1130 lines
54 KiB
Python
1130 lines
54 KiB
Python
# Copyright (c) 2025 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team.
|
|
# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
|
|
# SPDX-License-Identifier: Apache-2.0
|
|
#
|
|
# This file has been modified by Bytedance Ltd. and/or its affiliates on September 15, 2025.
|
|
#
|
|
# Original file was released under Apache License 2.0, with the full license text
|
|
# available at https://github.com/huggingface/finetrainers/blob/main/LICENSE.
|
|
#
|
|
# This modified file is released under the same license.
|
|
|
|
|
|
import inspect
|
|
import math
|
|
import random
|
|
from typing import Any, Callable, Dict, List, Optional, Tuple, Union, Literal
|
|
|
|
import PIL
|
|
import torch
|
|
from transformers import T5EncoderModel, T5Tokenizer
|
|
|
|
from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
|
|
from diffusers.image_processor import PipelineImageInput
|
|
from diffusers.loaders import CogVideoXLoraLoaderMixin
|
|
from diffusers.models import AutoencoderKLCogVideoX, CogVideoXTransformer3DModel
|
|
# from diffusers.models.embeddings import get_3d_rotary_pos_embed # Use MOT version instead
|
|
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
|
|
from diffusers.schedulers import CogVideoXDDIMScheduler, CogVideoXDPMScheduler
|
|
from diffusers.utils import (
|
|
is_torch_xla_available,
|
|
logging,
|
|
replace_example_docstring,
|
|
)
|
|
from diffusers.utils.torch_utils import randn_tensor
|
|
from diffusers.video_processor import VideoProcessor
|
|
from diffusers.pipelines.cogvideo.pipeline_output import CogVideoXPipelineOutput
|
|
|
|
from .cogvideox_transformer_3d_mot import CogVideoXTransformer3DMOTModel
|
|
from .embeddings_mot import get_3d_rotary_pos_embed
|
|
|
|
if is_torch_xla_available():
|
|
import torch_xla.core.xla_model as xm
|
|
|
|
XLA_AVAILABLE = True
|
|
else:
|
|
XLA_AVAILABLE = False
|
|
|
|
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
|
|
|
def soft_empty_cache(force=False):
|
|
torch.cuda.empty_cache()
|
|
torch.cuda.ipc_collect()
|
|
|
|
EXAMPLE_DOC_STRING = """
|
|
Examples:
|
|
```py
|
|
>>> import torch
|
|
>>> from diffusers import CogVideoXImageToVideoPipeline
|
|
>>> from diffusers.utils import export_to_video, load_image
|
|
|
|
>>> pipe = CogVideoXImageToVideoPipeline.from_pretrained("THUDM/CogVideoX-5b-I2V", torch_dtype=torch.bfloat16)
|
|
>>> pipe.to("cuda")
|
|
|
|
>>> prompt = "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot."
|
|
>>> image = load_image(
|
|
... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
|
|
... )
|
|
>>> video = pipe(image, prompt, use_dynamic_cfg=True)
|
|
>>> export_to_video(video.frames[0], "output.mp4", fps=8)
|
|
```
|
|
"""
|
|
|
|
|
|
# Similar to diffusers.pipelines.hunyuandit.pipeline_hunyuandit.get_resize_crop_region_for_grid
|
|
def get_resize_crop_region_for_grid(src, tgt_width, tgt_height):
|
|
tw = tgt_width
|
|
th = tgt_height
|
|
h, w = src
|
|
r = h / w
|
|
if r > (th / tw):
|
|
resize_height = th
|
|
resize_width = int(round(th / h * w))
|
|
else:
|
|
resize_width = tw
|
|
resize_height = int(round(tw / w * h))
|
|
|
|
crop_top = int(round((th - resize_height) / 2.0))
|
|
crop_left = int(round((tw - resize_width) / 2.0))
|
|
|
|
return (crop_top, crop_left), (crop_top + resize_height, crop_left + resize_width)
|
|
|
|
|
|
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
|
|
def retrieve_timesteps(
|
|
scheduler,
|
|
num_inference_steps: Optional[int] = None,
|
|
device: Optional[Union[str, torch.device]] = None,
|
|
timesteps: Optional[List[int]] = None,
|
|
sigmas: Optional[List[float]] = None,
|
|
**kwargs,
|
|
):
|
|
r"""
|
|
Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
|
|
custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
|
|
|
|
Args:
|
|
scheduler (`SchedulerMixin`):
|
|
The scheduler to get timesteps from.
|
|
num_inference_steps (`int`):
|
|
The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
|
|
must be `None`.
|
|
device (`str` or `torch.device`, *optional*):
|
|
The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
|
|
timesteps (`List[int]`, *optional*):
|
|
Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
|
|
`num_inference_steps` and `sigmas` must be `None`.
|
|
sigmas (`List[float]`, *optional*):
|
|
Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
|
|
`num_inference_steps` and `timesteps` must be `None`.
|
|
|
|
Returns:
|
|
`Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
|
|
second element is the number of inference steps.
|
|
"""
|
|
if timesteps is not None and sigmas is not None:
|
|
raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
|
|
if timesteps is not None:
|
|
accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
|
|
if not accepts_timesteps:
|
|
raise ValueError(
|
|
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
|
|
f" timestep schedules. Please check whether you are using the correct scheduler."
|
|
)
|
|
scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
|
|
timesteps = scheduler.timesteps
|
|
num_inference_steps = len(timesteps)
|
|
elif sigmas is not None:
|
|
accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
|
|
if not accept_sigmas:
|
|
raise ValueError(
|
|
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
|
|
f" sigmas schedules. Please check whether you are using the correct scheduler."
