mirror of
https://github.com/huggingface/diffusers.git
synced 2026-01-27 17:22:53 +03:00
Add STG to community pipelines (#10960)
* Support STG for video pipelines * Update README.md * Update README.md * Update README.md * Update README.md * Update README.md * Update README.md * Update pipeline_stg_cogvideox.py * Update pipeline_stg_hunyuan_video.py * Update pipeline_stg_ltx.py * Update pipeline_stg_ltx_image2video.py * Update pipeline_stg_mochi.py * Update pipeline_stg_hunyuan_video.py * Update pipeline_stg_ltx.py * Update pipeline_stg_ltx_image2video.py * Update pipeline_stg_mochi.py * update * remove rescaling * Apply style fixes --------- Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
This commit is contained in:
Kinam Kim
GitHub
@@ -10,6 +10,7 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
|
||||
|
||||
| Example | Description | Code Example | Colab | Author |
|
||||
|:--------------------------------------------------------------------------------------------------------------------------------------|:---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:------------------------------------------------------------------------------------------|:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------:|
|
||||
|Spatiotemporal Skip Guidance (STG)|[Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling](https://arxiv.org/abs/2411.18664) (CVPR 2025) enhances video diffusion models by generating a weaker model through layer skipping and using it as guidance, improving fidelity in models like HunyuanVideo, LTXVideo, and Mochi.|[Spatiotemporal Skip Guidance](#spatiotemporal-skip-guidance)|-|[Junha Hyung](https://junhahyung.github.io/), [Kinam Kim](https://kinam0252.github.io/)|
|
||||
|Adaptive Mask Inpainting|Adaptive Mask Inpainting algorithm from [Beyond the Contact: Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models](https://github.com/snuvclab/coma) (ECCV '24, Oral) provides a way to insert human inside the scene image without altering the background, by inpainting with adapting mask.|[Adaptive Mask Inpainting](#adaptive-mask-inpainting)|-|[Hyeonwoo Kim](https://sshowbiz.xyz),[Sookwan Han](https://jellyheadandrew.github.io)|
|
||||
|Flux with CFG|[Flux with CFG](https://github.com/ToTheBeginning/PuLID/blob/main/docs/pulid_for_flux.md) provides an implementation of using CFG in [Flux](https://blackforestlabs.ai/announcing-black-forest-labs/).|[Flux with CFG](#flux-with-cfg)|[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/flux_with_cfg.ipynb)|[Linoy Tsaban](https://github.com/linoytsaban), [Apolinário](https://github.com/apolinario), and [Sayak Paul](https://github.com/sayakpaul)|
|
||||
|Differential Diffusion|[Differential Diffusion](https://github.com/exx8/differential-diffusion) modifies an image according to a text prompt, and according to a map that specifies the amount of change in each region.|[Differential Diffusion](#differential-diffusion)|[](https://huggingface.co/spaces/exx8/differential-diffusion) [](https://colab.research.google.com/github/exx8/differential-diffusion/blob/main/examples/SD2.ipynb)|[Eran Levin](https://github.com/exx8) and [Ohad Fried](https://www.ohadf.com/)|
|
||||
@@ -93,6 +94,55 @@ pipe = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion
|
||||
|
||||
## Example usages
|
||||
|
||||
### Spatiotemporal Skip Guidance
|
||||
|
||||
**Junha Hyung\*, Kinam Kim\*, Susung Hong, Min-Jung Kim, Jaegul Choo**
|
||||
|
||||
**KAIST AI, University of Washington**
|
||||
|
||||
[*Spatiotemporal Skip Guidance (STG) for Enhanced Video Diffusion Sampling*](https://arxiv.org/abs/2411.18664) (CVPR 2025) is a simple training-free sampling guidance method for enhancing transformer-based video diffusion models. STG employs an implicit weak model via self-perturbation, avoiding the need for external models or additional training. By selectively skipping spatiotemporal layers, STG produces an aligned, degraded version of the original model to boost sample quality without compromising diversity or dynamic degree.
|
||||
|
||||
Following is the example video of STG applied to Mochi.
|
||||
|
||||
|
||||
https://github.com/user-attachments/assets/148adb59-da61-4c50-9dfa-425dcb5c23b3
|
||||
|
||||
More examples and information can be found on the [GitHub repository](https://github.com/junhahyung/STGuidance) and the [Project website](https://junhahyung.github.io/STGuidance/).
|
||||
|
||||
#### Usage example
|
||||
```python
|
||||
import torch
|
||||
from pipeline_stg_mochi import MochiSTGPipeline
|
||||
from diffusers.utils import export_to_video
|
||||
|
||||
# Load the pipeline
|
||||
pipe = MochiSTGPipeline.from_pretrained("genmo/mochi-1-preview", variant="bf16", torch_dtype=torch.bfloat16)
|
||||
|
||||
# Enable memory savings
|
||||
pipe = pipe.to("cuda")
|
||||
|
||||
#--------Option--------#
|
||||
prompt = "A close-up of a beautiful woman's face with colored powder exploding around her, creating an abstract splash of vibrant hues, realistic style."
|
||||
stg_applied_layers_idx = [34]
|
||||
stg_mode = "STG"
|
||||
stg_scale = 1.0 # 0.0 for CFG
|
||||
#----------------------#
|
||||
|
||||
# Generate video frames
|
||||
frames = pipe(
|
||||
prompt,
|
||||
height=480,
|
||||
width=480,
|
||||
num_frames=81,
|
||||
stg_applied_layers_idx=stg_applied_layers_idx,
|
||||
stg_scale=stg_scale,
|
||||
generator = torch.Generator().manual_seed(42),
|
||||
do_rescaling=do_rescaling,
|
||||
).frames[0]
|
||||
|
||||
export_to_video(frames, "output.mp4", fps=30)
|
||||
```
|
||||
|
||||
### Adaptive Mask Inpainting
|
||||
|
||||
**Hyeonwoo Kim\*, Sookwan Han\*, Patrick Kwon, Hanbyul Joo**
|
||||
|
||||
876
examples/community/pipeline_stg_cogvideox.py
Normal file
876
examples/community/pipeline_stg_cogvideox.py
Normal file
@@ -0,0 +1,876 @@
|
||||
# Copyright 2024 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team.
|
||||
# All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
import inspect
|
||||
import math
|
||||
import types
|
||||
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
|
||||
|
||||
import torch
|
||||
from transformers import T5EncoderModel, T5Tokenizer
|
||||
|
||||
from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
|
||||
from diffusers.loaders import CogVideoXLoraLoaderMixin
|
||||
from diffusers.models import AutoencoderKLCogVideoX, CogVideoXTransformer3DModel
|
||||
from diffusers.models.embeddings import get_3d_rotary_pos_embed
|
||||
from diffusers.pipelines.cogvideo.pipeline_output import CogVideoXPipelineOutput
|
||||
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
|
||||
|
||||
|
||||
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
|
||||
|
||||
|
||||
EXAMPLE_DOC_STRING = """
|
||||
Examples:
|
||||
```python
|
||||
>>> import torch
|
||||
>>> from diffusers.utils import export_to_video
|
||||
>>> from examples.community.pipeline_stg_cogvideox import CogVideoXSTGPipeline
|
||||
|
||||
>>> # Models: "THUDM/CogVideoX-2b" or "THUDM/CogVideoX-5b"
|
||||
>>> pipe = CogVideoXSTGPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch.float16).to("cuda")
|
||||
>>> prompt = (
|
||||
... "A father and son building a treehouse together, their hands covered in sawdust and smiles on their faces, realistic style."
|
||||
... )
|
||||
>>> pipe.transformer.to(memory_format=torch.channels_last)
|
||||
|
||||
>>> # Configure STG mode options
|
||||
>>> stg_applied_layers_idx = [11] # Layer indices from 0 to 41
|
||||
>>> stg_scale = 1.0 # Set to 0.0 for CFG
|
||||
>>> do_rescaling = False
|
||||
|
||||
>>> # Generate video frames with STG parameters
|
||||
>>> frames = pipe(
|
||||
... prompt=prompt,
|
||||
... stg_applied_layers_idx=stg_applied_layers_idx,
|
||||
... stg_scale=stg_scale,
|
||||
... do_rescaling=do_rescaling,
|
||||
>>> ).frames[0]
|
||||
>>> export_to_video(frames, "output.mp4", fps=8)
|
||||
```
|
||||
"""
|
||||
|
||||
|
||||
def forward_with_stg(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
encoder_hidden_states: torch.Tensor,
|
||||
temb: torch.Tensor,
|
||||
image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
|
||||
) -> torch.Tensor:
|
||||
hidden_states_ptb = hidden_states[2:]
|
||||
encoder_hidden_states_ptb = encoder_hidden_states[2:]
|
||||
|
||||
text_seq_length = encoder_hidden_states.size(1)
|
||||
|
||||
# norm & modulate
|
||||
norm_hidden_states, norm_encoder_hidden_states, gate_msa, enc_gate_msa = self.norm1(
|
||||
hidden_states, encoder_hidden_states, temb
|
||||
)
|
||||
|
||||
# attention
|
||||
attn_hidden_states, attn_encoder_hidden_states = self.attn1(
|
||||
hidden_states=norm_hidden_states,
|
||||
encoder_hidden_states=norm_encoder_hidden_states,
|
||||
image_rotary_emb=image_rotary_emb,
|
||||
)
|
||||
|
||||
hidden_states = hidden_states + gate_msa * attn_hidden_states
|
||||
encoder_hidden_states = encoder_hidden_states + enc_gate_msa * attn_encoder_hidden_states
|
||||
|
||||
# norm & modulate
|
||||
norm_hidden_states, norm_encoder_hidden_states, gate_ff, enc_gate_ff = self.norm2(
|
||||
hidden_states, encoder_hidden_states, temb
|
||||
)
|
||||
|
||||
# feed-forward
|
||||
norm_hidden_states = torch.cat([norm_encoder_hidden_states, norm_hidden_states], dim=1)
|
||||
ff_output = self.ff(norm_hidden_states)
|
||||
|
||||
hidden_states = hidden_states + gate_ff * ff_output[:, text_seq_length:]
|
||||
encoder_hidden_states = encoder_hidden_states + enc_gate_ff * ff_output[:, :text_seq_length]
|
||||
|
||||
hidden_states[2:] = hidden_states_ptb
|
||||
encoder_hidden_states[2:] = encoder_hidden_states_ptb
|
||||
|
||||
return hidden_states, encoder_hidden_states
|
||||
|
||||
|
||||
# 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
|
||||
|
||||
|
||||
class CogVideoXSTGPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
|
||||
r"""
|
||||
Pipeline for text-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: CogVideoXTransformer3DModel,
|
||||
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)
|
||||
|
||||
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
|
||||
|
||||
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, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=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."
|
||||
)
|
||||
|
||||
shape = (
|
||||
batch_size,
|
||||
(num_frames - 1) // self.vae_scale_factor_temporal + 1,
|
||||
num_channels_latents,
|
||||
height // self.vae_scale_factor_spatial,
|
||||
width // self.vae_scale_factor_spatial,
|
||||
)
|
||||
|
||||
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
|
||||
return 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.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://arxiv.org/abs/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
|
||||
|
||||
# Copied from diffusers.pipelines.latte.pipeline_latte.LattePipeline.check_inputs
|
||||
def check_inputs(
|
||||
self,
|
||||
prompt,
|
||||
height,
|
||||
width,
|
||||
negative_prompt,
|
||||
callback_on_step_end_tensor_inputs,
|
||||
prompt_embeds=None,
|
||||
negative_prompt_embeds=None,
|
||||
):
|
||||
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}."
|
||||
)
|
||||
|
||||
def fuse_qkv_projections(self) -> None:
|
||||
r"""Enables fused QKV projections."""
|
||||
self.fusing_transformer = True
|
||||
self.transformer.fuse_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
|
||||
|
||||
def _prepare_rotary_positional_embeddings(
|
||||
self,
|
||||
height: int,
|
||||
width: int,
|
||||
num_frames: int,
|
||||
device: torch.device,
|
||||
) -> 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,
|
||||
)
|
||||
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,
|
||||
)
|
||||
|
||||
return freqs_cos, freqs_sin
|
||||
|
||||
@property
|
||||
def guidance_scale(self):
|
||||
return self._guidance_scale
|
||||
|
||||
@property
|
||||
def do_spatio_temporal_guidance(self):
|
||||
return self._stg_scale > 0.0
|
||||
|
||||
@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,
|
||||
prompt: Optional[Union[str, List[str]]] = None,
|
||||
negative_prompt: Optional[Union[str, List[str]]] = None,
|
||||
height: Optional[int] = None,
|
||||
width: Optional[int] = None,
|
||||
num_frames: Optional[int] = None,
|
||||
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,
|
||||
stg_applied_layers_idx: Optional[List[int]] = [11],
|
||||
stg_scale: Optional[float] = 0.0,
|
||||
do_rescaling: Optional[bool] = False,
|
||||
) -> Union[CogVideoXPipelineOutput, Tuple]:
|
||||
"""
|
||||
Function invoked when calling the pipeline for generation.
|
||||
|
||||
Args:
|
||||
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://arxiv.org/abs/2207.12598).
