Updated for Torch Compiling fix
Browse files- pipeline.py +222 -4
pipeline.py
CHANGED
@@ -7,6 +7,9 @@ import PIL
|
|
7 |
import torch
|
8 |
from packaging import version
|
9 |
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
|
|
|
|
|
|
|
10 |
|
11 |
from diffusers import DiffusionPipeline
|
12 |
from diffusers.configuration_utils import FrozenDict
|
@@ -792,7 +795,7 @@ class StableDiffusionLongPromptWeightingPipeline(
|
|
792 |
|
793 |
def decode_latents(self, latents):
|
794 |
latents = 1 / self.vae.config.scaling_factor * latents
|
795 |
-
image = self.vae.decode(latents).sample
|
796 |
image = (image / 2 + 0.5).clamp(0, 1)
|
797 |
# we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
|
798 |
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
|
@@ -1063,7 +1066,9 @@ class StableDiffusionLongPromptWeightingPipeline(
|
|
1063 |
t,
|
1064 |
encoder_hidden_states=prompt_embeds,
|
1065 |
cross_attention_kwargs=cross_attention_kwargs,
|
1066 |
-
|
|
|
|
|
1067 |
|
1068 |
# perform guidance
|
1069 |
if do_classifier_free_guidance:
|
@@ -1071,7 +1076,7 @@ class StableDiffusionLongPromptWeightingPipeline(
|
|
1071 |
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
1072 |
|
1073 |
# compute the previous noisy sample x_t -> x_t-1
|
1074 |
-
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
|
1075 |
|
1076 |
if mask is not None:
|
1077 |
# masking
|
@@ -1467,4 +1472,217 @@ class StableDiffusionLongPromptWeightingPipeline(
|
|
1467 |
is_cancelled_callback=is_cancelled_callback,
|
1468 |
callback_steps=callback_steps,
|
1469 |
cross_attention_kwargs=cross_attention_kwargs,
|
1470 |
-
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
7 |
import torch
|
8 |
from packaging import version
|
9 |
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
|
10 |
+
import random
|
11 |
+
import sys
|
12 |
+
from tqdm.auto import tqdm
|
13 |
|
14 |
from diffusers import DiffusionPipeline
|
15 |
from diffusers.configuration_utils import FrozenDict
|
|
|
795 |
|
796 |
def decode_latents(self, latents):
|
797 |
latents = 1 / self.vae.config.scaling_factor * latents
|
798 |
+
image = self.vae.decode(latents, return_dict=False)[0] #).sample
|
799 |
image = (image / 2 + 0.5).clamp(0, 1)
|
800 |
# we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
|
801 |
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
|
|
|
1066 |
t,
|
1067 |
encoder_hidden_states=prompt_embeds,
|
1068 |
cross_attention_kwargs=cross_attention_kwargs,
|
1069 |
+
return_dict=False,
|
1070 |
+
)[0]
|
1071 |
+
#).sample
|
1072 |
|
1073 |
# perform guidance
|
1074 |
if do_classifier_free_guidance:
|
|
|
1076 |
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
1077 |
|
1078 |
# compute the previous noisy sample x_t -> x_t-1
|
1079 |
+
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] #).prev_sample
|
1080 |
|
1081 |
if mask is not None:
|
1082 |
# masking
|
|
|
1472 |
is_cancelled_callback=is_cancelled_callback,
|
1473 |
callback_steps=callback_steps,
|
1474 |
cross_attention_kwargs=cross_attention_kwargs,
|
1475 |
+
)
|
1476 |
+
|
1477 |
+
|
1478 |
+
# Borrowed from https://github.com/csaluski/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py
|
1479 |
+
def get_text_latent_space(self, prompt, guidance_scale = 7.5):
|
1480 |
+
# get prompt text embeddings
|
1481 |
+
text_input = self.tokenizer(
|
1482 |
+
prompt,
|
1483 |
+
padding="max_length",
|
1484 |
+
max_length=self.tokenizer.model_max_length,
|
1485 |
+
truncation=True,
|
1486 |
+
return_tensors="pt",
|
1487 |
+
)
|
1488 |
+
text_embeddings = self.text_encoder(text_input.input_ids.to(self.device))[0]