|
|
)
|
|
scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
|
|
timesteps = scheduler.timesteps
|
|
num_inference_steps = len(timesteps)
|
|
else:
|
|
scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
|
|
timesteps = scheduler.timesteps
|
|
return timesteps, num_inference_steps
|
|
|
|
|
|
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
|
|
def retrieve_latents(
|
|
encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
|
|
):
|
|
if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
|
|
return encoder_output.latent_dist.sample(generator)
|
|
elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
|
|
return encoder_output.latent_dist.mode()
|
|
elif hasattr(encoder_output, "latents"):
|
|
return encoder_output.latents
|
|
else:
|
|
raise AttributeError("Could not access latents of provided encoder_output")
|
|
|
|
|
|
class CogVideoXImageToVideoMOTPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
|
|
r"""
|
|
Pipeline for image-to-video generation using CogVideoX.
|
|
|
|
This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
|
|
library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
|
|
|
|
Args:
|
|
vae ([`AutoencoderKL`]):
|
|
Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
|
|
text_encoder ([`T5EncoderModel`]):
|
|
Frozen text-encoder. CogVideoX uses
|
|
[T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel); specifically the
|
|
[t5-v1_1-xxl](https://huggingface.co/PixArt-alpha/PixArt-alpha/tree/main/t5-v1_1-xxl) variant.
|
|
tokenizer (`T5Tokenizer`):
|
|
Tokenizer of class
|
|
[T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
|
|
transformer ([`CogVideoXTransformer3DModel`]):
|
|
A text conditioned `CogVideoXTransformer3DModel` to denoise the encoded video latents.
|
|
scheduler ([`SchedulerMixin`]):
|
|
A scheduler to be used in combination with `transformer` to denoise the encoded video latents.
|
|
"""
|
|
|
|
_optional_components = []
|
|
model_cpu_offload_seq = "text_encoder->transformer->vae"
|
|
|
|
_callback_tensor_inputs = [
|
|
"latents",
|
|
"prompt_embeds",
|
|
"negative_prompt_embeds",
|
|
]
|
|
|
|
def __init__(
|
|
self,
|
|
tokenizer: T5Tokenizer,
|
|
text_encoder: T5EncoderModel,
|
|
vae: AutoencoderKLCogVideoX,
|
|
transformer: CogVideoXTransformer3DMOTModel,
|
|
scheduler: Union[CogVideoXDDIMScheduler, CogVideoXDPMScheduler],
|
|
):
|
|
super().__init__()
|
|
|
|
self.register_modules(
|
|
tokenizer=tokenizer,
|
|
text_encoder=text_encoder,
|
|
vae=vae,
|
|
transformer=transformer,
|
|
scheduler=scheduler,
|
|
)
|
|
self.vae_scale_factor_spatial = (
|
|
2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
|
|
)
|
|
self.vae_scale_factor_temporal = (
|
|
self.vae.config.temporal_compression_ratio if getattr(self, "vae", None) else 4
|
|
)
|
|
self.vae_scaling_factor_image = self.vae.config.scaling_factor if getattr(self, "vae", None) else 0.7
|
|
|
|
self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
|
|
|
|
# Copied from diffusers.pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipeline._get_t5_prompt_embeds
|
|
def _get_t5_prompt_embeds(
|
|
self,
|
|
prompt: Union[str, List[str]] = None,
|
|
num_videos_per_prompt: int = 1,
|
|
max_sequence_length: int = 226,
|
|
device: Optional[torch.device] = None,
|
|
dtype: Optional[torch.dtype] = None,
|
|
):
|
|
device = device or self._execution_device
|
|
dtype = dtype or self.text_encoder.dtype
|
|
|
|
prompt = [prompt] if isinstance(prompt, str) else prompt
|
|
batch_size = len(prompt)
|
|
|
|
text_inputs = self.tokenizer(
|
|
prompt,
|
|
padding="max_length",
|
|
max_length=max_sequence_length,
|
|
truncation=True,
|
|
add_special_tokens=True,
|
|
return_tensors="pt",
|
|
)
|
|
text_input_ids = text_inputs.input_ids
|
|
untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
|
|
|
|
if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
|
|
removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
|
|
logger.warning(
|
|
"The following part of your input was truncated because `max_sequence_length` is set to "
|
|
f" {max_sequence_length} tokens: {removed_text}"
|
|
)
|
|
|
|
prompt_embeds = self.text_encoder(text_input_ids.to(device))[0]
|
|
prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
|
|
|
|
# duplicate text embeddings for each generation per prompt, using mps friendly method
|
|
_, seq_len, _ = prompt_embeds.shape
|
|
prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
|
|
prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
|
|
|
|
return prompt_embeds
|
|
|
|
# Copied from diffusers.pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipeline.encode_prompt
|
|
def encode_prompt(
|
|
self,
|
|
prompt: Union[str, List[str]],
|
|
negative_prompt: Optional[Union[str, List[str]]] = None,
|
|
do_classifier_free_guidance: bool = True,
|
|
num_videos_per_prompt: int = 1,
|
|
prompt_embeds: Optional[torch.Tensor] = None,
|
|
negative_prompt_embeds: Optional[torch.Tensor] = None,
|
|
max_sequence_length: int = 226,
|
|
device: Optional[torch.device] = None,
|
|
dtype: Optional[torch.dtype] = None,
|
|
):
|
|
r"""
|
|
Encodes the prompt into text encoder hidden states.
|
|
|
|
Args:
|
|
prompt (`str` or `List[str]`, *optional*):
|
|
prompt to be encoded
|
|
negative_prompt (`str` or `List[str]`, *optional*):
|
|
The prompt or prompts not to guide the image generation. If not defined, one has to pass
|
|
`negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
|
|
less than `1`).
|
|
do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
|
|
Whether to use classifier free guidance or not.
|
|
num_videos_per_prompt (`int`, *optional*, defaults to 1):
|
|
Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
|
|
prompt_embeds (`torch.Tensor`, *optional*):
|
|
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
|
|
provided, text embeddings will be generated from `prompt` input argument.
|
|
negative_prompt_embeds (`torch.Tensor`, *optional*):
|
|
Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
|
|
weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
|
|
argument.