|
||||
`guidance_scale` is defined as `w` of equation 2. of [Imagen
|
||||
Paper](https://arxiv.org/pdf/2205.11487.pdf). 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_cogvideox.CogVideoXPipelineOutput`] or `tuple`:
|
||||
[`~pipelines.cogvideo.pipeline_cogvideox.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
|
||||
|
||||
# 1. Check inputs. Raise error if not correct
|
||||
self.check_inputs(
|
||||
prompt,
|
||||
height,
|
||||
width,
|
||||
negative_prompt,
|
||||
callback_on_step_end_tensor_inputs,
|
||||
prompt_embeds,
|
||||
negative_prompt_embeds,
|
||||
)
|
||||
self._stg_scale = stg_scale
|
||||
self._guidance_scale = guidance_scale
|
||||
self._attention_kwargs = attention_kwargs
|
||||
self._current_timestep = None
|
||||
self._interrupt = False
|
||||
|
||||
if self.do_spatio_temporal_guidance:
|
||||
for i in stg_applied_layers_idx:
|
||||
self.transformer.transformer_blocks[i].forward = types.MethodType(
|
||||
forward_with_stg, self.transformer.transformer_blocks[i]
|
||||
)
|
||||
|
||||
# 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
|
||||
|
||||
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
||||
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
||||
# corresponds to doing no classifier free guidance.
|
||||
do_classifier_free_guidance = guidance_scale > 1.0
|
||||
|
||||
# 3. Encode input prompt
|
||||
prompt_embeds, negative_prompt_embeds = self.encode_prompt(
|
||||
prompt,
|
||||
negative_prompt,
|
||||
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 and not self.do_spatio_temporal_guidance:
|
||||
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
|
||||
elif do_classifier_free_guidance and self.do_spatio_temporal_guidance:
|
||||
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds, prompt_embeds], dim=0)
|
||||
|
||||
# 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
|
||||
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
|
||||
|
||||
latent_channels = self.transformer.config.in_channels
|
||||
latents = self.prepare_latents(
|
||||
batch_size * num_videos_per_prompt,
|
||||
latent_channels,
|
||||
num_frames,
|
||||
height,
|
||||
width,
|
||||
prompt_embeds.dtype,
|
||||
device,
|
||||
generator,
|
||||
latents,
|
||||
)
|
||||
|
||||
# 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
|
||||
)
|
||||
|
||||
# 8. Denoising loop
|
||||
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
|
||||
|
||||
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
|
||||
if do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
|
||||
latent_model_input = torch.cat([latents] * 2)
|
||||
elif do_classifier_free_guidance and self.do_spatio_temporal_guidance:
|
||||
latent_model_input = torch.cat([latents] * 3)
|
||||
else:
|
||||
latent_model_input = latents
|
||||
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
||||
|
||||
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
||||
timestep = t.expand(latent_model_input.shape[0])
|
||||
|
||||
# predict noise model_output
|
||||
noise_pred = self.transformer(
|
||||
hidden_states=latent_model_input,
|
||||
encoder_hidden_states=prompt_embeds,
|
||||
timestep=timestep,
|
||||
image_rotary_emb=image_rotary_emb,
|
||||
attention_kwargs=attention_kwargs,
|
||||
return_dict=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 and not self.do_spatio_temporal_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)
|
||||
elif do_classifier_free_guidance and self.do_spatio_temporal_guidance:
|
||||
noise_pred_uncond, noise_pred_text, noise_pred_perturb = noise_pred.chunk(3)
|
||||
noise_pred = (
|
||||
noise_pred_uncond
|
||||
+ self.guidance_scale * (noise_pred_text - noise_pred_uncond)
|
||||
+ self._stg_scale * (noise_pred_text - noise_pred_perturb)
|
||||
)
|
||||
|
||||
if do_rescaling:
|
||||
rescaling_scale = 0.7
|
||||
factor = noise_pred_text.std() / noise_pred.std()
|
||||
factor = rescaling_scale * factor + (1 - rescaling_scale)
|
||||
noise_pred = noise_pred * factor
|
||||
|
||||
# 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)
|
||||
|
||||
# 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
|
||||
|
||||
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)
|
||||
794
examples/community/pipeline_stg_hunyuan_video.py
Normal file
794
examples/community/pipeline_stg_hunyuan_video.py
Normal file
@@ -0,0 +1,794 @@
|
||||
# Copyright 2024 The HunyuanVideo Team and The HuggingFace Team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
import inspect
|
||||
import types
|
||||
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
from transformers import CLIPTextModel, CLIPTokenizer, LlamaModel, LlamaTokenizerFast
|
||||
|
||||
from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
|
||||
from diffusers.loaders import HunyuanVideoLoraLoaderMixin
|
||||
from diffusers.models import AutoencoderKLHunyuanVideo, HunyuanVideoTransformer3DModel
|
||||
from diffusers.pipelines.hunyuan_video.pipeline_output import HunyuanVideoPipelineOutput
|
||||
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
|
||||
from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
|
||||
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
|
||||
|
||||
|
||||
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
|
||||
|
||||
|
||||
EXAMPLE_DOC_STRING = """
|
||||
Examples:
|
||||
```python
|
||||
>>> import torch
|
||||
>>> from diffusers.utils import export_to_video
|
||||
>>> from diffusers import HunyuanVideoTransformer3DModel
|
||||
>>> from examples.community.pipeline_stg_hunyuan_video import HunyuanVideoSTGPipeline
|
||||
|
||||
>>> model_id = "hunyuanvideo-community/HunyuanVideo"
|
||||
>>> transformer = HunyuanVideoTransformer3DModel.from_pretrained(
|
||||
... model_id, subfolder="transformer", torch_dtype=torch.bfloat16
|
||||
... )
|
||||
>>> pipe = HunyuanVideoSTGPipeline.from_pretrained(model_id, transformer=transformer, torch_dtype=torch.float16)
|
||||
>>> pipe.vae.enable_tiling()
|
||||
>>> pipe.to("cuda")
|
||||
|
||||
>>> # Configure STG mode options
|
||||
>>> stg_applied_layers_idx = [2] # Layer indices from 0 to 41
|
||||
>>> stg_scale = 1.0 # Set 0.0 for CFG
|
||||
|
||||
>>> output = pipe(
|
||||
... prompt="A wolf howling at the moon, with the moon subtly resembling a giant clock face, realistic style.",
|
||||
... height=320,
|
||||
... width=512,
|
||||
... num_frames=61,
|
||||
... num_inference_steps=30,
|
||||
... stg_applied_layers_idx=stg_applied_layers_idx,
|
||||
... stg_scale=stg_scale,
|
||||
>>> ).frames[0]
|
||||
>>> export_to_video(output, "output.mp4", fps=15)
|
||||
```
|
||||
"""
|
||||
|
||||
|
||||
DEFAULT_PROMPT_TEMPLATE = {
|
||||
"template": (
|
||||
"<|start_header_id|>system<|end_header_id|>\n\nDescribe the video by detailing the following aspects: "
|
||||
"1. The main content and theme of the video."
|
||||
"2. The color, shape, size, texture, quantity, text, and spatial relationships of the objects."
|
||||
"3. Actions, events, behaviors temporal relationships, physical movement changes of the objects."
|
||||
"4. background environment, light, style and atmosphere."
|
||||
"5. camera angles, movements, and transitions used in the video:<|eot_id|>"
|
||||
"<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|>"
|
||||
),
|
||||
"crop_start": 95,
|
||||
}
|
||||
|
||||
|
||||
def forward_with_stg(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
encoder_hidden_states: torch.Tensor,
|
||||
temb: torch.Tensor,
|
||||
attention_mask: Optional[torch.Tensor] = None,
|
||||
freqs_cis: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
|
||||
) -> Tuple[torch.Tensor, torch.Tensor]:
|
||||
return hidden_states, encoder_hidden_states
|
||||
|
||||
|
||||
def forward_without_stg(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
encoder_hidden_states: torch.Tensor,
|
||||
temb: torch.Tensor,
|
||||
attention_mask: Optional[torch.Tensor] = None,
|
||||
freqs_cis: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
|
||||
) -> Tuple[torch.Tensor, torch.Tensor]:
|
||||
# 1. Input normalization
|
||||
norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(hidden_states, emb=temb)
|
||||
norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = self.norm1_context(
|
||||
encoder_hidden_states, emb=temb
|
||||
)
|
||||
|
||||
# 2. Joint attention
|
||||
attn_output, context_attn_output = self.attn(
|
||||
hidden_states=norm_hidden_states,
|
||||
encoder_hidden_states=norm_encoder_hidden_states,
|
||||
attention_mask=attention_mask,
|
||||
image_rotary_emb=freqs_cis,
|
||||
)
|
||||
|
||||
# 3. Modulation and residual connection
|
||||
hidden_states = hidden_states + attn_output * gate_msa.unsqueeze(1)
|
||||
encoder_hidden_states = encoder_hidden_states + context_attn_output * c_gate_msa.unsqueeze(1)
|
||||
|
||||
norm_hidden_states = self.norm2(hidden_states)
|
||||
norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states)
|
||||
|
||||
norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
|
||||
norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None]
|
||||
|
||||
# 4. Feed-forward
|
||||
ff_output = self.ff(norm_hidden_states)
|
||||
context_ff_output = self.ff_context(norm_encoder_hidden_states)
|
||||
|
||||
hidden_states = hidden_states + gate_mlp.unsqueeze(1) * ff_output
|
||||
encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * context_ff_output
|
||||
|
||||
return hidden_states, encoder_hidden_states
|
||||
|
||||
|
||||
# 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
|
||||
|
||||
|
||||
class HunyuanVideoSTGPipeline(DiffusionPipeline, HunyuanVideoLoraLoaderMixin):
|
||||
r"""
|
||||
Pipeline for text-to-video generation using HunyuanVideo.
|
||||
|
||||
This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
|
||||
implemented for all pipelines (downloading, saving, running on a particular device, etc.).
|
||||
|
||||
Args:
|
||||
text_encoder ([`LlamaModel`]):
|
||||
[Llava Llama3-8B](https://huggingface.co/xtuner/llava-llama-3-8b-v1_1-transformers).
|
||||
tokenizer (`LlamaTokenizer`):
|
||||
Tokenizer from [Llava Llama3-8B](https://huggingface.co/xtuner/llava-llama-3-8b-v1_1-transformers).
|
||||
transformer ([`HunyuanVideoTransformer3DModel`]):
|
||||
Conditional Transformer to denoise the encoded image latents.
|
||||
scheduler ([`FlowMatchEulerDiscreteScheduler`]):
|
||||
A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
|
||||
vae ([`AutoencoderKLHunyuanVideo`]):
|
||||
Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
|
||||
text_encoder_2 ([`CLIPTextModel`]):
|
||||
[CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
|
||||
the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
|
||||
tokenizer_2 (`CLIPTokenizer`):
|
||||
Tokenizer of class
|
||||
[CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
|
||||
"""
|
||||
|
||||
model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae"
|
||||
_callback_tensor_inputs = ["latents", "prompt_embeds"]
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
text_encoder: LlamaModel,
|
||||
tokenizer: LlamaTokenizerFast,
|
||||
transformer: HunyuanVideoTransformer3DModel,
|
||||
vae: AutoencoderKLHunyuanVideo,
|
||||
scheduler: FlowMatchEulerDiscreteScheduler,
|
||||
text_encoder_2: CLIPTextModel,
|
||||
tokenizer_2: CLIPTokenizer,
|
||||
):
|
||||
super().__init__()
|
||||
|
||||
self.register_modules(
|
||||
vae=vae,
|
||||
text_encoder=text_encoder,
|
||||
tokenizer=tokenizer,
|
||||
transformer=transformer,
|
||||
scheduler=scheduler,
|
||||
text_encoder_2=text_encoder_2,
|
||||
tokenizer_2=tokenizer_2,
|
||||
)
|
||||
|
||||
self.vae_scale_factor_temporal = self.vae.temporal_compression_ratio if getattr(self, "vae", None) else 4
|
||||
self.vae_scale_factor_spatial = self.vae.spatial_compression_ratio if getattr(self, "vae", None) else 8
|
||||
self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
|
||||
|
||||
def _get_llama_prompt_embeds(
|
||||
self,
|
||||
prompt: Union[str, List[str]],
|
||||
prompt_template: Dict[str, Any],
|
||||
num_videos_per_prompt: int = 1,
|
||||
device: Optional[torch.device] = None,
|
||||
dtype: Optional[torch.dtype] = None,
|
||||
max_sequence_length: int = 256,
|
||||
num_hidden_layers_to_skip: int = 2,
|
||||
) -> Tuple[torch.Tensor, torch.Tensor]:
|
||||
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)
|
||||
|
||||
prompt = [prompt_template["template"].format(p) for p in prompt]
|
||||
|
||||
crop_start = prompt_template.get("crop_start", None)
|
||||
if crop_start is None:
|
||||
prompt_template_input = self.tokenizer(
|
||||
prompt_template["template"],
|
||||
padding="max_length",
|
||||
return_tensors="pt",
|
||||
return_length=False,
|
||||
return_overflowing_tokens=False,
|
||||
return_attention_mask=False,
|
||||
)
|
||||
crop_start = prompt_template_input["input_ids"].shape[-1]
|
||||
# Remove <|eot_id|> token and placeholder {}
|
||||
crop_start -= 2
|
||||
|
||||
max_sequence_length += crop_start
|
||||
text_inputs = self.tokenizer(
|
||||
prompt,
|
||||
max_length=max_sequence_length,
|
||||
padding="max_length",
|
||||
truncation=True,
|
||||
return_tensors="pt",
|
||||
return_length=False,
|
||||
return_overflowing_tokens=False,
|
||||
return_attention_mask=True,
|
||||
)
|
||||
text_input_ids = text_inputs.input_ids.to(device=device)
|
||||
prompt_attention_mask = text_inputs.attention_mask.to(device=device)
|
||||
|
||||
prompt_embeds = self.text_encoder(
|
||||
input_ids=text_input_ids,
|
||||
attention_mask=prompt_attention_mask,
|
||||
output_hidden_states=True,
|
||||
).hidden_states[-(num_hidden_layers_to_skip + 1)]
|
||||
prompt_embeds = prompt_embeds.to(dtype=dtype)
|
||||
|
||||
if crop_start is not None and crop_start > 0:
|
||||
prompt_embeds = prompt_embeds[:, crop_start:]
|
||||
prompt_attention_mask = prompt_attention_mask[:, crop_start:]
|
||||
|
||||
# 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)
|
||||
prompt_attention_mask = prompt_attention_mask.repeat(1, num_videos_per_prompt)
|
||||
prompt_attention_mask = prompt_attention_mask.view(batch_size * num_videos_per_prompt, seq_len)
|
||||
|
||||
return prompt_embeds, prompt_attention_mask
|
||||
|
||||
def _get_clip_prompt_embeds(
|
||||
self,
|
||||
prompt: Union[str, List[str]],
|
||||
num_videos_per_prompt: int = 1,
|
||||
device: Optional[torch.device] = None,
|
||||
dtype: Optional[torch.dtype] = None,
|
||||
max_sequence_length: int = 77,
|
||||
) -> torch.Tensor:
|
||||
device = device or self._execution_device
|
||||
dtype = dtype or self.text_encoder_2.dtype
|
||||
|
||||
prompt = [prompt] if isinstance(prompt, str) else prompt
|
||||
batch_size = len(prompt)
|
||||
|
||||
text_inputs = self.tokenizer_2(
|
||||
prompt,
|
||||
padding="max_length",
|
||||
max_length=max_sequence_length,
|
||||
truncation=True,
|
||||
return_tensors="pt",
|
||||
)
|
||||
|
||||
text_input_ids = text_inputs.input_ids
|
||||
untruncated_ids = self.tokenizer_2(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_2.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
|
||||
logger.warning(
|
||||
"The following part of your input was truncated because CLIP can only handle sequences up to"
|
||||
f" {max_sequence_length} tokens: {removed_text}"
|
||||
)
|
||||
|
||||
prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False).pooler_output
|
||||
|
||||
# duplicate text embeddings for each generation per prompt, using mps friendly method
|
||||
prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt)
|
||||
prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, -1)
|
||||
|
||||
return prompt_embeds
|
||||
|
||||
def encode_prompt(
|
||||
self,
|
||||
prompt: Union[str, List[str]],
|
||||
prompt_2: Union[str, List[str]] = None,
|
||||
prompt_template: Dict[str, Any] = DEFAULT_PROMPT_TEMPLATE,
|
||||
num_videos_per_prompt: int = 1,
|
||||
prompt_embeds: Optional[torch.Tensor] = None,
|
||||
pooled_prompt_embeds: Optional[torch.Tensor] = None,
|
||||
prompt_attention_mask: Optional[torch.Tensor] = None,
|
||||
device: Optional[torch.device] = None,
|
||||
dtype: Optional[torch.dtype] = None,
|
||||
max_sequence_length: int = 256,
|
||||
):
|
||||
if prompt_embeds is None:
|
||||
prompt_embeds, prompt_attention_mask = self._get_llama_prompt_embeds(
|
||||
prompt,
|
||||
prompt_template,
|
||||
num_videos_per_prompt,
|
||||
device=device,
|
||||
dtype=dtype,
|
||||
max_sequence_length=max_sequence_length,
|
||||
)
|
||||
|
||||
if pooled_prompt_embeds is None:
|
||||
if prompt_2 is None and pooled_prompt_embeds is None:
|
||||
prompt_2 = prompt
|
||||
pooled_prompt_embeds = self._get_clip_prompt_embeds(
|
||||
prompt,
|
||||
num_videos_per_prompt,
|
||||
device=device,
|
||||
dtype=dtype,
|
||||
max_sequence_length=77,
|
||||
)
|
||||
|
||||
return prompt_embeds, pooled_prompt_embeds, prompt_attention_mask
|
||||
|
||||
def check_inputs(
|
||||
self,
|
||||
prompt,
|
||||
prompt_2,
|
||||
height,
|
||||
width,
|
||||
prompt_embeds=None,
|
||||
callback_on_step_end_tensor_inputs=None,
|
||||
prompt_template=None,
|
||||
):
|
||||
if height % 16 != 0 or width % 16 != 0:
|
||||
raise ValueError(f"`height` and `width` have to be divisible by 16 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_2 is not None and prompt_embeds is not None:
|
||||
raise ValueError(
|
||||
f"Cannot forward both `prompt_2`: {prompt_2} 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)}")
|
||||
elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
|
||||
raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
|
||||
|
||||
if prompt_template is not None:
|
||||
if not isinstance(prompt_template, dict):
|
||||
raise ValueError(f"`prompt_template` has to be of type `dict` but is {type(prompt_template)}")
|
||||
if "template" not in prompt_template:
|
||||
raise ValueError(
|
||||
f"`prompt_template` has to contain a key `template` but only found {prompt_template.keys()}"
|
||||
)
|
||||
|
||||
def prepare_latents(
|
||||
self,
|
||||
batch_size: int,
|
||||
num_channels_latents: 32,
|
||||
height: int = 720,
|
||||
width: int = 1280,
|
||||
num_frames: int = 129,
|
||||
dtype: Optional[torch.dtype] = None,
|
||||
device: Optional[torch.device] = None,
|
||||
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
||||
latents: Optional[torch.Tensor] = None,
|
||||
) -> torch.Tensor:
|
||||
if latents is not None:
|
||||
return latents.to(device=device, dtype=dtype)
|
||||
|
||||
shape = (
|
||||
batch_size,
|
||||
num_channels_latents,
|
||||
num_frames,
|
||||
int(height) // self.vae_scale_factor_spatial,
|
||||
int(width) // self.vae_scale_factor_spatial,
|
||||
)
|
||||
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."