|
1489 |
+
|
1490 |
+
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
1491 |
+
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
1492 |
+
# corresponds to doing no classifier free guidance.
|
1493 |
+
do_classifier_free_guidance = guidance_scale > 1.0
|
1494 |
+
# get unconditional embeddings for classifier free guidance
|
1495 |
+
if do_classifier_free_guidance:
|
1496 |
+
max_length = text_input.input_ids.shape[-1]
|
1497 |
+
uncond_input = self.tokenizer(
|
1498 |
+
[""], padding="max_length", max_length=max_length, return_tensors="pt"
|
1499 |
+
)
|
1500 |
+
uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
|
1501 |
+
|
1502 |
+
# For classifier free guidance, we need to do two forward passes.
|
1503 |
+
# Here we concatenate the unconditional and text embeddings into a single batch
|
1504 |
+
# to avoid doing two forward passes
|
1505 |
+
text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
|
1506 |
+
|
1507 |
+
return text_embeddings
|
1508 |
+
|
1509 |
+
def slerp(self, t, v0, v1, DOT_THRESHOLD=0.9995):
|
1510 |
+
""" helper function to spherically interpolate two arrays v1 v2
|
1511 |
+
from https://gist.github.com/karpathy/00103b0037c5aaea32fe1da1af553355
|
1512 |
+
this should be better than lerping for moving between noise spaces """
|
1513 |
+
|
1514 |
+
if not isinstance(v0, np.ndarray):
|
1515 |
+
inputs_are_torch = True
|
1516 |
+
input_device = v0.device
|
1517 |
+
v0 = v0.cpu().numpy()
|
1518 |
+
v1 = v1.cpu().numpy()
|
1519 |
+
|
1520 |
+
dot = np.sum(v0 * v1 / (np.linalg.norm(v0) * np.linalg.norm(v1)))
|
1521 |
+
if np.abs(dot) > DOT_THRESHOLD:
|
1522 |
+
v2 = (1 - t) * v0 + t * v1
|
1523 |
+
else:
|
1524 |
+
theta_0 = np.arccos(dot)
|
1525 |
+
sin_theta_0 = np.sin(theta_0)
|
1526 |
+
theta_t = theta_0 * t
|
1527 |
+
sin_theta_t = np.sin(theta_t)
|
1528 |
+
s0 = np.sin(theta_0 - theta_t) / sin_theta_0
|
1529 |
+
s1 = sin_theta_t / sin_theta_0
|
1530 |
+
v2 = s0 * v0 + s1 * v1
|
1531 |
+
|
1532 |
+
if inputs_are_torch:
|
1533 |
+
v2 = torch.from_numpy(v2).to(input_device)
|
1534 |
+
|
1535 |
+
return v2
|
1536 |
+
|
1537 |
+
def lerp_between_prompts(self, first_prompt, second_prompt, seed = None, length = 10, save=False, guidance_scale: Optional[float] = 7.5, **kwargs):
|
1538 |
+
first_embedding = self.get_text_latent_space(first_prompt)
|
1539 |
+
second_embedding = self.get_text_latent_space(second_prompt)
|
1540 |
+
if not seed:
|
1541 |
+
seed = random.randint(0, sys.maxsize)
|
1542 |
+
generator = torch.Generator(self.device)
|
1543 |
+
generator.manual_seed(seed)
|
1544 |
+
generator_state = generator.get_state()
|
1545 |
+
lerp_embed_points = []
|
1546 |
+
for i in range(length):
|
1547 |
+
weight = i / length
|
1548 |
+
tensor_lerp = torch.lerp(first_embedding, second_embedding, weight)
|
1549 |
+
lerp_embed_points.append(tensor_lerp)