|
|
device: (`torch.device`, *optional*):
|
|
torch device
|
|
dtype: (`torch.dtype`, *optional*):
|
|
torch dtype
|
|
"""
|
|
device = device or self._execution_device
|
|
|
|
prompt = [prompt] if isinstance(prompt, str) else prompt
|
|
if prompt is not None:
|
|
batch_size = len(prompt)
|
|
else:
|
|
batch_size = prompt_embeds.shape[0]
|
|
|
|
if prompt_embeds is None:
|
|
prompt_embeds = self._get_t5_prompt_embeds(
|
|
prompt=prompt,
|
|
num_videos_per_prompt=num_videos_per_prompt,
|
|
max_sequence_length=max_sequence_length,
|
|
device=device,
|
|
dtype=dtype,
|
|
)
|
|
|
|
if do_classifier_free_guidance and negative_prompt_embeds is None:
|
|
negative_prompt = negative_prompt or ""
|
|
negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
|
|
|
|
if prompt is not None and type(prompt) is not type(negative_prompt):
|
|
raise TypeError(
|
|
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
|
|
f" {type(prompt)}."
|
|
)
|
|
elif batch_size != len(negative_prompt):
|
|
raise ValueError(
|
|
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
|
|
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
|
|
" the batch size of `prompt`."
|
|
)
|
|
|
|
negative_prompt_embeds = self._get_t5_prompt_embeds(
|
|
prompt=negative_prompt,
|
|
num_videos_per_prompt=num_videos_per_prompt,
|
|
max_sequence_length=max_sequence_length,
|
|
device=device,
|
|
dtype=dtype,
|
|
)
|
|
|
|
return prompt_embeds, negative_prompt_embeds
|
|
|
|
def prepare_latents(
|
|
self,
|
|
image: torch.Tensor,
|
|
batch_size: int = 1,
|
|
num_channels_latents: int = 16,
|
|
num_frames: int = 13,
|
|
height: int = 60,
|
|
width: int = 90,
|
|
dtype: Optional[torch.dtype] = None,
|
|
device: Optional[torch.device] = None,
|
|
generator: Optional[torch.Generator] = None,
|
|
latents: Optional[torch.Tensor] = None,
|
|
ref_videos: Optional[List[torch.Tensor]] = None,
|
|
ref_video_first_frames: Optional[torch.Tensor] = None,
|
|
):
|
|
if isinstance(generator, list) and len(generator) != batch_size:
|
|
raise ValueError(
|
|
f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
|
|
f" size of {batch_size}. Make sure the batch size matches the length of the generators."
|
|
)
|
|
|
|
num_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
|
|
shape = (
|
|
batch_size,
|
|
num_frames,
|
|
num_channels_latents,
|
|
height // self.vae_scale_factor_spatial,
|
|
width // self.vae_scale_factor_spatial,
|
|
)
|
|
|
|
# For CogVideoX1.5, the latent should add 1 for padding (Not use)
|
|
if self.transformer.config.patch_size_t is not None:
|
|
shape = shape[:1] + (shape[1] + shape[1] % self.transformer.config.patch_size_t,) + shape[2:]
|
|
|
|
image = image.unsqueeze(2) # [B, C, F, H, W]
|
|
|
|
if isinstance(generator, list):
|
|
image_latents = [
|
|
retrieve_latents(self.vae.encode(image[i].unsqueeze(0)), generator[i]) for i in range(batch_size)
|
|
]
|
|
else:
|
|
image_latents = [retrieve_latents(self.vae.encode(img.unsqueeze(0)), generator) for img in image]
|
|
|
|
image_latents = torch.cat(image_latents, dim=0).to(dtype).permute(0, 2, 1, 3, 4) # [B, F, C, H, W]
|
|
|
|
if not self.vae.config.invert_scale_latents:
|
|
image_latents = self.vae_scaling_factor_image * image_latents
|
|
else:
|
|
# This is awkward but required because the CogVideoX team forgot to multiply the
|
|
# scaling factor during training :)
|
|
# image_latents = 1 / self.vae_scaling_factor_image * image_latents
|
|
image_latents = 1.0 * image_latents
|
|
|
|
padding_shape = (
|
|
batch_size,
|
|
num_frames - 1,
|
|
num_channels_latents,
|
|
height // self.vae_scale_factor_spatial,
|
|
width // self.vae_scale_factor_spatial,
|
|
)
|
|
|
|
latent_padding = torch.zeros(padding_shape, device=device, dtype=dtype)
|
|
image_latents = torch.cat([image_latents, latent_padding], dim=1)
|
|
|
|
# Select the first frame along the second dimension
|
|
if self.transformer.config.patch_size_t is not None:
|
|
first_frame = image_latents[:, : image_latents.size(1) % self.transformer.config.patch_size_t, ...]
