Spaces:
Running
on
A10G
Running
on
A10G
Linoy Tsaban
commited on
Commit
•
c71b83b
1
Parent(s):
8f22003
Update pipeline_semantic_stable_diffusion_img2img_solver.py
Browse files
pipeline_semantic_stable_diffusion_img2img_solver.py
CHANGED
@@ -36,20 +36,21 @@ class AttentionStore():
|
|
36 |
|
37 |
def __call__(self, attn, is_cross: bool, place_in_unet: str, editing_prompts, PnP=False):
|
38 |
# attn.shape = batch_size * head_size, seq_len query, seq_len_key
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
|
|
48 |
|
49 |
def forward(self, attn, is_cross: bool, place_in_unet: str):
|
50 |
key = f"{place_in_unet}_{'cross' if is_cross else 'self'}"
|
51 |
-
|
52 |
-
|
53 |
|
54 |
def between_steps(self, store_step=True):
|
55 |
if store_step:
|
@@ -95,13 +96,12 @@ class AttentionStore():
|
|
95 |
out = out.sum(1) / out.shape[1]
|
96 |
return out
|
97 |
|
98 |
-
def __init__(self, average: bool, batch_size=1
|
99 |
self.step_store = self.get_empty_store()
|
100 |
self.attention_store = []
|
101 |
self.cur_step = 0
|
102 |
self.average = average
|
103 |
self.batch_size = batch_size
|
104 |
-
self.max_size = max_resolution ** 2
|
105 |
|
106 |
|
107 |
class CrossAttnProcessor:
|
@@ -433,10 +433,10 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
433 |
|
434 |
# Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
|
435 |
def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, latents):
|
436 |
-
|
437 |
|
438 |
-
|
439 |
-
|
440 |
|
441 |
latents = latents.to(device)
|
442 |
|
@@ -456,7 +456,7 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
456 |
else:
|
457 |
continue
|
458 |
|
459 |
-
if "attn2" in name
|
460 |
attn_procs[name] = CrossAttnProcessor(
|
461 |
attention_store=attention_store,
|
462 |
place_in_unet=place_in_unet,
|
@@ -470,8 +470,16 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
470 |
@torch.no_grad()
|
471 |
def __call__(
|
472 |
self,
|
473 |
-
|
|
|
|
|
|
|
|
|
474 |
negative_prompt: Optional[Union[str, List[str]]] = None,
|
|
|
|
|
|
|
|
|
475 |
output_type: Optional[str] = "pil",
|
476 |
return_dict: bool = True,
|
477 |
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
|
@@ -480,10 +488,12 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
480 |
editing_prompt_embeddings: Optional[torch.Tensor] = None,
|
481 |
reverse_editing_direction: Optional[Union[bool, List[bool]]] = False,
|
482 |
edit_guidance_scale: Optional[Union[float, List[float]]] = 5,
|
483 |
-
edit_warmup_steps: Optional[Union[int, List[int]]] =
|
484 |
edit_cooldown_steps: Optional[Union[int, List[int]]] = None,
|
485 |
edit_threshold: Optional[Union[float, List[float]]] = 0.9,
|
486 |
user_mask: Optional[torch.FloatTensor] = None,
|
|
|
|
|
487 |
edit_weights: Optional[List[float]] = None,
|
488 |
sem_guidance: Optional[List[torch.Tensor]] = None,
|
489 |
verbose=True,
|
@@ -494,7 +504,7 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
494 |
use_intersect_mask: bool = False,
|
495 |
init_latents = None,
|
496 |
zs = None,
|
497 |
-
|
498 |
):
|
499 |
r"""
|
500 |
Function invoked when calling the pipeline for generation.
|
@@ -589,7 +599,7 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
589 |
second element is a list of `bool`s denoting whether the corresponding generated image likely represents
|
590 |
"not-safe-for-work" (nsfw) content, according to the `safety_checker`.
|
591 |
"""
|
592 |
-
eta =
|
593 |
num_images_per_prompt = 1
|
594 |
# latents = self.init_latents
|
595 |
latents = init_latents
|
@@ -604,7 +614,18 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
604 |
if use_cross_attn_mask:
|
605 |
self.smoothing = GaussianSmoothing(self.device)
|
606 |
|
607 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
608 |
|
609 |
# 2. Define call parameters
|
610 |
batch_size = self.batch_size
|
@@ -621,6 +642,35 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
621 |
self.enabled_editing_prompts = 0
|
622 |
enable_edit_guidance = False
|
623 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
624 |
if enable_edit_guidance:
|
625 |
# get safety text embeddings
|
626 |
if editing_prompt_embeddings is None:
|
@@ -663,47 +713,54 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
663 |
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
664 |
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
665 |
# corresponds to doing no classifier free guidance.
|
|
|
666 |
# get unconditional embeddings for classifier free guidance
|
667 |
|
668 |
-
|
669 |
-
|
670 |
-
|
671 |
-
|
672 |
-
|
673 |
-
|
674 |
-
|
675 |
-
|
676 |
-
|
677 |
-
|
678 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
679 |
)
|
680 |
-
|
681 |
-
uncond_tokens = negative_prompt
|
682 |
|
683 |
-
|
684 |
-
|
685 |
-
|
686 |
-
|
687 |
-
max_length=max_length,
|
688 |
-
truncation=True,
|
689 |
-
return_tensors="pt",
|
690 |
-
)
|
691 |
-
uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
|
692 |
|
693 |
-
|
694 |
-
|
695 |
-
|
696 |
-
|
697 |
-
|
698 |
-
|
699 |
-
|
700 |
-
|
701 |
-
|
702 |
-
|
703 |
-
self.text_cross_attention_maps = \
|
704 |
-
([editing_prompt] if isinstance(editing_prompt, str) else editing_prompt)
|
705 |
-
else:
|
706 |
-
text_embeddings = torch.cat([uncond_embeddings])
|
707 |
|
708 |
# 4. Prepare timesteps
|
709 |
#self.scheduler.set_timesteps(num_inference_steps, device=self.device)