|
||||
)
|
||||
|
||||
latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
|
||||
return latents
|
||||
|
||||
def enable_vae_slicing(self):
|
||||
r"""
|
||||
Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
|
||||
compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
|
||||
"""
|
||||
self.vae.enable_slicing()
|
||||
|
||||
def disable_vae_slicing(self):
|
||||
r"""
|
||||
Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
|
||||
computing decoding in one step.
|
||||
"""
|
||||
self.vae.disable_slicing()
|
||||
|
||||
def enable_vae_tiling(self):
|
||||
r"""
|
||||
Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
|
||||
compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
|
||||
processing larger images.
|
||||
"""
|
||||
self.vae.enable_tiling()
|
||||
|
||||
def disable_vae_tiling(self):
|
||||
r"""
|
||||
Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
|
||||
computing decoding in one step.
|
||||
"""
|
||||
self.vae.disable_tiling()
|
||||
|
||||
@property
|
||||
def guidance_scale(self):
|
||||
return self._guidance_scale
|
||||
|
||||
@property
|
||||
def do_spatio_temporal_guidance(self):
|
||||
return self._stg_scale > 0.0
|
||||
|
||||
@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,
|
||||
prompt: Union[str, List[str]] = None,
|
||||
prompt_2: Union[str, List[str]] = None,
|
||||
height: int = 720,
|
||||
width: int = 1280,
|
||||
num_frames: int = 129,
|
||||
num_inference_steps: int = 50,
|
||||
sigmas: List[float] = None,
|
||||
guidance_scale: float = 6.0,
|
||||
num_videos_per_prompt: Optional[int] = 1,
|
||||
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
||||
latents: Optional[torch.Tensor] = None,
|
||||
prompt_embeds: Optional[torch.Tensor] = None,
|
||||
pooled_prompt_embeds: Optional[torch.Tensor] = None,
|
||||
prompt_attention_mask: Optional[torch.Tensor] = None,
|
||||
output_type: Optional[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"],
|
||||
prompt_template: Dict[str, Any] = DEFAULT_PROMPT_TEMPLATE,
|
||||
max_sequence_length: int = 256,
|
||||
stg_applied_layers_idx: Optional[List[int]] = [2],
|
||||
stg_scale: Optional[float] = 0.0,
|
||||
):
|
||||
r"""
|
||||
The call function to the pipeline for generation.
|
||||
|
||||
Args:
|
||||
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.
|
||||
prompt_2 (`str` or `List[str]`, *optional*):
|
||||
The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
|
||||
will be used instead.
|
||||
height (`int`, defaults to `720`):
|
||||
The height in pixels of the generated image.
|
||||
width (`int`, defaults to `1280`):
|
||||
The width in pixels of the generated image.
|
||||
num_frames (`int`, defaults to `129`):
|
||||
The number of frames in the generated video.
|
||||
num_inference_steps (`int`, defaults to `50`):
|
||||
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
|
||||
expense of slower inference.
|
||||
sigmas (`List[float]`, *optional*):
|
||||
Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
|
||||
their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
|
||||
will be used.
|
||||
guidance_scale (`float`, defaults to `6.0`):
|
||||
Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
|
||||
`guidance_scale` is defined as `w` of equation 2. of [Imagen
|
||||
Paper](https://arxiv.org/pdf/2205.11487.pdf). 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. Note that the only available HunyuanVideo model is
|
||||
CFG-distilled, which means that traditional guidance between unconditional and conditional latent is
|
||||
not applied.
|
||||
num_videos_per_prompt (`int`, *optional*, defaults to 1):
|
||||
The number of images to generate per prompt.
|
||||
generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
|
||||
A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
|
||||
generation deterministic.
|
||||
latents (`torch.Tensor`, *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 is generated by sampling using the supplied random `generator`.
|
||||
prompt_embeds (`torch.Tensor`, *optional*):
|
||||
Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
|
||||
provided, text embeddings are generated from the `prompt` input argument.
|
||||
output_type (`str`, *optional*, defaults to `"pil"`):
|
||||
The output format of the generated image. Choose between `PIL.Image` or `np.array`.
|
||||
return_dict (`bool`, *optional*, defaults to `True`):
|
||||
Whether or not to return a [`HunyuanVideoPipelineOutput`] 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).
|
||||
clip_skip (`int`, *optional*):
|
||||
Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
|
||||
the output of the pre-final layer will be used for computing the prompt embeddings.
|
||||
callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
|
||||
A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
|
||||
each denoising step during the inference. 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.
|
||||
|
||||
Examples:
|
||||
|
||||
Returns:
|
||||
[`~HunyuanVideoPipelineOutput`] or `tuple`:
|
||||
If `return_dict` is `True`, [`HunyuanVideoPipelineOutput`] is returned, otherwise a `tuple` is returned
|
||||
where the first element is a list with the generated images and the second element is a list of `bool`s
|
||||
indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content.
|
||||
"""
|
||||
|
||||
if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
|
||||
callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
|
||||
|
||||
# 1. Check inputs. Raise error if not correct
|
||||
self.check_inputs(
|
||||
prompt,
|
||||
prompt_2,
|
||||
height,
|
||||
width,
|
||||
prompt_embeds,
|
||||
callback_on_step_end_tensor_inputs,
|
||||
prompt_template,
|
||||
)
|
||||
|
||||
self._stg_scale = stg_scale
|
||||
self._guidance_scale = guidance_scale
|
||||
self._attention_kwargs = attention_kwargs
|
||||
self._current_timestep = None
|
||||
self._interrupt = False
|
||||
|
||||
device = self._execution_device
|
||||
|
||||
# 2. Define 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]
|
||||
|
||||
# 3. Encode input prompt
|
||||
prompt_embeds, pooled_prompt_embeds, prompt_attention_mask = self.encode_prompt(
|
||||
prompt=prompt,
|
||||
prompt_2=prompt_2,
|
||||
prompt_template=prompt_template,
|
||||
num_videos_per_prompt=num_videos_per_prompt,
|
||||
prompt_embeds=prompt_embeds,
|
||||
pooled_prompt_embeds=pooled_prompt_embeds,
|
||||
prompt_attention_mask=prompt_attention_mask,
|
||||
device=device,
|
||||
max_sequence_length=max_sequence_length,
|
||||
)
|
||||
|
||||
transformer_dtype = self.transformer.dtype
|
||||
prompt_embeds = prompt_embeds.to(transformer_dtype)
|
||||
prompt_attention_mask = prompt_attention_mask.to(transformer_dtype)
|
||||
if pooled_prompt_embeds is not None:
|
||||
pooled_prompt_embeds = pooled_prompt_embeds.to(transformer_dtype)
|
||||
|
||||
# 4. Prepare timesteps
|
||||
sigmas = np.linspace(1.0, 0.0, num_inference_steps + 1)[:-1] if sigmas is None else sigmas
|
||||
timesteps, num_inference_steps = retrieve_timesteps(
|
||||
self.scheduler,
|
||||
num_inference_steps,
|
||||
device,
|
||||
sigmas=sigmas,
|
||||
)
|
||||
|
||||
# 5. Prepare latent variables
|
||||
num_channels_latents = self.transformer.config.in_channels
|
||||
num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
|
||||
latents = self.prepare_latents(
|
||||
batch_size * num_videos_per_prompt,
|
||||
num_channels_latents,
|
||||
height,
|
||||
width,
|
||||
num_latent_frames,
|
||||
torch.float32,
|
||||
device,
|
||||
generator,
|
||||
latents,
|
||||
)
|
||||
|
||||
# 6. Prepare guidance condition
|
||||
guidance = torch.tensor([guidance_scale] * latents.shape[0], dtype=transformer_dtype, device=device) * 1000.0
|
||||
|
||||
# 7. Denoising loop
|
||||
num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
|
||||
self._num_timesteps = len(timesteps)
|
||||
|
||||
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
||||
for i, t in enumerate(timesteps):
|
||||
if self.interrupt:
|
||||
continue
|
||||
|
||||
self._current_timestep = t
|
||||
latent_model_input = latents.to(transformer_dtype)
|
||||
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
||||
timestep = t.expand(latents.shape[0]).to(latents.dtype)
|
||||
|
||||
if self.do_spatio_temporal_guidance:
|
||||
for i in stg_applied_layers_idx:
|
||||
self.transformer.transformer_blocks[i].forward = types.MethodType(
|
||||
forward_without_stg, self.transformer.transformer_blocks[i]
|
||||
)
|
||||
|
||||
noise_pred = self.transformer(
|
||||
hidden_states=latent_model_input,
|
||||
timestep=timestep,
|
||||
encoder_hidden_states=prompt_embeds,
|
||||
encoder_attention_mask=prompt_attention_mask,
|
||||
pooled_projections=pooled_prompt_embeds,
|
||||
guidance=guidance,
|
||||
attention_kwargs=attention_kwargs,
|
||||
return_dict=False,
|
||||
)[0]
|
||||
|
||||
if self.do_spatio_temporal_guidance:
|
||||
for i in stg_applied_layers_idx:
|
||||
self.transformer.transformer_blocks[i].forward = types.MethodType(
|
||||
forward_with_stg, self.transformer.transformer_blocks[i]
|
||||
)
|
||||
|
||||
noise_pred_perturb = self.transformer(
|
||||
hidden_states=latent_model_input,
|
||||
timestep=timestep,
|
||||
encoder_hidden_states=prompt_embeds,
|
||||
encoder_attention_mask=prompt_attention_mask,
|
||||
pooled_projections=pooled_prompt_embeds,
|
||||
guidance=guidance,
|
||||
attention_kwargs=attention_kwargs,
|
||||
return_dict=False,
|
||||
)[0]
|
||||
noise_pred = noise_pred + self._stg_scale * (noise_pred - noise_pred_perturb)
|
||||
|
||||
# compute the previous noisy sample x_t -> x_t-1
|
||||
latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
|
||||
|
||||
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)
|
||||
|
||||
# call the callback, if provided
|
||||
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
|
||||
|
||||
if not output_type == "latent":
|
||||
latents = latents.to(self.vae.dtype) / self.vae.config.scaling_factor
|
||||
video = self.vae.decode(latents, return_dict=False)[0]
|
||||
video = self.video_processor.postprocess_video(video, output_type=output_type)
|
||||
else:
|
||||
video = latents
|
||||
|
||||
# Offload all models
|
||||
self.maybe_free_model_hooks()
|
||||
|
||||
if not return_dict:
|
||||
return (video,)
|
||||
|
||||
return HunyuanVideoPipelineOutput(frames=video)
|
||||
886
examples/community/pipeline_stg_ltx.py
Normal file
886
examples/community/pipeline_stg_ltx.py
Normal file
@@ -0,0 +1,886 @@
|
||||
# Copyright 2024 Lightricks and The HuggingFace Team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
import inspect
|
||||
import types
|
||||
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
from transformers import T5EncoderModel, T5TokenizerFast
|
||||
|
||||
from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
|
||||
from diffusers.loaders import FromSingleFileMixin, LTXVideoLoraLoaderMixin
|
||||
from diffusers.models.autoencoders import AutoencoderKLLTXVideo
|
||||
from diffusers.models.transformers import LTXVideoTransformer3DModel
|
||||
from diffusers.pipelines.ltx.pipeline_output import LTXPipelineOutput
|
||||
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
|
||||
from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
|
||||
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
|
||||
|
||||
|
||||
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
|
||||
|
||||
EXAMPLE_DOC_STRING = """
|
||||
Examples:
|
||||
```py
|
||||
>>> import torch
|
||||
>>> from diffusers.utils import export_to_video
|
||||
>>> from examples.community.pipeline_stg_ltx import LTXSTGPipeline
|
||||
|
||||
>>> pipe = LTXSTGPipeline.from_pretrained("Lightricks/LTX-Video", torch_dtype=torch.bfloat16)
|
||||
>>> pipe.to("cuda")
|
||||
|
||||
>>> prompt = "A woman with light skin, wearing a blue jacket and a black hat with a veil, looks down and to her right, then back up as she speaks; she has brown hair styled in an updo, light brown eyebrows, and is wearing a white collared shirt under her jacket; the camera remains stationary on her face as she speaks; the background is out of focus, but shows trees and people in period clothing; the scene is captured in real-life footage."
