Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -66,7 +66,7 @@ from diffusers.utils import (
|
|
66 |
)
|
67 |
from diffusers.utils.torch_utils import randn_tensor
|
68 |
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
|
69 |
-
from
|
70 |
|
71 |
|
72 |
if is_invisible_watermark_available():
|
@@ -428,10 +428,11 @@ class LEditsPPPipelineStableDiffusionXL(
|
|
428 |
editing_prompt: Optional[str] = None,
|
429 |
editing_prompt_embeds: Optional[torch.Tensor] = None,
|
430 |
editing_pooled_prompt_embeds: Optional[torch.Tensor] = None,
|
431 |
-
avg_diff
|
432 |
-
|
433 |
-
correlation_weight_factor
|
434 |
scale=2,
|
|
|
435 |
) -> object:
|
436 |
r"""
|
437 |
Encodes the prompt into text encoder hidden states.
|
@@ -551,9 +552,8 @@ class LEditsPPPipelineStableDiffusionXL(
|
|
551 |
negative_pooled_prompt_embeds = negative_prompt_embeds[0]
|
552 |
negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
|
553 |
|
554 |
-
if avg_diff is not None
|
555 |
-
#scale=3
|
556 |
-
print("SHALOM neg")
|
557 |
normed_prompt_embeds = negative_prompt_embeds / negative_prompt_embeds.norm(dim=-1, keepdim=True)
|
558 |
sims = normed_prompt_embeds[0] @ normed_prompt_embeds[0].T
|
559 |
if j == 0:
|
@@ -562,15 +562,26 @@ class LEditsPPPipelineStableDiffusionXL(
|
|
562 |
standard_weights = torch.ones_like(weights)
|
563 |
|
564 |
weights = standard_weights + (weights - standard_weights) * correlation_weight_factor
|
565 |
-
edit_concepts_embeds = negative_prompt_embeds + (
|
|
|
|
|
|
|
|
|
|
|
|
|
566 |
else:
|
567 |
weights = sims[toks.argmax(), :][None, :, None].repeat(1, 1, 1280)
|
568 |
|
569 |
standard_weights = torch.ones_like(weights)
|
570 |
|
571 |
weights = standard_weights + (weights - standard_weights) * correlation_weight_factor
|
572 |
-
edit_concepts_embeds = negative_prompt_embeds + (
|
|
|
573 |
|
|
|
|
|
|
|
|
|
574 |
|
575 |
negative_prompt_embeds_list.append(negative_prompt_embeds)
|
576 |
j+=1
|
@@ -858,8 +869,8 @@ class LEditsPPPipelineStableDiffusionXL(
|
|
858 |
|
859 |
self.unet.set_attn_processor(attn_procs)
|
860 |
|
861 |
-
@spaces.GPU
|
862 |
@torch.no_grad()
|
|
|
863 |
@replace_example_docstring(EXAMPLE_DOC_STRING)
|
864 |
def __call__(
|
865 |
self,
|
@@ -892,10 +903,12 @@ class LEditsPPPipelineStableDiffusionXL(
|
|
892 |
clip_skip: Optional[int] = None,
|
893 |
callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
|
894 |
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
|
895 |
-
avg_diff
|
896 |
-
|
897 |
-
correlation_weight_factor
|
898 |
scale=2,
|
|
|
|
|
899 |
init_latents: [torch.Tensor] = None,
|
900 |
zs: [torch.Tensor] = None,
|
901 |
**kwargs,
|
@@ -1102,9 +1115,10 @@ class LEditsPPPipelineStableDiffusionXL(
|
|
1102 |
editing_prompt_embeds=editing_prompt_embeddings,
|
1103 |
editing_pooled_prompt_embeds=editing_pooled_prompt_embeds,
|
1104 |
avg_diff = avg_diff,
|
1105 |
-
|
1106 |
correlation_weight_factor = correlation_weight_factor,
|
1107 |
scale=scale,
|
|
|
1108 |
)
|
1109 |
|
1110 |
# 4. Prepare timesteps
|
@@ -1475,7 +1489,6 @@ class LEditsPPPipelineStableDiffusionXL(
|
|
1475 |
|
1476 |
@torch.no_grad()
|
1477 |
# Modified from diffusers.pipelines.ledits_pp.pipeline_leditspp_stable_diffusion.LEditsPPPipelineStableDiffusion.encode_image
|
1478 |
-
@spaces.GPU
|
1479 |
def encode_image(self, image, dtype=None, height=None, width=None, resize_mode="default", crops_coords=None):
|
1480 |
image = self.image_processor.preprocess(
|
1481 |
image=image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords
|
@@ -1504,8 +1517,8 @@ class LEditsPPPipelineStableDiffusionXL(
|
|
1504 |
x0 = self.vae.config.scaling_factor * x0
|
1505 |
return x0, resized
|
1506 |
|
1507 |
-
