpa = config_path.replace("\\", "\\\\")
     json_file.write(f'You can modify your "{cpa}" using the below keys, formats, and examples.\n'
                     f'Do not modify this file. Modifications in this file will not take effect.\n'
                     f'This file is a tutorial and example. Please edit "{cpa}" to really change any settings.\n'
-                    f'Remember to split the paths with "\\\\" rather than "\\".\n\n\n')
+                    + 'Remember to split the paths with "\\\\" rather than "\\", '
+                      'and there is no "," before the last "}". \n\n\n')
     json.dump({k: config_dict[k] for k in visited_keys}, json_file, indent=4)
 
 
@@ -264,9 +378,6 @@ os.makedirs(path_outputs, exist_ok=True)
 model_filenames = []
 lora_filenames = []
 
-available_aspect_ratios = [x.replace('*', '×') for x in available_aspect_ratios]
-default_aspect_ratio = default_aspect_ratio.replace('*', '×')
-
 
 def get_model_filenames(folder_path, name_filter=None):
     return get_files_from_folder(folder_path, ['.pth', '.ckpt', '.bin', '.safetensors', '.fooocus.patch'], name_filter)
@@ -280,7 +391,7 @@ def update_all_model_names():
 
 
 def downloading_inpaint_models(v):
-    assert v in ['v1', 'v2.5']
+    assert v in modules.flags.inpaint_engine_versions
 
     load_file_from_url(
         url='https://huggingface.co/lllyasviel/fooocus_inpaint/resolve/main/fooocus_inpaint_head.pth',
@@ -306,9 +417,26 @@ def downloading_inpaint_models(v):
         )
         patch_file = os.path.join(path_inpaint, 'inpaint_v25.fooocus.patch')
 
+    if v == 'v2.6':
+        load_file_from_url(
+            url='https://huggingface.co/lllyasviel/fooocus_inpaint/resolve/main/inpaint_v26.fooocus.patch',
+            model_dir=path_inpaint,
+            file_name='inpaint_v26.fooocus.patch'
+        )
+        patch_file = os.path.join(path_inpaint, 'inpaint_v26.fooocus.patch')
+
     return head_file, patch_file
 
 
+def downloading_sdxl_lcm_lora():
+    load_file_from_url(
+        url='https://huggingface.co/lllyasviel/misc/resolve/main/sdxl_lcm_lora.safetensors',
+        model_dir=path_loras,
+        file_name='sdxl_lcm_lora.safetensors'
+    )
+    return 'sdxl_lcm_lora.safetensors'
+
+
 def downloading_controlnet_canny():
     load_file_from_url(
         url='https://huggingface.co/lllyasviel/misc/resolve/main/control-lora-canny-rank128.safetensors',
@@ -327,7 +455,9 @@ def downloading_controlnet_cpds():
     return os.path.join(path_controlnet, 'fooocus_xl_cpds_128.safetensors')
 
 
-def downloading_ip_adapters():
+def downloading_ip_adapters(v):
+    assert v in ['ip', 'face']
+
     results = []
 
     load_file_from_url(
@@ -344,6 +474,7 @@ def downloading_ip_adapters():
     )
     results += [os.path.join(path_controlnet, 'fooocus_ip_negative.safetensors')]
 
+    if v == 'ip':
         load_file_from_url(
             url='https://huggingface.co/lllyasviel/misc/resolve/main/ip-adapter-plus_sdxl_vit-h.bin',
             model_dir=path_controlnet,
@@ -351,6 +482,14 @@ def downloading_ip_adapters():
         )
         results += [os.path.join(path_controlnet, 'ip-adapter-plus_sdxl_vit-h.bin')]
 
+    if v == 'face':
+        load_file_from_url(
+            url='https://huggingface.co/lllyasviel/misc/resolve/main/ip-adapter-plus-face_sdxl_vit-h.bin',
+            model_dir=path_controlnet,
+            file_name='ip-adapter-plus-face_sdxl_vit-h.bin'
+        )
+        results += [os.path.join(path_controlnet, 'ip-adapter-plus-face_sdxl_vit-h.bin')]
+
     return results
 
 

modules/core.py

@@ -16,6 +16,7 @@ import fcbh.controlnet
 import modules.sample_hijack
 import fcbh.samplers
 import fcbh.latent_formats
+import modules.advanced_parameters
 
 from fcbh.sd import load_checkpoint_guess_config
 from nodes import VAEDecode, EmptyLatentImage, VAEEncode, VAEEncodeTiled, VAEDecodeTiled, \
@@ -23,9 +24,10 @@ from nodes import VAEDecode, EmptyLatentImage, VAEEncode, VAEEncodeTiled, VAEDec
 from fcbh_extras.nodes_freelunch import FreeU_V2
 from fcbh.sample import prepare_mask
 from modules.patch import patched_sampler_cfg_function
-from fcbh.lora import model_lora_keys_unet, model_lora_keys_clip, load_lora
+from modules.lora import match_lora
+from fcbh.lora import model_lora_keys_unet, model_lora_keys_clip
 from modules.config import path_embeddings
-from modules.lora import load_dangerous_lora
+from fcbh_extras.nodes_model_advanced import ModelSamplingDiscrete
 
 
 opEmptyLatentImage = EmptyLatentImage()
@@ -35,6 +37,7 @@ opVAEDecodeTiled = VAEDecodeTiled()
 opVAEEncodeTiled = VAEEncodeTiled()
 opControlNetApplyAdvanced = ControlNetApplyAdvanced()
 opFreeU = FreeU_V2()
+opModelSamplingDiscrete = ModelSamplingDiscrete()
 
 
 class StableDiffusionModel:
@@ -47,13 +50,17 @@ class StableDiffusionModel:
         self.unet_with_lora = unet
         self.clip_with_lora = clip
         self.visited_loras = ''
-        self.lora_key_map = {}
 
-        if self.unet is not None and self.clip is not None:
-            self.lora_key_map = model_lora_keys_unet(self.unet.model, self.lora_key_map)
-            self.lora_key_map = model_lora_keys_clip(self.clip.cond_stage_model, self.lora_key_map)
-            self.lora_key_map.update({x: x for x in self.unet.model.state_dict().keys()})
-            self.lora_key_map.update({x: x for x in self.clip.cond_stage_model.state_dict().keys()})
+        self.lora_key_map_unet = {}
+        self.lora_key_map_clip = {}
+
+        if self.unet is not None:
+            self.lora_key_map_unet = model_lora_keys_unet(self.unet.model, self.lora_key_map_unet)
+            self.lora_key_map_unet.update({x: x for x in self.unet.model.state_dict().keys()})
+
+        if self.clip is not None:
+            self.lora_key_map_clip = model_lora_keys_clip(self.clip.cond_stage_model, self.lora_key_map_clip)
+            self.lora_key_map_clip.update({x: x for x in self.clip.cond_stage_model.state_dict().keys()})
 
     @torch.no_grad()
     @torch.inference_mode()
@@ -64,13 +71,14 @@ class StableDiffusionModel:
             return
 
         self.visited_loras = str(loras)
-        loras_to_load = []
 
         if self.unet is None:
             return
 
         print(f'Request to load LoRAs {str(loras)} for model [{self.filename}].')
 
+        loras_to_load = []
+
         for name, weight in loras:
             if name == 'None':
                 continue
@@ -90,27 +98,33 @@ class StableDiffusionModel:
         self.clip_with_lora = self.clip.clone() if self.clip is not None else None
 
         for lora_filename, weight in loras_to_load:
-            lora = fcbh.utils.load_torch_file(lora_filename, safe_load=False)
-            lora_items = load_dangerous_lora(lora, self.lora_key_map)
+            lora_unmatch = fcbh.utils.load_torch_file(lora_filename, safe_load=False)
+            lora_unet, lora_unmatch = match_lora(lora_unmatch, self.lora_key_map_unet)
+            lora_clip, lora_unmatch = match_lora(lora_unmatch, self.lora_key_map_clip)
 
-            if len(lora_items) == 0:
+            if len(lora_unmatch) > 12:
+                # model mismatch
                 continue
 
-            print(f'Loaded LoRA [{lora_filename}] for model [{self.filename}] with {len(lora_items)} keys at weight {weight}.')
+            if len(lora_unmatch) > 0:
+                print(f'Loaded LoRA [{lora_filename}] for model [{self.filename}] '
+                      f'with unmatched keys {list(lora_unmatch.keys())}')
 
-            if self.unet_with_lora is not None:
-                loaded_unet_keys = self.unet_with_lora.add_patches(lora_items, weight)
-            else:
-                loaded_unet_keys = []
+            if self.unet_with_lora is not None and len(lora_unet) > 0:
+                loaded_keys = self.unet_with_lora.add_patches(lora_unet, weight)
+                print(f'Loaded LoRA [{lora_filename}] for UNet [{self.filename}] '
+                      f'with {len(loaded_keys)} keys at weight {weight}.')
+                for item in lora_unet:
+                    if item not in loaded_keys:
+                        print("UNet LoRA key skipped: ", item)
 
-            if self.clip_with_lora is not None:
-                loaded_clip_keys = self.clip_with_lora.add_patches(lora_items, weight)
-            else:
-                loaded_clip_keys = []
-
-            for item in lora_items:
-                if item not in set(list(loaded_unet_keys) + list(loaded_clip_keys)):
-                    print("LoRA key skipped: ", item)
+            if self.clip_with_lora is not None and len(lora_clip) > 0:
+                loaded_keys = self.clip_with_lora.add_patches(lora_clip, weight)
+                print(f'Loaded LoRA [{lora_filename}] for CLIP [{self.filename}] '
+                      f'with {len(loaded_keys)} keys at weight {weight}.')
+                for item in lora_clip:
+                    if item not in loaded_keys:
+                        print("CLIP LoRA key skipped: ", item)
 
 
 @torch.no_grad()
@@ -140,36 +154,6 @@ def load_model(ckpt_filename):
     return StableDiffusionModel(unet=unet, clip=clip, vae=vae, clip_vision=clip_vision, filename=ckpt_filename)
 
 
-@torch.no_grad()
-@torch.inference_mode()
-def load_sd_lora(model, lora_filename, strength_model=1.0, strength_clip=1.0):
-    if strength_model == 0 and strength_clip == 0:
-        return model
-
-    lora = fcbh.utils.load_torch_file(lora_filename, safe_load=False)
-
-    if lora_filename.lower().endswith('.fooocus.patch'):
-        loaded = lora
-    else:
-        key_map = model_lora_keys_unet(model.unet.model)
-        key_map = model_lora_keys_clip(model.clip.cond_stage_model, key_map)
-        loaded = load_lora(lora, key_map)
-
-    new_unet = model.unet.clone()
-    loaded_unet_keys = new_unet.add_patches(loaded, strength_model)
-
-    new_clip = model.clip.clone()
-    loaded_clip_keys = new_clip.add_patches(loaded, strength_clip)
-
-    loaded_keys = set(list(loaded_unet_keys) + list(loaded_clip_keys))
-
-    for x in loaded:
-        if x not in loaded_keys:
-            print("Lora key not loaded: ", x)
-
-    return StableDiffusionModel(unet=new_unet, clip=new_clip, vae=model.vae, clip_vision=model.clip_vision)
-
-
 @torch.no_grad()
 @torch.inference_mode()
 def generate_empty_latent(width=1024, height=1024, batch_size=1):
@@ -317,7 +301,7 @@ def ksampler(model, positive, negative, latent, seed=None, steps=30, cfg=7.0, sa
     def callback(step, x0, x, total_steps):
         fcbh.model_management.throw_exception_if_processing_interrupted()
         y = None
-        if previewer is not None:
+        if previewer is not None and not modules.advanced_parameters.disable_preview:
             y = previewer(x0, previewer_start + step, previewer_end)
         if callback_function is not None:
             callback_function(previewer_start + step, x0, x, previewer_end, y)

modules/default_pipeline.py

@@ -102,6 +102,26 @@ def refresh_refiner_model(name):
     return
 
 
+@torch.no_grad()
+@torch.inference_mode()
+def synthesize_refiner_model():
+    global model_base, model_refiner
+
+    print('Synthetic Refiner Activated')
+    model_refiner = core.StableDiffusionModel(
+        unet=model_base.unet,
+        vae=model_base.vae,
+        clip=model_base.clip,
+        clip_vision=model_base.clip_vision,
+        filename=model_base.filename
+    )
+    model_refiner.vae = None
+    model_refiner.clip = None
+    model_refiner.clip_vision = None
+
+    return
+
+
 @torch.no_grad()
 @torch.inference_mode()
 def refresh_loras(loras, base_model_additional_loras=None):
@@ -132,6 +152,25 @@ def clip_encode_single(clip, text, verbose=False):
     return result
 