|
||||
>>> negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted"
|
||||
|
||||
>>> # Configure STG mode options
|
||||
>>> stg_applied_layers_idx = [19] # Layer indices from 0 to 41
|
||||
>>> stg_scale = 1.0 # Set 0.0 for CFG
|
||||
>>> do_rescaling = False
|
||||
|
||||
>>> video = pipe(
|
||||
... prompt=prompt,
|
||||
... negative_prompt=negative_prompt,
|
||||
... width=704,
|
||||
... height=480,
|
||||
... num_frames=161,
|
||||
... num_inference_steps=50,
|
||||
... stg_applied_layers_idx=stg_applied_layers_idx,
|
||||
... stg_scale=stg_scale,
|
||||
... do_rescaling=do_rescaling,
|
||||
>>> ).frames[0]
|
||||
>>> export_to_video(video, "output.mp4", fps=24)
|
||||
```
|
||||
"""
|
||||
|
||||
|
||||
def forward_with_stg(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
encoder_hidden_states: torch.Tensor,
|
||||
temb: torch.Tensor,
|
||||
image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
|
||||
encoder_attention_mask: Optional[torch.Tensor] = None,
|
||||
) -> torch.Tensor:
|
||||
hidden_states_ptb = hidden_states[2:]
|
||||
encoder_hidden_states_ptb = encoder_hidden_states[2:]
|
||||
|
||||
batch_size = hidden_states.size(0)
|
||||
norm_hidden_states = self.norm1(hidden_states)
|
||||
|
||||
num_ada_params = self.scale_shift_table.shape[0]
|
||||
ada_values = self.scale_shift_table[None, None] + temb.reshape(batch_size, temb.size(1), num_ada_params, -1)
|
||||
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ada_values.unbind(dim=2)
|
||||
norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
|
||||
|
||||
attn_hidden_states = self.attn1(
|
||||
hidden_states=norm_hidden_states,
|
||||
encoder_hidden_states=None,
|
||||
image_rotary_emb=image_rotary_emb,
|
||||
)
|
||||
hidden_states = hidden_states + attn_hidden_states * gate_msa
|
||||
|
||||
attn_hidden_states = self.attn2(
|
||||
hidden_states,
|
||||
encoder_hidden_states=encoder_hidden_states,
|
||||
image_rotary_emb=None,
|
||||
attention_mask=encoder_attention_mask,
|
||||
)
|
||||
hidden_states = hidden_states + attn_hidden_states
|
||||
norm_hidden_states = self.norm2(hidden_states) * (1 + scale_mlp) + shift_mlp
|
||||
|
||||
ff_output = self.ff(norm_hidden_states)
|
||||
hidden_states = hidden_states + ff_output * gate_mlp
|
||||
|
||||
hidden_states[2:] = hidden_states_ptb
|
||||
encoder_hidden_states[2:] = encoder_hidden_states_ptb
|
||||
|
||||
return hidden_states
|
||||
|
||||
|
||||
# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift
|
||||
def calculate_shift(
|
||||
image_seq_len,
|
||||
base_seq_len: int = 256,
|
||||
max_seq_len: int = 4096,
|
||||
base_shift: float = 0.5,
|
||||
max_shift: float = 1.16,
|
||||
):
|
||||
m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
|
||||
b = base_shift - m * base_seq_len
|
||||
mu = image_seq_len * m + b
|
||||
return mu
|
||||
|
||||
|
||||
# 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
|
||||
|
||||
|
||||
class LTXSTGPipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoLoraLoaderMixin):
|
||||
r"""
|
||||
Pipeline for text-to-video generation.
|
||||
|
||||
Reference: https://github.com/Lightricks/LTX-Video
|
||||
|
||||
Args:
|
||||
transformer ([`LTXVideoTransformer3DModel`]):
|
||||
Conditional Transformer architecture to denoise the encoded video latents.
|
||||
scheduler ([`FlowMatchEulerDiscreteScheduler`]):
|
||||
A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
|
||||
vae ([`AutoencoderKLLTXVideo`]):
|
||||
Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
|
||||
text_encoder ([`T5EncoderModel`]):
|
||||
[T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
|
||||
the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant.
|
||||
tokenizer (`CLIPTokenizer`):
|
||||
Tokenizer of class
|
||||
[CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
|
||||
tokenizer (`T5TokenizerFast`):
|
||||
Second Tokenizer of class
|
||||
[T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast).
|
||||
"""
|
||||
|
||||
model_cpu_offload_seq = "text_encoder->transformer->vae"
|
||||
_optional_components = []
|
||||
_callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
scheduler: FlowMatchEulerDiscreteScheduler,
|
||||
vae: AutoencoderKLLTXVideo,
|
||||
text_encoder: T5EncoderModel,
|
||||
tokenizer: T5TokenizerFast,
|
||||
transformer: LTXVideoTransformer3DModel,
|
||||
):
|
||||
super().__init__()
|
||||
|
||||
self.register_modules(
|
||||
vae=vae,
|
||||
text_encoder=text_encoder,
|
||||
tokenizer=tokenizer,
|
||||
transformer=transformer,
|
||||
scheduler=scheduler,
|
||||
)
|
||||
|
||||
self.vae_spatial_compression_ratio = (
|
||||
self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 32
|
||||
)
|
||||
self.vae_temporal_compression_ratio = (
|
||||
self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 8
|
||||
)
|
||||
self.transformer_spatial_patch_size = (
|
||||
self.transformer.config.patch_size if getattr(self, "transformer", None) is not None else 1
|
||||
)
|
||||
self.transformer_temporal_patch_size = (
|
||||
self.transformer.config.patch_size_t if getattr(self, "transformer") is not None else 1
|
||||
)
|
||||
|
||||
self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio)
|
||||
self.tokenizer_max_length = (
|
||||
self.tokenizer.model_max_length if getattr(self, "tokenizer", None) is not None else 128
|
||||
)
|
||||
|
||||
def _get_t5_prompt_embeds(
|
||||
self,
|
||||
prompt: Union[str, List[str]] = None,
|
||||
num_videos_per_prompt: int = 1,
|
||||
max_sequence_length: int = 128,
|
||||
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
|
||||
prompt_attention_mask = text_inputs.attention_mask
|
||||
prompt_attention_mask = prompt_attention_mask.bool().to(device)
|
||||
|
||||
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)
|
||||
|
||||
prompt_attention_mask = prompt_attention_mask.view(batch_size, -1)
|
||||
prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1)
|
||||
|
||||
return prompt_embeds, prompt_attention_mask
|
||||
|
||||
# Copied from diffusers.pipelines.mochi.pipeline_mochi.MochiPipeline.encode_prompt with 256->128
|
||||
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,
|
||||
prompt_attention_mask: Optional[torch.Tensor] = None,
|
||||
negative_prompt_attention_mask: Optional[torch.Tensor] = None,
|
||||
max_sequence_length: int = 128,
|
||||
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, prompt_attention_mask = 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, negative_prompt_attention_mask = 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, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
|
||||
|
||||
def check_inputs(
|
||||
self,
|
||||
prompt,
|
||||
height,
|
||||
width,
|
||||
callback_on_step_end_tensor_inputs=None,
|
||||
prompt_embeds=None,
|
||||
negative_prompt_embeds=None,
|
||||
prompt_attention_mask=None,
|
||||
negative_prompt_attention_mask=None,
|
||||
):
|
||||
if height % 32 != 0 or width % 32 != 0:
|
||||
raise ValueError(f"`height` and `width` have to be divisible by 32 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_embeds is not None and prompt_attention_mask is None:
|
||||
raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.")
|
||||
|
||||
if negative_prompt_embeds is not None and negative_prompt_attention_mask is None:
|
||||
raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.")
|
||||
|
||||
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}."
|
||||
)
|
||||
if prompt_attention_mask.shape != negative_prompt_attention_mask.shape:
|
||||
raise ValueError(
|
||||
"`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but"
|
||||
f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`"
|
||||
f" {negative_prompt_attention_mask.shape}."
|
||||
)
|
||||
|
||||
@staticmethod
|
||||
def _pack_latents(latents: torch.Tensor, patch_size: int = 1, patch_size_t: int = 1) -> torch.Tensor:
|
||||
# Unpacked latents of shape are [B, C, F, H, W] are patched into tokens of shape [B, C, F // p_t, p_t, H // p, p, W // p, p].
|
||||
# The patch dimensions are then permuted and collapsed into the channel dimension of shape:
|
||||
# [B, F // p_t * H // p * W // p, C * p_t * p * p] (an ndim=3 tensor).
|
||||
# dim=0 is the batch size, dim=1 is the effective video sequence length, dim=2 is the effective number of input features
|
||||
batch_size, num_channels, num_frames, height, width = latents.shape
|
||||
post_patch_num_frames = num_frames // patch_size_t
|
||||
post_patch_height = height // patch_size
|
||||
post_patch_width = width // patch_size
|
||||
latents = latents.reshape(
|
||||
batch_size,
|
||||
-1,
|
||||
post_patch_num_frames,
|
||||
patch_size_t,
|
||||
post_patch_height,
|
||||
patch_size,
|
||||
post_patch_width,
|
||||
patch_size,
|
||||
)
|
||||
latents = latents.permute(0, 2, 4, 6, 1, 3, 5, 7).flatten(4, 7).flatten(1, 3)
|
||||
return latents
|
||||
|
||||
@staticmethod
|
||||
def _unpack_latents(
|
||||
latents: torch.Tensor, num_frames: int, height: int, width: int, patch_size: int = 1, patch_size_t: int = 1
|
||||
) -> torch.Tensor:
|
||||
# Packed latents of shape [B, S, D] (S is the effective video sequence length, D is the effective feature dimensions)
|
||||
# are unpacked and reshaped into a video tensor of shape [B, C, F, H, W]. This is the inverse operation of
|
||||
# what happens in the `_pack_latents` method.
|
||||
batch_size = latents.size(0)
|
||||
latents = latents.reshape(batch_size, num_frames, height, width, -1, patch_size_t, patch_size, patch_size)
|
||||
latents = latents.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(2, 3)
|
||||
return latents
|
||||
|
||||
@staticmethod
|
||||
def _normalize_latents(
|
||||
latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0
|
||||
) -> torch.Tensor:
|
||||
# Normalize latents across the channel dimension [B, C, F, H, W]
|
||||
latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
|
||||
latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
|
||||
latents = (latents - latents_mean) * scaling_factor / latents_std
|
||||
return latents
|
||||
|
||||
@staticmethod
|
||||
def _denormalize_latents(
|
||||
latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0
|
||||
) -> torch.Tensor:
|
||||
# Denormalize latents across the channel dimension [B, C, F, H, W]
|
||||
latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
|
||||
latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
|
||||
latents = latents * latents_std / scaling_factor + latents_mean
|
||||
return latents
|
||||
|
||||
def prepare_latents(
|
||||
self,
|
||||
batch_size: int = 1,
|
||||
num_channels_latents: int = 128,
|
||||
height: int = 512,
|
||||
width: int = 704,
|
||||
num_frames: int = 161,
|
||||
dtype: Optional[torch.dtype] = None,
|
||||
device: Optional[torch.device] = None,
|
||||
generator: Optional[torch.Generator] = None,
|
||||
latents: Optional[torch.Tensor] = None,
|
||||
) -> torch.Tensor:
|
||||
if latents is not None:
|
||||
return latents.to(device=device, dtype=dtype)
|
||||
|
||||
height = height // self.vae_spatial_compression_ratio
|
||||
width = width // self.vae_spatial_compression_ratio
|
||||
num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1
|
||||
|
||||
shape = (batch_size, num_channels_latents, num_frames, height, width)
|
||||
|
||||
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."
|
||||
)
|
||||
|
||||
latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
|
||||
latents = self._pack_latents(
|
||||
latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
|
||||
)
|
||||
return latents
|
||||
|
||||
@property
|
||||
def guidance_scale(self):
|
||||
return self._guidance_scale
|
||||
|
||||
@property
|
||||
def do_classifier_free_guidance(self):
|
||||
return self._guidance_scale > 1.0
|
||||
|
||||
@property
|
||||
def do_spatio_temporal_guidance(self):
|
||||
return self._stg_scale > 0.0
|
||||
|
||||
@property
|
||||
def num_timesteps(self):
|
||||
return self._num_timesteps
|
||||
|
||||
@property
|
||||
def attention_kwargs(self):
|
||||
return self._attention_kwargs
|
||||
|
||||
@property
|
||||
def interrupt(self):
|
||||
return self._interrupt
|
||||
|
||||
@torch.no_grad()
|
||||
@replace_example_docstring(EXAMPLE_DOC_STRING)
|
||||
def __call__(
|
||||
self,
|
||||
prompt: Union[str, List[str]] = None,
|
||||
negative_prompt: Optional[Union[str, List[str]]] = None,
|
||||
height: int = 512,
|
||||
width: int = 704,
|
||||
num_frames: int = 161,
|
||||
frame_rate: int = 25,
|
||||
num_inference_steps: int = 50,
|
||||
timesteps: List[int] = None,
|
||||
guidance_scale: float = 3,
|
||||
num_videos_per_prompt: Optional[int] = 1,
|
||||
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
||||
latents: Optional[torch.Tensor] = None,
|
||||
prompt_embeds: Optional[torch.Tensor] = None,
|
||||
prompt_attention_mask: Optional[torch.Tensor] = None,
|
||||
negative_prompt_embeds: Optional[torch.Tensor] = None,
|
||||
negative_prompt_attention_mask: Optional[torch.Tensor] = None,
|
||||
decode_timestep: Union[float, List[float]] = 0.0,
|
||||
decode_noise_scale: Optional[Union[float, List[float]]] = None,
|
||||
output_type: Optional[str] = "pil",
|
||||
return_dict: bool = True,
|
||||
attention_kwargs: Optional[Dict[str, Any]] = None,
|
||||
callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
|
||||
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
|
||||
max_sequence_length: int = 128,
|
||||
stg_applied_layers_idx: Optional[List[int]] = [19],
|
||||
stg_scale: Optional[float] = 1.0,
|
||||
do_rescaling: Optional[bool] = False,
|
||||
):
|
||||
r"""
|
||||
Function invoked when calling the pipeline for generation.
|
||||
|
||||
Args:
|
||||
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.
|
||||
height (`int`, defaults to `512`):
|
||||
The height in pixels of the generated image. This is set to 480 by default for the best results.
|
||||
width (`int`, defaults to `704`):
|
||||
The width in pixels of the generated image. This is set to 848 by default for the best results.
|
||||
num_frames (`int`, defaults to `161`):
|
||||
The number of video frames to generate
|
||||
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`, defaults to `3 `):
|
||||
Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
|
||||
`guidance_scale` is defined as `w` of equation 2. of [Imagen
|
||||
Paper](https://arxiv.org/pdf/2205.11487.pdf). 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.Tensor`, *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.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.
|
||||
prompt_attention_mask (`torch.Tensor`, *optional*):
|
||||
Pre-generated attention mask for text embeddings.
|
||||
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
|
||||
Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not
|
||||
provided, negative_prompt_embeds will be generated from `negative_prompt` input argument.
|
||||
negative_prompt_attention_mask (`torch.FloatTensor`, *optional*):
|
||||
Pre-generated attention mask for negative text embeddings.
|
||||
decode_timestep (`float`, defaults to `0.0`):
|
||||
The timestep at which generated video is decoded.
|
||||
decode_noise_scale (`float`, defaults to `None`):
|
||||
The interpolation factor between random noise and denoised latents at the decode timestep.
|
||||
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.ltx.LTXPipelineOutput`] 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 `128 `):
|
||||
Maximum sequence length to use with the `prompt`.