|
|
image_latents = torch.cat([first_frame, image_latents], dim=1)
|
|
|
|
if latents is None:
|
|
latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
|
|
else:
|
|
latents = latents.to(device)
|
|
|
|
# scale the initial noise by the standard deviation required by the scheduler
|
|
latents = latents * self.scheduler.init_noise_sigma
|
|
|
|
|
|
ref_video_latents_list = []
|
|
for ref_video in ref_videos:
|
|
if isinstance(generator, list):
|
|
init_latents = [
|
|
retrieve_latents(self.vae.encode(ref_video[i].unsqueeze(0)), generator[i]) for i in range(batch_size)
|
|
]
|
|
else:
|
|
init_latents = [retrieve_latents(self.vae.encode(vid.unsqueeze(0)), generator) for vid in ref_video]
|
|
|
|
init_latents = torch.cat(init_latents, dim=0).to(dtype).permute(0, 2, 1, 3, 4) # [B, F, C, H, W]
|
|
ref_video_latents = self.vae_scaling_factor_image * init_latents
|
|
|
|
if self.transformer.config.patch_size_t is not None:
|
|
additional_frames = self.transformer.config.patch_size_t - (ref_video_latents.size(1) % self.transformer.config.patch_size_t)
|
|
if additional_frames > 0:
|
|
last_frame = ref_video_latents[:, -1:]
|
|
padding_frames = last_frame.expand(-1, additional_frames, -1, -1, -1)
|
|
ref_video_latents = torch.cat([ref_video_latents, padding_frames], dim=1)
|
|
ref_video_latents_list.append(ref_video_latents)
|
|
|
|
ref_video_first_frame_latents_list = []
|
|
for ref_video_first_frame in ref_video_first_frames:
|
|
|
|
ref_video_first_frame = ref_video_first_frame.unsqueeze(2) # [B, C, F, H, W]
|
|
|
|
if isinstance(generator, list):
|
|
ref_video_first_frame_latents = [
|
|
retrieve_latents(self.vae.encode(ref_video_first_frame[i].unsqueeze(0)), generator[i]) for i in range(batch_size)
|
|
]
|
|
else:
|
|
ref_video_first_frame_latents = [retrieve_latents(self.vae.encode(img.unsqueeze(0)), generator) for img in ref_video_first_frame]
|
|
|
|
ref_video_first_frame_latents = torch.cat(ref_video_first_frame_latents, dim=0).to(dtype).permute(0, 2, 1, 3, 4) # [B, F, C, H, W]
|
|
|
|
if not self.vae.config.invert_scale_latents:
|
|
ref_video_first_frame_latents = self.vae_scaling_factor_image * ref_video_first_frame_latents
|
|
else:
|
|
# # This is awkward but required because the CogVideoX team forgot to multiply the
|
|
# # scaling factor during training :)
|
|
# ref_video_first_frame_latents = 1 / self.vae_scaling_factor_image * ref_video_first_frame_latents
|
|
ref_video_first_frame_latents = 1.0 * ref_video_first_frame_latents
|
|
|
|
padding_shape = (
|
|
batch_size,
|
|
num_frames - 1,
|
|
num_channels_latents,
|
|
height // self.vae_scale_factor_spatial,
|
|
width // self.vae_scale_factor_spatial,
|
|
)
|
|
|
|
latent_padding = torch.zeros(padding_shape, device=device, dtype=dtype)
|
|
ref_video_first_frame_latents = torch.cat([ref_video_first_frame_latents, latent_padding], dim=1)
|
|
|
|
# Select the first frame along the second dimension
|
|
if self.transformer.config.patch_size_t is not None:
|
|
first_frame = ref_video_first_frame_latents[:, : ref_video_first_frame_latents.size(1) % self.transformer.config.patch_size_t, ...]
|
|
ref_video_first_frame_latents = torch.cat([first_frame, ref_video_first_frame_latents], dim=1)
|
|
|
|
ref_video_first_frame_latents_list.append(ref_video_first_frame_latents)
|
|
|
|
|
|
ref_video_latents = torch.cat(ref_video_latents_list, dim=1)
|
|
ref_video_image_latents = torch.cat(ref_video_first_frame_latents_list, dim=1)
|
|
|
|
return latents, image_latents, ref_video_latents, ref_video_image_latents
|
|
|
|
# Copied from diffusers.pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipeline.decode_latents
|
|
def decode_latents(self, latents: torch.Tensor) -> torch.Tensor:
|
|
latents = latents.permute(0, 2, 1, 3, 4) # [batch_size, num_channels, num_frames, height, width]
|
|
latents = 1 / self.vae_scaling_factor_image * latents
|
|
|
|
frames = self.vae.decode(latents).sample
|
|
return frames
|
|
|
|
# Copied from diffusers.pipelines.animatediff.pipeline_animatediff_video2video.AnimateDiffVideoToVideoPipeline.get_timesteps
|
|
def get_timesteps(self, num_inference_steps, timesteps, strength, device):
|
|
# get the original timestep using init_timestep
|
|
init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
|
|
|
|
t_start = max(num_inference_steps - init_timestep, 0)
|
|
timesteps = timesteps[t_start * self.scheduler.order :]
|
|
|
|
return timesteps, num_inference_steps - t_start
|
|
|
|
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
|
|
def prepare_extra_step_kwargs(self, generator, eta):
|
|
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
|
|
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
|
|
# eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502
|
|
# and should be between [0, 1]
|
|
|
|
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
|
extra_step_kwargs = {}
|
|
if accepts_eta:
|
|
extra_step_kwargs["eta"] = eta
|
|
|
|
# check if the scheduler accepts generator
|
|
accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
|
if accepts_generator:
|
|
extra_step_kwargs["generator"] = generator
|
|
return extra_step_kwargs
|
|
|
|
def check_inputs(
|
|
self,
|
|
image,
|
|
prompt,
|
|
height,
|
|
width,
|
|
negative_prompt,
|
|
callback_on_step_end_tensor_inputs,
|
|
latents=None,
|
|
prompt_embeds=None,
|
|
negative_prompt_embeds=None,
|
|
):
|
|
if (
|
|
not isinstance(image, torch.Tensor)
|
|
and not isinstance(image, PIL.Image.Image)
|
|
and not isinstance(image, list)
|
|
):
|
|
raise ValueError(
|
|
"`image` has to be of type `torch.Tensor` or `PIL.Image.Image` or `List[PIL.Image.Image]` but is"
|
|
f" {type(image)}"
|
|
)
|
|
|
|
if height % 8 != 0 or width % 8 != 0:
|
|
raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
|
|
|
|
if callback_on_step_end_tensor_inputs is not None and not all(
|
|
k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
|
|
):
|
|
raise ValueError(
|
|
f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
|
|
)
|
|
if prompt is not None and prompt_embeds is not None:
|
|
raise ValueError(
|
|
f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
|
|
" only forward one of the two."