 
+@torch.no_grad()
+@torch.inference_mode()
+def clone_cond(conds):
+    results = []
+
+    for c, p in conds:
+        p = p["pooled_output"]
+
+        if isinstance(c, torch.Tensor):
+            c = c.clone()
+
+        if isinstance(p, torch.Tensor):
+            p = p.clone()
+
+        results.append([c, {"pooled_output": p}])
+
+    return results
+
+
 @torch.no_grad()
 @torch.inference_mode()
 def clip_encode(texts, pool_top_k=1):
@@ -175,7 +214,8 @@ def prepare_text_encoder(async_call=True):
 
 @torch.no_grad()
 @torch.inference_mode()
-def refresh_everything(refiner_model_name, base_model_name, loras, base_model_additional_loras=None):
+def refresh_everything(refiner_model_name, base_model_name, loras,
+                       base_model_additional_loras=None, use_synthetic_refiner=False):
     global final_unet, final_clip, final_vae, final_refiner_unet, final_refiner_vae, final_expansion
 
     final_unet = None
@@ -184,8 +224,14 @@ def refresh_everything(refiner_model_name, base_model_name, loras, base_model_ad
     final_refiner_unet = None
     final_refiner_vae = None
 
+    if use_synthetic_refiner and refiner_model_name == 'None':
+        print('Synthetic Refiner Activated')
+        refresh_base_model(base_model_name)
+        synthesize_refiner_model()
+    else:
         refresh_refiner_model(refiner_model_name)
         refresh_base_model(base_model_name)
+
     refresh_loras(loras, base_model_additional_loras=base_model_additional_loras)
     assert_model_integrity()
 
@@ -193,14 +239,9 @@ def refresh_everything(refiner_model_name, base_model_name, loras, base_model_ad
     final_clip = model_base.clip_with_lora
     final_vae = model_base.vae
 
-    final_unet.model.diffusion_model.in_inpaint = False
-
     final_refiner_unet = model_refiner.unet_with_lora
     final_refiner_vae = model_refiner.vae
 
-    if final_refiner_unet is not None:
-        final_refiner_unet.model.diffusion_model.in_inpaint = False
-
     if final_expansion is None:
         final_expansion = FooocusExpansion()
 
@@ -212,13 +253,7 @@ def refresh_everything(refiner_model_name, base_model_name, loras, base_model_ad
 refresh_everything(
     refiner_model_name=modules.config.default_refiner_model_name,
     base_model_name=modules.config.default_base_model_name,
-    loras=[
-        (modules.config.default_lora_name, modules.config.default_lora_weight),
-        ('None', modules.config.default_lora_weight),
-        ('None', modules.config.default_lora_weight),
-        ('None', modules.config.default_lora_weight),
-        ('None', modules.config.default_lora_weight)
-    ]
+    loras=modules.config.default_loras
 )
 
 
@@ -263,32 +298,52 @@ def calculate_sigmas(sampler, model, scheduler, steps, denoise):
 
 @torch.no_grad()
 @torch.inference_mode()
-def process_diffusion(positive_cond, negative_cond, steps, switch, width, height, image_seed, callback, sampler_name, scheduler_name, latent=None, denoise=1.0, tiled=False, cfg_scale=7.0, refiner_swap_method='joint'):
-    global final_unet, final_refiner_unet, final_vae, final_refiner_vae
+def get_candidate_vae(steps, switch, denoise=1.0, refiner_swap_method='joint'):
+    assert refiner_swap_method in ['joint', 'separate', 'vae']
 
-    assert refiner_swap_method in ['joint', 'separate', 'vae', 'upscale']
-
-    refiner_use_different_vae = final_refiner_vae is not None and final_refiner_unet is not None
-
-    if refiner_swap_method == 'upscale':
-        if not refiner_use_different_vae:
-            refiner_swap_method = 'joint'
+    if final_refiner_vae is not None and final_refiner_unet is not None:
+        if denoise > 0.9:
+            return final_vae, final_refiner_vae
         else:
-        if refiner_use_different_vae:
-            if denoise > 0.95:
+            if denoise > (float(steps - switch) / float(steps)) ** 0.834:  # karras 0.834
+                return final_vae, None
+            else:
+                return final_refiner_vae, None
+
+    return final_vae, final_refiner_vae
+
+
+@torch.no_grad()
+@torch.inference_mode()
+def process_diffusion(positive_cond, negative_cond, steps, switch, width, height, image_seed, callback, sampler_name, scheduler_name, latent=None, denoise=1.0, tiled=False, cfg_scale=7.0, refiner_swap_method='joint'):
+    target_unet, target_vae, target_refiner_unet, target_refiner_vae, target_clip \
+        = final_unet, final_vae, final_refiner_unet, final_refiner_vae, final_clip
+
+    assert refiner_swap_method in ['joint', 'separate', 'vae']
+
+    if final_refiner_vae is not None and final_refiner_unet is not None:
+        # Refiner Use Different VAE (then it is SD15)
+        if denoise > 0.9:
             refiner_swap_method = 'vae'
         else:
-                # VAE swap only support full denoise
-                # Disable refiner to avoid SD15 in joint/separate swap
-                final_refiner_unet = None
-                final_refiner_vae = None
+            refiner_swap_method = 'joint'
+            if denoise > (float(steps - switch) / float(steps)) ** 0.834:  # karras 0.834
+                target_unet, target_vae, target_refiner_unet, target_refiner_vae \
+                    = final_unet, final_vae, None, None
+                print(f'[Sampler] only use Base because of partial denoise.')
+            else:
+                positive_cond = clip_separate(positive_cond, target_model=final_refiner_unet.model, target_clip=final_clip)
+                negative_cond = clip_separate(negative_cond, target_model=final_refiner_unet.model, target_clip=final_clip)
+                target_unet, target_vae, target_refiner_unet, target_refiner_vae \
+                    = final_refiner_unet, final_refiner_vae, None, None
+                print(f'[Sampler] only use Refiner because of partial denoise.')
 
     print(f'[Sampler] refiner_swap_method = {refiner_swap_method}')
 
     if latent is None:
-        empty_latent = core.generate_empty_latent(width=width, height=height, batch_size=1)
+        initial_latent = core.generate_empty_latent(width=width, height=height, batch_size=1)
     else:
-        empty_latent = latent
+        initial_latent = latent
 
     minmax_sigmas = calculate_sigmas(sampler=sampler_name, scheduler=scheduler_name, model=final_unet.model, steps=steps, denoise=denoise)
     sigma_min, sigma_max = minmax_sigmas[minmax_sigmas > 0].min(), minmax_sigmas.max()
@@ -297,18 +352,18 @@ def process_diffusion(positive_cond, negative_cond, steps, switch, width, height
     print(f'[Sampler] sigma_min = {sigma_min}, sigma_max = {sigma_max}')
 
     modules.patch.BrownianTreeNoiseSamplerPatched.global_init(
-        empty_latent['samples'].to(fcbh.model_management.get_torch_device()),
+        initial_latent['samples'].to(fcbh.model_management.get_torch_device()),
         sigma_min, sigma_max, seed=image_seed, cpu=False)
 
     decoded_latent = None
 
     if refiner_swap_method == 'joint':
         sampled_latent = core.ksampler(
-            model=final_unet,
-            refiner=final_refiner_unet,
+            model=target_unet,
+            refiner=target_refiner_unet,
             positive=positive_cond,
             negative=negative_cond,
-            latent=empty_latent,
+            latent=initial_latent,
             steps=steps, start_step=0, last_step=steps, disable_noise=False, force_full_denoise=True,
             seed=image_seed,
             denoise=denoise,
@@ -320,32 +375,14 @@ def process_diffusion(positive_cond, negative_cond, steps, switch, width, height
             previewer_start=0,
             previewer_end=steps,
         )
-        decoded_latent = core.decode_vae(vae=final_vae, latent_image=sampled_latent, tiled=tiled)
-
-    if refiner_swap_method == 'upscale':
-        sampled_latent = core.ksampler(
-            model=final_refiner_unet,
-            positive=clip_separate(positive_cond, target_model=final_refiner_unet.model, target_clip=final_clip),
-            negative=clip_separate(negative_cond, target_model=final_refiner_unet.model, target_clip=final_clip),
-            latent=empty_latent,
-            steps=steps, start_step=0, last_step=steps, disable_noise=False, force_full_denoise=True,
-            seed=image_seed,
-            denoise=denoise,
-            callback_function=callback,
-            cfg=cfg_scale,
-            sampler_name=sampler_name,
-            scheduler=scheduler_name,
-            previewer_start=0,
-            previewer_end=steps,
-        )
-        decoded_latent = core.decode_vae(vae=final_refiner_vae, latent_image=sampled_latent, tiled=tiled)
+        decoded_latent = core.decode_vae(vae=target_vae, latent_image=sampled_latent, tiled=tiled)
 
     if refiner_swap_method == 'separate':
         sampled_latent = core.ksampler(
-            model=final_unet,
+            model=target_unet,
             positive=positive_cond,
             negative=negative_cond,
-            latent=empty_latent,
+            latent=initial_latent,
             steps=steps, start_step=0, last_step=switch, disable_noise=False, force_full_denoise=False,
             seed=image_seed,
             denoise=denoise,
@@ -358,15 +395,15 @@ def process_diffusion(positive_cond, negative_cond, steps, switch, width, height
         )
         print('Refiner swapped by changing ksampler. Noise preserved.')
 