|
||||
|
||||
Examples:
|
||||
|
||||
Returns:
|
||||
[`~pipelines.ltx.LTXPipelineOutput`] or `tuple`:
|
||||
If `return_dict` is `True`, [`~pipelines.ltx.LTXPipelineOutput`] is returned, otherwise a `tuple` is
|
||||
returned where 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
|
||||
|
||||
# 1. Check inputs. Raise error if not correct
|
||||
self.check_inputs(
|
||||
prompt=prompt,
|
||||
height=height,
|
||||
width=width,
|
||||
callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
|
||||
prompt_embeds=prompt_embeds,
|
||||
negative_prompt_embeds=negative_prompt_embeds,
|
||||
prompt_attention_mask=prompt_attention_mask,
|
||||
negative_prompt_attention_mask=negative_prompt_attention_mask,
|
||||
)
|
||||
|
||||
self._stg_scale = stg_scale
|
||||
self._guidance_scale = guidance_scale
|
||||
self._attention_kwargs = attention_kwargs
|
||||
self._interrupt = False
|
||||
|
||||
if self.do_spatio_temporal_guidance:
|
||||
for i in stg_applied_layers_idx:
|
||||
self.transformer.transformer_blocks[i].forward = types.MethodType(
|
||||
forward_with_stg, self.transformer.transformer_blocks[i]
|
||||
)
|
||||
|
||||
# 2. Define 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
|
||||
|
||||
# 3. Prepare text embeddings
|
||||
(
|
||||
prompt_embeds,
|
||||
prompt_attention_mask,
|
||||
negative_prompt_embeds,
|
||||
negative_prompt_attention_mask,
|
||||
) = self.encode_prompt(
|
||||
prompt=prompt,
|
||||
negative_prompt=negative_prompt,
|
||||
do_classifier_free_guidance=self.do_classifier_free_guidance,
|
||||
num_videos_per_prompt=num_videos_per_prompt,
|
||||
prompt_embeds=prompt_embeds,
|
||||
negative_prompt_embeds=negative_prompt_embeds,
|
||||
prompt_attention_mask=prompt_attention_mask,
|
||||
negative_prompt_attention_mask=negative_prompt_attention_mask,
|
||||
max_sequence_length=max_sequence_length,
|
||||
device=device,
|
||||
)
|
||||
if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
|
||||
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
|
||||
prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0)
|
||||
elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
|
||||
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds, prompt_embeds], dim=0)
|
||||
prompt_attention_mask = torch.cat(
|
||||
[negative_prompt_attention_mask, prompt_attention_mask, prompt_attention_mask], dim=0
|
||||
)
|
||||
|
||||
# 4. Prepare latent variables
|
||||
num_channels_latents = self.transformer.config.in_channels
|
||||
latents = self.prepare_latents(
|
||||
batch_size * num_videos_per_prompt,
|
||||
num_channels_latents,
|
||||
height,
|
||||
width,
|
||||
num_frames,
|
||||
torch.float32,
|
||||
device,
|
||||
generator,
|
||||
latents,
|
||||
)
|
||||
|
||||
# 5. Prepare timesteps
|
||||
latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1
|
||||
latent_height = height // self.vae_spatial_compression_ratio
|
||||
latent_width = width // self.vae_spatial_compression_ratio
|
||||
video_sequence_length = latent_num_frames * latent_height * latent_width
|
||||
sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
|
||||
mu = calculate_shift(
|
||||
video_sequence_length,
|
||||
self.scheduler.config.get("base_image_seq_len", 256),
|
||||
self.scheduler.config.get("max_image_seq_len", 4096),
|
||||
self.scheduler.config.get("base_shift", 0.5),
|
||||
self.scheduler.config.get("max_shift", 1.16),
|
||||
)
|
||||
timesteps, num_inference_steps = retrieve_timesteps(
|
||||
self.scheduler,
|
||||
num_inference_steps,
|
||||
device,
|
||||
timesteps,
|
||||
sigmas=sigmas,
|
||||
mu=mu,
|
||||
)
|
||||
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
|
||||
self._num_timesteps = len(timesteps)
|
||||
|
||||
# 6. Prepare micro-conditions
|
||||
latent_frame_rate = frame_rate / self.vae_temporal_compression_ratio
|
||||
rope_interpolation_scale = (
|
||||
1 / latent_frame_rate,
|
||||
self.vae_spatial_compression_ratio,
|
||||
self.vae_spatial_compression_ratio,
|
||||
)
|
||||
|
||||
# 7. Denoising loop
|
||||
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
||||
for i, t in enumerate(timesteps):
|
||||
if self.interrupt:
|
||||
continue
|
||||
|
||||
if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
|
||||
latent_model_input = torch.cat([latents] * 2)
|
||||
elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
|
||||
latent_model_input = torch.cat([latents] * 3)
|
||||
else:
|
||||
latent_model_input = latents
|
||||
|
||||
latent_model_input = latent_model_input.to(prompt_embeds.dtype)
|
||||
|
||||
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
||||
timestep = t.expand(latent_model_input.shape[0])
|
||||
|
||||
noise_pred = self.transformer(
|
||||
hidden_states=latent_model_input,
|
||||
encoder_hidden_states=prompt_embeds,
|
||||
timestep=timestep,
|
||||
encoder_attention_mask=prompt_attention_mask,
|
||||
num_frames=latent_num_frames,
|
||||
height=latent_height,
|
||||
width=latent_width,
|
||||
rope_interpolation_scale=rope_interpolation_scale,
|
||||
attention_kwargs=attention_kwargs,
|
||||
return_dict=False,
|
||||
)[0]
|
||||
noise_pred = noise_pred.float()
|
||||
|
||||
if self.do_classifier_free_guidance and not self.do_spatio_temporal_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)
|
||||
elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
|
||||
noise_pred_uncond, noise_pred_text, noise_pred_perturb = noise_pred.chunk(3)
|
||||
noise_pred = (
|
||||
noise_pred_uncond
|
||||
+ self.guidance_scale * (noise_pred_text - noise_pred_uncond)
|
||||
+ self._stg_scale * (noise_pred_text - noise_pred_perturb)
|
||||
)
|
||||
|
||||
if do_rescaling:
|
||||
rescaling_scale = 0.7
|
||||
factor = noise_pred_text.std() / noise_pred.std()
|
||||
factor = rescaling_scale * factor + (1 - rescaling_scale)
|
||||
noise_pred = noise_pred * factor
|
||||
|
||||
# compute the previous noisy sample x_t -> x_t-1
|
||||
latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
|
||||
|
||||
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)
|
||||
|
||||
# call the callback, if provided
|
||||
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()
|
||||
|
||||
if output_type == "latent":
|
||||
video = latents
|
||||
else:
|
||||
latents = self._unpack_latents(
|
||||
latents,
|
||||
latent_num_frames,
|
||||
latent_height,
|
||||
latent_width,
|
||||
self.transformer_spatial_patch_size,
|
||||
self.transformer_temporal_patch_size,
|
||||
)
|
||||
latents = self._denormalize_latents(
|
||||
latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor
|
||||
)
|
||||
latents = latents.to(prompt_embeds.dtype)
|
||||
|
||||
if not self.vae.config.timestep_conditioning:
|
||||
timestep = None
|
||||
else:
|
||||
noise = randn_tensor(latents.shape, generator=generator, device=device, dtype=latents.dtype)
|
||||
if not isinstance(decode_timestep, list):
|
||||
decode_timestep = [decode_timestep] * batch_size
|
||||
if decode_noise_scale is None:
|
||||
decode_noise_scale = decode_timestep
|
||||
elif not isinstance(decode_noise_scale, list):
|
||||
decode_noise_scale = [decode_noise_scale] * batch_size
|
||||
|
||||
timestep = torch.tensor(decode_timestep, device=device, dtype=latents.dtype)
|
||||
decode_noise_scale = torch.tensor(decode_noise_scale, device=device, dtype=latents.dtype)[
|
||||
:, None, None, None, None
|
||||
]
|
||||
latents = (1 - decode_noise_scale) * latents + decode_noise_scale * noise
|
||||
|
||||
video = self.vae.decode(latents, timestep, return_dict=False)[0]
|
||||
video = self.video_processor.postprocess_video(video, output_type=output_type)
|
||||
|
||||
# Offload all models
|
||||
self.maybe_free_model_hooks()
|
||||
|
||||
if not return_dict:
|
||||
return (video,)
|
||||
|
||||
return LTXPipelineOutput(frames=video)
|
||||
985
examples/community/pipeline_stg_ltx_image2video.py
Normal file
985
examples/community/pipeline_stg_ltx_image2video.py
Normal file
@@ -0,0 +1,985 @@
|
||||
# Copyright 2024 Lightricks and The HuggingFace Team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
import inspect
|
||||
import types
|
||||
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
from transformers import T5EncoderModel, T5TokenizerFast
|
||||
|
||||
from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
|
||||
from diffusers.image_processor import PipelineImageInput
|
||||
from diffusers.loaders import FromSingleFileMixin, LTXVideoLoraLoaderMixin
|
||||
from diffusers.models.autoencoders import AutoencoderKLLTXVideo
|
||||
from diffusers.models.transformers import LTXVideoTransformer3DModel
|
||||
from diffusers.pipelines.ltx.pipeline_output import LTXPipelineOutput
|
||||
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
|
||||
from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
|
||||
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
|
||||
|
||||
|
||||
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
|
||||
|
||||
EXAMPLE_DOC_STRING = """
|
||||
Examples:
|
||||
```py
|
||||
>>> import torch
|
||||
>>> from diffusers.utils import export_to_video, load_image
|
||||
>>> from examples.community.pipeline_stg_ltx_image2video import LTXImageToVideoSTGPipeline
|
||||
|
||||
>>> pipe = LTXImageToVideoSTGPipeline.from_pretrained("Lightricks/LTX-Video", torch_dtype=torch.bfloat16)
|
||||
>>> pipe.to("cuda")
|
||||
|
||||
>>> image = load_image(
|
||||
... "https://huggingface.co/datasets/a-r-r-o-w/tiny-meme-dataset-captioned/resolve/main/images/11.png"
|
||||
>>> )
|
||||
>>> prompt = "A medieval fantasy scene featuring a rugged man with shoulder-length brown hair and a beard. He wears a dark leather tunic over a maroon shirt with intricate metal details. His facial expression is serious and intense, and he is making a gesture with his right hand, forming a small circle with his thumb and index finger. The warm golden lighting casts dramatic shadows on his face. The background includes an ornate stone arch and blurred medieval-style decor, creating an epic atmosphere."
|
||||
>>> negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted"
|
||||
|
||||
>>> # Configure STG mode options
|
||||
>>> stg_applied_layers_idx = [19] # Layer indices from 0 to 41
|
||||
>>> stg_scale = 1.0 # Set 0.0 for CFG
|
||||
>>> do_rescaling = False
|
||||
|
||||
>>> video = pipe(
|
||||
... image=image,
|
||||
... prompt=prompt,
|
||||
... negative_prompt=negative_prompt,
|
||||
... width=704,
|
||||
... height=480,
|
||||
... num_frames=161,
|
||||
... num_inference_steps=50,
|
||||
... stg_applied_layers_idx=stg_applied_layers_idx,
|
||||
... stg_scale=stg_scale,
|
||||
... do_rescaling=do_rescaling,
|
||||
>>> ).frames[0]
|
||||
>>> export_to_video(video, "output.mp4", fps=24)
|
||||
```
|
||||
"""
|
||||
|
||||
|
||||
def forward_with_stg(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
encoder_hidden_states: torch.Tensor,
|
||||
temb: torch.Tensor,
|
||||
image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
|
||||
encoder_attention_mask: Optional[torch.Tensor] = None,
|
||||
) -> torch.Tensor:
|
||||
hidden_states_ptb = hidden_states[2:]
|
||||
encoder_hidden_states_ptb = encoder_hidden_states[2:]
|
||||
|
||||
batch_size = hidden_states.size(0)
|
||||
norm_hidden_states = self.norm1(hidden_states)
|
||||
|
||||
num_ada_params = self.scale_shift_table.shape[0]
|
||||
ada_values = self.scale_shift_table[None, None] + temb.reshape(batch_size, temb.size(1), num_ada_params, -1)
|
||||
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ada_values.unbind(dim=2)
|
||||
norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
|
||||
|
||||
attn_hidden_states = self.attn1(
|
||||
hidden_states=norm_hidden_states,
|
||||
encoder_hidden_states=None,
|
||||
image_rotary_emb=image_rotary_emb,
|
||||
)
|
||||
hidden_states = hidden_states + attn_hidden_states * gate_msa
|
||||
|
||||
attn_hidden_states = self.attn2(
|
||||
hidden_states,
|
||||
encoder_hidden_states=encoder_hidden_states,
|
||||
image_rotary_emb=None,
|
||||
attention_mask=encoder_attention_mask,
|
||||
)
|
||||
hidden_states = hidden_states + attn_hidden_states
|
||||
norm_hidden_states = self.norm2(hidden_states) * (1 + scale_mlp) + shift_mlp
|
||||
|
||||
ff_output = self.ff(norm_hidden_states)
|
||||
hidden_states = hidden_states + ff_output * gate_mlp
|
||||
|
||||
hidden_states[2:] = hidden_states_ptb
|
||||
encoder_hidden_states[2:] = encoder_hidden_states_ptb
|
||||
|
||||
return hidden_states
|
||||
|
||||
|
||||
# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift
|
||||
def calculate_shift(
|
||||
image_seq_len,
|
||||
base_seq_len: int = 256,
|
||||
max_seq_len: int = 4096,
|
||||
base_shift: float = 0.5,
|
||||
max_shift: float = 1.16,
|
||||
):
|
||||
m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
|
||||
b = base_shift - m * base_seq_len
|
||||
mu = image_seq_len * m + b
|
||||
return mu
|
||||
|
||||
|
||||
# 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 LTXImageToVideoSTGPipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoLoraLoaderMixin):
|
||||
r"""
|
||||
Pipeline for image-to-video generation.
|
||||
|
||||
Reference: https://github.com/Lightricks/LTX-Video
|
||||
|
||||
Args:
|
||||
transformer ([`LTXVideoTransformer3DModel`]):
|
||||
Conditional Transformer architecture to denoise the encoded video latents.
|
||||
scheduler ([`FlowMatchEulerDiscreteScheduler`]):
|
||||
A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
|
||||
vae ([`AutoencoderKLLTXVideo`]):
|
||||
Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
|
||||
text_encoder ([`T5EncoderModel`]):
|
||||
[T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
|
||||
the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant.
|
||||
tokenizer (`CLIPTokenizer`):
|
||||
Tokenizer of class
|
||||
[CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
|
||||
tokenizer (`T5TokenizerFast`):
|
||||
Second Tokenizer of class
|
||||
[T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast).