|
|
)
|
|
elif prompt is None and prompt_embeds is None:
|
|
raise ValueError(
|
|
"Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
|
|
)
|
|
elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
|
|
raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
|
|
|
|
if prompt is not None and negative_prompt_embeds is not None:
|
|
raise ValueError(
|
|
f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:"
|
|
f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
|
|
)
|
|
|
|
if negative_prompt is not None and negative_prompt_embeds is not None:
|
|
raise ValueError(
|
|
f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
|
|
f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
|
|
)
|
|
|
|
if prompt_embeds is not None and negative_prompt_embeds is not None:
|
|
if prompt_embeds.shape != negative_prompt_embeds.shape:
|
|
raise ValueError(
|
|
"`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
|
|
f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
|
|
f" {negative_prompt_embeds.shape}."
|
|
)
|
|
|
|
# Copied from diffusers.pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipeline.fuse_qkv_projections
|
|
def fuse_qkv_projections(self) -> None:
|
|
r"""Enables fused QKV projections."""
|
|
self.fusing_transformer = True
|
|
self.transformer.fuse_qkv_projections()
|
|
|
|
# Copied from diffusers.pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipeline.unfuse_qkv_projections
|
|
def unfuse_qkv_projections(self) -> None:
|
|
r"""Disable QKV projection fusion if enabled."""
|
|
if not self.fusing_transformer:
|
|
logger.warning("The Transformer was not initially fused for QKV projections. Doing nothing.")
|
|
else:
|
|
self.transformer.unfuse_qkv_projections()
|
|
self.fusing_transformer = False
|
|
|
|
# Copied from diffusers.pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipeline._prepare_rotary_positional_embeddings
|
|
def _prepare_rotary_positional_embeddings(
|
|
self,
|
|
height: int,
|
|
width: int,
|
|
num_frames: int,
|
|
device: torch.device,
|
|
mot_num: int = 0,
|
|
ref_type: str = "continous_negative",
|
|
) -> Tuple[torch.Tensor, torch.Tensor]:
|
|
grid_height = height // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
|
|
grid_width = width // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
|
|
|
|
p = self.transformer.config.patch_size
|
|
p_t = self.transformer.config.patch_size_t
|
|
|
|
base_size_width = self.transformer.config.sample_width // p
|
|
base_size_height = self.transformer.config.sample_height // p
|
|
|
|
if p_t is None:
|
|
# CogVideoX 1.0
|
|
grid_crops_coords = get_resize_crop_region_for_grid(
|
|
(grid_height, grid_width), base_size_width, base_size_height
|
|
)
|
|
freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
|
|
embed_dim=self.transformer.config.attention_head_dim,
|
|
crops_coords=grid_crops_coords,
|
|
grid_size=(grid_height, grid_width),
|
|
temporal_size=num_frames,
|
|
device=device,
|
|
mot_num=mot_num,
|
|
ref_type=ref_type,
|
|
)
|
|
else:
|
|
# CogVideoX 1.5
|
|
base_num_frames = (num_frames + p_t - 1) // p_t
|
|
|
|
freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
|
|
embed_dim=self.transformer.config.attention_head_dim,
|
|
crops_coords=None,
|
|
grid_size=(grid_height, grid_width),
|
|
temporal_size=base_num_frames,
|
|
grid_type="slice",
|
|
max_size=(base_size_height, base_size_width),
|
|
device=device,
|
|
mot_num=mot_num,
|
|
)
|
|
|
|
return freqs_cos, freqs_sin
|
|
|
|
@property
|
|
def guidance_scale(self):
|
|
return self._guidance_scale
|
|
|
|
@property
|
|
def num_timesteps(self):
|
|
return self._num_timesteps
|
|
|
|
@property
|
|
def attention_kwargs(self):
|
|
return self._attention_kwargs
|
|
|
|
@property
|
|
def current_timestep(self):
|
|
return self._current_timestep
|
|
|
|
@property
|
|
def interrupt(self):
|
|
return self._interrupt
|
|
|
|
@torch.no_grad()
|
|
@replace_example_docstring(EXAMPLE_DOC_STRING)
|
|
def __call__(
|
|
self,
|
|
image: PipelineImageInput,
|
|
prompt: Optional[Union[str, List[str]]] = None,
|
|
negative_prompt: Optional[Union[str, List[str]]] = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards",
|
|
height: Optional[int] = None,
|
|
width: Optional[int] = None,
|
|
num_frames: int = 49,
|
|
num_inference_steps: int = 50,
|
|
timesteps: Optional[List[int]] = None,
|
|
guidance_scale: float = 6,
|
|
use_dynamic_cfg: bool = False,
|
|
num_videos_per_prompt: int = 1,
|
|
eta: float = 0.0,
|
|
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
|
latents: Optional[torch.FloatTensor] = None,
|
|
prompt_embeds: Optional[torch.FloatTensor] = None,
|
|
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
|
|
output_type: str = "pil",
|
|
return_dict: bool = True,
|
|
attention_kwargs: Optional[Dict[str, Any]] = None,
|
|
callback_on_step_end: Optional[
|
|
Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
|
|
] = None,
|
|
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
|
|
max_sequence_length: int = 226,
|
|
# mot
|
|
ref_videos: Optional[List[List[PIL.Image.Image]]] = None,
|
|
prompt_mot_ref: Optional[Union[str, List[str]]] = None,
|
|
negative_prompt_mot_ref: Optional[Union[str, List[str]]] = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards",
|
|
empty_caption_mot_ref: Optional[str] = None,
|
|
caption_guidance_scale: float = 0,
|
|
effect_types: Optional[List[str]] = None,
|
|
reference_train_mode: Optional[str] = None,
|
|
random_refer_noise: Optional[bool] = False,
|
|
frames_selection: Literal["first", "evenly", "random"] = "evenly",
|
|
ref_type: Optional[str] = "continous_negative",
|
|
# ablation
|
|
ablation_single_branch: bool = False,
|
|
# baseline
|
|
baseline_single_condition: Optional[str] = None,
|
|
**kwargs,
|
|
) -> Union[CogVideoXPipelineOutput, Tuple]:
|
|
"""
|
|
Function invoked when calling the pipeline for generation.