@spaces.GPU
|
1508 |
@torch.no_grad()
|
|
|
1509 |
def invert(
|
1510 |
self,
|
1511 |
image: PipelineImageInput,
|
@@ -1669,20 +1682,17 @@ class LEditsPPPipelineStableDiffusionXL(
|
|
1669 |
t_to_idx = {int(v): k for k, v in enumerate(timesteps)}
|
1670 |
xts = torch.zeros(size=variance_noise_shape, device=self.device, dtype=negative_prompt_embeds.dtype)
|
1671 |
|
1672 |
-
print("pre loop 1")
|
1673 |
for t in reversed(timesteps):
|
1674 |
idx = num_inversion_steps - t_to_idx[int(t)] - 1
|
1675 |
noise = randn_tensor(shape=x0.shape, generator=generator, device=self.device, dtype=x0.dtype)
|
1676 |
xts[idx] = self.scheduler.add_noise(x0, noise, t.unsqueeze(0))
|
1677 |
xts = torch.cat([x0.unsqueeze(0), xts], dim=0)
|
1678 |
-
|
1679 |
-
|
1680 |
# noise maps
|
1681 |
zs = torch.zeros(size=variance_noise_shape, device=self.device, dtype=negative_prompt_embeds.dtype)
|
1682 |
|
1683 |
self.scheduler.set_timesteps(len(self.scheduler.timesteps))
|
1684 |
|
1685 |
-
print("pre loop 2")
|
1686 |
for t in self.progress_bar(timesteps):
|
1687 |
idx = num_inversion_steps - t_to_idx[int(t)] - 1
|
1688 |
# 1. predict noise residual
|
@@ -1714,21 +1724,18 @@ class LEditsPPPipelineStableDiffusionXL(
|
|
1714 |
|
1715 |
# correction to avoid error accumulation
|
1716 |
xts[idx] = xtm1_corrected
|
1717 |
-
print("post loop 2")
|
1718 |
|
1719 |
-
|
1720 |
zs = zs.flip(0)
|
1721 |
-
|
1722 |
if num_zero_noise_steps > 0:
|
1723 |
zs[-num_zero_noise_steps:] = torch.zeros_like(zs[-num_zero_noise_steps:])
|
1724 |
-
|
1725 |
-
#self.zs = zs
|
1726 |
#return LEditsPPInversionPipelineOutput(images=resized, vae_reconstruction_images=image_rec)
|
1727 |
return xts[-1], zs
|
1728 |
|
1729 |
|
1730 |
# Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.rescale_noise_cfg
|
1731 |
-
@spaces.GPU
|
1732 |
def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
|
1733 |
"""
|
1734 |
Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
|
@@ -1783,7 +1790,6 @@ def compute_noise_ddim(scheduler, prev_latents, latents, timestep, noise_pred, e
|
|
1783 |
|
1784 |
|
1785 |
# Copied from diffusers.pipelines.ledits_pp.pipeline_leditspp_stable_diffusion.compute_noise_sde_dpm_pp_2nd
|
1786 |
-
@spaces.GPU
|
1787 |
def compute_noise_sde_dpm_pp_2nd(scheduler, prev_latents, latents, timestep, noise_pred, eta):
|
1788 |
def first_order_update(model_output, sample): # timestep, prev_timestep, sample):
|
1789 |
sigma_t, sigma_s = scheduler.sigmas[scheduler.step_index + 1], scheduler.sigmas[scheduler.step_index]
|
@@ -1869,7 +1875,6 @@ def compute_noise_sde_dpm_pp_2nd(scheduler, prev_latents, latents, timestep, noi
|
|
1869 |
|
1870 |
|
1871 |
# Copied from diffusers.pipelines.ledits_pp.pipeline_leditspp_stable_diffusion.compute_noise
|
1872 |
-
@spaces.GPU
|
1873 |
def compute_noise(scheduler, *args):
|
1874 |
if isinstance(scheduler, DDIMScheduler):
|
1875 |
return compute_noise_ddim(scheduler, *args)
|
|
|
66 |
)
|
67 |
from diffusers.utils.torch_utils import randn_tensor
|
68 |
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
|
69 |
+
from .pipeline_output import LEditsPPDiffusionPipelineOutput, LEditsPPInversionPipelineOutput
|
70 |
|
71 |
|
72 |
if is_invisible_watermark_available():
|
|
|
428 |
editing_prompt: Optional[str] = None,
|
429 |
editing_prompt_embeds: Optional[torch.Tensor] = None,
|
430 |
editing_pooled_prompt_embeds: Optional[torch.Tensor] = None,
|
431 |
+
avg_diff=None, # [0] -> text encoder 1,[1] ->text encoder 2
|
432 |
+
avg_diff_2nd=None, # text encoder 1,2
|
433 |
+
correlation_weight_factor=0.7,
|
434 |
scale=2,
|
435 |
+
scale_2nd=2,
|
436 |
) -> object:
|
437 |
r"""
|
438 |
Encodes the prompt into text encoder hidden states.