|
||||
"""
|
||||
|
||||
model_cpu_offload_seq = "text_encoder->transformer->vae"
|
||||
_optional_components = []
|
||||
_callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
scheduler: FlowMatchEulerDiscreteScheduler,
|
||||
vae: AutoencoderKLLTXVideo,
|
||||
text_encoder: T5EncoderModel,
|
||||
tokenizer: T5TokenizerFast,
|
||||
transformer: LTXVideoTransformer3DModel,
|
||||
):
|
||||
super().__init__()
|
||||
|
||||
self.register_modules(
|
||||
vae=vae,
|
||||
text_encoder=text_encoder,
|
||||
tokenizer=tokenizer,
|
||||
transformer=transformer,
|
||||
scheduler=scheduler,
|
||||
)
|
||||
|
||||
self.vae_spatial_compression_ratio = (
|
||||
self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 32
|
||||
)
|
||||
self.vae_temporal_compression_ratio = (
|
||||
self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 8
|
||||
)
|
||||
self.transformer_spatial_patch_size = (
|
||||
self.transformer.config.patch_size if getattr(self, "transformer", None) is not None else 1
|
||||
)
|
||||
self.transformer_temporal_patch_size = (
|
||||
self.transformer.config.patch_size_t if getattr(self, "transformer") is not None else 1
|
||||
)
|
||||
|
||||
self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio)
|
||||
self.tokenizer_max_length = (
|
||||
self.tokenizer.model_max_length if getattr(self, "tokenizer", None) is not None else 128
|
||||
)
|
||||
|
||||
self.default_height = 512
|
||||
self.default_width = 704
|
||||
self.default_frames = 121
|
||||
|
||||
def _get_t5_prompt_embeds(
|
||||
self,
|
||||
prompt: Union[str, List[str]] = None,
|
||||
num_videos_per_prompt: int = 1,
|
||||
max_sequence_length: int = 128,
|
||||
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
|
||||
prompt_attention_mask = text_inputs.attention_mask
|
||||
prompt_attention_mask = prompt_attention_mask.bool().to(device)
|
||||
|
||||
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)
|
||||
|
||||
prompt_attention_mask = prompt_attention_mask.view(batch_size, -1)
|
||||
prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1)
|
||||
|
||||
return prompt_embeds, prompt_attention_mask
|
||||
|
||||
# Copied from diffusers.pipelines.mochi.pipeline_mochi.MochiPipeline.encode_prompt with 256->128
|
||||
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,
|
||||
prompt_attention_mask: Optional[torch.Tensor] = None,
|
||||
negative_prompt_attention_mask: Optional[torch.Tensor] = None,
|
||||
max_sequence_length: int = 128,
|
||||
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, prompt_attention_mask = 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, negative_prompt_attention_mask = 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, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
|
||||
|
||||
# Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline.check_inputs
|
||||
def check_inputs(
|
||||
self,
|
||||
prompt,
|
||||
height,
|
||||
width,
|
||||
callback_on_step_end_tensor_inputs=None,
|
||||
prompt_embeds=None,
|
||||
negative_prompt_embeds=None,
|
||||
prompt_attention_mask=None,
|
||||
negative_prompt_attention_mask=None,
|
||||
):
|
||||
if height % 32 != 0 or width % 32 != 0:
|
||||
raise ValueError(f"`height` and `width` have to be divisible by 32 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_embeds is not None and prompt_attention_mask is None:
|
||||
raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.")
|
||||
|
||||
if negative_prompt_embeds is not None and negative_prompt_attention_mask is None:
|
||||
raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.")
|
||||
|
||||
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}."
|
||||
)
|
||||
if prompt_attention_mask.shape != negative_prompt_attention_mask.shape:
|
||||
raise ValueError(
|
||||
"`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but"
|
||||
f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`"
|
||||
f" {negative_prompt_attention_mask.shape}."
|
||||
)
|
||||
|
||||
@staticmethod
|
||||
# Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._pack_latents
|
||||
def _pack_latents(latents: torch.Tensor, patch_size: int = 1, patch_size_t: int = 1) -> torch.Tensor:
|
||||
# Unpacked latents of shape are [B, C, F, H, W] are patched into tokens of shape [B, C, F // p_t, p_t, H // p, p, W // p, p].
|
||||
# The patch dimensions are then permuted and collapsed into the channel dimension of shape:
|
||||
# [B, F // p_t * H // p * W // p, C * p_t * p * p] (an ndim=3 tensor).
|
||||
# dim=0 is the batch size, dim=1 is the effective video sequence length, dim=2 is the effective number of input features
|
||||
batch_size, num_channels, num_frames, height, width = latents.shape
|
||||
post_patch_num_frames = num_frames // patch_size_t
|
||||
post_patch_height = height // patch_size
|
||||
post_patch_width = width // patch_size
|
||||
latents = latents.reshape(
|
||||
batch_size,
|
||||
-1,
|
||||
post_patch_num_frames,
|
||||
patch_size_t,
|
||||
post_patch_height,
|
||||
patch_size,
|
||||
post_patch_width,
|
||||
patch_size,
|
||||
)
|
||||
latents = latents.permute(0, 2, 4, 6, 1, 3, 5, 7).flatten(4, 7).flatten(1, 3)
|
||||
return latents
|
||||
|
||||
@staticmethod
|
||||
# Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._unpack_latents
|
||||
def _unpack_latents(
|
||||
latents: torch.Tensor, num_frames: int, height: int, width: int, patch_size: int = 1, patch_size_t: int = 1
|
||||
) -> torch.Tensor:
|
||||
# Packed latents of shape [B, S, D] (S is the effective video sequence length, D is the effective feature dimensions)
|
||||
# are unpacked and reshaped into a video tensor of shape [B, C, F, H, W]. This is the inverse operation of
|
||||
# what happens in the `_pack_latents` method.
|
||||
batch_size = latents.size(0)
|
||||
latents = latents.reshape(batch_size, num_frames, height, width, -1, patch_size_t, patch_size, patch_size)
|
||||
latents = latents.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(2, 3)
|
||||
return latents
|
||||
|
||||
@staticmethod
|
||||
# Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._normalize_latents
|
||||
def _normalize_latents(
|
||||
latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0
|
||||
) -> torch.Tensor:
|
||||
# Normalize latents across the channel dimension [B, C, F, H, W]
|
||||
latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
|
||||
latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
|
||||
latents = (latents - latents_mean) * scaling_factor / latents_std
|
||||
return latents
|
||||
|
||||
@staticmethod
|
||||
# Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._denormalize_latents
|
||||
def _denormalize_latents(
|
||||
latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0
|
||||
) -> torch.Tensor:
|
||||
# Denormalize latents across the channel dimension [B, C, F, H, W]
|
||||
latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
|
||||
latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
|
||||
latents = latents * latents_std / scaling_factor + latents_mean
|
||||
return latents
|
||||
|
||||
def prepare_latents(
|
||||
self,
|
||||
image: Optional[torch.Tensor] = None,
|
||||
batch_size: int = 1,
|
||||
num_channels_latents: int = 128,
|
||||
height: int = 512,
|
||||
width: int = 704,
|
||||
num_frames: int = 161,
|
||||
dtype: Optional[torch.dtype] = None,
|
||||
device: Optional[torch.device] = None,
|
||||
generator: Optional[torch.Generator] = None,
|
||||
latents: Optional[torch.Tensor] = None,
|
||||
) -> torch.Tensor:
|
||||
height = height // self.vae_spatial_compression_ratio
|
||||
width = width // self.vae_spatial_compression_ratio
|
||||
num_frames = (
|
||||
(num_frames - 1) // self.vae_temporal_compression_ratio + 1 if latents is None else latents.size(2)
|
||||
)
|
||||
|
||||
shape = (batch_size, num_channels_latents, num_frames, height, width)
|
||||
mask_shape = (batch_size, 1, num_frames, height, width)
|
||||
|
||||
if latents is not None:
|
||||
conditioning_mask = latents.new_zeros(shape)
|
||||
conditioning_mask[:, :, 0] = 1.0
|
||||
conditioning_mask = self._pack_latents(
|
||||
conditioning_mask, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
|
||||
)
|
||||
return latents.to(device=device, dtype=dtype), conditioning_mask
|
||||
|
||||
if isinstance(generator, list):
|
||||
if 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."
|
||||
)
|
||||
|
||||
init_latents = [
|
||||
retrieve_latents(self.vae.encode(image[i].unsqueeze(0).unsqueeze(2)), generator[i])
|
||||
for i in range(batch_size)
|
||||
]
|
||||
else:
|
||||
init_latents = [
|
||||
retrieve_latents(self.vae.encode(img.unsqueeze(0).unsqueeze(2)), generator) for img in image
|
||||
]
|
||||
|
||||
init_latents = torch.cat(init_latents, dim=0).to(dtype)
|
||||
init_latents = self._normalize_latents(init_latents, self.vae.latents_mean, self.vae.latents_std)
|
||||
init_latents = init_latents.repeat(1, 1, num_frames, 1, 1)
|
||||
conditioning_mask = torch.zeros(mask_shape, device=device, dtype=dtype)
|
||||
conditioning_mask[:, :, 0] = 1.0
|
||||
|
||||
noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
|
||||
latents = init_latents * conditioning_mask + noise * (1 - conditioning_mask)
|
||||
|
||||
conditioning_mask = self._pack_latents(
|
||||
conditioning_mask, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
|
||||
).squeeze(-1)
|
||||
latents = self._pack_latents(
|
||||
latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
|
||||
)
|
||||
|
||||
return latents, conditioning_mask
|
||||
|
||||
@property
|
||||
def guidance_scale(self):
|
||||
return self._guidance_scale
|
||||
|
||||
@property
|
||||
def do_classifier_free_guidance(self):
|
||||
return self._guidance_scale > 1.0
|
||||
|
||||
@property
|
||||
def do_spatio_temporal_guidance(self):
|
||||
return self._stg_scale > 0.0
|
||||
|
||||
@property
|
||||
def num_timesteps(self):
|
||||
return self._num_timesteps
|
||||
|
||||
@property
|
||||
def attention_kwargs(self):
|
||||
return self._attention_kwargs
|
||||
|
||||
@property
|
||||
def interrupt(self):
|
||||
return self._interrupt
|
||||
|
||||
@torch.no_grad()
|
||||
@replace_example_docstring(EXAMPLE_DOC_STRING)
|
||||
def __call__(
|
||||
self,
|
||||
image: PipelineImageInput = None,
|
||||
prompt: Union[str, List[str]] = None,
|
||||
negative_prompt: Optional[Union[str, List[str]]] = None,
|
||||
height: int = 512,
|
||||
width: int = 704,
|
||||
num_frames: int = 161,
|
||||
frame_rate: int = 25,
|
||||
num_inference_steps: int = 50,
|
||||
timesteps: List[int] = None,
|
||||
guidance_scale: float = 3,
|
||||
num_videos_per_prompt: Optional[int] = 1,
|
||||
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
||||
latents: Optional[torch.Tensor] = None,
|
||||
prompt_embeds: Optional[torch.Tensor] = None,
|
||||
prompt_attention_mask: Optional[torch.Tensor] = None,
|
||||
negative_prompt_embeds: Optional[torch.Tensor] = None,
|
||||
negative_prompt_attention_mask: Optional[torch.Tensor] = None,
|
||||
decode_timestep: Union[float, List[float]] = 0.0,
|
||||
decode_noise_scale: Optional[Union[float, List[float]]] = None,
|
||||
output_type: Optional[str] = "pil",
|
||||
return_dict: bool = True,
|
||||
attention_kwargs: Optional[Dict[str, Any]] = None,
|
||||
callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
|
||||
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
|
||||
max_sequence_length: int = 128,
|
||||
stg_applied_layers_idx: Optional[List[int]] = [19],
|
||||
stg_scale: Optional[float] = 1.0,
|
||||
do_rescaling: Optional[bool] = False,
|
||||
):
|
||||
r"""
|
||||
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.
|
||||
height (`int`, defaults to `512`):
|
||||
The height in pixels of the generated image. This is set to 480 by default for the best results.
|
||||
width (`int`, defaults to `704`):
|
||||
The width in pixels of the generated image. This is set to 848 by default for the best results.
|
||||
num_frames (`int`, defaults to `161`):
|
||||
The number of video frames to generate
|
||||
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`, defaults to `3 `):
|
||||
Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
|
||||
`guidance_scale` is defined as `w` of equation 2. of [Imagen
|
||||
Paper](https://arxiv.org/pdf/2205.11487.pdf). 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.Tensor`, *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.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.
|
||||
prompt_attention_mask (`torch.Tensor`, *optional*):
|
||||
Pre-generated attention mask for text embeddings.
|
||||
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
|
||||
Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not
|
||||
provided, negative_prompt_embeds will be generated from `negative_prompt` input argument.
|
||||
negative_prompt_attention_mask (`torch.FloatTensor`, *optional*):
|
||||
Pre-generated attention mask for negative text embeddings.
|
||||
decode_timestep (`float`, defaults to `0.0`):
|
||||
The timestep at which generated video is decoded.
|
||||
decode_noise_scale (`float`, defaults to `None`):
|
||||
The interpolation factor between random noise and denoised latents at the decode timestep.
|
||||
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.ltx.LTXPipelineOutput`] 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 `128 `):
|
||||
Maximum sequence length to use with the `prompt`.