|
@@ -721,8 +778,8 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
721 |
latents = self.prepare_latents(
|
722 |
batch_size * num_images_per_prompt,
|
723 |
num_channels_latents,
|
724 |
-
|
725 |
-
|
726 |
text_embeddings.dtype,
|
727 |
self.device,
|
728 |
latents,
|
@@ -731,16 +788,23 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
731 |
# 6. Prepare extra step kwargs.
|
732 |
extra_step_kwargs = self.prepare_extra_step_kwargs(eta)
|
733 |
|
|
|
|
|
|
|
734 |
self.uncond_estimates = None
|
|
|
735 |
self.edit_estimates = None
|
736 |
self.sem_guidance = None
|
737 |
self.activation_mask = None
|
738 |
|
739 |
for i, t in enumerate(self.progress_bar(timesteps, verbose=verbose)):
|
|
|
|
|
|
|
740 |
# expand the latents if we are doing classifier free guidance
|
741 |
|
742 |
-
if
|
743 |
-
latent_model_input = torch.cat([latents] * (
|
744 |
else:
|
745 |
latent_model_input = latents
|
746 |
|
@@ -751,219 +815,256 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
751 |
# predict the noise residual
|
752 |
noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embed_input).sample
|
753 |
|
|
|
|
|
754 |
|
755 |
-
|
756 |
-
|
757 |
-
|
758 |
|
759 |
-
|
760 |
-
|
761 |
|
762 |
-
|
763 |
-
|
764 |
-
|
765 |
|
766 |
-
|
767 |
-
|
768 |
-
|
769 |
|
770 |
-
|
771 |
-
|
772 |
-
self.activation_mask = torch.zeros(
|
773 |
-
(len(timesteps), len(noise_pred_edit_concepts), *noise_pred_edit_concepts[0].shape)
|
774 |
-
)
|
775 |
-
if self.edit_estimates is None and enable_edit_guidance:
|
776 |
-
self.edit_estimates = torch.zeros(
|
777 |
-
(len(timesteps), len(noise_pred_edit_concepts), *noise_pred_edit_concepts[0].shape)
|
778 |
-
)
|
779 |
|
780 |
-
if
|
781 |
-
|
|
|
|
|
782 |
|
783 |
-
|
784 |
-
|
785 |
-
|
786 |
-
|
787 |
-
|
788 |
-
|
789 |
-
|
790 |
-
|
791 |
-
dtype=noise_guidance.dtype,
|
792 |
-
)
|
793 |
-
warmup_inds = []
|
794 |
-
# noise_guidance_edit = torch.zeros_like(noise_guidance)
|
795 |
-
for c, noise_pred_edit_concept in enumerate(noise_pred_edit_concepts):
|
796 |
-
self.edit_estimates[i, c] = noise_pred_edit_concept
|
797 |
-
if isinstance(edit_warmup_steps, list):
|
798 |
-
edit_warmup_steps_c = edit_warmup_steps[c]
|
799 |
-
else:
|
800 |
-
edit_warmup_steps_c = edit_warmup_steps
|
801 |
-
if i >= edit_warmup_steps_c:
|
802 |
-
warmup_inds.append(c)
|
803 |
-
else:
|
804 |
-
continue
|
805 |
-
|
806 |
-
if isinstance(edit_guidance_scale, list):
|
807 |
-
edit_guidance_scale_c = edit_guidance_scale[c]
|
808 |
-
else:
|
809 |
-
edit_guidance_scale_c = edit_guidance_scale
|
810 |
-
|
811 |
-
if isinstance(edit_threshold, list):
|
812 |
-
edit_threshold_c = edit_threshold[c]
|
813 |
-
else:
|
814 |
-
edit_threshold_c = edit_threshold
|
815 |
-
if isinstance(reverse_editing_direction, list):
|
816 |
-
reverse_editing_direction_c = reverse_editing_direction[c]
|
817 |
-
else:
|
818 |
-
reverse_editing_direction_c = reverse_editing_direction
|
819 |
-
if edit_weights:
|
820 |
-
edit_weight_c = edit_weights[c]
|
821 |
-
else:
|
822 |
-
edit_weight_c = 1.0
|
823 |
-
|
824 |
-
if isinstance(edit_cooldown_steps, list):
|
825 |
-
edit_cooldown_steps_c = edit_cooldown_steps[c]
|
826 |
-
elif edit_cooldown_steps is None:
|
827 |
-
edit_cooldown_steps_c = i + 1
|
828 |
-
else:
|
829 |
-
edit_cooldown_steps_c = edit_cooldown_steps
|
830 |
-
|
831 |
-
if i >= edit_cooldown_steps_c:
|
832 |
-
noise_guidance_edit[c, :, :, :, :] = torch.zeros_like(noise_pred_edit_concept)
|
833 |
-
continue
|
834 |
-
|
835 |
-
noise_guidance_edit_tmp = noise_pred_edit_concept - noise_pred_uncond
|
836 |
-
# tmp_weights = (noise_pred_text - noise_pred_edit_concept).sum(dim=(1, 2, 3))
|
837 |
-
tmp_weights = (noise_guidance - noise_pred_edit_concept).sum(dim=(1, 2, 3))
|
838 |
-
|
839 |
-
tmp_weights = torch.full_like(tmp_weights, edit_weight_c) # * (1 / enabled_editing_prompts)
|
840 |
-
if reverse_editing_direction_c:
|
841 |
-
noise_guidance_edit_tmp = noise_guidance_edit_tmp * -1
|
842 |
-
concept_weights[c, :] = tmp_weights
|
843 |
-
|
844 |
-
noise_guidance_edit_tmp = noise_guidance_edit_tmp * edit_guidance_scale_c
|
845 |
-
|
846 |
-
if user_mask is not None:
|
847 |
-
noise_guidance_edit_tmp = noise_guidance_edit_tmp * user_mask
|
848 |
-
|
849 |
-
if use_cross_attn_mask:
|
850 |
-
out = self.attention_store.aggregate_attention(
|
851 |
-
attention_maps=self.attention_store.step_store,
|
852 |
-
prompts=self.text_cross_attention_maps,
|
853 |
-
res=16,
|
854 |
-
from_where=["up", "down"],
|
855 |
-
is_cross=True,