|
1550 |
+
images = []
|
1551 |
+
for idx, latent_point in enumerate(lerp_embed_points):
|
1552 |
+
generator.set_state(generator_state)
|
1553 |
+
image = self.diffuse_from_inits(latent_point, **kwargs)["image"][0]
|
1554 |
+
images.append(image)
|
1555 |
+
if save:
|
1556 |
+
image.save(f"{first_prompt}-{second_prompt}-{idx:02d}.png", "PNG")
|
1557 |
+
return {"images": images, "latent_points": lerp_embed_points,"generator_state": generator_state}
|
1558 |
+
|
1559 |
+
def slerp_through_seeds(self,
|
1560 |
+
prompt,
|
1561 |
+
height: Optional[int] = 512,
|
1562 |
+
width: Optional[int] = 512,
|
1563 |
+
save = False,
|
1564 |
+
seed = None, steps = 10, **kwargs):
|
1565 |
+
|
1566 |
+
if not seed:
|
1567 |
+
seed = random.randint(0, sys.maxsize)
|
1568 |
+
generator = torch.Generator(self.device)
|
1569 |
+
generator.manual_seed(seed)
|
1570 |
+
init_start = torch.randn(
|
1571 |
+
(1, self.unet.in_channels, height // 8, width // 8),
|
1572 |
+
generator = generator, device = self.device)
|
1573 |
+
init_end = torch.randn(
|
1574 |
+
(1, self.unet.in_channels, height // 8, width // 8),
|
1575 |
+
generator = generator, device = self.device)
|
1576 |
+
generator_state = generator.get_state()
|
1577 |
+
slerp_embed_points = []
|
1578 |
+
# weight from 0 to 1/(steps - 1), add init_end specifically so that we
|
1579 |
+
# have len(images) = steps
|
1580 |
+
for i in range(steps - 1):
|
1581 |
+
weight = i / steps
|
1582 |
+
tensor_slerp = self.slerp(weight, init_start, init_end)
|
1583 |
+
slerp_embed_points.append(tensor_slerp)
|
1584 |
+
slerp_embed_points.append(init_end)
|
1585 |
+
images = []
|
1586 |
+
embed_point = self.get_text_latent_space(prompt)
|
1587 |
+
for idx, noise_point in enumerate(slerp_embed_points):
|
1588 |
+
generator.set_state(generator_state)
|
1589 |
+
image = self.diffuse_from_inits(embed_point, init = noise_point, **kwargs)["image"][0]
|
1590 |
+
images.append(image)
|
1591 |
+
if save:
|
1592 |
+
image.save(f"{seed}-{idx:02d}.png", "PNG")
|
1593 |
+
return {"images": images, "noise_samples": slerp_embed_points,"generator_state": generator_state}
|
1594 |
+
|
1595 |
+
@torch.no_grad()
|
1596 |
+
def diffuse_from_inits(self, text_embeddings,
|
1597 |
+
init = None,
|
1598 |
+
height: Optional[int] = 512,
|
1599 |
+
width: Optional[int] = 512,
|
1600 |
+
num_inference_steps: Optional[int] = 50,
|
1601 |
+
guidance_scale: Optional[float] = 7.5,
|
1602 |
+
eta: Optional[float] = 0.0,
|
1603 |
+
generator: Optional[torch.Generator] = None,
|
1604 |
+
output_type: Optional[str] = "pil",
|
1605 |
+
**kwargs,):
|
1606 |
+
|
1607 |
+
from diffusers.schedulers import LMSDiscreteScheduler
|
1608 |
+
batch_size = 1
|
1609 |
+
|
1610 |
+
if generator == None:
|
1611 |
+
generator = torch.Generator("cuda")
|
1612 |
+
generator_state = generator.get_state()