|
||||
|
||||
Examples:
|
||||
|
||||
Returns:
|
||||
[`~pipelines.ltx.LTXPipelineOutput`] or `tuple`:
|
||||
If `return_dict` is `True`, [`~pipelines.ltx.LTXPipelineOutput`] is returned, otherwise a `tuple` is
|
||||
returned where 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
|
||||
|
||||
# 1. Check inputs. Raise error if not correct
|
||||
self.check_inputs(
|
||||
prompt=prompt,
|
||||
height=height,
|
||||
width=width,
|
||||
callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
|
||||
prompt_embeds=prompt_embeds,
|
||||
negative_prompt_embeds=negative_prompt_embeds,
|
||||
prompt_attention_mask=prompt_attention_mask,
|
||||
negative_prompt_attention_mask=negative_prompt_attention_mask,
|
||||
)
|
||||
|
||||
self._stg_scale = stg_scale
|
||||
self._guidance_scale = guidance_scale
|
||||
self._attention_kwargs = attention_kwargs
|
||||
self._interrupt = False
|
||||
|
||||
if self.do_spatio_temporal_guidance:
|
||||
for i in stg_applied_layers_idx:
|
||||
self.transformer.transformer_blocks[i].forward = types.MethodType(
|
||||
forward_with_stg, self.transformer.transformer_blocks[i]
|
||||
)
|
||||
|
||||
# 2. Define 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
|
||||
|
||||
# 3. Prepare text embeddings
|
||||
(
|
||||
prompt_embeds,
|
||||
prompt_attention_mask,
|
||||
negative_prompt_embeds,
|
||||
negative_prompt_attention_mask,
|
||||
) = self.encode_prompt(
|
||||
prompt=prompt,
|
||||
negative_prompt=negative_prompt,
|
||||
do_classifier_free_guidance=self.do_classifier_free_guidance,
|
||||
num_videos_per_prompt=num_videos_per_prompt,
|
||||
prompt_embeds=prompt_embeds,
|
||||
negative_prompt_embeds=negative_prompt_embeds,
|
||||
prompt_attention_mask=prompt_attention_mask,
|
||||
negative_prompt_attention_mask=negative_prompt_attention_mask,
|
||||
max_sequence_length=max_sequence_length,
|
||||
device=device,
|
||||
)
|
||||
if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
|
||||
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
|
||||
prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0)
|
||||
elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
|
||||
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds, prompt_embeds], dim=0)
|
||||
prompt_attention_mask = torch.cat(
|
||||
[negative_prompt_attention_mask, prompt_attention_mask, prompt_attention_mask], dim=0
|
||||
)
|
||||
|
||||
# 4. Prepare latent variables
|
||||
if latents is None:
|
||||
image = self.video_processor.preprocess(image, height=height, width=width)
|
||||
image = image.to(device=device, dtype=prompt_embeds.dtype)
|
||||
|
||||
num_channels_latents = self.transformer.config.in_channels
|
||||
latents, conditioning_mask = self.prepare_latents(
|
||||
image,
|
||||
batch_size * num_videos_per_prompt,
|
||||
num_channels_latents,
|
||||
height,
|
||||
width,
|
||||
num_frames,
|
||||
torch.float32,
|
||||
device,
|
||||
generator,
|
||||
latents,
|
||||
)
|
||||
|
||||
if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
|
||||
conditioning_mask = torch.cat([conditioning_mask, conditioning_mask])
|
||||
elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
|
||||
conditioning_mask = torch.cat([conditioning_mask, conditioning_mask, conditioning_mask])
|
||||
|
||||
# 5. Prepare timesteps
|
||||
latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1
|
||||
latent_height = height // self.vae_spatial_compression_ratio
|
||||
latent_width = width // self.vae_spatial_compression_ratio
|
||||
video_sequence_length = latent_num_frames * latent_height * latent_width
|
||||
sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
|
||||
mu = calculate_shift(
|
||||
video_sequence_length,
|
||||
self.scheduler.config.get("base_image_seq_len", 256),
|
||||
self.scheduler.config.get("max_image_seq_len", 4096),
|
||||
self.scheduler.config.get("base_shift", 0.5),
|
||||
self.scheduler.config.get("max_shift", 1.16),
|
||||
)
|
||||
timesteps, num_inference_steps = retrieve_timesteps(
|
||||
self.scheduler,
|
||||
num_inference_steps,
|
||||
device,
|
||||
timesteps,
|
||||
sigmas=sigmas,
|
||||
mu=mu,
|
||||
)
|
||||
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
|
||||
self._num_timesteps = len(timesteps)
|
||||
|
||||
# 6. Prepare micro-conditions
|
||||
latent_frame_rate = frame_rate / self.vae_temporal_compression_ratio
|
||||
rope_interpolation_scale = (
|
||||
1 / latent_frame_rate,
|
||||
self.vae_spatial_compression_ratio,
|
||||
self.vae_spatial_compression_ratio,
|
||||
)
|
||||
|
||||
# 7. Denoising loop
|
||||
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
||||
for i, t in enumerate(timesteps):
|
||||
if self.interrupt:
|
||||
continue
|
||||
|
||||
if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
|
||||
latent_model_input = torch.cat([latents] * 2)
|
||||
elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
|
||||
latent_model_input = torch.cat([latents] * 3)
|
||||
else:
|
||||
latent_model_input = latents
|
||||
|
||||
latent_model_input = latent_model_input.to(prompt_embeds.dtype)
|
||||
|
||||
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
||||
timestep = t.expand(latent_model_input.shape[0])
|
||||
timestep = timestep.unsqueeze(-1) * (1 - conditioning_mask)
|
||||
|
||||
noise_pred = self.transformer(
|
||||
hidden_states=latent_model_input,
|
||||
encoder_hidden_states=prompt_embeds,
|
||||
timestep=timestep,
|
||||
encoder_attention_mask=prompt_attention_mask,
|
||||
num_frames=latent_num_frames,
|
||||
height=latent_height,
|
||||
width=latent_width,
|
||||
rope_interpolation_scale=rope_interpolation_scale,
|
||||
attention_kwargs=attention_kwargs,
|
||||
return_dict=False,
|
||||
)[0]
|
||||
noise_pred = noise_pred.float()
|
||||
|
||||
if self.do_classifier_free_guidance and not self.do_spatio_temporal_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)
|
||||
timestep, _ = timestep.chunk(2)
|
||||
elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
|
||||
noise_pred_uncond, noise_pred_text, noise_pred_perturb = noise_pred.chunk(3)
|
||||
noise_pred = (
|
||||
noise_pred_uncond
|
||||
+ self.guidance_scale * (noise_pred_text - noise_pred_uncond)
|
||||
+ self._stg_scale * (noise_pred_text - noise_pred_perturb)
|
||||
)
|
||||
timestep, _, _ = timestep.chunk(3)
|
||||
|
||||
if do_rescaling:
|
||||
rescaling_scale = 0.7
|
||||
factor = noise_pred_text.std() / noise_pred.std()
|
||||
factor = rescaling_scale * factor + (1 - rescaling_scale)
|
||||
noise_pred = noise_pred * factor
|
||||
|
||||
# compute the previous noisy sample x_t -> x_t-1
|
||||
noise_pred = self._unpack_latents(
|
||||
noise_pred,
|
||||
latent_num_frames,
|
||||
latent_height,
|
||||
latent_width,
|
||||
self.transformer_spatial_patch_size,
|
||||
self.transformer_temporal_patch_size,
|
||||
)
|
||||
latents = self._unpack_latents(
|
||||
latents,
|
||||
latent_num_frames,
|
||||
latent_height,
|
||||
latent_width,
|
||||
self.transformer_spatial_patch_size,
|
||||
self.transformer_temporal_patch_size,
|
||||
)
|
||||
|
||||
noise_pred = noise_pred[:, :, 1:]
|
||||
noise_latents = latents[:, :, 1:]
|
||||
pred_latents = self.scheduler.step(noise_pred, t, noise_latents, return_dict=False)[0]
|
||||
|
||||
latents = torch.cat([latents[:, :, :1], pred_latents], dim=2)
|
||||
latents = self._pack_latents(
|
||||
latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
|
||||
)
|
||||
|
||||
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)
|
||||
|
||||
# call the callback, if provided
|
||||
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()
|
||||
|
||||
if output_type == "latent":
|
||||
video = latents
|
||||
else:
|
||||
latents = self._unpack_latents(
|
||||
latents,
|
||||
latent_num_frames,
|
||||
latent_height,
|
||||
latent_width,
|
||||
self.transformer_spatial_patch_size,
|
||||
self.transformer_temporal_patch_size,
|
||||
)
|
||||
latents = self._denormalize_latents(
|
||||
latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor
|
||||
)
|
||||
latents = latents.to(prompt_embeds.dtype)
|
||||
|
||||
if not self.vae.config.timestep_conditioning:
|
||||
timestep = None
|
||||
else:
|
||||
noise = torch.randn(latents.shape, generator=generator, device=device, dtype=latents.dtype)
|
||||
if not isinstance(decode_timestep, list):
|
||||
decode_timestep = [decode_timestep] * batch_size
|
||||
if decode_noise_scale is None:
|
||||
decode_noise_scale = decode_timestep
|
||||
elif not isinstance(decode_noise_scale, list):
|
||||
decode_noise_scale = [decode_noise_scale] * batch_size
|
||||
|
||||
timestep = torch.tensor(decode_timestep, device=device, dtype=latents.dtype)
|
||||
decode_noise_scale = torch.tensor(decode_noise_scale, device=device, dtype=latents.dtype)[
|
||||
:, None, None, None, None
|
||||
]
|
||||
latents = (1 - decode_noise_scale) * latents + decode_noise_scale * noise
|
||||
|
||||
video = self.vae.decode(latents, timestep, return_dict=False)[0]
|
||||
video = self.video_processor.postprocess_video(video, output_type=output_type)
|
||||
|
||||
# Offload all models
|
||||
self.maybe_free_model_hooks()
|
||||
|
||||
if not return_dict:
|
||||
return (video,)
|
||||
|
||||
return LTXPipelineOutput(frames=video)
|
||||
843
examples/community/pipeline_stg_mochi.py
Normal file
843
examples/community/pipeline_stg_mochi.py
Normal file
@@ -0,0 +1,843 @@
|
||||
# Copyright 2024 Genmo and The HuggingFace Team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
import inspect
|
||||
import types
|
||||
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
from transformers import T5EncoderModel, T5TokenizerFast
|
||||
|
||||
from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
|
||||
from diffusers.loaders import Mochi1LoraLoaderMixin
|
||||
from diffusers.models import AutoencoderKLMochi, MochiTransformer3DModel
|
||||
from diffusers.pipelines.mochi.pipeline_output import MochiPipelineOutput
|
||||
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
|
||||
from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
|
||||
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
|
||||
|
||||
|
||||
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
|
||||
|
||||
EXAMPLE_DOC_STRING = """
|
||||
Examples:
|
||||
```py
|
||||
>>> import torch
|
||||
>>> from diffusers.utils import export_to_video
|
||||
>>> from examples.community.pipeline_stg_mochi import MochiSTGPipeline
|
||||
|
||||
>>> pipe = MochiSTGPipeline.from_pretrained("genmo/mochi-1-preview", torch_dtype=torch.bfloat16)
|
||||
>>> pipe.enable_model_cpu_offload()
|
||||
>>> pipe.enable_vae_tiling()
|
||||
>>> prompt = "A close-up of a beautiful woman's face with colored powder exploding around her, creating an abstract splash of vibrant hues, realistic style."
|
||||
|
||||
>>> # Configure STG mode options
|
||||
>>> stg_applied_layers_idx = [34] # Layer indices from 0 to 41
|
||||
>>> stg_scale = 1.0 # Set 0.0 for CFG
|
||||
>>> do_rescaling = False
|
||||
|
||||
>>> frames = pipe(
|
||||
... prompt=prompt,
|
||||
... num_inference_steps=28,
|
||||
... guidance_scale=3.5,
|
||||
... stg_applied_layers_idx=stg_applied_layers_idx,
|
||||
... stg_scale=stg_scale,
|
||||
... do_rescaling=do_rescaling).frames[0]
|
||||
>>> export_to_video(frames, "mochi.mp4")
|
||||
```
|
||||
"""
|
||||
|
||||
|
||||
def forward_with_stg(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
encoder_hidden_states: torch.Tensor,
|
||||
temb: torch.Tensor,
|
||||
encoder_attention_mask: torch.Tensor,
|
||||
image_rotary_emb: Optional[torch.Tensor] = None,
|
||||
) -> Tuple[torch.Tensor, torch.Tensor]:
|
||||
hidden_states_ptb = hidden_states[2:]
|
||||
encoder_hidden_states_ptb = encoder_hidden_states[2:]
|
||||
norm_hidden_states, gate_msa, scale_mlp, gate_mlp = self.norm1(hidden_states, temb)
|
||||
|
||||
if not self.context_pre_only:
|
||||
norm_encoder_hidden_states, enc_gate_msa, enc_scale_mlp, enc_gate_mlp = self.norm1_context(
|
||||
encoder_hidden_states, temb
|
||||
)
|
||||
else:
|
||||
norm_encoder_hidden_states = self.norm1_context(encoder_hidden_states, temb)
|
||||
|
||||
attn_hidden_states, context_attn_hidden_states = self.attn1(
|
||||
hidden_states=norm_hidden_states,
|
||||
encoder_hidden_states=norm_encoder_hidden_states,
|
||||
image_rotary_emb=image_rotary_emb,
|
||||
attention_mask=encoder_attention_mask,
|
||||
)
|
||||
|
||||
hidden_states = hidden_states + self.norm2(attn_hidden_states, torch.tanh(gate_msa).unsqueeze(1))
|
||||
norm_hidden_states = self.norm3(hidden_states, (1 + scale_mlp.unsqueeze(1).to(torch.float32)))
|
||||
ff_output = self.ff(norm_hidden_states)
|
||||
hidden_states = hidden_states + self.norm4(ff_output, torch.tanh(gate_mlp).unsqueeze(1))
|
||||
|
||||
if not self.context_pre_only:
|
||||
encoder_hidden_states = encoder_hidden_states + self.norm2_context(
|
||||
context_attn_hidden_states, torch.tanh(enc_gate_msa).unsqueeze(1)
|
||||
)
|
||||
norm_encoder_hidden_states = self.norm3_context(
|
||||
encoder_hidden_states, (1 + enc_scale_mlp.unsqueeze(1).to(torch.float32))
|
||||
)
|
||||
context_ff_output = self.ff_context(norm_encoder_hidden_states)
|
||||
encoder_hidden_states = encoder_hidden_states + self.norm4_context(
|
||||
context_ff_output, torch.tanh(enc_gate_mlp).unsqueeze(1)
|
||||
)
|
||||
|
||||
hidden_states[2:] = hidden_states_ptb
|
||||
encoder_hidden_states[2:] = encoder_hidden_states_ptb
|
||||
|
||||
return hidden_states, encoder_hidden_states
|
||||
|
||||
|
||||
# from: https://github.com/genmoai/models/blob/075b6e36db58f1242921deff83a1066887b9c9e1/src/mochi_preview/infer.py#L77
|
||||
def linear_quadratic_schedule(num_steps, threshold_noise, linear_steps=None):
|
||||
if linear_steps is None:
|
||||
linear_steps = num_steps // 2
|
||||
linear_sigma_schedule = [i * threshold_noise / linear_steps for i in range(linear_steps)]
|
||||
threshold_noise_step_diff = linear_steps - threshold_noise * num_steps
|
||||
quadratic_steps = num_steps - linear_steps
|
||||
quadratic_coef = threshold_noise_step_diff / (linear_steps * quadratic_steps**2)
|
||||
linear_coef = threshold_noise / linear_steps - 2 * threshold_noise_step_diff / (quadratic_steps**2)
|
||||
const = quadratic_coef * (linear_steps**2)
|
||||
quadratic_sigma_schedule = [
|
||||
quadratic_coef * (i**2) + linear_coef * i + const for i in range(linear_steps, num_steps)
|
||||
]
|
||||
sigma_schedule = linear_sigma_schedule + quadratic_sigma_schedule
|
||||
sigma_schedule = [1.0 - x for x in sigma_schedule]
|
||||
return sigma_schedule
|
||||
|
||||
|
||||
# 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 value")
|
||||
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
|
||||
|
||||
|
||||
class MochiSTGPipeline(DiffusionPipeline, Mochi1LoraLoaderMixin):
|
||||
r"""
|
||||
The mochi pipeline for text-to-video generation.
|
||||
|
||||
Reference: https://github.com/genmoai/models
|
||||
|
||||
Args:
|
||||
transformer ([`MochiTransformer3DModel`]):
|
||||
Conditional Transformer architecture to denoise the encoded video latents.
|
||||
scheduler ([`FlowMatchEulerDiscreteScheduler`]):
|
||||
A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
|
||||
vae ([`AutoencoderKLMochi`]):
|
||||
Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
|
||||
text_encoder ([`T5EncoderModel`]):
|
||||
[T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
|
||||
the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant.
|
||||
tokenizer (`CLIPTokenizer`):
|
||||
Tokenizer of class
|
||||
[CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
|
||||
tokenizer (`T5TokenizerFast`):
|
||||
Second Tokenizer of class
|
||||
[T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast).