|
856 |
-
select=self.text_cross_attention_maps.index(editing_prompt[c]),
|
857 |
)
|
858 |
-
attn_map = out[:, :, :, 1:1 + num_edit_tokens[c]] # 0 -> startoftext
|
859 |
|
860 |
-
|
861 |
-
|
862 |
-
attn_map = torch.sum(attn_map, dim=3)
|
863 |
|
864 |
-
|
865 |
-
|
866 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
867 |
|
868 |
-
|
869 |
-
|
870 |
-
tmp = torch.quantile(attn_map.flatten(start_dim=1), edit_threshold_c, dim=1)
|
871 |
else:
|
872 |
-
|
873 |
-
|
874 |
-
|
875 |
-
# resolution must match latent space dimension
|
876 |
-
attn_mask = F.interpolate(
|
877 |
-
attn_mask.unsqueeze(1),
|
878 |
-
noise_guidance_edit_tmp.shape[-2:] # 64,64
|
879 |
-
).repeat(1, 4, 1, 1)
|
880 |
-
self.activation_mask[i, c] = attn_mask.detach().cpu()
|
881 |
-
if not use_intersect_mask:
|
882 |
-
noise_guidance_edit_tmp = noise_guidance_edit_tmp * attn_mask
|
883 |
-
|
884 |
-
if use_intersect_mask:
|
885 |
-
noise_guidance_edit_tmp_quantile = torch.abs(noise_guidance_edit_tmp)
|
886 |
-
noise_guidance_edit_tmp_quantile = torch.sum(noise_guidance_edit_tmp_quantile, dim=1,
|
887 |
-
keepdim=True)
|
888 |
-
noise_guidance_edit_tmp_quantile = noise_guidance_edit_tmp_quantile.repeat(1, 4, 1, 1)
|
889 |
-
|
890 |
-
# torch.quantile function expects float32
|
891 |
-
if noise_guidance_edit_tmp_quantile.dtype == torch.float32:
|
892 |
-
tmp = torch.quantile(
|
893 |
-
noise_guidance_edit_tmp_quantile.flatten(start_dim=2),
|
894 |
-
edit_threshold_c,
|
895 |
-
dim=2,
|
896 |
-
keepdim=False,
|
897 |
-
)
|
898 |
else:
|
899 |
-
|
900 |
-
|
901 |
-
|
902 |
-
|
903 |
-
|
904 |
-
|
905 |
-
|
906 |
-
|
907 |
-
|
908 |
-
|
909 |
-
|
910 |
-
|
911 |
-
|
912 |
-
|
913 |
-
|
914 |
-
noise_guidance_edit_tmp = noise_guidance_edit_tmp * intersect_mask
|
915 |
-
|
916 |
-
elif not use_cross_attn_mask:
|
917 |
-
# calculate quantile
|
918 |
-
noise_guidance_edit_tmp_quantile = torch.abs(noise_guidance_edit_tmp)
|
919 |
-
noise_guidance_edit_tmp_quantile = torch.sum(noise_guidance_edit_tmp_quantile, dim=1,
|
920 |
-
keepdim=True)
|
921 |
-
noise_guidance_edit_tmp_quantile = noise_guidance_edit_tmp_quantile.repeat(1, 4, 1, 1)
|
922 |
-
|
923 |
-
# torch.quantile function expects float32
|
924 |
-
if noise_guidance_edit_tmp_quantile.dtype == torch.float32:
|
925 |
-
tmp = torch.quantile(
|
926 |
-
noise_guidance_edit_tmp_quantile.flatten(start_dim=2),
|
927 |
-
edit_threshold_c,
|
928 |
-
dim=2,
|
929 |
-
keepdim=False,
|
930 |
-
)
|
931 |
else:
|
932 |
-
|
933 |
-
|
934 |
-
|
935 |
-
|
936 |
-
|
937 |
-
|
938 |
-
|
939 |
-
|
940 |
-
|
941 |
-
|
942 |
-
|
943 |
-
|
944 |
-
|
945 |
-
|
946 |
-
|
947 |
-
|
948 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
949 |
)
|
|
|
|
|
950 |
|
951 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
952 |
|
953 |
-
|
954 |
-
concept_weights = torch.index_select(concept_weights, 0, warmup_inds)
|
955 |
-
concept_weights = torch.where(
|
956 |
-
concept_weights < 0, torch.zeros_like(concept_weights), concept_weights
|
957 |
-
)
|
958 |
|
959 |
-
|
|
|
|
|
|
|
960 |
|
961 |
-
|
|
|
|
|
962 |
|
963 |
-
|
964 |
-
self.sem_guidance[i] = noise_guidance_edit.detach().cpu()
|
965 |
|
966 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
967 |
|
968 |
# compute the previous noisy sample x_t -> x_t-1
|
969 |
if use_ddpm:
|
@@ -971,7 +1072,7 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
971 |
latents = self.scheduler.step(noise_pred, t, latents, variance_noise=zs[idx],
|
972 |
**extra_step_kwargs).prev_sample
|
973 |
|
974 |
-
else:
|
975 |
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
|
976 |
|
977 |
# step callback
|
@@ -1031,7 +1132,7 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
1031 |
source_prompt: str = "",
|
1032 |
source_guidance_scale=3.5,
|
1033 |
num_inversion_steps: int = 30,
|
1034 |
-
skip:
|
1035 |
eta: float = 1.0,
|
1036 |
generator: Optional[torch.Generator] = None,
|
1037 |
verbose=True,
|
@@ -1048,7 +1149,7 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
1048 |
# self.eta = eta
|
1049 |
# assert (self.eta > 0)
|
1050 |
skip = skip/100
|
1051 |
-
|
1052 |
train_steps = self.scheduler.config.num_train_timesteps
|
1053 |
timesteps = torch.from_numpy(
|
1054 |
np.linspace(train_steps - skip * train_steps - 1, 1, num_inversion_steps).astype(np.int64)).to(self.device)
|
@@ -1057,8 +1158,11 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
1057 |
self.num_inversion_steps = timesteps.shape[0]
|
1058 |
self.scheduler.num_inference_steps = timesteps.shape[0]
|
