Foreign-Mirrors/Fooocus
1
0
Fork 0
mirror of https://github.com/lllyasviel/Fooocus.git synced 2026-08-16 13:13:16 +02:00
Code Issues Packages Projects Releases Wiki Activity

Update Backend

Browse Source 
Update Backend
This commit is contained in:
lllyasviel
2023-10-25 09:40:13 -07:00
parent bb965067e0
commit 38e70cebcc
11 changed files with 288 additions and 252 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/sample_hijack.py
+48 -24
View File
@@ -3,10 +3,10 @@ import fcbh.samplers
import fcbh.model_management
from fcbh.model_base import SDXLRefiner, SDXL
from fcbh.conds import CONDRegular
from fcbh.sample import get_additional_models, get_models_from_cond, cleanup_additional_models
from fcbh.samplers import resolve_areas_and_cond_masks, wrap_model, calculate_start_end_timesteps, \
create_cond_with_same_area_if_none, pre_run_control, apply_empty_x_to_equal_area, encode_adm, \
encode_cond
create_cond_with_same_area_if_none, pre_run_control, apply_empty_x_to_equal_area, encode_model_conds
current_refiner = None
@@ -15,15 +15,13 @@ refiner_switch_step = -1
@torch.no_grad()
@torch.inference_mode()
def clip_separate(cond, target_model=None, target_clip=None):
c, p = cond[0]
def clip_separate_inner(c, p, target_model=None, target_clip=None):
if target_model is None or isinstance(target_model, SDXLRefiner):
c = c[..., -1280:].clone()
p = {"pooled_output": p["pooled_output"].clone()}
elif isinstance(target_model, SDXL):
c = c.clone()
p = {"pooled_output": p["pooled_output"].clone()}
else:
p = None
c = c[..., :768].clone()
final_layer_norm = target_clip.cond_stage_model.clip_l.transformer.text_model.final_layer_norm
@@ -43,9 +41,42 @@ def clip_separate(cond, target_model=None, target_clip=None):
final_layer_norm.to(device=final_layer_norm_origin_device, dtype=final_layer_norm_origin_dtype)
c = c.to(device=c_origin_device, dtype=c_origin_dtype)
return c, p
p = {}
return [[c, p]]
@torch.no_grad()
@torch.inference_mode()
def clip_separate(cond, target_model=None, target_clip=None):
results = []
for c, px in cond:
p = px.get('pooled_output', None)
c, p = clip_separate_inner(c, p, target_model=target_model, target_clip=target_clip)
p = {} if p is None else {'pooled_output': p.clone()}
results.append([c, p])
return results
@torch.no_grad()
@torch.inference_mode()
def clip_separate_after_preparation(cond, target_model=None, target_clip=None):
results = []
for x in cond:
p = x.get('pooled_output', None)
c = x['model_conds']['c_crossattn'].cond
c, p = clip_separate_inner(c, p, target_model=target_model, target_clip=target_clip)
result = {'model_conds': {'c_crossattn': CONDRegular(c)}}
if p is not None:
result['pooled_output'] = p.clone()
results.append(result)
return results
@torch.no_grad()
@@ -73,31 +104,24 @@ def sample_hacked(model, noise, positive, negative, cfg, device, sampler, sigmas
# pre_run_control(model_wrap, negative + positive)
pre_run_control(model_wrap, positive) # negative is not necessary in Fooocus, 0.5s faster.
apply_empty_x_to_equal_area(list(filter(lambda c: c[1].get('control_apply_to_uncond', False) == True, positive)), negative, 'control', lambda cond_cnets, x: cond_cnets[x])
apply_empty_x_to_equal_area(list(filter(lambda c: c.get('control_apply_to_uncond', False) == True, positive)), negative, 'control', lambda cond_cnets, x: cond_cnets[x])
apply_empty_x_to_equal_area(positive, negative, 'gligen', lambda cond_cnets, x: cond_cnets[x])
if latent_image is not None:
latent_image = model.process_latent_in(latent_image)
if model.is_adm():
positive = encode_adm(model, positive, noise.shape[0], noise.shape[3], noise.shape[2], device, "positive")
negative = encode_adm(model, negative, noise.shape[0], noise.shape[3], noise.shape[2], device, "negative")
if hasattr(model, 'cond_concat'):
positive = encode_cond(model.cond_concat, "concat", positive, device, noise=noise, latent_image=latent_image, denoise_mask=denoise_mask)
negative = encode_cond(model.cond_concat, "concat", negative, device, noise=noise, latent_image=latent_image, denoise_mask=denoise_mask)
if hasattr(model, 'extra_conds'):
positive = encode_model_conds(model.extra_conds, positive, noise, device, "positive", latent_image=latent_image, denoise_mask=denoise_mask)
negative = encode_model_conds(model.extra_conds, negative, noise, device, "negative", latent_image=latent_image, denoise_mask=denoise_mask)
extra_args = {"cond":positive, "uncond":negative, "cond_scale": cfg, "model_options": model_options, "seed":seed}
if current_refiner is not None and current_refiner.model.is_adm():
positive_refiner = clip_separate(positive, target_model=current_refiner.model)
negative_refiner = clip_separate(negative, target_model=current_refiner.model)
if current_refiner is not None and hasattr(current_refiner.model, 'extra_conds'):
positive_refiner = clip_separate_after_preparation(positive, target_model=current_refiner.model)
negative_refiner = clip_separate_after_preparation(negative, target_model=current_refiner.model)
positive_refiner = encode_adm(current_refiner.model, positive_refiner, noise.shape[0], noise.shape[3], noise.shape[2], device, "positive")
negative_refiner = encode_adm(current_refiner.model, negative_refiner, noise.shape[0], noise.shape[3], noise.shape[2], device, "negative")
positive_refiner[0][1]['adm_encoded'].to(positive[0][1]['adm_encoded'])
negative_refiner[0][1]['adm_encoded'].to(negative[0][1]['adm_encoded'])
positive_refiner = encode_model_conds(current_refiner.model.extra_conds, positive_refiner, noise, device, "positive", latent_image=latent_image, denoise_mask=denoise_mask)
negative_refiner = encode_model_conds(current_refiner.model.extra_conds, negative_refiner, noise, device, "negative", latent_image=latent_image, denoise_mask=denoise_mask)
def refiner_switch():
cleanup_additional_models(set(get_models_from_cond(positive, "control") + get_models_from_cond(negative, "control")))