|
||||
"""
|
||||
|
||||
model_cpu_offload_seq = "text_encoder->transformer->vae"
|
||||
_optional_components = []
|
||||
_callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
scheduler: FlowMatchEulerDiscreteScheduler,
|
||||
vae: AutoencoderKLMochi,
|
||||
text_encoder: T5EncoderModel,
|
||||
tokenizer: T5TokenizerFast,
|
||||
transformer: MochiTransformer3DModel,
|
||||
force_zeros_for_empty_prompt: bool = False,
|
||||
):
|
||||
super().__init__()
|
||||
|
||||
self.register_modules(
|
||||
vae=vae,
|
||||
text_encoder=text_encoder,
|
||||
tokenizer=tokenizer,
|
||||
transformer=transformer,
|
||||
scheduler=scheduler,
|
||||
)
|
||||
# TODO: determine these scaling factors from model parameters
|
||||
self.vae_spatial_scale_factor = 8
|
||||
self.vae_temporal_scale_factor = 6
|
||||
self.patch_size = 2
|
||||
|
||||
self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_scale_factor)
|
||||
self.tokenizer_max_length = (
|
||||
self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 256
|
||||
)
|
||||
self.default_height = 480
|
||||
self.default_width = 848
|
||||
self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
|
||||
|
||||
def _get_t5_prompt_embeds(
|
||||
self,
|
||||
prompt: Union[str, List[str]] = None,
|
||||
num_videos_per_prompt: int = 1,
|
||||
max_sequence_length: int = 256,
|
||||
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
|
||||
prompt_attention_mask = text_inputs.attention_mask
|
||||
prompt_attention_mask = prompt_attention_mask.bool().to(device)
|
||||
|
||||
# The original Mochi implementation zeros out empty negative prompts
|
||||
# but this can lead to overflow when placing the entire pipeline under the autocast context
|
||||
# adding this here so that we can enable zeroing prompts if necessary
|
||||
if self.config.force_zeros_for_empty_prompt and (prompt == "" or prompt[-1] == ""):
|
||||
text_input_ids = torch.zeros_like(text_input_ids, device=device)
|
||||
prompt_attention_mask = torch.zeros_like(prompt_attention_mask, dtype=torch.bool, device=device)
|
||||
|
||||
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), attention_mask=prompt_attention_mask)[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)
|
||||
|
||||
prompt_attention_mask = prompt_attention_mask.view(batch_size, -1)
|
||||
prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1)
|
||||
|
||||
return prompt_embeds, prompt_attention_mask
|
||||
|
||||
# Adapted 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,
|
||||
prompt_attention_mask: Optional[torch.Tensor] = None,
|
||||
negative_prompt_attention_mask: Optional[torch.Tensor] = None,
|
||||
max_sequence_length: int = 256,
|
||||
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, prompt_attention_mask = 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, negative_prompt_attention_mask = 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, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
|
||||
|
||||
def check_inputs(
|
||||
self,
|
||||
prompt,
|
||||
height,
|
||||
width,
|
||||
callback_on_step_end_tensor_inputs=None,
|
||||
prompt_embeds=None,
|
||||
negative_prompt_embeds=None,
|
||||
prompt_attention_mask=None,
|
||||
negative_prompt_attention_mask=None,
|
||||
):
|
||||
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_embeds is not None and prompt_attention_mask is None:
|
||||
raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.")
|
||||
|
||||
if negative_prompt_embeds is not None and negative_prompt_attention_mask is None:
|
||||
raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.")
|
||||
|
||||
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}."
|
||||
)
|
||||
if prompt_attention_mask.shape != negative_prompt_attention_mask.shape:
|
||||
raise ValueError(
|
||||
"`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but"
|
||||
f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`"
|
||||
f" {negative_prompt_attention_mask.shape}."
|
||||
)
|
||||
|
||||
def enable_vae_slicing(self):
|
||||
r"""
|
||||
Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
|
||||
compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
|
||||
"""
|
||||
self.vae.enable_slicing()
|
||||
|
||||
def disable_vae_slicing(self):
|
||||
r"""
|
||||
Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
|
||||
computing decoding in one step.
|
||||
"""
|
||||
self.vae.disable_slicing()
|
||||
|
||||
def enable_vae_tiling(self):
|
||||
r"""
|
||||
Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
|
||||
compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
|
||||
processing larger images.
|
||||
"""
|
||||
self.vae.enable_tiling()
|
||||
|
||||
def disable_vae_tiling(self):
|
||||
r"""
|
||||
Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
|
||||
computing decoding in one step.
|
||||
"""
|
||||
self.vae.disable_tiling()
|
||||
|
||||
def prepare_latents(
|
||||
self,
|
||||
batch_size,
|
||||
num_channels_latents,
|
||||
height,
|
||||
width,
|
||||
num_frames,
|
||||
dtype,
|
||||
device,
|
||||
generator,
|
||||
latents=None,
|
||||
):
|
||||
height = height // self.vae_spatial_scale_factor
|
||||
width = width // self.vae_spatial_scale_factor
|
||||
num_frames = (num_frames - 1) // self.vae_temporal_scale_factor + 1
|
||||
|
||||
shape = (batch_size, num_channels_latents, num_frames, height, width)
|
||||
|
||||
if latents is not None:
|
||||
return latents.to(device=device, dtype=dtype)
|
||||
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."
|
||||
)
|
||||
|
||||
latents = randn_tensor(shape, generator=generator, device=device, dtype=torch.float32)
|
||||
latents = latents.to(dtype)
|
||||
return latents
|
||||
|
||||
@property
|
||||
def guidance_scale(self):
|
||||
return self._guidance_scale
|
||||
|
||||
@property
|
||||
def do_classifier_free_guidance(self):
|
||||
return self._guidance_scale > 1.0
|
||||
|
||||
@property
|
||||
def do_spatio_temporal_guidance(self):
|
||||
return self._stg_scale > 0.0
|
||||
|
||||
@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,
|
||||
prompt: Union[str, List[str]] = None,
|
||||
negative_prompt: Optional[Union[str, List[str]]] = None,
|
||||
height: Optional[int] = None,
|
||||
width: Optional[int] = None,
|
||||
num_frames: int = 19,
|
||||
num_inference_steps: int = 64,
|
||||
timesteps: List[int] = None,
|
||||
guidance_scale: float = 4.5,
|
||||
num_videos_per_prompt: Optional[int] = 1,
|
||||
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
||||
latents: Optional[torch.Tensor] = None,
|
||||
prompt_embeds: Optional[torch.Tensor] = None,
|
||||
prompt_attention_mask: Optional[torch.Tensor] = None,
|
||||
negative_prompt_embeds: Optional[torch.Tensor] = None,
|
||||
negative_prompt_attention_mask: Optional[torch.Tensor] = None,
|
||||
output_type: Optional[str] = "pil",
|
||||
return_dict: bool = True,
|
||||
attention_kwargs: Optional[Dict[str, Any]] = None,
|
||||
callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
|
||||
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
|
||||
max_sequence_length: int = 256,
|
||||
stg_applied_layers_idx: Optional[List[int]] = [34],
|
||||
stg_scale: Optional[float] = 0.0,
|
||||
do_rescaling: Optional[bool] = False,
|
||||
):
|
||||
r"""
|
||||
Function invoked when calling the pipeline for generation.
|
||||
|
||||
Args:
|
||||
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.
|
||||
height (`int`, *optional*, defaults to `self.default_height`):
|
||||
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.default_width`):
|
||||
The width in pixels of the generated image. This is set to 848 by default for the best results.
|
||||
num_frames (`int`, defaults to `19`):
|
||||
The number of video frames to generate
|
||||
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`, defaults to `4.5`):
|
||||
Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
|
||||
`guidance_scale` is defined as `w` of equation 2. of [Imagen
|
||||
Paper](https://arxiv.org/pdf/2205.11487.pdf). 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.Tensor`, *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.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.
|
||||
prompt_attention_mask (`torch.Tensor`, *optional*):
|
||||
Pre-generated attention mask for text embeddings.
|
||||
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
|
||||
Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not
|
||||
provided, negative_prompt_embeds will be generated from `negative_prompt` input argument.
|
||||
negative_prompt_attention_mask (`torch.FloatTensor`, *optional*):
|
||||
Pre-generated attention mask for negative text embeddings.
|
||||
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.mochi.MochiPipelineOutput`] 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 `256`):
|
||||
Maximum sequence length to use with the `prompt`.
|
||||
|
||||
Examples:
|
||||
|
||||
Returns:
|
||||
[`~pipelines.mochi.MochiPipelineOutput`] or `tuple`:
|
||||
If `return_dict` is `True`, [`~pipelines.mochi.MochiPipelineOutput`] is returned, otherwise a `tuple`
|
||||
is returned where 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.default_height
|
||||
width = width or self.default_width
|
||||
|
||||
# 1. Check inputs. Raise error if not correct
|
||||
self.check_inputs(
|
||||
prompt=prompt,
|
||||
height=height,
|
||||
width=width,
|
||||
callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
|
||||
prompt_embeds=prompt_embeds,
|
||||
negative_prompt_embeds=negative_prompt_embeds,
|
||||
prompt_attention_mask=prompt_attention_mask,
|
||||
negative_prompt_attention_mask=negative_prompt_attention_mask,
|
||||
)
|
||||
|
||||
self._guidance_scale = guidance_scale
|
||||
self._stg_scale = stg_scale
|
||||
self._attention_kwargs = attention_kwargs
|
||||
self._current_timestep = None
|
||||
self._interrupt = False
|
||||
|
||||
if self.do_spatio_temporal_guidance:
|
||||
for i in stg_applied_layers_idx:
|
||||
self.transformer.transformer_blocks[i].forward = types.MethodType(
|
||||
forward_with_stg, self.transformer.transformer_blocks[i]
|
||||
)
|
||||
|
||||
# 2. Define 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
|
||||
# 3. Prepare text embeddings
|
||||
(
|
||||
prompt_embeds,
|
||||
prompt_attention_mask,
|
||||
negative_prompt_embeds,
|
||||
negative_prompt_attention_mask,
|
||||
) = self.encode_prompt(
|
||||
prompt=prompt,
|
||||
negative_prompt=negative_prompt,
|
||||
do_classifier_free_guidance=self.do_classifier_free_guidance,
|
||||
num_videos_per_prompt=num_videos_per_prompt,
|
||||
prompt_embeds=prompt_embeds,
|
||||
negative_prompt_embeds=negative_prompt_embeds,
|
||||
prompt_attention_mask=prompt_attention_mask,
|
||||
negative_prompt_attention_mask=negative_prompt_attention_mask,
|
||||
max_sequence_length=max_sequence_length,
|
||||
device=device,
|
||||
)
|
||||
# 4. Prepare latent variables
|
||||
num_channels_latents = self.transformer.config.in_channels
|
||||
latents = self.prepare_latents(
|
||||
batch_size * num_videos_per_prompt,
|
||||
num_channels_latents,
|
||||
height,
|
||||
width,
|
||||
num_frames,
|
||||
prompt_embeds.dtype,
|
||||
device,
|
||||
generator,
|
||||
latents,
|
||||
)
|
||||
|
||||
if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
|
||||
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
|
||||
prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0)
|
||||
elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
|
||||
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds, prompt_embeds], dim=0)
|
||||
prompt_attention_mask = torch.cat(
|
||||
[negative_prompt_attention_mask, prompt_attention_mask, prompt_attention_mask], dim=0
|
||||
)
|
||||
|
||||
# 5. Prepare timestep
|
||||
# from https://github.com/genmoai/models/blob/075b6e36db58f1242921deff83a1066887b9c9e1/src/mochi_preview/infer.py#L77
|
||||
threshold_noise = 0.025
|
||||
sigmas = linear_quadratic_schedule(num_inference_steps, threshold_noise)
|
||||
sigmas = np.array(sigmas)
|
||||
|
||||
timesteps, num_inference_steps = retrieve_timesteps(
|
||||
self.scheduler,
|
||||
num_inference_steps,
|
||||
device,
|
||||
timesteps,
|
||||
sigmas,
|
||||
)
|
||||
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
|
||||
self._num_timesteps = len(timesteps)
|
||||
|
||||
# 6. Denoising loop
|
||||
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
||||
for i, t in enumerate(timesteps):
|
||||
if self.interrupt:
|
||||
continue
|
||||
|
||||
# Note: Mochi uses reversed timesteps. To ensure compatibility with methods like FasterCache, we need
|
||||
# to make sure we're using the correct non-reversed timestep value.
|
||||
self._current_timestep = 1000 - t
|
||||
if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
|
||||
latent_model_input = torch.cat([latents] * 2)
|
||||
elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
|
||||
latent_model_input = torch.cat([latents] * 3)
|
||||
else:
|
||||
latent_model_input = latents
|
||||
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
||||
timestep = t.expand(latent_model_input.shape[0]).to(latents.dtype)
|
||||
|
||||
noise_pred = self.transformer(
|
||||
hidden_states=latent_model_input,
|
||||
encoder_hidden_states=prompt_embeds,
|
||||
timestep=timestep,
|
||||
encoder_attention_mask=prompt_attention_mask,
|
||||
attention_kwargs=attention_kwargs,
|
||||
return_dict=False,
|
||||
)[0]
|
||||
# Mochi CFG + Sampling runs in FP32
|
||||
noise_pred = noise_pred.to(torch.float32)
|
||||
|
||||
if self.do_classifier_free_guidance and not self.do_spatio_temporal_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)
|
||||
elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
|
||||
noise_pred_uncond, noise_pred_text, noise_pred_perturb = noise_pred.chunk(3)
|
||||
noise_pred = (
|
||||
noise_pred_uncond
|
||||
+ self.guidance_scale * (noise_pred_text - noise_pred_uncond)
|
||||
+ self._stg_scale * (noise_pred_text - noise_pred_perturb)
|
||||
)
|
||||
|
||||
if do_rescaling:
|
||||
rescaling_scale = 0.7
|
||||
factor = noise_pred_text.std() / noise_pred.std()
|
||||
factor = rescaling_scale * factor + (1 - rescaling_scale)
|
||||
noise_pred = noise_pred * factor
|
||||
|
||||
# compute the previous noisy sample x_t -> x_t-1
|
||||
latents_dtype = latents.dtype
|
||||
latents = self.scheduler.step(noise_pred, t, latents.to(torch.float32), return_dict=False)[0]
|
||||
latents = latents.to(latents_dtype)
|
||||
|
||||
if latents.dtype != latents_dtype:
|
||||
if torch.backends.mps.is_available():
|
||||
# some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
|
||||
latents = latents.to(latents_dtype)
|
||||
|
||||
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)
|
||||
|
||||
# call the callback, if provided
|
||||
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
|
||||
|
||||
if output_type == "latent":
|
||||
video = latents
|
||||
else:
|
||||
# unscale/denormalize the latents
|
||||
# denormalize with the mean and std if available and not None
|
||||
has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
|
||||
has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
|
||||
if has_latents_mean and has_latents_std:
|
||||
latents_mean = (
|
||||
torch.tensor(self.vae.config.latents_mean).view(1, 12, 1, 1, 1).to(latents.device, latents.dtype)
|
||||
)
|
||||
latents_std = (
|
||||
torch.tensor(self.vae.config.latents_std).view(1, 12, 1, 1, 1).to(latents.device, latents.dtype)
|
||||
)
|
||||
latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
|
||||
else:
|
||||
latents = latents / self.vae.config.scaling_factor
|
||||
|
||||
video = self.vae.decode(latents, return_dict=False)[0]
|
||||
video = self.video_processor.postprocess_video(video, output_type=output_type)
|
||||
|
||||
# Offload all models
|
||||
self.maybe_free_model_hooks()
|
||||
|
||||
if not return_dict:
|
||||
return (video,)
|
||||
|
||||
return MochiPipelineOutput(frames=video)
|
||||
Reference in New Issue
Block a user