1059 |
self.scheduler.timesteps = timesteps
|
1060 |
-
|
|
|
|
|
1061 |
self.unet.set_attn_processor(AttnProcessor())
|
|
|
1062 |
# 1. get embeddings
|
1063 |
|
1064 |
uncond_embedding = self.encode_text("")
|
@@ -1073,7 +1177,6 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
1073 |
# autoencoder reconstruction
|
1074 |
# image_rec = self.vae.decode(x0 / self.vae.config.scaling_factor, return_dict=False)[0]
|
1075 |
# image_rec = self.image_processor.postprocess(image_rec, output_type="pil")
|
1076 |
-
|
1077 |
# 3. find zs and xts
|
1078 |
variance_noise_shape = (
|
1079 |
self.num_inversion_steps,
|
@@ -1123,8 +1226,8 @@ class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
|
|
1123 |
# self.zs = zs
|
1124 |
|
1125 |
|
1126 |
-
|
1127 |
return zs, xts
|
|
|
1128 |
|
1129 |
@torch.no_grad()
|
1130 |
def encode_image(self, image_path, dtype=None):
|
|
|
36 |
|
37 |
def __call__(self, attn, is_cross: bool, place_in_unet: str, editing_prompts, PnP=False):
|
38 |
# attn.shape = batch_size * head_size, seq_len query, seq_len_key
|
39 |
+
bs = 2 + int(PnP) + editing_prompts
|
40 |
+
skip = 2 if PnP else 1 # skip PnP & unconditional
|
41 |
+
|
42 |
+
head_size = int(attn.shape[0] / self.batch_size)
|
43 |
+
attn = torch.stack(attn.split(self.batch_size)).permute(1, 0, 2, 3)
|
44 |
+
source_batch_size = int(attn.shape[1] // bs)
|
45 |
+
self.forward(
|
46 |
+
attn[:, skip * source_batch_size:],
|
47 |
+
is_cross,
|
48 |
+
place_in_unet)
|
49 |
|
50 |
def forward(self, attn, is_cross: bool, place_in_unet: str):
|
51 |
key = f"{place_in_unet}_{'cross' if is_cross else 'self'}"
|
52 |
+
if attn.shape[1] <= 32 ** 2: # avoid memory overhead
|
53 |
+
self.step_store[key].append(attn)
|
54 |
|
55 |
def between_steps(self, store_step=True):
|
56 |
if store_step:
|
|
|
96 |
out = out.sum(1) / out.shape[1]
|
97 |
return out
|
98 |
|
99 |
+
def __init__(self, average: bool, batch_size=1):
|
100 |
self.step_store = self.get_empty_store()
|
101 |
self.attention_store = []
|
102 |
self.cur_step = 0
|
103 |
self.average = average
|
104 |
self.batch_size = batch_size
|
|
|
105 |
|
106 |
|
107 |
class CrossAttnProcessor:
|
|
|
433 |
|
434 |
# Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
|
435 |
def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, latents):
|
436 |
+
shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
|
437 |
|
438 |
+
if latents.shape != shape:
|
439 |
+
raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
|
440 |
|
441 |
latents = latents.to(device)
|
442 |
|
|
|
456 |
else:
|
457 |
continue
|
458 |
|
459 |
+
if "attn2" in name:
|
460 |
attn_procs[name] = CrossAttnProcessor(
|
461 |
attention_store=attention_store,
|
462 |
place_in_unet=place_in_unet,
|
|
|
470 |
@torch.no_grad()
|
471 |
def __call__(
|
472 |
self,
|
473 |
+
prompt: Union[str, List[str]] = "",
|
474 |
+
height: Optional[int] = None,
|
475 |
+
width: Optional[int] = None,
|
476 |
+
# num_inference_steps: int = 50,
|
477 |
+
guidance_scale: float = 7.5,
|
478 |
negative_prompt: Optional[Union[str, List[str]]] = None,
|
479 |
+
# num_images_per_prompt: int = 1,
|
480 |
+
eta: float = 1.0,
|
481 |
+
# generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
482 |
+
# latents: Optional[torch.FloatTensor] = None,
|
483 |
output_type: Optional[str] = "pil",
|
484 |
return_dict: bool = True,
|
485 |
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
|
|
|
488 |
editing_prompt_embeddings: Optional[torch.Tensor] = None,
|
489 |
reverse_editing_direction: Optional[Union[bool, List[bool]]] = False,
|
490 |
edit_guidance_scale: Optional[Union[float, List[float]]] = 5,
|
491 |
+
edit_warmup_steps: Optional[Union[int, List[int]]] = 10,
|
492 |
edit_cooldown_steps: Optional[Union[int, List[int]]] = None,
|
493 |
edit_threshold: Optional[Union[float, List[float]]] = 0.9,
|
494 |
user_mask: Optional[torch.FloatTensor] = None,
|
495 |
+
edit_momentum_scale: Optional[float] = 0.1,
|
496 |
+
edit_mom_beta: Optional[float] = 0.4,
|
497 |
edit_weights: Optional[List[float]] = None,
|
498 |
sem_guidance: Optional[List[torch.Tensor]] = None,
|
499 |
verbose=True,
|
|
|
504 |
use_intersect_mask: bool = False,
|
505 |
init_latents = None,
|
506 |
zs = None,
|
507 |
+
|
508 |
):
|
509 |
r"""
|
510 |
Function invoked when calling the pipeline for generation.
|
|
|
599 |
second element is a list of `bool`s denoting whether the corresponding generated image likely represents
|
600 |
"not-safe-for-work" (nsfw) content, according to the `safety_checker`.
