Spaces:
Configuration error
Configuration error
Commit
•
345d7b4
1
Parent(s):
45e73ca
Update pipeline_semantic_stable_diffusion_img2img_solver.py (#9)
Browse files- Update pipeline_semantic_stable_diffusion_img2img_solver.py (4065064f7aab311c2f9705f66c0b0aa7669cfdac)
- Update app.py (24b22ad95d5e32ad4374733cf2b0acb5c0e13f26)
Co-authored-by: Linoy Tsaban <LinoyTsaban@users.noreply.huggingface.co>
- app.py +17 -14
- pipeline_semantic_stable_diffusion_img2img_solver.py +6 -5
app.py
CHANGED
@@ -35,9 +35,9 @@ def caption_image(input_image):
|
|
35 |
generated_caption = blip_processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
|
36 |
return generated_caption, generated_caption
|
37 |
|
38 |
-
def sample(zs, wts, attention_store, prompt_tar="", cfg_scale_tar=15, skip=36, eta=1):
|
39 |
latents = wts[-1].expand(1, -1, -1, -1)
|
40 |
-
img, attention_store = pipe(
|
41 |
prompt=prompt_tar,
|
42 |
init_latents=latents,
|
43 |
guidance_scale=cfg_scale_tar,
|
@@ -45,10 +45,10 @@ def sample(zs, wts, attention_store, prompt_tar="", cfg_scale_tar=15, skip=36, e
|
|
45 |
# num_inference_steps=steps,
|
46 |
# use_ddpm=True,
|
47 |
# wts=wts.value,
|
48 |
-
attention_store = attention_store,
|
49 |
zs=zs,
|
50 |
)
|
51 |
-
return img.images[0], attention_store
|
52 |
|
53 |
|
54 |
def reconstruct(
|
@@ -59,6 +59,7 @@ def reconstruct(
|
|
59 |
wts,
|
60 |
zs,
|
61 |
attention_store,
|
|
|
62 |
do_reconstruction,
|
63 |
reconstruction,
|
64 |
reconstruct_button,
|
@@ -79,8 +80,8 @@ def reconstruct(
|
|
79 |
): # if image caption was not changed, run actual reconstruction
|
80 |
tar_prompt = ""
|
81 |
latents = wts[-1].expand(1, -1, -1, -1)
|
82 |
-
reconstruction, attention_store = sample(
|
83 |
-
zs, wts, attention_store=attention_store, prompt_tar=tar_prompt, skip=skip, cfg_scale_tar=tar_cfg_scale
|
84 |
)
|
85 |
do_reconstruction = False
|
86 |
return (
|
@@ -130,7 +131,7 @@ def load_and_invert(
|
|
130 |
## SEGA ##
|
131 |
|
132 |
def edit(input_image,
|
133 |
-
wts, zs, attention_store,
|
134 |
tar_prompt,
|
135 |
image_caption,
|
136 |
steps,
|
@@ -197,27 +198,27 @@ def edit(input_image,
|
|
197 |
)
|
198 |
|
199 |
latnets = wts[-1].expand(1, -1, -1, -1)
|
200 |
-
sega_out, attention_store = pipe(prompt=tar_prompt,
|
201 |
init_latents=latnets,
|
202 |
guidance_scale = tar_cfg_scale,
|
203 |
# num_images_per_prompt=1,
|
204 |
# num_inference_steps=steps,
|
205 |
# use_ddpm=True,
|
206 |
# wts=wts.value,
|
207 |
-
zs=zs, attention_store=attention_store, **editing_args)
|
208 |
|
209 |
-
return sega_out.images[0], gr.update(visible=True), do_reconstruction, reconstruction, wts, zs, attention_store, do_inversion, show_share_button
|
210 |
|
211 |
|
212 |
else: # if sega concepts were not added, performs regular ddpm sampling
|
213 |
|
214 |
if do_reconstruction: # if ddpm sampling wasn't computed
|
215 |
-
pure_ddpm_img, attention_store = sample(zs, wts, attention_store=attention_store, prompt_tar=tar_prompt, skip=skip, cfg_scale_tar=tar_cfg_scale)
|
216 |
reconstruction = pure_ddpm_img
|
217 |
do_reconstruction = False
|
218 |
-
return pure_ddpm_img, gr.update(visible=False), do_reconstruction, reconstruction, wts, zs, attention_store, do_inversion, show_share_button
|
219 |
|
220 |
-
return reconstruction, gr.update(visible=False), do_reconstruction, reconstruction, wts, zs, attention_store, do_inversion, show_share_button
|
221 |
|
222 |
|
223 |
def randomize_seed_fn(seed, is_random):
|
@@ -461,6 +462,7 @@ with gr.Blocks(css="style.css") as demo:
|
|
461 |
wts = gr.State()
|
462 |
zs = gr.State()
|
463 |
attention_store=gr.State()
|
|
|
464 |
reconstruction = gr.State()
|
465 |
do_inversion = gr.State(value=True)
|
466 |
do_reconstruction = gr.State(value=True)
|
@@ -697,6 +699,7 @@ with gr.Blocks(css="style.css") as demo:
|
|
697 |
fn=edit,
|
698 |
inputs=[input_image,
|
699 |
wts, zs, attention_store,
|
|
|
700 |
tar_prompt,
|
701 |
image_caption,
|
702 |
steps,