|
|
|
552 |
negative_pooled_prompt_embeds = negative_prompt_embeds[0]
|
553 |
negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
|
554 |
|
555 |
+
if avg_diff is not None:
|
556 |
+
# scale=3
|
|
|
557 |
normed_prompt_embeds = negative_prompt_embeds / negative_prompt_embeds.norm(dim=-1, keepdim=True)
|
558 |
sims = normed_prompt_embeds[0] @ normed_prompt_embeds[0].T
|
559 |
if j == 0:
|
|
|
562 |
standard_weights = torch.ones_like(weights)
|
563 |
|
564 |
weights = standard_weights + (weights - standard_weights) * correlation_weight_factor
|
565 |
+
edit_concepts_embeds = negative_prompt_embeds + (
|
566 |
+
weights * avg_diff[0][None, :].repeat(1, tokenizer.model_max_length, 1) * scale)
|
567 |
+
|
568 |
+
if avg_diff_2nd is not None:
|
569 |
+
edit_concepts_embeds += (weights * avg_diff_2nd[0][None, :].repeat(1,
|
570 |
+
self.pipe.tokenizer.model_max_length,
|
571 |
+
1) * scale_2nd)
|
572 |
else:
|
573 |
weights = sims[toks.argmax(), :][None, :, None].repeat(1, 1, 1280)
|
574 |
|
575 |
standard_weights = torch.ones_like(weights)
|
576 |
|
577 |
weights = standard_weights + (weights - standard_weights) * correlation_weight_factor
|
578 |
+
edit_concepts_embeds = negative_prompt_embeds + (
|
579 |
+
weights * avg_diff[1][None, :].repeat(1, tokenizer.model_max_length, 1) * scale)
|
580 |
|
581 |
+
if avg_diff_2nd is not None:
|
582 |
+
edit_concepts_embeds += (weights * avg_diff_2nd[1][None, :].repeat(1,
|
583 |
+
self.pipe.tokenizer_2.model_max_length,
|
584 |
+
1) * scale_2nd)
|
585 |
|
586 |
negative_prompt_embeds_list.append(negative_prompt_embeds)
|
587 |
j+=1
|
|
|
869 |
|
870 |
self.unet.set_attn_processor(attn_procs)
|
871 |
|
|
|
872 |
@torch.no_grad()
|
873 |
+
@spaces.GPU
|
874 |
@replace_example_docstring(EXAMPLE_DOC_STRING)
|
875 |
def __call__(
|
876 |
self,
|
|
|
903 |
clip_skip: Optional[int] = None,
|
904 |
callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
|
905 |
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
|
906 |
+
avg_diff=None, # [0] -> text encoder 1,[1] ->text encoder 2
|
907 |
+
avg_diff_2nd=None, # text encoder 1,2
|
908 |
+
correlation_weight_factor=0.7,
|
909 |
scale=2,
|
910 |
+
scale_2nd=2,
|
911 |
+
correlation_weight_factor = 0.7,
|
912 |
init_latents: [torch.Tensor] = None,
|
913 |
zs: [torch.Tensor] = None,
|
914 |
**kwargs,
|
|
|
1115 |
editing_prompt_embeds=editing_prompt_embeddings,
|
1116 |
editing_pooled_prompt_embeds=editing_pooled_prompt_embeds,
|
1117 |
avg_diff = avg_diff,
|
1118 |
+
avg_diff_2nd = avg_diff_2nd,
|
1119 |
correlation_weight_factor = correlation_weight_factor,
|
1120 |
scale=scale,
|
1121 |
+
scale_2nd=scale_2nd
|
1122 |
)
|
1123 |
|
1124 |
# 4. Prepare timesteps
|
|
|
1489 |
|
1490 |
@torch.no_grad()
|
1491 |
# Modified from diffusers.pipelines.ledits_pp.pipeline_leditspp_stable_diffusion.LEditsPPPipelineStableDiffusion.encode_image
|