|
|
|
|
Args:
|
|
image (`PipelineImageInput`):
|
|
The input image to condition the generation on. Must be an image, a list of images or a `torch.Tensor`.
|
|
prompt (`str` or `List[str]`, *optional*):
|
|
The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
|
|
instead.
|
|
negative_prompt (`str` or `List[str]`, *optional*):
|
|
The prompt or prompts not to guide the image generation. If not defined, one has to pass
|
|
`negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
|
|
less than `1`).
|
|
height (`int`, *optional*, defaults to self.transformer.config.sample_height * self.vae_scale_factor_spatial):
|
|
The height in pixels of the generated image. This is set to 480 by default for the best results.
|
|
width (`int`, *optional*, defaults to self.transformer.config.sample_height * self.vae_scale_factor_spatial):
|
|
The width in pixels of the generated image. This is set to 720 by default for the best results.
|
|
num_frames (`int`, defaults to `48`):
|
|
Number of frames to generate. Must be divisible by self.vae_scale_factor_temporal. Generated video will
|
|
contain 1 extra frame because CogVideoX is conditioned with (num_seconds * fps + 1) frames where
|
|
num_seconds is 6 and fps is 8. However, since videos can be saved at any fps, the only condition that
|
|
needs to be satisfied is that of divisibility mentioned above.
|
|
num_inference_steps (`int`, *optional*, defaults to 50):
|
|
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
|
|
expense of slower inference.
|
|
timesteps (`List[int]`, *optional*):
|
|
Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
|
|
in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
|
|
passed will be used. Must be in descending order.
|
|
guidance_scale (`float`, *optional*, defaults to 7.0):
|
|
Guidance scale as defined in [Classifier-Free Diffusion
|
|
Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
|
|
of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting
|
|
`guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to
|
|
the text `prompt`, usually at the expense of lower image quality.
|
|
num_videos_per_prompt (`int`, *optional*, defaults to 1):
|
|
The number of videos to generate per prompt.
|
|
generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
|
|
One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
|
|
to make generation deterministic.
|
|
latents (`torch.FloatTensor`, *optional*):
|
|
Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
|
|
generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
|
|
tensor will ge generated by sampling using the supplied random `generator`.
|
|
prompt_embeds (`torch.FloatTensor`, *optional*):
|
|
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
|
|
provided, text embeddings will be generated from `prompt` input argument.
|
|
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
|
|
Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
|
|
weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
|
|
argument.
|
|
output_type (`str`, *optional*, defaults to `"pil"`):
|
|
The output format of the generate image. Choose between
|
|
[PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
|
|
return_dict (`bool`, *optional*, defaults to `True`):
|
|
Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
|
|
of a plain tuple.
|
|
attention_kwargs (`dict`, *optional*):
|
|
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
|
|
`self.processor` in
|
|
[diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
|
|
callback_on_step_end (`Callable`, *optional*):
|
|
A function that calls at the end of each denoising steps during the inference. The function is called
|
|
with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
|
|
callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
|
|
`callback_on_step_end_tensor_inputs`.
|
|
callback_on_step_end_tensor_inputs (`List`, *optional*):
|
|
The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
|
|
will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
|
|
`._callback_tensor_inputs` attribute of your pipeline class.
|
|
max_sequence_length (`int`, defaults to `226`):
|
|
Maximum sequence length in encoded prompt. Must be consistent with
|
|
`self.transformer.config.max_text_seq_length` otherwise may lead to poor results.
|
|
|
|
Examples:
|
|
|
|
Returns:
|
|
[`~pipelines.cogvideo.pipeline_output.CogVideoXPipelineOutput`] or `tuple`:
|
|
[`~pipelines.cogvideo.pipeline_output.CogVideoXPipelineOutput`] if `return_dict` is True, otherwise a
|
|
`tuple`. When returning a tuple, the first element is a list with the generated images.
|
|
"""
|
|
|
|
if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
|
|
callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
|
|
|
|
height = height or self.transformer.config.sample_height * self.vae_scale_factor_spatial
|
|
width = width or self.transformer.config.sample_width * self.vae_scale_factor_spatial
|
|
num_frames = num_frames or self.transformer.config.sample_frames
|
|
|
|
num_videos_per_prompt = 1
|
|
|
|
#kiki
|
|
update = kwargs.get('update_func', lambda *args, **kwargs: None)
|
|
|
|
# 1. Check inputs. Raise error if not correct
|
|
self.check_inputs(
|
|
image=image,
|
|
prompt=prompt,
|
|
height=height,
|
|
width=width,
|
|
negative_prompt=negative_prompt,
|
|
callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
|
|
latents=latents,
|
|
prompt_embeds=prompt_embeds,
|
|
negative_prompt_embeds=negative_prompt_embeds,
|
|
)
|
|
self._guidance_scale = guidance_scale
|
|
self._current_timestep = None
|
|
self._attention_kwargs = attention_kwargs
|
|
self._interrupt = False
|
|
|
|
# 2. Default call parameters
|
|
if prompt is not None and isinstance(prompt, str):
|
|
batch_size = 1
|
|
elif prompt is not None and isinstance(prompt, list):
|
|
batch_size = len(prompt)
|
|
else:
|
|
batch_size = prompt_embeds.shape[0]