|
601 |
"""
|
602 |
+
# eta = self.eta
|
603 |
num_images_per_prompt = 1
|
604 |
# latents = self.init_latents
|
605 |
latents = init_latents
|
|
|
614 |
if use_cross_attn_mask:
|
615 |
self.smoothing = GaussianSmoothing(self.device)
|
616 |
|
617 |
+
# 0. Default height and width to unet
|
618 |
+
height = height or self.unet.config.sample_size * self.vae_scale_factor
|
619 |
+
width = width or self.unet.config.sample_size * self.vae_scale_factor
|
620 |
+
|
621 |
+
# 1. Check inputs. Raise error if not correct
|
622 |
+
self.check_inputs(prompt, height, width, callback_steps)
|
623 |
+
|
624 |
+
org_prompt = prompt
|
625 |
+
if isinstance(prompt, list):
|
626 |
+
assert len(prompt) == self.batch_size
|
627 |
+
elif isinstance(prompt, str):
|
628 |
+
prompt = list(repeat(prompt, self.batch_size))
|
629 |
|
630 |
# 2. Define call parameters
|
631 |
batch_size = self.batch_size
|
|
|
642 |
self.enabled_editing_prompts = 0
|
643 |
enable_edit_guidance = False
|
644 |
|
645 |
+
# get prompt text embeddings
|
646 |
+
text_inputs = self.tokenizer(
|
647 |
+
prompt,
|
648 |
+
padding="max_length",
|
649 |
+
max_length=self.tokenizer.model_max_length,
|
650 |
+
truncation=True,
|
651 |
+
return_tensors="pt",
|
652 |
+
)
|
653 |
+
text_input_ids = text_inputs.input_ids
|
654 |
+
untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
|
655 |
+
|
656 |
+
if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
|
657 |
+
text_input_ids, untruncated_ids
|
658 |
+
):
|
659 |
+
removed_text = self.tokenizer.batch_decode(
|
660 |
+
untruncated_ids[:, self.tokenizer.model_max_length - 1: -1]
|
661 |
+
)
|
662 |
+
logger.warning(
|
663 |
+
"The following part of your input was truncated because CLIP can only handle sequences up to"
|
664 |
+
f" {self.tokenizer.model_max_length} tokens: {removed_text}"
|
665 |
+
)
|
666 |
+
|
667 |
+
text_embeddings = self.text_encoder(text_input_ids.to(self.device))[0]
|
668 |
+
|
669 |
+
# duplicate text embeddings for each generation per prompt, using mps friendly method
|
670 |
+
bs_embed, seq_len, _ = text_embeddings.shape
|
671 |
+
text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
|
672 |
+
text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
|
673 |
+
|
674 |
if enable_edit_guidance:
|
675 |
# get safety text embeddings
|
676 |
if editing_prompt_embeddings is None:
|
|
|
713 |
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
714 |
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
715 |
# corresponds to doing no classifier free guidance.
|
716 |
+
do_classifier_free_guidance = guidance_scale > 1.0
|
717 |
# get unconditional embeddings for classifier free guidance
|
718 |
|
719 |
+
if do_classifier_free_guidance:
|
720 |
+
uncond_tokens: List[str]
|
721 |
+
if negative_prompt is None:
|
722 |
+
uncond_tokens = [""]
|
723 |
+
elif type(prompt) is not type(negative_prompt):
|
724 |
+
raise TypeError(
|
725 |
+
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
|
726 |
+
f" {type(prompt)}."
|
727 |
+
)
|
728 |
+
elif isinstance(negative_prompt, str):
|
729 |
+
uncond_tokens = [negative_prompt]
|
730 |
+
elif batch_size != len(negative_prompt):
|
731 |
+
raise ValueError(
|
732 |
+
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
|
733 |
+
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
|
734 |
+
" the batch size of `prompt`."
|
735 |
+
)
|
736 |
+
else:
|
737 |
+
uncond_tokens = negative_prompt
|
738 |
+
|
739 |
+
max_length = text_input_ids.shape[-1]
|
740 |
+
uncond_input = self.tokenizer(
|
741 |
+
uncond_tokens,
|
742 |
+
padding="max_length",
|
743 |
+
max_length=max_length,
|
744 |
+
truncation=True,
|
745 |
+
return_tensors="pt",
|
746 |
)
|
747 |
+
uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
|
|
|
748 |
|
749 |
+
# duplicate unconditional embeddings for each generation per prompt, using mps friendly method
|
750 |
+
seq_len = uncond_embeddings.shape[1]
|
751 |
+
uncond_embeddings = uncond_embeddings.repeat(batch_size, num_images_per_prompt, 1)
|
752 |
+
uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1)
|
|
|
|
|
|
|
|
|
|
|
753 |
|
754 |
+
# For classifier free guidance, we need to do two forward passes.
|
755 |
+
# Here we concatenate the unconditional and text embeddings into a single batch
|
756 |
+
# to avoid doing two forward passes
|
757 |
+
self.text_cross_attention_maps = [org_prompt] if isinstance(org_prompt, str) else org_prompt
|
758 |
+
if enable_edit_guidance:
|
759 |
+
text_embeddings = torch.cat([uncond_embeddings, text_embeddings, edit_concepts])
|
760 |
+
self.text_cross_attention_maps += \
|
761 |
+
([editing_prompt] if isinstance(editing_prompt, str) else editing_prompt)
|
762 |
+
else:
|
763 |
+
text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
|
|
|
|
|
|
|
|
|
764 |
|
765 |
# 4. Prepare timesteps
|
766 |
#self.scheduler.set_timesteps(num_inference_steps, device=self.device)