-        target_model = final_refiner_unet
+        target_model = target_refiner_unet
         if target_model is None:
-            target_model = final_unet
+            target_model = target_unet
             print('Use base model to refine itself - this may because of developer mode.')
 
         sampled_latent = core.ksampler(
             model=target_model,
-            positive=clip_separate(positive_cond, target_model=target_model.model, target_clip=final_clip),
-            negative=clip_separate(negative_cond, target_model=target_model.model, target_clip=final_clip),
+            positive=clip_separate(positive_cond, target_model=target_model.model, target_clip=target_clip),
+            negative=clip_separate(negative_cond, target_model=target_model.model, target_clip=target_clip),
             latent=sampled_latent,
             steps=steps, start_step=switch, last_step=steps, disable_noise=True, force_full_denoise=True,
             seed=image_seed,
@@ -379,9 +416,9 @@ def process_diffusion(positive_cond, negative_cond, steps, switch, width, height
             previewer_end=steps,
         )
 
-        target_model = final_refiner_vae
+        target_model = target_refiner_vae
         if target_model is None:
-            target_model = final_vae
+            target_model = target_vae
         decoded_latent = core.decode_vae(vae=target_model, latent_image=sampled_latent, tiled=tiled)
 
     if refiner_swap_method == 'vae':
@@ -391,10 +428,10 @@ def process_diffusion(positive_cond, negative_cond, steps, switch, width, height
             modules.inpaint_worker.current_task.unswap()
 
         sampled_latent = core.ksampler(
-            model=final_unet,
+            model=target_unet,
             positive=positive_cond,
             negative=negative_cond,
-            latent=empty_latent,
+            latent=initial_latent,
             steps=steps, start_step=0, last_step=switch, disable_noise=False, force_full_denoise=True,
             seed=image_seed,
             denoise=denoise,
@@ -407,9 +444,9 @@ def process_diffusion(positive_cond, negative_cond, steps, switch, width, height
         )
         print('Fooocus VAE-based swap.')
 
-        target_model = final_refiner_unet
+        target_model = target_refiner_unet
         if target_model is None:
-            target_model = final_unet
+            target_model = target_unet
             print('Use base model to refine itself - this may because of developer mode.')
 
         sampled_latent = vae_parse(sampled_latent)
@@ -429,8 +466,8 @@ def process_diffusion(positive_cond, negative_cond, steps, switch, width, height
 
         sampled_latent = core.ksampler(
             model=target_model,
-            positive=clip_separate(positive_cond, target_model=target_model.model, target_clip=final_clip),
-            negative=clip_separate(negative_cond, target_model=target_model.model, target_clip=final_clip),
+            positive=clip_separate(positive_cond, target_model=target_model.model, target_clip=target_clip),
+            negative=clip_separate(negative_cond, target_model=target_model.model, target_clip=target_clip),
             latent=sampled_latent,
             steps=len_sigmas, start_step=0, last_step=len_sigmas, disable_noise=False, force_full_denoise=True,
             seed=image_seed+1,
@@ -445,9 +482,9 @@ def process_diffusion(positive_cond, negative_cond, steps, switch, width, height
             noise_mean=noise_mean
         )
 
-        target_model = final_refiner_vae
+        target_model = target_refiner_vae
         if target_model is None:
-            target_model = final_vae
+            target_model = target_vae
         decoded_latent = core.decode_vae(vae=target_model, latent_image=sampled_latent, tiled=tiled)
 
     images = core.pytorch_to_numpy(decoded_latent)

modules/flags.py

@@ -10,23 +10,32 @@ uov_list = [
     disabled, subtle_variation, strong_variation, upscale_15, upscale_2, upscale_fast
 ]
 
-KSAMPLER_NAMES = ["euler", "euler_ancestral", "heun", "dpm_2", "dpm_2_ancestral",
+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"]
 
-SCHEDULER_NAMES = ["normal", "karras", "exponential", "sgm_uniform", "simple", "ddim_uniform"]
+SCHEDULER_NAMES = ["normal", "karras", "exponential", "sgm_uniform", "simple", "ddim_uniform", "lcm"]
 SAMPLER_NAMES = KSAMPLER_NAMES + ["ddim", "uni_pc", "uni_pc_bh2"]
 
 sampler_list = SAMPLER_NAMES
 scheduler_list = SCHEDULER_NAMES
 
-cn_ip = "Image Prompt"
+cn_ip = "ImagePrompt"
+cn_ip_face = "FaceSwap"
 cn_canny = "PyraCanny"
 cn_cpds = "CPDS"
 
-ip_list = [cn_ip, cn_canny, cn_cpds]
+ip_list = [cn_ip, cn_canny, cn_cpds, cn_ip_face]
 default_ip = cn_ip
 
 default_parameters = {
-    cn_ip: (0.5, 0.6), cn_canny: (0.5, 1.0), cn_cpds: (0.5, 1.0)
+    cn_ip: (0.5, 0.6), cn_ip_face: (0.9, 0.75), cn_canny: (0.5, 1.0), cn_cpds: (0.5, 1.0)
 }  # stop, weight
+
+inpaint_engine_versions = ['None', 'v1', 'v2.5', 'v2.6']
+performance_selections = ['Speed', 'Quality', 'Extreme Speed']
+
+inpaint_option_default = 'Inpaint or Outpaint (default)'
+inpaint_option_detail = 'Improve Detail (face, hand, eyes, etc.)'
+inpaint_option_modify = 'Modify Content (add objects, change background, etc.)'
+inpaint_options = [inpaint_option_default, inpaint_option_detail, inpaint_option_modify]

modules/html.py

@@ -100,6 +100,18 @@ progress::after {
   overflow: auto !important;
 }
 
+.aspect_ratios label {
+    width: 140px !important;
+}
+
+.aspect_ratios label span {
+    white-space: nowrap !important;
+}
+
+.aspect_ratios label input {
+    margin-left: -5px !important;
+}
+
 '''
 progress_html = '''
 <div class="loader-container">

modules/inpaint_worker.py

@@ -1,12 +1,12 @@
 import torch
 import numpy as np
-import modules.default_pipeline as pipeline
 
 from PIL import Image, ImageFilter
-from modules.util import resample_image, set_image_shape_ceil
+from modules.util import resample_image, set_image_shape_ceil, get_image_shape_ceil
+from modules.upscaler import perform_upscale
 
 
-inpaint_head = None
+inpaint_head_model = None
 
 
 class InpaintHead(torch.nn.Module):
@@ -77,29 +77,32 @@ def regulate_abcd(x, a, b, c, d):
 
 def compute_initial_abcd(x):
     indices = np.where(x)
-    a = np.min(indices[0]) - 64
-    b = np.max(indices[0]) + 65
-    c = np.min(indices[1]) - 64
-    d = np.max(indices[1]) + 65
+    a = np.min(indices[0])
+    b = np.max(indices[0])
+    c = np.min(indices[1])
+    d = np.max(indices[1])
     abp = (b + a) // 2
     abm = (b - a) // 2
     cdp = (d + c) // 2
     cdm = (d - c) // 2
-    l = max(abm, cdm)
+    l = int(max(abm, cdm) * 1.15)
     a = abp - l
-    b = abp + l
+    b = abp + l + 1
     c = cdp - l
-    d = cdp + l
+    d = cdp + l + 1
     a, b, c, d = regulate_abcd(x, a, b, c, d)
     return a, b, c, d
 
 
-def solve_abcd(x, a, b, c, d, outpaint):
+def solve_abcd(x, a, b, c, d, k):
+    k = float(k)
+    assert 0.0 <= k <= 1.0
+
     H, W = x.shape[:2]
-    if outpaint:
+    if k == 1.0:
         return 0, H, 0, W
     while True:
-        if b - a > H * 0.618 and d - c > W * 0.618:
+        if b - a >= H * k and d - c >= W * k:
             break
 
         add_h = (b - a) < (d - c)
@@ -138,20 +141,29 @@ def fooocus_fill(image, mask):
 
 
 class InpaintWorker:
-    def __init__(self, image, mask, is_outpaint):
+    def __init__(self, image, mask, use_fill=True, k=0.618):
         a, b, c, d = compute_initial_abcd(mask > 0)
-        a, b, c, d = solve_abcd(mask, a, b, c, d, outpaint=is_outpaint)
+        a, b, c, d = solve_abcd(mask, a, b, c, d, k=k)
 
         # interested area
         self.interested_area = (a, b, c, d)
         self.interested_mask = mask[a:b, c:d]
         self.interested_image = image[a:b, c:d]
 
+        # super resolution
+        if get_image_shape_ceil(self.interested_image) < 1024:
+            self.interested_image = perform_upscale(self.interested_image)
+
         # resize to make images ready for diffusion
         self.interested_image = set_image_shape_ceil(self.interested_image, 1024)
+        self.interested_fill = self.interested_image.copy()
         H, W, C = self.interested_image.shape
 
+        # process mask
         self.interested_mask = up255(resample_image(self.interested_mask, W, H), t=127)
+
+        # compute filling
+        if use_fill:
             self.interested_fill = fooocus_fill(self.interested_image, self.interested_mask)
 
         # soft pixels
@@ -166,34 +178,36 @@ class InpaintWorker:
         self.inpaint_head_feature = None
         return
 
-    def load_latent(self,
-                    latent_fill,
-                    latent_inpaint,
-                    latent_mask,
-                    latent_swap=None,
-                    inpaint_head_model_path=None):
-
-        global inpaint_head
-        assert inpaint_head_model_path is not None
-
+    def load_latent(self, latent_fill, latent_mask, latent_swap=None):
         self.latent = latent_fill
         self.latent_mask = latent_mask
         self.latent_after_swap = latent_swap
+        return
 
-        if inpaint_head is None:
-            inpaint_head = InpaintHead()
+    def patch(self, inpaint_head_model_path, inpaint_latent, inpaint_latent_mask, model):
+        global inpaint_head_model
+
+        if inpaint_head_model is None:
+            inpaint_head_model = InpaintHead()
             sd = torch.load(inpaint_head_model_path, map_location='cpu')
-            inpaint_head.load_state_dict(sd)
+            inpaint_head_model.load_state_dict(sd)
 
         feed = torch.cat([
-            latent_mask,
-            pipeline.final_unet.model.process_latent_in(latent_inpaint)
+            inpaint_latent_mask,
+            model.model.process_latent_in(inpaint_latent)
         ], dim=1)
 
-        inpaint_head.to(device=feed.device, dtype=feed.dtype)
-        self.inpaint_head_feature = inpaint_head(feed)
+        inpaint_head_model.to(device=feed.device, dtype=feed.dtype)
+        inpaint_head_feature = inpaint_head_model(feed)
 
-        return
+        def input_block_patch(h, transformer_options):
+            if transformer_options["block"][1] == 0:
+                h = h + inpaint_head_feature.to(h)
+            return h
+
+        m = model.clone()
+        m.set_model_input_block_patch(input_block_patch)
+        return m
 
     def swap(self):
         if self.swapped:
@@ -239,5 +253,5 @@ class InpaintWorker:
         return result
 
     def visualize_mask_processing(self):
-        return [self.interested_fill, self.interested_mask, self.image, self.mask]
+        return [self.interested_fill, self.interested_mask, self.interested_image]
 

modules/localization.py

@@ -2,29 +2,30 @@ import json
 import os
 
 
+current_translation = {}
 localization_root = os.path.join(os.path.dirname(os.path.dirname(__file__)), 'language')
 
 
 def localization_js(filename):
-    data = {}
+    global current_translation
 
     if isinstance(filename, str):
         full_name = os.path.abspath(os.path.join(localization_root, filename + '.json'))
         if os.path.exists(full_name):
             try:
                 with open(full_name, encoding='utf-8') as f:
-                    data = json.load(f)
-                    assert isinstance(data, dict)
-                    for k, v in data.items():
+                    current_translation = json.load(f)
+                    assert isinstance(current_translation, dict)
+                    for k, v in current_translation.items():
                         assert isinstance(k, str)
                         assert isinstance(v, str)
             except Exception as e:
                 print(str(e))
                 print(f'Failed to load localization file {full_name}')
 
-    # data = {k: 'XXX' for k in data.keys()}  # use this to see if all texts are covered
+    # current_translation = {k: 'XXX' for k in current_translation.keys()}  # use this to see if all texts are covered
 