|
@@ -716,7 +719,7 @@ with gr.Blocks(css="style.css") as demo:
|
|
716 |
|
717 |
|
718 |
],
|
719 |
-
outputs=[sega_edited_image, reconstruct_button, do_reconstruction, reconstruction, wts, zs,attention_store, do_inversion, share_btn_container])
|
720 |
# .success(fn=update_gallery_display, inputs= [prev_output_image, sega_edited_image], outputs = [gallery, gallery, prev_output_image])
|
721 |
|
722 |
|
|
|
35 |
generated_caption = blip_processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
|
36 |
return generated_caption, generated_caption
|
37 |
|
38 |
+
def sample(zs, wts, attention_store, text_cross_attention_maps, prompt_tar="", cfg_scale_tar=15, skip=36, eta=1):
|
39 |
latents = wts[-1].expand(1, -1, -1, -1)
|
40 |
+
img, attention_store, text_cross_attention_maps = pipe(
|
41 |
prompt=prompt_tar,
|
42 |
init_latents=latents,
|
43 |
guidance_scale=cfg_scale_tar,
|
|
|
45 |
# num_inference_steps=steps,
|
46 |
# use_ddpm=True,
|
47 |
# wts=wts.value,
|
48 |
+
attention_store = attention_store, text_cross_attention_maps=text_cross_attention_maps,
|
49 |
zs=zs,
|
50 |
)
|
51 |
+
return img.images[0], attention_store, text_cross_attention_maps
|
52 |
|
53 |
|
54 |
def reconstruct(
|
|
|
59 |
wts,
|
60 |
zs,
|
61 |
attention_store,
|
62 |
+
text_cross_attention_maps,
|
63 |
do_reconstruction,
|
64 |
reconstruction,
|
65 |
reconstruct_button,
|
|
|
80 |
): # if image caption was not changed, run actual reconstruction
|
81 |
tar_prompt = ""
|
82 |
latents = wts[-1].expand(1, -1, -1, -1)
|
83 |
+
reconstruction, attention_store, text_cross_attention_maps = sample(
|
84 |
+
zs, wts, attention_store=attention_store, text_cross_attention_maps=text_cross_attention_maps,prompt_tar=tar_prompt, skip=skip, cfg_scale_tar=tar_cfg_scale
|
85 |
)
|
86 |
do_reconstruction = False
|
87 |
return (
|
|
|
131 |
## SEGA ##
|
132 |
|
133 |
def edit(input_image,
|
134 |
+
wts, zs, attention_store, text_cross_attention_maps,
|
135 |
tar_prompt,
|
136 |
image_caption,
|
137 |
steps,
|
|
|
198 |
)
|
199 |
|
200 |
latnets = wts[-1].expand(1, -1, -1, -1)
|
201 |
+
sega_out, attention_store, text_cross_attention_maps = pipe(prompt=tar_prompt,
|
202 |
init_latents=latnets,
|
203 |
guidance_scale = tar_cfg_scale,
|
204 |
# num_images_per_prompt=1,
|
205 |
# num_inference_steps=steps,
|
206 |
# use_ddpm=True,
|
207 |
# wts=wts.value,
|
208 |
+
zs=zs, attention_store=attention_store, text_cross_attention_maps=text_cross_attention_maps, **editing_args)
|
209 |
|
210 |
+
return sega_out.images[0], gr.update(visible=True), do_reconstruction, reconstruction, wts, zs, attention_store, text_cross_attention_maps, do_inversion, show_share_button
|
211 |
|
212 |
|
213 |
else: # if sega concepts were not added, performs regular ddpm sampling
|
214 |
|
215 |
if do_reconstruction: # if ddpm sampling wasn't computed
|
216 |
+
pure_ddpm_img, attention_store, text_cross_attention_maps = sample(zs, wts, attention_store=attention_store, text_cross_attention_maps=text_cross_attention_maps, prompt_tar=tar_prompt, skip=skip, cfg_scale_tar=tar_cfg_scale)
|
217 |
reconstruction = pure_ddpm_img
|
218 |
do_reconstruction = False
|
219 |
+
return pure_ddpm_img, gr.update(visible=False), do_reconstruction, reconstruction, wts, zs, attention_store, text_cross_attention_maps, do_inversion, show_share_button
|
220 |
|
221 |
+
return reconstruction, gr.update(visible=False), do_reconstruction, reconstruction, wts, zs, attention_store, text_cross_attention_maps, do_inversion, show_share_button
|
222 |
|
223 |
|
224 |
def randomize_seed_fn(seed, is_random):
|
|
|
462 |
wts = gr.State()
|
463 |
zs = gr.State()
|
464 |
attention_store=gr.State()
|
465 |
+
text_cross_attention_maps = gr.State()
|
466 |
reconstruction = gr.State()
|
467 |
do_inversion = gr.State(value=True)
|
468 |
do_reconstruction = gr.State(value=True)
|
|
|
699 |
fn=edit,
|
700 |
inputs=[input_image,
|
701 |
wts, zs, attention_store,
|
702 |
+