|
1613 |
+
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
1614 |
+
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
1615 |
+
# corresponds to doing no classifier free guidance.
|
1616 |
+
do_classifier_free_guidance = guidance_scale > 1.0
|
1617 |
+
# get the intial random noise
|
1618 |
+
latents = init if init is not None else torch.randn(
|
1619 |
+
(batch_size, self.unet.in_channels, height // 8, width // 8),
|
1620 |
+
generator=generator,
|
1621 |
+
device=self.device,)
|
1622 |
+
|
1623 |
+
# set timesteps
|
1624 |
+
accepts_offset = "offset" in set(inspect.signature(self.scheduler.set_timesteps).parameters.keys())
|
1625 |
+
extra_set_kwargs = {}
|
1626 |
+
if accepts_offset:
|
1627 |
+
extra_set_kwargs["offset"] = 1
|
1628 |
+
|
1629 |
+
self.scheduler.set_timesteps(num_inference_steps, **extra_set_kwargs)
|
1630 |
+
|
1631 |
+
# if we use LMSDiscreteScheduler, let's make sure latents are mulitplied by sigmas
|
1632 |
+
if isinstance(self.scheduler, LMSDiscreteScheduler):
|
1633 |
+
latents = latents * self.scheduler.sigmas[0]
|
1634 |
+
|
1635 |
+
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
|
1636 |
+
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
|
1637 |
+
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
|
1638 |
+
# and should be between [0, 1]
|
1639 |
+
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
1640 |
+
extra_step_kwargs = {}
|
1641 |
+
if accepts_eta:
|
1642 |
+
extra_step_kwargs["eta"] = eta
|
1643 |
+
|
1644 |
+
for i, t in tqdm(enumerate(self.scheduler.timesteps)):
|
1645 |
+
# expand the latents if we are doing classifier free guidance
|
1646 |
+
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
|
1647 |
+
if isinstance(self.scheduler, LMSDiscreteScheduler):
|
1648 |
+
sigma = self.scheduler.sigmas[i]
|
1649 |
+
latent_model_input = latent_model_input / ((sigma**2 + 1) ** 0.5)
|
1650 |
+
|
1651 |
+
# predict the noise residual
|
1652 |
+
noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings, return_dict=False)[0] #).sample
|
1653 |
+
|
1654 |
+
# perform guidance
|
1655 |
+
if do_classifier_free_guidance:
|
1656 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
1657 |
+
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
1658 |
+
|
1659 |
+
# compute the previous noisy sample x_t -> x_t-1
|
1660 |
+
if isinstance(self.scheduler, LMSDiscreteScheduler):
|
1661 |
+
latents = self.scheduler.step(noise_pred, i, latents, **extra_step_kwargs, return_dict=False)[0] #).prev_sample
|
1662 |
+
else:
|
1663 |
+
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] #).prev_sample
|
1664 |
+
|
1665 |
+
# scale and decode the image latents with vae
|
1666 |
+
latents = 1 / 0.18215 * latents
|
1667 |
+
image = self.vae.decode(latents)
|
1668 |
+
|
1669 |
+
image = (image / 2 + 0.5).clamp(0, 1)
|
1670 |
+
image = image.cpu().permute(0, 2, 3, 1).numpy()
|
1671 |
+
|
1672 |
+
if output_type == "pil":
|
1673 |
+
image = self.numpy_to_pil(image)
|
1674 |
+
|
1675 |
+
return {"image": image, "generator_state": generator_state}
|
1676 |
+
|
1677 |
+
def variation(self, text_embeddings, generator_state, variation_magnitude = 100, **kwargs):
|
1678 |
+
# random vector to move in latent space
|
1679 |
+
rand_t = (torch.rand(text_embeddings.shape, device = self.device) * 2) - 1
|
1680 |
+
rand_mag = torch.sum(torch.abs(rand_t)) / variation_magnitude
|
1681 |
+
scaled_rand_t = rand_t / rand_mag
|
1682 |
+
variation_embedding = text_embeddings + scaled_rand_t
|
1683 |
+
|
1684 |
+
generator = torch.Generator("cuda")
|
1685 |
+
generator.set_state(generator_state)
|
1686 |
+
result = self.diffuse_from_inits(variation_embedding, generator=generator, **kwargs)
|
1687 |
+
result.update({"latent_point": variation_embedding})
|
1688 |
+
return result
|