-    return f"window.localization = {json.dumps(data)}"
+    return f"window.localization = {json.dumps(current_translation)}"
 
 
 def dump_english_config(components):

modules/lora.py

@@ -1,4 +1,4 @@
-def load_dangerous_lora(lora, to_load):
+def match_lora(lora, to_load):
     patch_dict = {}
     loaded_keys = set()
     for x in to_load:
@@ -136,7 +136,5 @@ def load_dangerous_lora(lora, to_load):
             patch_dict["{}.bias".format(to_load[x][:-len(".weight")])] = (diff_bias,)
             loaded_keys.add(diff_bias_name)
 
-    for x in lora.keys():
-        if x not in loaded_keys:
-            return {}
-    return patch_dict
+    remaining_dict = {x: y for x, y in lora.items() if x not in loaded_keys}
+    return patch_dict, remaining_dict

modules/patch.py

@@ -1,6 +1,8 @@
 import os
 import torch
+import math
 import time
+import numpy as np
 import fcbh.model_base
 import fcbh.ldm.modules.diffusionmodules.openaimodel
 import fcbh.samplers
@@ -22,8 +24,10 @@ import warnings
 import safetensors.torch
 import modules.constants as constants
 
+from einops import repeat
 from fcbh.k_diffusion.sampling import BatchedBrownianTree
-from fcbh.ldm.modules.diffusionmodules.openaimodel import forward_timestep_embed, apply_control, timestep_embedding
+from fcbh.ldm.modules.diffusionmodules.openaimodel import forward_timestep_embed, apply_control
+from fcbh.ldm.modules.diffusionmodules.util import make_beta_schedule
 
 
 sharpness = 2.0
@@ -300,15 +304,17 @@ def encode_token_weights_patched_with_a1111_method(self, token_weight_pairs):
 
 def patched_KSamplerX0Inpaint_forward(self, x, sigma, uncond, cond, cond_scale, denoise_mask, model_options={}, seed=None):
     if inpaint_worker.current_task is not None:
+        latent_processor = self.inner_model.inner_model.process_latent_in
+        inpaint_latent = latent_processor(inpaint_worker.current_task.latent).to(x)
+        inpaint_mask = inpaint_worker.current_task.latent_mask.to(x)
+
         if getattr(self, 'energy_generator', None) is None:
             # avoid bad results by using different seeds.
             self.energy_generator = torch.Generator(device='cpu').manual_seed((seed + 1) % constants.MAX_SEED)
 
-        latent_processor = self.inner_model.inner_model.process_latent_in
-        inpaint_latent = latent_processor(inpaint_worker.current_task.latent).to(x)
-        inpaint_mask = inpaint_worker.current_task.latent_mask.to(x)
         energy_sigma = sigma.reshape([sigma.shape[0]] + [1] * (len(x.shape) - 1))
-        current_energy = torch.randn(x.size(), dtype=x.dtype, generator=self.energy_generator, device="cpu").to(x) * energy_sigma
+        current_energy = torch.randn(
+            x.size(), dtype=x.dtype, generator=self.energy_generator, device="cpu").to(x) * energy_sigma
         x = x * inpaint_mask + (inpaint_latent + current_energy) * (1.0 - inpaint_mask)
 
         out = self.inner_model(x, sigma,
@@ -338,8 +344,27 @@ def timed_adm(y, timesteps):
     return y
 
 
+def patched_timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False):
+    # Consistent with Kohya to reduce differences between model training and inference.
+
+    if not repeat_only:
+        half = dim // 2
+        freqs = torch.exp(
+            -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
+        ).to(device=timesteps.device)
+        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 patched_cldm_forward(self, x, hint, timesteps, context, y=None, **kwargs):
-    t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).to(self.dtype)
+    t_emb = fcbh.ldm.modules.diffusionmodules.openaimodel.timestep_embedding(
+        timesteps, self.model_channels, repeat_only=False).to(self.dtype)
+
     emb = self.time_embed(t_emb)
 
     guided_hint = self.input_hint_block(hint, emb, context)
@@ -380,10 +405,6 @@ def patched_unet_forward(self, x, timesteps=None, context=None, y=None, control=
     self.current_step = 1.0 - timesteps.to(x) / 999.0
     global_diffusion_progress = float(self.current_step.detach().cpu().numpy().tolist()[0])
 
-    inpaint_fix = None
-    if getattr(self, 'in_inpaint', False) and inpaint_worker.current_task is not None:
-        inpaint_fix = inpaint_worker.current_task.inpaint_head_feature
-
     transformer_options["original_shape"] = list(x.shape)
     transformer_options["current_index"] = 0
     transformer_patches = transformer_options.get("patches", {})
@@ -391,7 +412,8 @@ def patched_unet_forward(self, x, timesteps=None, context=None, y=None, control=
     y = timed_adm(y, timesteps)
 
     hs = []
-    t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).to(self.dtype)
+    t_emb = fcbh.ldm.modules.diffusionmodules.openaimodel.timestep_embedding(
+        timesteps, self.model_channels, repeat_only=False).to(self.dtype)
     emb = self.time_embed(t_emb)
 
     if self.num_classes is not None:
@@ -402,14 +424,17 @@ def patched_unet_forward(self, x, timesteps=None, context=None, y=None, control=
     for id, module in enumerate(self.input_blocks):
         transformer_options["block"] = ("input", id)
         h = forward_timestep_embed(module, h, emb, context, transformer_options)
-
-        if inpaint_fix is not None:
-            if int(h.shape[1]) == int(inpaint_fix.shape[1]):
-                h = h + inpaint_fix.to(h)
-                inpaint_fix = None
-
         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)
@@ -439,6 +464,31 @@ def patched_unet_forward(self, x, timesteps=None, context=None, y=None, control=
         return self.out(h)
 
 
+def patched_register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000,
+                          linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+    # Consistent with Kohya to reduce differences between model training and inference.
+
+    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 = np.cumprod(alphas, axis=0)
+    timesteps, = betas.shape
+    self.num_timesteps = int(timesteps)
+    self.linear_start = linear_start
+    self.linear_end = linear_end
+    sigmas = torch.tensor(((1 - alphas_cumprod) / alphas_cumprod) ** 0.5, dtype=torch.float32)
+    self.set_sigmas(sigmas)
+    return
+
+
 def patched_load_models_gpu(*args, **kwargs):
     execution_start_time = time.perf_counter()
     y = fcbh.model_management.load_models_gpu_origin(*args, **kwargs)
@@ -494,6 +544,8 @@ def patch_all():
     fcbh.sd1_clip.ClipTokenWeightEncoder.encode_token_weights = encode_token_weights_patched_with_a1111_method
     fcbh.samplers.KSamplerX0Inpaint.forward = patched_KSamplerX0Inpaint_forward
     fcbh.k_diffusion.sampling.BrownianTreeNoiseSampler = BrownianTreeNoiseSamplerPatched
+    fcbh.ldm.modules.diffusionmodules.openaimodel.timestep_embedding = patched_timestep_embedding
+    fcbh.model_base.ModelSamplingDiscrete._register_schedule = patched_register_schedule
 
     warnings.filterwarnings(action='ignore', module='torchsde')
 

modules/private_logger.py

@@ -1,4 +1,5 @@
 import os
+import args_manager
 import modules.config
 
 from PIL import Image
@@ -16,6 +17,9 @@ def get_current_html_path():
 
 
 def log(img, dic, single_line_number=3):
+    if args_manager.args.disable_image_log:
+        return
+
     date_string, local_temp_filename, only_name = generate_temp_filename(folder=modules.config.path_outputs, extension='png')
     os.makedirs(os.path.dirname(local_temp_filename), exist_ok=True)
     Image.fromarray(img).save(local_temp_filename)
@@ -40,7 +44,7 @@ def log(img, dic, single_line_number=3):
                 item += f"<p>{k}: <b>{v}</b>, "
             else:
                 item += f"{k}: <b>{v}</b></p>\n"
-    item += f"<p><img src=\"{only_name}\" width=512 onerror=\"document.getElementById('{div_name}').style.display = 'none';\"></img></p><hr></div>\n"
+    item += f"<p><img src=\"{only_name}\" width=auto height=100% loading=lazy style=\"height:auto;max-width:512px\" onerror=\"document.getElementById('{div_name}').style.display = 'none';\"></img></p><hr></div>\n"
     existing_log = item + existing_log
 
     with open(html_name, 'w', encoding='utf-8') as f:

modules/sample_hijack.py

@@ -133,8 +133,7 @@ def sample_hacked(model, noise, positive, negative, cfg, device, sampler, sigmas
         extra_args['model_options'] = {k: {} if k == 'transformer_options' else v for k, v in extra_args['model_options'].items()}
 
         models, inference_memory = get_additional_models(positive_refiner, negative_refiner, current_refiner.model_dtype())
-        fcbh.model_management.load_models_gpu([current_refiner] + models, fcbh.model_management.batch_area_memory(
-            noise.shape[0] * noise.shape[2] * noise.shape[3]) + inference_memory)
+        fcbh.model_management.load_models_gpu([current_refiner] + models, current_refiner.memory_required(noise.shape) + inference_memory)
 
         model_wrap.inner_model = current_refiner.model
         print('Refiner Swapped')

modules/style_sorter.py

@@ -0,0 +1,56 @@
+import os
+import gradio as gr
+import modules.localization as localization
+import json
+
+
+all_styles = []
+
+
+def try_load_sorted_styles(style_names, default_selected):
+    global all_styles
+
+    all_styles = style_names
+
+    try:
+        if os.path.exists('sorted_styles.json'):
+            with open('sorted_styles.json', 'rt', encoding='utf-8') as fp:
+                sorted_styles = json.load(fp)
+                if len(sorted_styles) == len(all_styles):
+                    if all(x in all_styles for x in sorted_styles):
+                        if all(x in sorted_styles for x in all_styles):
+                            all_styles = sorted_styles
+    except Exception as e:
+        print('Load style sorting failed.')
+        print(e)
+
+    unselected = [y for y in all_styles if y not in default_selected]
+    all_styles = default_selected + unselected
+
+    return
+
+
+def sort_styles(selected):
+    global all_styles
+    unselected = [y for y in all_styles if y not in selected]
+    sorted_styles = selected + unselected
+    try:
+        with open('sorted_styles.json', 'wt', encoding='utf-8') as fp:
+            json.dump(sorted_styles, fp, indent=4)
+    except Exception as e:
+        print('Write style sorting failed.')
+        print(e)
+    all_styles = sorted_styles
+    return gr.CheckboxGroup.update(choices=sorted_styles)
+
+
+def localization_key(x):
+    return x + localization.current_translation.get(x, '')
+
+
+def search_styles(selected, query):
+    unselected = [y for y in all_styles if y not in selected]
+    matched = [y for y in unselected if query.lower() in localization_key(y).lower()] if len(query.replace(' ', '')) > 0 else []
+    unmatched = [y for y in unselected if y not in matched]
+    sorted_styles = matched + selected + unmatched
+    return gr.CheckboxGroup.update(choices=sorted_styles)

modules/ui_gradio_extensions.py

@@ -28,12 +28,20 @@ def javascript_html():
     localization_js_path = webpath('javascript/localization.js')
     zoom_js_path = webpath('javascript/zoom.js')
     edit_attention_js_path = webpath('javascript/edit-attention.js')
+    viewer_js_path = webpath('javascript/viewer.js')
+    image_viewer_js_path = webpath('javascript/imageviewer.js')
     head = f'<script type="text/javascript">{localization_js(args_manager.args.language)}</script>\n'
     head += f'<script type="text/javascript" src="{script_js_path}"></script>\n'
     head += f'<script type="text/javascript" src="{context_menus_js_path}"></script>\n'
     head += f'<script type="text/javascript" src="{localization_js_path}"></script>\n'
     head += f'<script type="text/javascript" src="{zoom_js_path}"></script>\n'
     head += f'<script type="text/javascript" src="{edit_attention_js_path}"></script>\n'
+    head += f'<script type="text/javascript" src="{viewer_js_path}"></script>\n'
+    head += f'<script type="text/javascript" src="{image_viewer_js_path}"></script>\n'
+
+    if args_manager.args.theme:
+        head += f'<script type="text/javascript">set_theme(\"{args_manager.args.theme}\");</script>\n'
+
     return head
 