|
|
|
778 |
latents = self.prepare_latents(
|
779 |
batch_size * num_images_per_prompt,
|
780 |
num_channels_latents,
|
781 |
+
height,
|
782 |
+
width,
|
783 |
text_embeddings.dtype,
|
784 |
self.device,
|
785 |
latents,
|
|
|
788 |
# 6. Prepare extra step kwargs.
|
789 |
extra_step_kwargs = self.prepare_extra_step_kwargs(eta)
|
790 |
|
791 |
+
# Initialize edit_momentum to None
|
792 |
+
edit_momentum = None
|
793 |
+
|
794 |
self.uncond_estimates = None
|
795 |
+
self.text_estimates = None
|
796 |
self.edit_estimates = None
|
797 |
self.sem_guidance = None
|
798 |
self.activation_mask = None
|
799 |
|
800 |
for i, t in enumerate(self.progress_bar(timesteps, verbose=verbose)):
|
801 |
+
idx = t_to_idx[int(t)]
|
802 |
+
|
803 |
+
|
804 |
# expand the latents if we are doing classifier free guidance
|
805 |
|
806 |
+
if do_classifier_free_guidance:
|
807 |
+
latent_model_input = torch.cat([latents] * (2 + self.enabled_editing_prompts))
|
808 |
else:
|
809 |
latent_model_input = latents
|
810 |
|
|
|
815 |
# predict the noise residual
|
816 |
noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embed_input).sample
|
817 |
|
818 |
+
# perform guidance
|
819 |
+
if do_classifier_free_guidance:
|
820 |
|
821 |
+
noise_pred_out = noise_pred.chunk(2 + self.enabled_editing_prompts) # [b,4, 64, 64]
|
822 |
+
noise_pred_uncond, noise_pred_text = noise_pred_out[0], noise_pred_out[1]
|
823 |
+
noise_pred_edit_concepts = noise_pred_out[2:]
|
824 |
|
825 |
+
# default text guidance
|
826 |
+
noise_guidance = guidance_scale * (noise_pred_text - noise_pred_uncond)
|
827 |
|
828 |
+
if self.uncond_estimates is None:
|
829 |
+
self.uncond_estimates = torch.zeros((len(timesteps), *noise_pred_uncond.shape))
|
830 |
+
self.uncond_estimates[i] = noise_pred_uncond.detach().cpu()
|
831 |
|
832 |
+
if self.text_estimates is None:
|
833 |
+
self.text_estimates = torch.zeros((len(timesteps), *noise_pred_text.shape))
|
834 |
+
self.text_estimates[i] = noise_pred_text.detach().cpu()
|
835 |
|
836 |
+
if edit_momentum is None:
|
837 |
+
edit_momentum = torch.zeros_like(noise_guidance)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
838 |
|
839 |
+
if sem_guidance is not None and len(sem_guidance) > i:
|
840 |
+
edit_guidance = sem_guidance[i].to(self.device)
|
841 |
+
edit_momentum = edit_mom_beta * edit_momentum + (1 - edit_mom_beta) * edit_guidance
|
842 |
+
noise_guidance = noise_guidance + edit_guidance
|
843 |
|
844 |
+
elif enable_edit_guidance:
|
845 |
+
if self.activation_mask is None:
|
846 |
+
self.activation_mask = torch.zeros(
|
847 |
+
(len(timesteps), len(noise_pred_edit_concepts), *noise_pred_edit_concepts[0].shape)
|
848 |
+
)
|
849 |
+
if self.edit_estimates is None and enable_edit_guidance:
|
850 |
+
self.edit_estimates = torch.zeros(
|
851 |
+
(len(timesteps), len(noise_pred_edit_concepts), *noise_pred_edit_concepts[0].shape)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
852 |
)
|
|
|
853 |
|
854 |
+
if self.sem_guidance is None:
|
855 |
+
self.sem_guidance = torch.zeros((len(timesteps), *noise_pred_text.shape))
|
|
|
856 |
|
857 |
+
concept_weights = torch.zeros(
|
858 |
+
(len(noise_pred_edit_concepts), noise_guidance.shape[0]),
|
859 |
+
device=self.device,
|
860 |
+
dtype=noise_guidance.dtype,
|
861 |
+
)
|
862 |
+
noise_guidance_edit = torch.zeros(
|
863 |
+
(len(noise_pred_edit_concepts), *noise_guidance.shape),
|
864 |
+
device=self.device,
|
865 |
+
dtype=noise_guidance.dtype,
|
866 |
+
)
|
867 |
+
# noise_guidance_edit = torch.zeros_like(noise_guidance)
|
868 |
+
warmup_inds = []
|
869 |
+
for c, noise_pred_edit_concept in enumerate(noise_pred_edit_concepts):
|
870 |
+
self.edit_estimates[i, c] = noise_pred_edit_concept
|
871 |
+
if isinstance(edit_guidance_scale, list):
|
872 |
+
edit_guidance_scale_c = edit_guidance_scale[c]
|
873 |
+
else:
|
874 |
+
edit_guidance_scale_c = edit_guidance_scale
|
875 |
|
876 |
+
if isinstance(edit_threshold, list):
|
877 |
+
edit_threshold_c = edit_threshold[c]
|
|
|
878 |
else:
|
879 |
+
edit_threshold_c = edit_threshold
|
880 |
+
if isinstance(reverse_editing_direction, list):
|
881 |
+
reverse_editing_direction_c = reverse_editing_direction[c]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
882 |
else:
|
883 |
+
reverse_editing_direction_c = reverse_editing_direction
|
884 |
+
if edit_weights:
|
885 |
+
edit_weight_c = edit_weights[c]
|
886 |
+
else:
|
887 |
+
edit_weight_c = 1.0
|
888 |
+
if isinstance(edit_warmup_steps, list):
|
889 |
+
edit_warmup_steps_c = edit_warmup_steps[c]
|
890 |
+
else:
|
891 |
+
edit_warmup_steps_c = edit_warmup_steps
|
892 |
+
|
893 |
+
if isinstance(edit_cooldown_steps, list):
|
894 |
+
edit_cooldown_steps_c = edit_cooldown_steps[c]
|
895 |
+
elif edit_cooldown_steps is None:
|
896 |
+
edit_cooldown_steps_c = i + 1
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
897 |
else:
|
898 |
+