|
|
|
|
device = self._execution_device
|
|
# kiki
|
|
# device = "cuda"
|
|
|
|
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
|
# of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1`
|
|
# corresponds to doing no classifier free guidance.
|
|
do_classifier_free_guidance = guidance_scale > 1.0
|
|
|
|
|
|
# kiki
|
|
# self.text_encoder.to(device)
|
|
# 3. Encode input prompt
|
|
prompt_embeds, negative_prompt_embeds = self.encode_prompt(
|
|
prompt=prompt,
|
|
negative_prompt=negative_prompt,
|
|
do_classifier_free_guidance=do_classifier_free_guidance,
|
|
num_videos_per_prompt=num_videos_per_prompt,
|
|
prompt_embeds=prompt_embeds,
|
|
negative_prompt_embeds=negative_prompt_embeds,
|
|
max_sequence_length=max_sequence_length,
|
|
device=device,
|
|
)
|
|
if do_classifier_free_guidance:
|
|
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
|
|
|
|
prompt_embeds_mot_ref_list = []
|
|
negative_prompt_embeds_mot_ref_list = []
|
|
for prompt_mot_ref_item in prompt_mot_ref:
|
|
prompt_embeds_mot_ref, negative_prompt_embeds_mot_ref = self.encode_prompt(
|
|
prompt=prompt_mot_ref_item,
|
|
negative_prompt=negative_prompt_mot_ref,
|
|
do_classifier_free_guidance=do_classifier_free_guidance,
|
|
num_videos_per_prompt=num_videos_per_prompt,
|
|
prompt_embeds=None,
|
|
negative_prompt_embeds=None,
|
|
max_sequence_length=max_sequence_length,
|
|
device=device,
|
|
)
|
|
prompt_embeds_mot_ref_list.append(prompt_embeds_mot_ref)
|
|
negative_prompt_embeds_mot_ref_list.append(negative_prompt_embeds_mot_ref)
|
|
prompt_embeds_mot_ref = torch.cat(prompt_embeds_mot_ref_list, dim=1)
|
|
negative_prompt_embeds_mot_ref = torch.cat(negative_prompt_embeds_mot_ref_list, dim=1)
|
|
if do_classifier_free_guidance:
|
|
prompt_embeds_mot_ref = torch.cat([negative_prompt_embeds_mot_ref, prompt_embeds_mot_ref], dim=0)
|
|
|
|
update()
|
|
# kiki
|
|
# print('----> step 3 done')
|
|
# self.text_encoder.to("cpu")
|
|
# kiki_debug_gpu('step 3')
|
|
# 4. Prepare timesteps
|
|
timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
|
|
self._num_timesteps = len(timesteps)
|
|
|
|
# 5. Prepare latents
|
|
latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
|
|
|
|
# For CogVideoX 1.5, the latent frames should be padded to make it divisible by patch_size_t
|
|
ori_num_frames = num_frames
|
|
patch_size_t = self.transformer.config.patch_size_t
|
|
additional_frames = 0
|
|
if patch_size_t is not None and latent_frames % patch_size_t != 0:
|
|
additional_frames = patch_size_t - latent_frames % patch_size_t
|
|
num_frames += additional_frames * self.vae_scale_factor_temporal
|
|
|
|
image = self.video_processor.preprocess(image, height=height, width=width).to(
|
|
device, dtype=prompt_embeds.dtype
|
|
)
|
|
|
|
if frames_selection == "first":
|
|
ref_videos = [ref_video[:num_frames] for ref_video in ref_videos]
|
|
elif frames_selection == "evenly":
|
|
indices = torch.linspace(0, len(ref_videos[0]) - 1, num_frames).long().tolist()
|
|
for ref_idx, ref_video in enumerate(ref_videos):
|
|
ref_videos[ref_idx] = [ref_videos[ref_idx][i] for i in indices]
|
|
elif frames_selection == "random":
|
|
for ref_idx, ref_video in enumerate(ref_videos):
|
|
start_index = random.randint(0, len(ref_videos[ref_idx]) - num_frames)
|
|
indices = torch.arange(start_index, start_index + num_frames).long().tolist()
|
|
ref_videos[ref_idx] = [ref_videos[ref_idx][i] for i in indices]
|
|
else:
|
|
raise ValueError(f"Invalid frames_selection: {frames_selection}. Choose from 'first' or 'evenly' or 'random'.")
|
|
|
|
tmp_ref_videos = ref_videos
|
|
|
|
ref_video_first_frames = [ref_video[0] for ref_video in ref_videos]
|
|
ref_video_first_frames = [self.video_processor.preprocess(ref_video_first_frame, height=height, width=width) for ref_video_first_frame in ref_video_first_frames]
|
|
ref_video_first_frames = [ref_video_first_frame.to(device=device, dtype=prompt_embeds.dtype) for ref_video_first_frame in ref_video_first_frames]
|
|
|
|
ref_videos = [self.video_processor.preprocess_video(ref_video, height=height, width=width).to(device=device, dtype=prompt_embeds.dtype) for ref_video in ref_videos]
|
|
|
|
# kiki
|
|
# self.vae.to(device)
|
|
latent_channels = self.transformer.config.in_channels // 2
|
|
latents, image_latents, ref_video_latents, ref_video_image_latents = self.prepare_latents(
|
|
image,
|
|
batch_size * num_videos_per_prompt,
|
|
latent_channels,
|
|
num_frames,
|
|
height,
|
|
width,
|
|
prompt_embeds.dtype,
|
|
device,
|
|
generator,
|
|
latents,
|
|
ref_videos=ref_videos,
|
|
ref_video_first_frames=ref_video_first_frames,
|
|
)
|
|
|
|
mot_num = int(ref_video_latents.shape[1] // latents.shape[1])
|
|
|
|
# kiki
|
|
update()
|
|
# print('----> step 5 done')
|
|
# self.vae.to("cpu")
|
|
# kiki_debug_gpu('step 5')
|
|
# 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
|
|
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
|
|
|
|
# 7. Create rotary embeds if required
|
|
image_rotary_emb = (
|
|
self._prepare_rotary_positional_embeddings(height, width, latents.size(1), device)
|
|
if self.transformer.config.use_rotary_positional_embeddings
|
|
else None
|
|
)
|
|
|
|
image_rotary_emb_mot_ref = (
|
|
self._prepare_rotary_positional_embeddings(height, width, latents.size(1), device, mot_num=mot_num, ref_type=ref_type)