 

modules/upscaler.py

@@ -1,5 +1,6 @@
 import os
 import torch
+import modules.core as core
 
 from fcbh_extras.chainner_models.architecture.RRDB import RRDBNet as ESRGAN
 from fcbh_extras.nodes_upscale_model import ImageUpscaleWithModel
@@ -13,6 +14,9 @@ model = None
 
 def perform_upscale(img):
     global model
+
+    print(f'Upscaling image with shape {str(img.shape)} ...')
+
     if model is None:
         sd = torch.load(model_filename)
         sdo = OrderedDict()
@@ -22,4 +26,9 @@ def perform_upscale(img):
         model = ESRGAN(sdo)
         model.cpu()
         model.eval()
-    return opImageUpscaleWithModel.upscale(model, img)[0]
+
+    img = core.numpy_to_pytorch(img)
+    img = opImageUpscaleWithModel.upscale(model, img)[0]
+    img = core.pytorch_to_numpy(img)[0]
+
+    return img

modules/util.py

@@ -79,7 +79,7 @@ def get_shape_ceil(h, w):
 
 
 def get_image_shape_ceil(im):
-    H, W, _ = im.shape
+    H, W = im.shape[:2]
     return get_shape_ceil(H, W)
 
 

presets/anime.json

@@ -1,12 +1,36 @@
 {
     "default_model": "bluePencilXL_v050.safetensors",
     "default_refiner": "DreamShaper_8_pruned.safetensors",
-    "default_lora": "sd_xl_offset_example-lora_1.0.safetensors",
     "default_refiner_switch": 0.667,
-    "default_lora_weight": 0.5,
+    "default_loras": [
+        [
+            "sd_xl_offset_example-lora_1.0.safetensors",
+            0.5
+        ],
+        [
+            "None",
+            1.0
+        ],
+        [
+            "None",
+            1.0
+        ],
+        [
+            "None",
+            1.0
+        ],
+        [
+            "None",
+            1.0
+        ]
+    ],
     "default_cfg_scale": 7.0,
+    "default_sample_sharpness": 2.0,
     "default_sampler": "dpmpp_2m_sde_gpu",
     "default_scheduler": "karras",
+    "default_performance": "Speed",
+    "default_prompt": "1girl, ",
+    "default_prompt_negative": "(embedding:unaestheticXLv31:0.8), low quality, watermark",
     "default_styles": [
         "Fooocus V2",
         "Fooocus Masterpiece",
@@ -15,8 +39,7 @@
         "SAI Enhance",
         "SAI Fantasy Art"
     ],
-    "default_prompt_negative": "(embedding:unaestheticXLv31:0.8), low quality, watermark",
-    "default_prompt": "1girl, ",
+    "default_aspect_ratio": "896*1152",
     "checkpoint_downloads": {
         "bluePencilXL_v050.safetensors": "https://huggingface.co/lllyasviel/fav_models/resolve/main/fav/bluePencilXL_v050.safetensors",
         "DreamShaper_8_pruned.safetensors": "https://huggingface.co/lllyasviel/fav_models/resolve/main/fav/DreamShaper_8_pruned.safetensors"
@@ -24,7 +47,6 @@
     "embeddings_downloads": {
         "unaestheticXLv31.safetensors": "https://huggingface.co/lllyasviel/fav_models/resolve/main/fav/unaestheticXLv31.safetensors"
     },
-    "default_aspect_ratio": "896*1152",
     "lora_downloads": {
         "sd_xl_offset_example-lora_1.0.safetensors": "https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_offset_example-lora_1.0.safetensors"
     }

presets/default.json

@@ -0,0 +1,47 @@
+{
+    "default_model": "juggernautXL_version6Rundiffusion.safetensors",
+    "default_refiner": "None",
+    "default_refiner_switch": 0.5,
+    "default_loras": [
+        [
+            "sd_xl_offset_example-lora_1.0.safetensors",
+            0.1
+        ],
+        [
+            "None",
+            1.0
+        ],
+        [
+            "None",
+            1.0
+        ],
+        [
+            "None",
+            1.0
+        ],
+        [
+            "None",
+            1.0
+        ]
+    ],
+    "default_cfg_scale": 4.0,
+    "default_sample_sharpness": 2.0,
+    "default_sampler": "dpmpp_2m_sde_gpu",
+    "default_scheduler": "karras",
+    "default_performance": "Speed",
+    "default_prompt": "",
+    "default_prompt_negative": "",
+    "default_styles": [
+        "Fooocus V2",
+        "Fooocus Enhance",
+        "Fooocus Sharp"
+    ],
+    "default_aspect_ratio": "1152*896",
+    "checkpoint_downloads": {
+        "juggernautXL_version6Rundiffusion.safetensors": "https://huggingface.co/lllyasviel/fav_models/resolve/main/fav/juggernautXL_version6Rundiffusion.safetensors"
+    },
+    "embeddings_downloads": {},
+    "lora_downloads": {
+        "sd_xl_offset_example-lora_1.0.safetensors": "https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_offset_example-lora_1.0.safetensors"
+    }
+}

presets/lcm.json

@@ -0,0 +1,45 @@
+{
+    "default_model": "juggernautXL_version6Rundiffusion.safetensors",
+    "default_refiner": "None",
+    "default_refiner_switch": 0.5,
+    "default_loras": [
+        [
+            "None",
+            1.0
+        ],
+        [
+            "None",
+            1.0
+        ],
+        [
+            "None",
+            1.0
+        ],
+        [
+            "None",
+            1.0
+        ],
+        [
+            "None",
+            1.0
+        ]
+    ],
+    "default_cfg_scale": 4.0,
+    "default_sample_sharpness": 2.0,
+    "default_sampler": "dpmpp_2m_sde_gpu",
+    "default_scheduler": "karras",
+    "default_performance": "Extreme Speed",
+    "default_prompt": "",
+    "default_prompt_negative": "",
+    "default_styles": [
+        "Fooocus V2",
+        "Fooocus Enhance",
+        "Fooocus Sharp"
+    ],
+    "default_aspect_ratio": "1152*896",
+    "checkpoint_downloads": {
+        "juggernautXL_version6Rundiffusion.safetensors": "https://huggingface.co/lllyasviel/fav_models/resolve/main/fav/juggernautXL_version6Rundiffusion.safetensors"
+    },
+    "embeddings_downloads": {},
+    "lora_downloads": {}
+}

presets/realistic.json

@@ -1,23 +1,46 @@
 {
     "default_model": "realisticStockPhoto_v10.safetensors",
     "default_refiner": "",
-    "default_lora": "SDXL_FILM_PHOTOGRAPHY_STYLE_BetaV0.4.safetensors",
-    "default_lora_weight": 0.25,
+    "default_refiner_switch": 0.5,
+    "default_loras": [
+        [
+            "SDXL_FILM_PHOTOGRAPHY_STYLE_BetaV0.4.safetensors",
+            0.25
+        ],
+        [
+            "None",
+            1.0
+        ],
+        [
+            "None",
+            1.0
+        ],
+        [
+            "None",
+            1.0
+        ],
+        [
+            "None",
+            1.0
+        ]
+    ],
     "default_cfg_scale": 3.0,
+    "default_sample_sharpness": 2.0,
     "default_sampler": "dpmpp_2m_sde_gpu",
     "default_scheduler": "karras",
+    "default_performance": "Speed",
+    "default_prompt": "",
+    "default_prompt_negative": "unrealistic, saturated, high contrast, big nose, painting, drawing, sketch, cartoon, anime, manga, render, CG, 3d, watermark, signature, label",
     "default_styles": [
         "Fooocus V2",
         "Fooocus Photograph",
         "Fooocus Negative"
     ],
-    "default_prompt_negative": "unrealistic, saturated, high contrast, big nose, painting, drawing, sketch, cartoon, anime, manga, render, CG, 3d, watermark, signature, label",
-    "default_prompt": "",
+    "default_aspect_ratio": "896*1152",
     "checkpoint_downloads": {
         "realisticStockPhoto_v10.safetensors": "https://huggingface.co/lllyasviel/fav_models/resolve/main/fav/realisticStockPhoto_v10.safetensors"
     },
     "embeddings_downloads": {},
-    "default_aspect_ratio": "896*1152",
     "lora_downloads": {
         "SDXL_FILM_PHOTOGRAPHY_STYLE_BetaV0.4.safetensors": "https://huggingface.co/lllyasviel/fav_models/resolve/main/fav/SDXL_FILM_PHOTOGRAPHY_STYLE_BetaV0.4.safetensors"
     }

presets/sai.json

@@ -0,0 +1,47 @@
+{
+    "default_model": "sd_xl_base_1.0_0.9vae.safetensors",
+    "default_refiner": "sd_xl_refiner_1.0_0.9vae.safetensors",
+    "default_refiner_switch": 0.75,
+    "default_loras": [
+        [
+            "sd_xl_offset_example-lora_1.0.safetensors",
+            0.5
+        ],
+        [
+            "None",
+            1.0
+        ],
+        [
+            "None",
+            1.0
+        ],
+        [
+            "None",
+            1.0
+        ],
+        [
+            "None",
+            1.0
+        ]
+    ],
+    "default_cfg_scale": 7.0,
+    "default_sample_sharpness": 2.0,
+    "default_sampler": "dpmpp_2m_sde_gpu",
+    "default_scheduler": "karras",
+    "default_performance": "Speed",
+    "default_prompt": "",
+    "default_prompt_negative": "",
+    "default_styles": [
+        "Fooocus V2",
+        "Fooocus Cinematic"
+    ],
+    "default_aspect_ratio": "1152*896",
+    "checkpoint_downloads": {
+        "sd_xl_base_1.0_0.9vae.safetensors": "https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_base_1.0_0.9vae.safetensors",
+        "sd_xl_refiner_1.0_0.9vae.safetensors": "https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0/resolve/main/sd_xl_refiner_1.0_0.9vae.safetensors"
+    },
+    "embeddings_downloads": {},
+    "lora_downloads": {
+        "sd_xl_offset_example-lora_1.0.safetensors": "https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_offset_example-lora_1.0.safetensors"
+    }
+}

readme.md

@@ -50,6 +50,7 @@ Using Fooocus is as easy as (probably easier than) Midjourney – but this does
 | Prompt Weights | You can use " I am (happy:1.5)". <br> Fooocus uses A1111's reweighting algorithm so that results are better than ComfyUI if users directly copy prompts from Civitai. (Because if prompts are written in ComfyUI's reweighting, users are less likely to copy prompt texts as they prefer dragging files) <br> To use embedding, you can use "(embedding:file_name:1.1)" |
 | --no | Advanced -> Negative Prompt |
 | --ar | Advanced -> Aspect Ratios |
+| InsightFace | Input Image -> Image Prompt -> Advanced -> FaceSwap |
 
 We also have a few things borrowed from the best parts of LeonardoAI:
 
@@ -67,7 +68,7 @@ Fooocus also developed many "fooocus-only" features for advanced users to get pe
 
 You can directly download Fooocus with:
 
-**[>>> Click here to download <<<](https://github.com/lllyasviel/Fooocus/releases/download/release/Fooocus_win64_2-1-754.7z)**
+**[>>> Click here to download <<<](https://github.com/lllyasviel/Fooocus/releases/download/release/Fooocus_win64_2-1-791.7z)**
 
 After you download the file, please uncompress it, and then run the "run.bat".
 