|
|
1492 |
def encode_image(self, image, dtype=None, height=None, width=None, resize_mode="default", crops_coords=None):
|
1493 |
image = self.image_processor.preprocess(
|
1494 |
image=image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords
|
|
|
1517 |
x0 = self.vae.config.scaling_factor * x0
|
1518 |
return x0, resized
|
1519 |
|
|
|
1520 |
@torch.no_grad()
|
1521 |
+
@spaces.GPU
|
1522 |
def invert(
|
1523 |
self,
|
1524 |
image: PipelineImageInput,
|
|
|
1682 |
t_to_idx = {int(v): k for k, v in enumerate(timesteps)}
|
1683 |
xts = torch.zeros(size=variance_noise_shape, device=self.device, dtype=negative_prompt_embeds.dtype)
|
1684 |
|
|
|
1685 |
for t in reversed(timesteps):
|
1686 |
idx = num_inversion_steps - t_to_idx[int(t)] - 1
|
1687 |
noise = randn_tensor(shape=x0.shape, generator=generator, device=self.device, dtype=x0.dtype)
|
1688 |
xts[idx] = self.scheduler.add_noise(x0, noise, t.unsqueeze(0))
|
1689 |
xts = torch.cat([x0.unsqueeze(0), xts], dim=0)
|
1690 |
+
|
|
|
1691 |
# noise maps
|
1692 |
zs = torch.zeros(size=variance_noise_shape, device=self.device, dtype=negative_prompt_embeds.dtype)
|
1693 |
|
1694 |
self.scheduler.set_timesteps(len(self.scheduler.timesteps))
|
1695 |
|
|
|
1696 |
for t in self.progress_bar(timesteps):
|
1697 |
idx = num_inversion_steps - t_to_idx[int(t)] - 1
|
1698 |
# 1. predict noise residual
|
|
|
1724 |
|
1725 |
# correction to avoid error accumulation
|
1726 |
xts[idx] = xtm1_corrected
|
|
|
1727 |
|
1728 |
+
self.init_latents = xts[-1]
|
1729 |
zs = zs.flip(0)
|
1730 |
+
|
1731 |
if num_zero_noise_steps > 0:
|
1732 |
zs[-num_zero_noise_steps:] = torch.zeros_like(zs[-num_zero_noise_steps:])
|
1733 |
+
self.zs = zs
|
|
|
1734 |
#return LEditsPPInversionPipelineOutput(images=resized, vae_reconstruction_images=image_rec)
|
1735 |
return xts[-1], zs
|
1736 |
|
1737 |
|
1738 |
# Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.rescale_noise_cfg
|
|
|
1739 |
def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
|
1740 |
"""
|
1741 |
Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
|
|
|
1790 |
|
1791 |
|
1792 |
# Copied from diffusers.pipelines.ledits_pp.pipeline_leditspp_stable_diffusion.compute_noise_sde_dpm_pp_2nd
|
|
|
1793 |
def compute_noise_sde_dpm_pp_2nd(scheduler, prev_latents, latents, timestep, noise_pred, eta):
|
1794 |
def first_order_update(model_output, sample): # timestep, prev_timestep, sample):
|
1795 |
sigma_t, sigma_s = scheduler.sigmas[scheduler.step_index + 1], scheduler.sigmas[scheduler.step_index]
|
|
|
1875 |
|
1876 |
|
1877 |
# Copied from diffusers.pipelines.ledits_pp.pipeline_leditspp_stable_diffusion.compute_noise
|
|
|
1878 |
def compute_noise(scheduler, *args):
|
1879 |
if isinstance(scheduler, DDIMScheduler):
|
1880 |
return compute_noise_ddim(scheduler, *args)
|