|
|
if self.transformer.config.use_rotary_positional_embeddings
|
|
else None
|
|
)
|
|
|
|
# 8. Create ofs embeds if required
|
|
ofs_emb = None if self.transformer.config.ofs_embed_dim is None else latents.new_full((1,), fill_value=2.0)
|
|
|
|
# 8. Denoising loop
|
|
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
|
|
|
|
|
|
if ablation_single_branch and not baseline_single_condition:
|
|
image_rotary_emb = (
|
|
torch.cat([image_rotary_emb[0], image_rotary_emb_mot_ref[0]], dim=0),
|
|
torch.cat([image_rotary_emb[1], image_rotary_emb_mot_ref[1]], dim=0)
|
|
)
|
|
|
|
# kiki
|
|
# self.transformer.to(device)
|
|
|
|
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
|
# for DPM-solver++
|
|
old_pred_original_sample = None
|
|
for i, t in enumerate(timesteps):
|
|
if self.interrupt:
|
|
continue
|
|
|
|
self._current_timestep = t
|
|
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
|
|
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
|
|
|
latent_image_input = torch.cat([image_latents] * 2) if do_classifier_free_guidance else image_latents
|
|
latent_model_input = torch.cat([latent_model_input, latent_image_input], dim=2)
|
|
|
|
latent_model_input_mot_ref = torch.cat([ref_video_latents] * 2) if do_classifier_free_guidance else ref_video_latents
|
|
latent_image_input_mot_ref = torch.cat([ref_video_image_latents] * 2) if do_classifier_free_guidance else ref_video_image_latents
|
|
latent_model_input_mot_ref = torch.cat([latent_model_input_mot_ref, latent_image_input_mot_ref], dim=2)
|
|
|
|
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
|
timestep = t.expand(latent_model_input.shape[0])
|
|
if not ablation_single_branch:
|
|
# predict noise model_output
|
|
noise_pred = self.transformer(
|
|
hidden_states=latent_model_input,
|
|
encoder_hidden_states=prompt_embeds,
|
|
timestep=timestep,
|
|
ofs=ofs_emb,
|
|
image_rotary_emb=image_rotary_emb,
|
|
attention_kwargs=attention_kwargs,
|
|
return_dict=False,
|
|
# mot
|
|
num_mot_ref=mot_num,
|
|
hidden_states_mot_ref=latent_model_input_mot_ref,
|
|
encoder_hidden_states_mot_ref=prompt_embeds_mot_ref,
|
|
image_rotary_emb_mot_ref=image_rotary_emb_mot_ref,
|
|
effect_types=effect_types,
|
|
timestep_list_mot_ref=clean_ref_video_latent_timesteps_list if random_refer_noise else None,
|
|
)[0]
|
|
else:
|
|
if not baseline_single_condition:
|
|
latent_model_input = torch.cat([latent_model_input, latent_model_input_mot_ref], dim=1)
|
|
|
|
noise_pred = self.transformer(
|
|
hidden_states=latent_model_input,
|
|
encoder_hidden_states=prompt_embeds,
|
|
timestep=timestep,
|
|
ofs=ofs_emb,
|
|
image_rotary_emb=image_rotary_emb,
|
|
attention_kwargs=attention_kwargs,
|
|
return_dict=False,
|
|
# ablation
|
|
ablation_single_branch=ablation_single_branch,
|
|
)[0]
|
|
|
|
noise_pred = noise_pred[:, :latents.shape[1], ...]
|
|
else:
|
|
|
|
noise_pred = self.transformer(
|
|
hidden_states=latent_model_input,
|
|
encoder_hidden_states=prompt_embeds,
|
|
timestep=timestep,
|
|
ofs=ofs_emb,
|
|
image_rotary_emb=image_rotary_emb,
|
|
attention_kwargs=attention_kwargs,
|
|
return_dict=False,
|
|
# ablation
|
|
ablation_single_branch=False,
|
|
)[0]
|
|
|
|
noise_pred = noise_pred.float()
|
|
|
|
# perform guidance
|
|
if use_dynamic_cfg:
|
|
self._guidance_scale = 1 + guidance_scale * (
|
|
(1 - math.cos(math.pi * ((num_inference_steps - t.item()) / num_inference_steps) ** 5.0)) / 2
|
|
)
|
|
if do_classifier_free_guidance:
|
|
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
|
noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
|
|
|
|
# compute the previous noisy sample x_t -> x_t-1
|
|
if not isinstance(self.scheduler, CogVideoXDPMScheduler):
|
|
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
|
|
else:
|
|
latents, old_pred_original_sample = self.scheduler.step(
|
|
noise_pred,
|
|
old_pred_original_sample,
|
|
t,
|
|
timesteps[i - 1] if i > 0 else None,
|
|
latents,
|
|
**extra_step_kwargs,
|
|
return_dict=False,
|
|
)
|
|
latents = latents.to(prompt_embeds.dtype)
|
|
|
|
#kiki
|
|
update()
|
|
|
|
# call the callback, if provided
|
|
if callback_on_step_end is not None:
|
|
callback_kwargs = {}
|
|
for k in callback_on_step_end_tensor_inputs:
|
|
callback_kwargs[k] = locals()[k]
|
|
callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
|
|
|
|
latents = callback_outputs.pop("latents", latents)
|
|
prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
|
|
negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
|
|
|
|
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
|
|
progress_bar.update()
|
|
|
|
if XLA_AVAILABLE:
|
|
xm.mark_step()
|
|
|
|
self._current_timestep = None
|
|
|
|
# kiki
|
|
# self.transformer.to("cpu")
|
|
# kiki_debug_gpu('step 8')
|
|
# self.vae.to(device)
|
|
# kiki_debug_gpu('vae reload')
|
|
|
|
if not output_type == "latent":
|
|
# Discard any padding frames that were added for CogVideoX 1.5
|
|
latents = latents[:, additional_frames:]
|
|
video = self.decode_latents(latents)
|
|
video = self.video_processor.postprocess_video(video=video, output_type=output_type)
|
|
else:
|
|
video = latents
|
|
|
|
# Offload all models
|
|
self.maybe_free_model_hooks()
|
|
|
|
if not return_dict:
|
|
return (video,)
|
|
|
|
return CogVideoXPipelineOutput(frames=video)
|