@@ -76,7 +77,7 @@ After you download the file, please uncompress it, and then run the "run.bat".
 In the first time you launch the software, it will automatically download models:
 
 1. It will download [default models](#models) to the folder "Fooocus\models\checkpoints" given different presets. You can download them in advance if you do not want automatic download.
-2. Note that if you use inpaint, at the first time you inpaint an image, it will download [Fooocus's own inpaint control model from here](https://huggingface.co/lllyasviel/fooocus_inpaint/resolve/main/inpaint.fooocus.patch) as the file "Fooocus\models\inpaint\inpaint.fooocus.patch" (the size of this file is 1.28GB).
+2. Note that if you use inpaint, at the first time you inpaint an image, it will download [Fooocus's own inpaint control model from here](https://huggingface.co/lllyasviel/fooocus_inpaint/resolve/main/inpaint_v26.fooocus.patch) as the file "Fooocus\models\inpaint\inpaint_v26.fooocus.patch" (the size of this file is 1.28GB).
 
 After Fooocus 2.1.60, you will also have `run_anime.bat` and `run_realistic.bat`. They are different model presets (and requires different models, but thet will be automatically downloaded). [Check here for more details](https://github.com/lllyasviel/Fooocus/discussions/679).
 
@@ -107,7 +108,7 @@ Please open an issue if you use similar devices but still cannot achieve accepta
 
 ### Colab
 
-(Last tested - 2023 Oct 10)
+(Last tested - 2023 Nov 15)
 
 | Colab | Info
 | --- | --- |
@@ -276,8 +277,7 @@ For example, an edited `Fooocus\config.txt` (this file will be generated after t
     "path_outputs": "D:\\Fooocus\\outputs",
     "default_model": "realisticStockPhoto_v10.safetensors",
     "default_refiner": "",
-    "default_lora": "",
-    "default_lora_weight": 0.25,
+    "default_loras": [["lora_filename_1.safetensors", 0.5], ["lora_filename_2.safetensors", 0.5]],
     "default_cfg_scale": 3.0,
     "default_sampler": "dpmpp_2m",
     "default_scheduler": "karras",

update_log.md

@@ -1,3 +1,117 @@
+**(2023 Nov 26) Hi all, the feature updating of Fooocus will be paused for about two or three weeks because we have some other workloads. See you soon and we will come back in mid December. However, you may still see updates if other collaborators are fixing bugs or solving problems.**
+
+# 2.1.823
+
+* Fix some potential problem when LoRAs has clip keys and user want to load those LoRAs to refiners.
+
+# 2.1.822
+
+* New inpaint system (inpaint beta test ends).
+
+# 2.1.821
+
+* New UI for LoRAs.
+* Improved preset system: normalized preset keys and file names.
+* Improved session system: now multiple users can use one Fooocus at the same time without seeing others' results.
+* Improved some computation related to model precision.
+* Improved config loading system with user-friendly prints.
+
+# 2.1.820
+
+* support "--disable-image-log" to prevent writing images and logs to hard drive.
+
+# 2.1.819
+
+* Allow disabling preview in dev tools.
+
+# 2.1.818
+
+* Fix preset lora failed to load when the weight is exactly one.
+
+# 2.1.817
+
+* support "--theme dark" and "--theme light".
+* added preset files "default" and "lcm", these presets exist but will not create launcher files (will not be exposed to users) to keep entry clean. The "--preset lcm" is equivalent to select "Extreme Speed" in UI, but will likely to make some online service deploying easier.
+
+# 2.1.815
+
+* Multiple loras in preset.
+
+# 2.1.814
+
+* Allow using previous preset of official SAI SDXL by modify the args to '--preset sai'. ~Note that this preset will set inpaint engine back to previous v1 to get same results like before. To change the inpaint engine to v2.6, use the dev tools -> inpaint engine -> v2.6.~ (update: it is not needed now after some tests.)
+
+# 2.1.813
+
+* Allow preset to set default inpaint engine.
+
+# 2.1.812
+
+* Allow preset to set default performance.
+* heunpp2 sampler.
+
+# 2.1.810
+
+* Added hints to config_modification_tutorial.txt
+* Removed user hacked aspect ratios in I18N english templates, but it will still be read like before.
+* fix some style sorting problem again (perhaps should try Gradio 4.0 later).
+* Refreshed I18N english templates with more keys.
+
+# 2.1.809
+
+* fix some sorting problem.
+
+# 2.1.808
+
+* Aspect ratios now show aspect ratios.
+* Added style search.
+* Added style sorting/ordering/favorites.
+
+# 2.1.807
+
+* Click on image to see it in full screen.
+
+# 2.1.806
+
+* Fix some lora problems related to clip.
+
+# 2.1.805
+
+* Responsive UI for small screens.
+* Added skip preprocessor in dev tools.
+
+# 2.1.802
+
+* Default inpaint engine changed to v2.6. You can still use inpaint engine v1 in dev tools.
+* Fix some VRAM problems.
+
+# 2.1.799
+
+* Added 'Extreme Speed' performance mode (based on LCM). The previous complicated settings are not needed now.
+
+# 2.1.798
+
+* added lcm scheduler - LCM may need to set both sampler and scheduler to "lcm". Other than that, see the description in 2.1.782 logs.
+
+# 2.1.797
+
+* fixed some dependency problems with facexlib and filterpy.
+
+# 2.1.793
+
+* Added many javascripts to improve user experience. Now users with small screen will always see full canvas without needing to scroll.
+
+# 2.1.790
+
+* Face swap (in line with Midjourney InsightFace): Input Image -> Image Prompt -> Advanced -> FaceSwap
+* The performance is super high. Use it carefully and never use it in any illegal things!
+* This implementation will crop faces for you and you do NOT need to crop faces before feeding images into Fooocus. (If you previously manually crop faces from images for other software, you do not need to do that now in Fooocus.)
+
+# 2.1.788
+
+* Fixed some math problems in previous versions.
+* Inpaint engine v2.6 join the beta test of Fooocus inpaint models. Use it in dev tools -> inpaint engine -> v2.6 .
+
 # 2.1.785
 
 * The `user_path_config.txt` is deprecated since 2.1.785. If you are using it right now, please use the new `config.txt` instead. See also the new documentation in the Readme.
@@ -16,7 +130,7 @@ Now when you load a lora, the following things will happen:
 
 In this way, Fooocus 2.1.782 can benefit from all models and loras from CivitAI with both SDXL and SD1.5 ecosystem, using the unique Fooocus swap algorithm, to achieve extremely high quality results (although the default setting is already very high quality), especially in some anime use cases, if users really want to play with all these things.
 
-Recently the community also developed LCM loras. Users can use it by setting the scheduler as 'LCM' and setting the forced overwrite of step as 4 to 8 in dev tools. If LCM's feedback in the Artists community is good (not the feedback in the programmer community of Stable Diffusion), fooocus may add some other shortcuts in the future.
+Recently the community also developed LCM loras. Users can use it by setting the sampler as 'LCM', scheduler as 'sgm_uniform' (Update in 2.1.798: scheduler should also be "lcm"), the forced overwrite of sampling step as 4 to 8, and CFG guidance as 1.0, in dev tools. Do not forget to change the LCM lora weight to 1.0 (many people forget this and report failure cases). Also, set refiner to None. If LCM's feedback in the artists community is good (not the feedback in the programmer community of Stable Diffusion), Fooocus may add some other shortcuts in the future.
 
 # 2.1.781
 

webui.py

@@ -11,7 +11,9 @@ import modules.constants as constants
 import modules.flags as flags
 import modules.gradio_hijack as grh
 import modules.advanced_parameters as advanced_parameters
+import modules.style_sorter as style_sorter
 import args_manager
+import copy
 
 from modules.sdxl_styles import legal_style_names
 from modules.private_logger import get_current_html_path
@@ -20,29 +22,33 @@ from modules.auth import auth_enabled, check_auth
 
 
 def generate_clicked(*args):
+    import fcbh.model_management as model_management
+
+    with model_management.interrupt_processing_mutex:
+        model_management.interrupt_processing = False
+
     # outputs=[progress_html, progress_window, progress_gallery, gallery]
 
     execution_start_time = time.perf_counter()
+    task = worker.AsyncTask(args=list(args))
+    finished = False
 
-    worker.outputs = []
-
-    yield gr.update(visible=True, value=modules.html.make_progress_html(1, 'Initializing ...')), \
+    yield gr.update(visible=True, value=modules.html.make_progress_html(1, 'Waiting for task to start ...')), \
         gr.update(visible=True, value=None), \
         gr.update(visible=False, value=None), \
         gr.update(visible=False)
 
-    worker.buffer.append(list(args))
-    finished = False
+    worker.async_tasks.append(task)
 
     while not finished:
         time.sleep(0.01)
-        if len(worker.outputs) > 0:
-            flag, product = worker.outputs.pop(0)
+        if len(task.yields) > 0:
+            flag, product = task.yields.pop(0)
             if flag == 'preview':
 
                 # help bad internet connection by skipping duplicated preview
-                if len(worker.outputs) > 0:  # if we have the next item
-                    if worker.outputs[0][0] == 'preview':   # if the next item is also a preview
+                if len(task.yields) > 0:  # if we have the next item
+                    if task.yields[0][0] == 'preview':   # if the next item is also a preview
                         # print('Skipped one preview for better internet connection.')
                         continue
 
@@ -70,18 +76,28 @@ def generate_clicked(*args):
 
 reload_javascript()
 
+title = f'Fooocus {fooocus_version.version}'
+
+if isinstance(args_manager.args.preset, str):
+    title += ' ' + args_manager.args.preset
+
 shared.gradio_root = gr.Blocks(
-    title=f'Fooocus {fooocus_version.version} ' + ('' if args_manager.args.preset is None else args_manager.args.preset),
+    title=title,
     css=modules.html.css).queue()
 
 with shared.gradio_root:
     with gr.Row():
         with gr.Column(scale=2):
             with gr.Row():
-                progress_window = grh.Image(label='Preview', show_label=True, height=640, visible=False)
-                progress_gallery = gr.Gallery(label='Finished Images', show_label=True, object_fit='contain', height=640, visible=False)
-            progress_html = gr.HTML(value=modules.html.make_progress_html(32, 'Progress 32%'), visible=False, elem_id='progress-bar', elem_classes='progress-bar')
-            gallery = gr.Gallery(label='Gallery', show_label=False, object_fit='contain', height=745, visible=True, elem_classes='resizable_area')
+                progress_window = grh.Image(label='Preview', show_label=True, visible=False, height=768,
+                                            elem_classes=['main_view'])
+                progress_gallery = gr.Gallery(label='Finished Images', show_label=True, object_fit='contain',
+                                              height=768, visible=False, elem_classes=['main_view', 'image_gallery'])
+            progress_html = gr.HTML(value=modules.html.make_progress_html(32, 'Progress 32%'), visible=False,
+                                    elem_id='progress-bar', elem_classes='progress-bar')
+            gallery = gr.Gallery(label='Gallery', show_label=False, object_fit='contain', visible=True, height=768,
+                                 elem_classes=['resizable_area', 'main_view', 'final_gallery', 'image_gallery'],
+                                 elem_id='final_gallery')
             with gr.Row(elem_classes='type_row'):
                 with gr.Column(scale=17):
                     prompt = gr.Textbox(show_label=False, placeholder="Type prompt here.", elem_id='positive_prompt',
@@ -108,8 +124,9 @@ with shared.gradio_root:
                         model_management.interrupt_current_processing()
                         return
 