text_cross_attention_maps,
|
703 |
tar_prompt,
|
704 |
image_caption,
|
705 |
steps,
|
|
|
719 |
|
720 |
|
721 |
],
|
722 |
+
outputs=[sega_edited_image, reconstruct_button, do_reconstruction, reconstruction, wts, zs,attention_store, text_cross_attention_maps, do_inversion, share_btn_container])
|
723 |
# .success(fn=update_gallery_display, inputs= [prev_output_image, sega_edited_image], outputs = [gallery, gallery, prev_output_image])
|
724 |
|
725 |
|
pipeline_semantic_stable_diffusion_img2img_solver.py
CHANGED
@@ -500,6 +500,7 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
500 |
use_cross_attn_mask: bool = False,
|
501 |
# Attention store (just for visualization purposes)
|
502 |
attention_store = None,
|
|
|
503 |
attn_store_steps: Optional[List[int]] = [],
|
504 |
store_averaged_over_steps: bool = True,
|
505 |
use_intersect_mask: bool = False,
|
@@ -755,10 +756,10 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
755 |
# For classifier free guidance, we need to do two forward passes.
|
756 |
# Here we concatenate the unconditional and text embeddings into a single batch
|
757 |
# to avoid doing two forward passes
|
758 |
-
|
759 |
if enable_edit_guidance:
|
760 |
text_embeddings = torch.cat([uncond_embeddings, text_embeddings, edit_concepts])
|
761 |
-
|
762 |
([editing_prompt] if isinstance(editing_prompt, str) else editing_prompt)
|
763 |
else:
|
764 |
text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
|
@@ -920,11 +921,11 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
920 |
if use_cross_attn_mask:
|
921 |
out = attention_store.aggregate_attention(
|
922 |
attention_maps=attention_store.step_store,
|
923 |
-
prompts=
|
924 |
res=16,
|
925 |
from_where=["up", "down"],
|
926 |
is_cross=True,
|
927 |
-
select=
|
928 |
)
|
929 |
attn_map = out[:, :, :, 1:1 + num_edit_tokens[c]] # 0 -> startoftext
|
930 |
|
@@ -1105,7 +1106,7 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
1105 |
if not return_dict:
|
1106 |
return (image, has_nsfw_concept), attention_store
|
1107 |
|
1108 |
-
return SemanticStableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept), attention_store
|
1109 |
|
1110 |
def encode_text(self, prompts):
|
1111 |
text_inputs = self.tokenizer(
|
|
|
500 |
use_cross_attn_mask: bool = False,
|
501 |
# Attention store (just for visualization purposes)
|
502 |
attention_store = None,
|
503 |
+
text_cross_attention_maps = None,
|
504 |
attn_store_steps: Optional[List[int]] = [],
|
505 |
store_averaged_over_steps: bool = True,
|
506 |
use_intersect_mask: bool = False,
|
|
|
756 |
# For classifier free guidance, we need to do two forward passes.
|
757 |
# Here we concatenate the unconditional and text embeddings into a single batch
|
758 |
# to avoid doing two forward passes
|
759 |
+
text_cross_attention_maps = [org_prompt] if isinstance(org_prompt, str) else org_prompt
|
760 |
if enable_edit_guidance:
|
761 |
text_embeddings = torch.cat([uncond_embeddings, text_embeddings, edit_concepts])
|
762 |
+
text_cross_attention_maps += \
|
763 |
([editing_prompt] if isinstance(editing_prompt, str) else editing_prompt)
|
764 |
else:
|
765 |
text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
|
|
|
921 |
if use_cross_attn_mask:
|
922 |
out = attention_store.aggregate_attention(
|
923 |
attention_maps=attention_store.step_store,
|
924 |
+
prompts=text_cross_attention_maps,
|
925 |
res=16,
|
926 |
from_where=["up", "down"],
|
927 |
is_cross=True,
|
928 |
+
select=text_cross_attention_maps.index(editing_prompt[c]),
|
929 |
)
|
930 |
attn_map = out[:, :, :, 1:1 + num_edit_tokens[c]] # 0 -> startoftext
|
931 |
|
|
|
1106 |
if not return_dict:
|
1107 |
return (image, has_nsfw_concept), attention_store
|
1108 |
|
1109 |
+
return SemanticStableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept), attention_store, text_cross_attention_maps
|
1110 |
|
1111 |
def encode_text(self, prompts):
|
1112 |
text_inputs = self.tokenizer(
|