edit_cooldown_steps_c = edit_cooldown_steps
|
899 |
+
if i >= edit_warmup_steps_c:
|
900 |
+
warmup_inds.append(c)
|
901 |
+
if i >= edit_cooldown_steps_c:
|
902 |
+
noise_guidance_edit[c, :, :, :, :] = torch.zeros_like(noise_pred_edit_concept)
|
903 |
+
continue
|
904 |
+
|
905 |
+
noise_guidance_edit_tmp = noise_pred_edit_concept - noise_pred_uncond
|
906 |
+
# tmp_weights = (noise_pred_text - noise_pred_edit_concept).sum(dim=(1, 2, 3))
|
907 |
+
tmp_weights = (noise_guidance - noise_pred_edit_concept).sum(dim=(1, 2, 3))
|
908 |
+
|
909 |
+
tmp_weights = torch.full_like(tmp_weights, edit_weight_c) # * (1 / enabled_editing_prompts)
|
910 |
+
if reverse_editing_direction_c:
|
911 |
+
noise_guidance_edit_tmp = noise_guidance_edit_tmp * -1
|
912 |
+
concept_weights[c, :] = tmp_weights
|
913 |
+
|
914 |
+
noise_guidance_edit_tmp = noise_guidance_edit_tmp * edit_guidance_scale_c
|
915 |
+
|
916 |
+
if user_mask is not None:
|
917 |
+
noise_guidance_edit_tmp = noise_guidance_edit_tmp * user_mask
|
918 |
+
|
919 |
+
if use_cross_attn_mask:
|
920 |
+
out = self.attention_store.aggregate_attention(
|
921 |
+
attention_maps=self.attention_store.step_store,
|
922 |
+
prompts=self.text_cross_attention_maps,
|
923 |
+
res=16,
|
924 |
+
from_where=["up", "down"],
|
925 |
+
is_cross=True,
|
926 |
+
select=self.text_cross_attention_maps.index(editing_prompt[c]),
|
927 |
+
)
|
928 |
+
attn_map = out[:, :, :, 1:1 + num_edit_tokens[c]] # 0 -> startoftext
|
929 |
+
|
930 |
+
# average over all tokens
|
931 |
+
assert (attn_map.shape[3] == num_edit_tokens[c])
|
932 |
+
attn_map = torch.sum(attn_map, dim=3)
|
933 |
+
|
934 |
+
# gaussian_smoothing
|
935 |
+
attn_map = F.pad(attn_map.unsqueeze(1), (1, 1, 1, 1), mode="reflect")
|
936 |
+
attn_map = self.smoothing(attn_map).squeeze(1)
|
937 |
+
|
938 |
+
# create binary mask
|
939 |
+
if attn_map.dtype == torch.float32:
|
940 |
+
tmp = torch.quantile(attn_map.flatten(start_dim=1), edit_threshold_c, dim=1)
|
941 |
+
else:
|
942 |
+
tmp = torch.quantile(attn_map.flatten(start_dim=1).to(torch.float32), edit_threshold_c, dim=1).to(attn_map.dtype)
|
943 |
+
attn_mask = torch.where(attn_map >= tmp.unsqueeze(1).unsqueeze(1).repeat(1,16,16), 1.0, 0.0)
|
944 |
+
|
945 |
+
# resolution must match latent space dimension
|
946 |
+
attn_mask = F.interpolate(
|
947 |
+
attn_mask.unsqueeze(1),
|
948 |
+
noise_guidance_edit_tmp.shape[-2:] # 64,64
|
949 |
+
).repeat(1, 4, 1, 1)
|
950 |
+
self.activation_mask[i, c] = attn_mask.detach().cpu()
|
951 |
+
if not use_intersect_mask:
|
952 |
+
noise_guidance_edit_tmp = noise_guidance_edit_tmp * attn_mask
|
953 |
+
|
954 |
+
if use_intersect_mask:
|
955 |
+
noise_guidance_edit_tmp_quantile = torch.abs(noise_guidance_edit_tmp)
|
956 |
+
noise_guidance_edit_tmp_quantile = torch.sum(noise_guidance_edit_tmp_quantile, dim=1,
|
957 |
+
keepdim=True)
|
958 |
+
noise_guidance_edit_tmp_quantile = noise_guidance_edit_tmp_quantile.repeat(1, 4, 1, 1)
|
959 |
+
|
960 |
+
# torch.quantile function expects float32
|
961 |
+
if noise_guidance_edit_tmp_quantile.dtype == torch.float32:
|
962 |
+
tmp = torch.quantile(
|
963 |
+
noise_guidance_edit_tmp_quantile.flatten(start_dim=2),
|
964 |
+
edit_threshold_c,
|
965 |
+
dim=2,
|
966 |
+
keepdim=False,
|
967 |
+
)
|
968 |
+
else:
|
969 |
+
tmp = torch.quantile(
|
970 |
+
noise_guidance_edit_tmp_quantile.flatten(start_dim=2).to(torch.float32),
|
971 |
+
edit_threshold_c,
|
972 |
+
dim=2,
|
973 |
+
keepdim=False,
|
974 |
+
).to(noise_guidance_edit_tmp_quantile.dtype)
|
975 |
+
|
976 |
+
intersect_mask = torch.where(
|
977 |
+
noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None],
|
978 |
+
torch.ones_like(noise_guidance_edit_tmp),
|
979 |
+
torch.zeros_like(noise_guidance_edit_tmp),
|
980 |
+
) * attn_mask
|
981 |
+
|
982 |
+
self.activation_mask[i, c] = intersect_mask.detach().cpu()
|
983 |
+
|
984 |
+
noise_guidance_edit_tmp = noise_guidance_edit_tmp * intersect_mask
|
985 |
+
|
986 |
+
elif not use_cross_attn_mask:
|
987 |
+
# calculate quantile
|
988 |
+
noise_guidance_edit_tmp_quantile = torch.abs(noise_guidance_edit_tmp)
|
989 |
+
noise_guidance_edit_tmp_quantile = torch.sum(noise_guidance_edit_tmp_quantile, dim=1,
|
990 |
+
keepdim=True)
|
991 |
+
noise_guidance_edit_tmp_quantile = noise_guidance_edit_tmp_quantile.repeat(1, 4, 1, 1)
|
992 |
+
|
993 |
+
# torch.quantile function expects float32
|
994 |
+
if noise_guidance_edit_tmp_quantile.dtype == torch.float32:
|
995 |
+
tmp = torch.quantile(
|
996 |
+
noise_guidance_edit_tmp_quantile.flatten(start_dim=2),
|
997 |
+
edit_threshold_c,
|
998 |
+
dim=2,
|
999 |
+
keepdim=False,
|
1000 |
+
)
|
1001 |
+
else:
|
1002 |
+
tmp = torch.quantile(
|
1003 |
+
noise_guidance_edit_tmp_quantile.flatten(start_dim=2).to(torch.float32),
|
1004 |
+
edit_threshold_c,
|
1005 |
+
dim=2,
|
1006 |
+
keepdim=False,
|
1007 |
+
).to(noise_guidance_edit_tmp_quantile.dtype)