-                    stop_button.click(stop_clicked, outputs=[skip_button, stop_button], queue=False, _js='cancelGenerateForever')
-                    skip_button.click(skip_clicked, queue=False)
+                    stop_button.click(stop_clicked, outputs=[skip_button, stop_button],
+                                      queue=False, show_progress=False, _js='cancelGenerateForever')
+                    skip_button.click(skip_clicked, queue=False, show_progress=False)
             with gr.Row(elem_classes='advanced_check_row'):
                 input_image_checkbox = gr.Checkbox(label='Input Image', value=False, container=False, elem_classes='min_check')
                 advanced_checkbox = gr.Checkbox(label='Advanced', value=modules.config.default_advanced_checkbox, container=False, elem_classes='min_check')
@@ -163,19 +180,25 @@ with shared.gradio_root:
                                 [flags.default_parameters[flags.default_ip][1]] * len(ip_weights)
 
                         ip_advanced.change(ip_advance_checked, inputs=ip_advanced,
-                                           outputs=ip_ad_cols + ip_types + ip_stops + ip_weights, queue=False)
+                                           outputs=ip_ad_cols + ip_types + ip_stops + ip_weights,
+                                           queue=False, show_progress=False)
 
-                    with gr.TabItem(label='Inpaint or Outpaint (beta)') as inpaint_tab:
+                    with gr.TabItem(label='Inpaint or Outpaint') as inpaint_tab:
                         inpaint_input_image = grh.Image(label='Drag above image to here', source='upload', type='numpy', tool='sketch', height=500, brush_color="#FFFFFF", elem_id='inpaint_canvas')
-                        gr.HTML('Outpaint Expansion Direction:')
-                        outpaint_selections = gr.CheckboxGroup(choices=['Left', 'Right', 'Top', 'Bottom'], value=[], label='Outpaint', show_label=False, container=False)
-                        gr.HTML('* Powered by Fooocus Inpaint Engine (beta) <a href="https://github.com/lllyasviel/Fooocus/discussions/414" target="_blank">\U0001F4D4 Document</a>')
+                        with gr.Row():
+                            inpaint_additional_prompt = gr.Textbox(placeholder="Describe what you want to inpaint.", elem_id='inpaint_additional_prompt', label='Inpaint Additional Prompt', visible=False)
+                            outpaint_selections = gr.CheckboxGroup(choices=['Left', 'Right', 'Top', 'Bottom'], value=[], label='Outpaint Direction')
+                            inpaint_mode = gr.Dropdown(choices=modules.flags.inpaint_options, value=modules.flags.inpaint_option_default, label='Method')
+                        example_inpaint_prompts = gr.Dataset(samples=modules.config.example_inpaint_prompts, label='Additional Prompt Quick List', components=[inpaint_additional_prompt], visible=False)
+                        gr.HTML('* Powered by Fooocus Inpaint Engine <a href="https://github.com/lllyasviel/Fooocus/discussions/414" target="_blank">\U0001F4D4 Document</a>')
+                        example_inpaint_prompts.click(lambda x: x[0], inputs=example_inpaint_prompts, outputs=inpaint_additional_prompt, show_progress=False, queue=False)
 
-            switch_js = "(x) => {if(x){setTimeout(() => window.scrollTo({ top: 850, behavior: 'smooth' }), 50);}else{setTimeout(() => window.scrollTo({ top: 0, behavior: 'smooth' }), 50);} return x}"
-            down_js = "() => {setTimeout(() => window.scrollTo({ top: 850, behavior: 'smooth' }), 50);}"
+            switch_js = "(x) => {if(x){viewer_to_bottom(100);viewer_to_bottom(500);}else{viewer_to_top();} return x;}"
+            down_js = "() => {viewer_to_bottom();}"
 
-            input_image_checkbox.change(lambda x: gr.update(visible=x), inputs=input_image_checkbox, outputs=image_input_panel, queue=False, _js=switch_js)
-            ip_advanced.change(lambda: None, queue=False, _js=down_js)
+            input_image_checkbox.change(lambda x: gr.update(visible=x), inputs=input_image_checkbox,
+                                        outputs=image_input_panel, queue=False, show_progress=False, _js=switch_js)
+            ip_advanced.change(lambda: None, queue=False, show_progress=False, _js=down_js)
 
             current_tab = gr.Textbox(value='uov', visible=False)
             uov_tab.select(lambda: 'uov', outputs=current_tab, queue=False, _js=down_js, show_progress=False)
@@ -184,9 +207,12 @@ with shared.gradio_root:
 
         with gr.Column(scale=1, visible=modules.config.default_advanced_checkbox) as advanced_column:
             with gr.Tab(label='Setting'):
-                performance_selection = gr.Radio(label='Performance', choices=['Speed', 'Quality'], value='Speed')
+                performance_selection = gr.Radio(label='Performance',
+                                                 choices=modules.flags.performance_selections,
+                                                 value=modules.config.default_performance)
                 aspect_ratios_selection = gr.Radio(label='Aspect Ratios', choices=modules.config.available_aspect_ratios,
-                                                   value=modules.config.default_aspect_ratio, info='width × height')
+                                                   value=modules.config.default_aspect_ratio, info='width × height',
+                                                   elem_classes='aspect_ratios')
                 image_number = gr.Slider(label='Image Number', minimum=1, maximum=32, step=1, value=modules.config.default_image_number)
                 negative_prompt = gr.Textbox(label='Negative Prompt', show_label=True, placeholder="Type prompt here.",
                                              info='Describing what you do not want to see.', lines=2,
@@ -210,16 +236,47 @@ with shared.gradio_root:
                             pass
                         return random.randint(constants.MIN_SEED, constants.MAX_SEED)
 
-                seed_random.change(random_checked, inputs=[seed_random], outputs=[image_seed], queue=False)
+                seed_random.change(random_checked, inputs=[seed_random], outputs=[image_seed],
+                                   queue=False, show_progress=False)
 
+                if not args_manager.args.disable_image_log:
                     gr.HTML(f'<a href="/file={get_current_html_path()}" target="_blank">\U0001F4DA History Log</a>')
 
             with gr.Tab(label='Style'):
+                style_sorter.try_load_sorted_styles(
+                    style_names=legal_style_names,
+                    default_selected=modules.config.default_styles)
+
+                style_search_bar = gr.Textbox(show_label=False, container=False,
+                                              placeholder="\U0001F50E Type here to search styles ...",
+                                              value="",
+                                              label='Search Styles')
                 style_selections = gr.CheckboxGroup(show_label=False, container=False,
-                                                    choices=legal_style_names,
-                                                    value=modules.config.default_styles,
-                                                    label='Image Style')
+                                                    choices=copy.deepcopy(style_sorter.all_styles),
+                                                    value=copy.deepcopy(modules.config.default_styles),
+                                                    label='Selected Styles',
+                                                    elem_classes=['style_selections'])
+                gradio_receiver_style_selections = gr.Textbox(elem_id='gradio_receiver_style_selections', visible=False)
+
+                shared.gradio_root.load(lambda: gr.update(choices=copy.deepcopy(style_sorter.all_styles)),
+                                        outputs=style_selections)
+
+                style_search_bar.change(style_sorter.search_styles,
+                                        inputs=[style_selections, style_search_bar],
+                                        outputs=style_selections,
+                                        queue=False,
+                                        show_progress=False).then(
+                    lambda: None, _js='()=>{refresh_style_localization();}')
+
+                gradio_receiver_style_selections.input(style_sorter.sort_styles,
+                                                       inputs=style_selections,
+                                                       outputs=style_selections,
+                                                       queue=False,
+                                                       show_progress=False).then(
+                    lambda: None, _js='()=>{refresh_style_localization();}')
+
             with gr.Tab(label='Model'):
+                with gr.Group():
                     with gr.Row():
                         base_model = gr.Dropdown(label='Base Model (SDXL only)', choices=modules.config.model_filenames, value=modules.config.default_base_model_name, show_label=True)
                         refiner_model = gr.Dropdown(label='Refiner (SDXL or SD 1.5)', choices=['None'] + modules.config.model_filenames, value=modules.config.default_refiner_model_name, show_label=True)
@@ -235,25 +292,31 @@ with shared.gradio_root:
                     refiner_model.change(lambda x: gr.update(visible=x != 'None'),
                                          inputs=refiner_model, outputs=refiner_switch, show_progress=False, queue=False)
 
-                with gr.Accordion(label='LoRAs (SDXL or SD 1.5)', open=True):
+                with gr.Group():
                     lora_ctrls = []
-                    for i in range(5):
+
+                    for i, (n, v) in enumerate(modules.config.default_loras):
                         with gr.Row():
-                            lora_model = gr.Dropdown(label=f'LoRA {i+1}', choices=['None'] + modules.config.lora_filenames, value=modules.config.default_lora_name if i == 0 else 'None')
-                            lora_weight = gr.Slider(label='Weight', minimum=-2, maximum=2, step=0.01, value=modules.config.default_lora_weight)
+                            lora_model = gr.Dropdown(label=f'LoRA {i + 1}',
+                                                     choices=['None'] + modules.config.lora_filenames, value=n)
+                            lora_weight = gr.Slider(label='Weight', minimum=-2, maximum=2, step=0.01, value=v,
+                                                    elem_classes='lora_weight')
                             lora_ctrls += [lora_model, lora_weight]
+
                 with gr.Row():
                     model_refresh = gr.Button(label='Refresh', value='\U0001f504 Refresh All Files', variant='secondary', elem_classes='refresh_button')
             with gr.Tab(label='Advanced'):
-                sharpness = gr.Slider(label='Sampling Sharpness', minimum=0.0, maximum=30.0, step=0.001, value=modules.config.default_sample_sharpness,
-                                      info='Higher value means image and texture are sharper.')
-                guidance_scale = gr.Slider(label='Guidance Scale', minimum=1.0, maximum=30.0, step=0.01, value=modules.config.default_cfg_scale,
+                guidance_scale = gr.Slider(label='Guidance Scale', minimum=1.0, maximum=30.0, step=0.01,
+                                           value=modules.config.default_cfg_scale,
                                            info='Higher value means style is cleaner, vivider, and more artistic.')
+                sharpness = gr.Slider(label='Image Sharpness', minimum=0.0, maximum=30.0, step=0.001,
+                                      value=modules.config.default_sample_sharpness,
+                                      info='Higher value means image and texture are sharper.')
                 gr.HTML('<a href="https://github.com/lllyasviel/Fooocus/discussions/117" target="_blank">\U0001F4D4 Document</a>')
                 dev_mode = gr.Checkbox(label='Developer Debug Mode', value=False, container=False)
 
                 with gr.Column(visible=False) as dev_tools:
-                    with gr.Tab(label='Developer Debug Tools'):
+                    with gr.Tab(label='Debug Tools'):
                         adm_scaler_positive = gr.Slider(label='Positive ADM Guidance Scaler', minimum=0.1, maximum=3.0,
                                                         step=0.001, value=1.5, info='The scaler multiplied to positive ADM (use 1.0 to disable). ')
                         adm_scaler_negative = gr.Slider(label='Negative ADM Guidance Scaler', minimum=0.1, maximum=3.0,
@@ -265,25 +328,26 @@ with shared.gradio_root:
                         refiner_swap_method = gr.Dropdown(label='Refiner swap method', value='joint',
                                                           choices=['joint', 'separate', 'vae'])
 