|
1008 |
+
|
1009 |
+
self.activation_mask[i, c] = torch.where(
|
1010 |
+
noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None],
|
1011 |
+
torch.ones_like(noise_guidance_edit_tmp),
|
1012 |
+
torch.zeros_like(noise_guidance_edit_tmp),
|
1013 |
+
).detach().cpu()
|
1014 |
+
|
1015 |
+
noise_guidance_edit_tmp = torch.where(
|
1016 |
+
noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None],
|
1017 |
+
noise_guidance_edit_tmp,
|
1018 |
+
torch.zeros_like(noise_guidance_edit_tmp),
|
1019 |
+
)
|
1020 |
+
|
1021 |
+
noise_guidance_edit[c, :, :, :, :] = noise_guidance_edit_tmp
|
1022 |
+
|
1023 |
+
warmup_inds = torch.tensor(warmup_inds).to(self.device)
|
1024 |
+
if len(noise_pred_edit_concepts) > warmup_inds.shape[0] > 0:
|
1025 |
+
concept_weights = concept_weights.to("cpu") # Offload to cpu
|
1026 |
+
noise_guidance_edit = noise_guidance_edit.to("cpu")
|
1027 |
+
|
1028 |
+
concept_weights_tmp = torch.index_select(concept_weights.to(self.device), 0, warmup_inds)
|
1029 |
+
concept_weights_tmp = torch.where(
|
1030 |
+
concept_weights_tmp < 0, torch.zeros_like(concept_weights_tmp), concept_weights_tmp
|
1031 |
)
|
1032 |
+
concept_weights_tmp = concept_weights_tmp / concept_weights_tmp.sum(dim=0)
|
1033 |
+
# concept_weights_tmp = torch.nan_to_num(concept_weights_tmp)
|
1034 |
|
1035 |
+
noise_guidance_edit_tmp = torch.index_select(
|
1036 |
+
noise_guidance_edit.to(self.device), 0, warmup_inds
|
1037 |
+
)
|
1038 |
+
noise_guidance_edit_tmp = torch.einsum(
|
1039 |
+
"cb,cbijk->bijk", concept_weights_tmp, noise_guidance_edit_tmp
|
1040 |
+
)
|
1041 |
+
noise_guidance_edit_tmp = noise_guidance_edit_tmp
|
1042 |
+
noise_guidance = noise_guidance + noise_guidance_edit_tmp
|
1043 |
|
1044 |
+
self.sem_guidance[i] = noise_guidance_edit_tmp.detach().cpu()
|
|
|
|
|
|
|
|
|
1045 |
|
1046 |
+
del noise_guidance_edit_tmp
|
1047 |
+
del concept_weights_tmp
|
1048 |
+
concept_weights = concept_weights.to(self.device)
|
1049 |
+
noise_guidance_edit = noise_guidance_edit.to(self.device)
|
1050 |
|
1051 |
+
concept_weights = torch.where(
|
1052 |
+
concept_weights < 0, torch.zeros_like(concept_weights), concept_weights
|
1053 |
+
)
|
1054 |
|
1055 |
+
concept_weights = torch.nan_to_num(concept_weights)
|
|
|
1056 |
|
1057 |
+
noise_guidance_edit = torch.einsum("cb,cbijk->bijk", concept_weights, noise_guidance_edit)
|
1058 |
+
|
1059 |
+
noise_guidance_edit = noise_guidance_edit + edit_momentum_scale * edit_momentum
|
1060 |
+
|
1061 |
+
edit_momentum = edit_mom_beta * edit_momentum + (1 - edit_mom_beta) * noise_guidance_edit
|
1062 |
+
|
1063 |
+
if warmup_inds.shape[0] == len(noise_pred_edit_concepts):
|
1064 |
+
noise_guidance = noise_guidance + noise_guidance_edit
|
1065 |
+
self.sem_guidance[i] = noise_guidance_edit.detach().cpu()
|
1066 |
+
|
1067 |
+
noise_pred = noise_pred_uncond + noise_guidance
|
1068 |
|
1069 |
# compute the previous noisy sample x_t -> x_t-1
|
1070 |
if use_ddpm:
|
|
|
1072 |
latents = self.scheduler.step(noise_pred, t, latents, variance_noise=zs[idx],
|
1073 |
**extra_step_kwargs).prev_sample
|
1074 |
|
1075 |
+
else: #if not use_ddpm:
|
1076 |
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
|
1077 |
|
1078 |
# step callback
|
|
|
1132 |
source_prompt: str = "",
|
1133 |
source_guidance_scale=3.5,
|
1134 |
num_inversion_steps: int = 30,
|
1135 |
+
skip: float = 0.15,
|
1136 |
eta: float = 1.0,
|
1137 |
generator: Optional[torch.Generator] = None,
|
1138 |
verbose=True,
|
|
|
1149 |
# self.eta = eta
|
1150 |
# assert (self.eta > 0)
|
1151 |
skip = skip/100
|
1152 |
+
print("YOOOOOOOOOOOOOOOOO ", skip, num_inversion_steps)
|
1153 |
train_steps = self.scheduler.config.num_train_timesteps
|
1154 |
timesteps = torch.from_numpy(
|
1155 |
np.linspace(train_steps - skip * train_steps - 1, 1, num_inversion_steps).astype(np.int64)).to(self.device)
|
|
|
1158 |
self.num_inversion_steps = timesteps.shape[0]
|
1159 |
self.scheduler.num_inference_steps = timesteps.shape[0]
|
1160 |
self.scheduler.timesteps = timesteps
|
1161 |
+
|
1162 |
+
# Reset attn processor, we do not want to store attn maps during inversion
|
1163 |
+
# self.unet.set_default_attn_processor()
|
1164 |
self.unet.set_attn_processor(AttnProcessor())
|
1165 |
+
|
1166 |
# 1. get embeddings
|
1167 |
|
1168 |
uncond_embedding = self.encode_text("")
|
|
|
1177 |
# autoencoder reconstruction
|
1178 |
# image_rec = self.vae.decode(x0 / self.vae.config.scaling_factor, return_dict=False)[0]
|
1179 |
# image_rec = self.image_processor.postprocess(image_rec, output_type="pil")
|
|
|
1180 |
# 3. find zs and xts
|
1181 |
variance_noise_shape = (
|
1182 |
self.num_inversion_steps,
|
|
|
1226 |
# self.zs = zs
|
1227 |
|
1228 |
|
|
|
1229 |
return zs, xts
|
1230 |
+
# return zs, xts, image_rec
|
1231 |
|
1232 |
@torch.no_grad()
|
1233 |
def encode_image(self, image_path, dtype=None):
|