-                        adaptive_cfg = gr.Slider(label='CFG Mimicking from TSNR', minimum=1.0, maximum=30.0, step=0.01, value=7.0,
+                        adaptive_cfg = gr.Slider(label='CFG Mimicking from TSNR', minimum=1.0, maximum=30.0, step=0.01,
+                                                 value=modules.config.default_cfg_tsnr,
                                                  info='Enabling Fooocus\'s implementation of CFG mimicking for TSNR '
                                                       '(effective when real CFG > mimicked CFG).')
                         sampler_name = gr.Dropdown(label='Sampler', choices=flags.sampler_list,
-                                                   value=modules.config.default_sampler,
-                                                   info='Only effective in non-inpaint mode.')
+                                                   value=modules.config.default_sampler)
                         scheduler_name = gr.Dropdown(label='Scheduler', choices=flags.scheduler_list,
-                                                     value=modules.config.default_scheduler,
-                                                     info='Scheduler of Sampler.')
+                                                     value=modules.config.default_scheduler)
 
                         generate_image_grid = gr.Checkbox(label='Generate Image Grid for Each Batch',
                                                           info='(Experimental) This may cause performance problems on some computers and certain internet conditions.',
                                                           value=False)
 
                         overwrite_step = gr.Slider(label='Forced Overwrite of Sampling Step',
-                                                   minimum=-1, maximum=200, step=1, value=-1,
+                                                   minimum=-1, maximum=200, step=1,
+                                                   value=modules.config.default_overwrite_step,
                                                    info='Set as -1 to disable. For developer debugging.')
                         overwrite_switch = gr.Slider(label='Forced Overwrite of Refiner Switch Step',
-                                                     minimum=-1, maximum=200, step=1, value=-1,
+                                                     minimum=-1, maximum=200, step=1,
+                                                     value=modules.config.default_overwrite_switch,
                                                      info='Set as -1 to disable. For developer debugging.')
                         overwrite_width = gr.Slider(label='Forced Overwrite of Generating Width',
                                                     minimum=-1, maximum=2048, step=1, value=-1,
@@ -299,12 +363,14 @@ with shared.gradio_root:
                         overwrite_upscale_strength = gr.Slider(label='Forced Overwrite of Denoising Strength of "Upscale"',
                                                                minimum=-1, maximum=1.0, step=0.001, value=-1,
                                                                info='Set as negative number to disable. For developer debugging.')
+                        disable_preview = gr.Checkbox(label='Disable Preview', value=False,
+                                                      info='Disable preview during generation.')
 
-                        inpaint_engine = gr.Dropdown(label='Inpaint Engine', value='v1', choices=['v1', 'v2.5'],
-                                                     info='Version of Fooocus inpaint model')
-
-                    with gr.Tab(label='Control Debug'):
-                        debugging_cn_preprocessor = gr.Checkbox(label='Debug Preprocessors', value=False)
+                    with gr.Tab(label='Control'):
+                        debugging_cn_preprocessor = gr.Checkbox(label='Debug Preprocessors', value=False,
+                                                                info='See the results from preprocessors.')
+                        skipping_cn_preprocessor = gr.Checkbox(label='Skip Preprocessors', value=False,
+                                                               info='Do not preprocess images. (Inputs are already canny/depth/cropped-face/etc.)')
 
                         mixing_image_prompt_and_vary_upscale = gr.Checkbox(label='Mixing Image Prompt and Vary/Upscale',
                                                                            value=False)
@@ -321,6 +387,27 @@ with shared.gradio_root:
                             canny_high_threshold = gr.Slider(label='Canny High Threshold', minimum=1, maximum=255,
                                                              step=1, value=128)
 
+                    with gr.Tab(label='Inpaint'):
+                        debugging_inpaint_preprocessor = gr.Checkbox(label='Debug Inpaint Preprocessing', value=False)
+                        inpaint_disable_initial_latent = gr.Checkbox(label='Disable initial latent in inpaint', value=False)
+                        inpaint_engine = gr.Dropdown(label='Inpaint Engine',
+                                                     value=modules.config.default_inpaint_engine_version,
+                                                     choices=flags.inpaint_engine_versions,
+                                                     info='Version of Fooocus inpaint model')
+                        inpaint_strength = gr.Slider(label='Inpaint Denoising Strength',
+                                                     minimum=0.0, maximum=1.0, step=0.001, value=1.0,
+                                                     info='Same as the denoising strength in A1111 inpaint. '
+                                                          'Only used in inpaint, not used in outpaint. '
+                                                          '(Outpaint always use 1.0)')
+                        inpaint_respective_field = gr.Slider(label='Inpaint Respective Field',
+                                                             minimum=0.0, maximum=1.0, step=0.001, value=0.618,
+                                                             info='The area to inpaint. '
+                                                                  'Value 0 is same as "Only Masked" in A1111. '
+                                                                  'Value 1 is same as "Whole Image" in A1111. '
+                                                                  'Only used in inpaint, not used in outpaint. '
+                                                                  '(Outpaint always use 1.0)')
+                        inpaint_ctrls = [debugging_inpaint_preprocessor, inpaint_disable_initial_latent, inpaint_engine, inpaint_strength, inpaint_respective_field]
+
                     with gr.Tab(label='FreeU'):
                         freeu_enabled = gr.Checkbox(label='Enabled', value=False)
                         freeu_b1 = gr.Slider(label='B1', minimum=0, maximum=2, step=0.01, value=1.01)
@@ -329,19 +416,21 @@ with shared.gradio_root:
                         freeu_s2 = gr.Slider(label='S2', minimum=0, maximum=4, step=0.01, value=0.95)
                         freeu_ctrls = [freeu_enabled, freeu_b1, freeu_b2, freeu_s1, freeu_s2]
 
-                adps = [adm_scaler_positive, adm_scaler_negative, adm_scaler_end, adaptive_cfg, sampler_name,
+                adps = [disable_preview, adm_scaler_positive, adm_scaler_negative, adm_scaler_end, adaptive_cfg, sampler_name,
                         scheduler_name, generate_image_grid, overwrite_step, overwrite_switch, overwrite_width, overwrite_height,
                         overwrite_vary_strength, overwrite_upscale_strength,
                         mixing_image_prompt_and_vary_upscale, mixing_image_prompt_and_inpaint,
-                        debugging_cn_preprocessor, controlnet_softness, canny_low_threshold, canny_high_threshold,
-                        inpaint_engine, refiner_swap_method]
+                        debugging_cn_preprocessor, skipping_cn_preprocessor, controlnet_softness,
+                        canny_low_threshold, canny_high_threshold, refiner_swap_method]
                 adps += freeu_ctrls
+                adps += inpaint_ctrls
 
                 def dev_mode_checked(r):
                     return gr.update(visible=r)
 
 
-                dev_mode.change(dev_mode_checked, inputs=[dev_mode], outputs=[dev_tools], queue=False)
+                dev_mode.change(dev_mode_checked, inputs=[dev_mode], outputs=[dev_tools],
+                                queue=False, show_progress=False)
 
                 def model_refresh_clicked():
                     modules.config.update_all_model_names()
@@ -351,9 +440,53 @@ with shared.gradio_root:
                         results += [gr.update(choices=['None'] + modules.config.lora_filenames), gr.update()]
                     return results
 
-                model_refresh.click(model_refresh_clicked, [], [base_model, refiner_model] + lora_ctrls, queue=False)
+                model_refresh.click(model_refresh_clicked, [], [base_model, refiner_model] + lora_ctrls,
+                                    queue=False, show_progress=False)
 
-        advanced_checkbox.change(lambda x: gr.update(visible=x), advanced_checkbox, advanced_column, queue=False)
+        performance_selection.change(lambda x: [gr.update(interactive=x != 'Extreme Speed')] * 11,
+                                     inputs=performance_selection,
+                                     outputs=[
+                                         guidance_scale, sharpness, adm_scaler_end, adm_scaler_positive,
+                                         adm_scaler_negative, refiner_switch, refiner_model, sampler_name,
+                                         scheduler_name, adaptive_cfg, refiner_swap_method
+                                     ], queue=False, show_progress=False)
+
+        advanced_checkbox.change(lambda x: gr.update(visible=x), advanced_checkbox, advanced_column,
+                                 queue=False, show_progress=False) \
+            .then(fn=lambda: None, _js='refresh_grid_delayed', queue=False, show_progress=False)
+
+        def inpaint_mode_change(mode):
+            assert mode in modules.flags.inpaint_options
+
+            # inpaint_additional_prompt, outpaint_selections, example_inpaint_prompts,
+            # inpaint_disable_initial_latent, inpaint_engine,
+            # inpaint_strength, inpaint_respective_field
+
+            if mode == modules.flags.inpaint_option_detail:
+                return [
+                    gr.update(visible=True), gr.update(visible=False, value=[]),
+                    gr.Dataset.update(visible=True, samples=modules.config.example_inpaint_prompts),
+                    False, 'None', 0.5, 0.0
+                ]
+
+            if mode == modules.flags.inpaint_option_modify:
+                return [
+                    gr.update(visible=True), gr.update(visible=False, value=[]),
+                    gr.Dataset.update(visible=False, samples=modules.config.example_inpaint_prompts),
+                    True, modules.config.default_inpaint_engine_version, 1.0, 0.0
+                ]
+
+            return [
+                gr.update(visible=False, value=''), gr.update(visible=True),
+                gr.Dataset.update(visible=False, samples=modules.config.example_inpaint_prompts),
+                False, modules.config.default_inpaint_engine_version, 1.0, 0.618
+            ]
+
+        inpaint_mode.input(inpaint_mode_change, inputs=inpaint_mode, outputs=[
+            inpaint_additional_prompt, outpaint_selections, example_inpaint_prompts,
+            inpaint_disable_initial_latent, inpaint_engine,
+            inpaint_strength, inpaint_respective_field
+        ], show_progress=False, queue=False)
 
         ctrls = [
             prompt, negative_prompt, style_selections,
@@ -363,7 +496,7 @@ with shared.gradio_root:
         ctrls += [base_model, refiner_model, refiner_switch] + lora_ctrls
         ctrls += [input_image_checkbox, current_tab]
         ctrls += [uov_method, uov_input_image]
-        ctrls += [outpaint_selections, inpaint_input_image]
+        ctrls += [outpaint_selections, inpaint_input_image, inpaint_additional_prompt]
         ctrls += ip_ctrls
 
         generate_button.click(lambda: (gr.update(visible=True, interactive=True), gr.update(visible=True, interactive=True), gr.update(visible=False), []), outputs=[stop_button, skip_button, generate_button, gallery]) \
@@ -371,7 +504,7 @@ with shared.gradio_root:
             .then(advanced_parameters.set_all_advanced_parameters, inputs=adps) \
             .then(fn=generate_clicked, inputs=ctrls, outputs=[progress_html, progress_window, progress_gallery, gallery]) \
             .then(lambda: (gr.update(visible=True), gr.update(visible=False), gr.update(visible=False)), outputs=[generate_button, stop_button, skip_button]) \
-            .then(fn=None, _js='playNotification')
+            .then(fn=lambda: None, _js='playNotification').then(fn=lambda: None, _js='refresh_grid_delayed')
 
         for notification_file in ['notification.ogg', 'notification.mp3']:
             if os.path.exists(notification_file):