This pipeline is for research purposes only.
VideoFusion: Decomposed Diffusion Models for High-Quality Video Generation by Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, Tieniu Tan.
The abstract of the paper is the following:
A diffusion probabilistic model (DPM), which constructs a forward diffusion process by gradually adding noise to data points and learns the reverse denoising process to generate new samples, has been shown to handle complex data distribution. Despite its recent success in image synthesis, applying DPMs to video generation is still challenging due to high-dimensional data spaces. Previous methods usually adopt a standard diffusion process, where frames in the same video clip are destroyed with independent noises, ignoring the content redundancy and temporal correlation. This work presents a decomposed diffusion process via resolving the per-frame noise into a base noise that is shared among all frames and a residual noise that varies along the time axis. The denoising pipeline employs two jointly-learned networks to match the noise decomposition accordingly. Experiments on various datasets confirm that our approach, termed as VideoFusion, surpasses both GAN-based and diffusion-based alternatives in high-quality video generation. We further show that our decomposed formulation can benefit from pre-trained image diffusion models and well-support text-conditioned video creation.
Resources:
Pipeline | Tasks | Demo |
---|---|---|
TextToVideoSDPipeline | Text-to-Video Generation | 🤗 Spaces |
VideoToVideoSDPipeline | Text-Guided Video-to-Video Generation | (TODO)🤗 Spaces |
text-to-video-ms-1.7b
Let’s start by generating a short video with the default length of 16 frames (2s at 8 fps):
import torch
from diffusers import DiffusionPipeline
from diffusers.utils import export_to_video
pipe = DiffusionPipeline.from_pretrained("damo-vilab/text-to-video-ms-1.7b", torch_dtype=torch.float16, variant="fp16")
pipe = pipe.to("cuda")
prompt = "Spiderman is surfing"
video_frames = pipe(prompt).frames
video_path = export_to_video(video_frames)
video_path
Diffusers supports different optimization techniques to improve the latency and memory footprint of a pipeline. Since videos are often more memory-heavy than images, we can enable CPU offloading and VAE slicing to keep the memory footprint at bay.
Let’s generate a video of 8 seconds (64 frames) on the same GPU using CPU offloading and VAE slicing:
import torch
from diffusers import DiffusionPipeline
from diffusers.utils import export_to_video
pipe = DiffusionPipeline.from_pretrained("damo-vilab/text-to-video-ms-1.7b", torch_dtype=torch.float16, variant="fp16")
pipe.enable_model_cpu_offload()
# memory optimization
pipe.enable_vae_slicing()
prompt = "Darth Vader surfing a wave"
video_frames = pipe(prompt, num_frames=64).frames
video_path = export_to_video(video_frames)
video_path
It just takes 7 GBs of GPU memory to generate the 64 video frames using PyTorch 2.0, “fp16” precision and the techniques mentioned above.
We can also use a different scheduler easily, using the same method we’d use for Stable Diffusion:
import torch
from diffusers import DiffusionPipeline, DPMSolverMultistepScheduler
from diffusers.utils import export_to_video
pipe = DiffusionPipeline.from_pretrained("damo-vilab/text-to-video-ms-1.7b", torch_dtype=torch.float16, variant="fp16")
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
pipe.enable_model_cpu_offload()
prompt = "Spiderman is surfing"
video_frames = pipe(prompt, num_inference_steps=25).frames
video_path = export_to_video(video_frames)
video_path
Here are some sample outputs:
![]() |
![]() |
cerspense/zeroscope_v2_576w
& cerspense/zeroscope_v2_XL
Zeroscope are watermark-free model and have been trained on specific sizes such as 576x320
and 1024x576
.
One should first generate a video using the lower resolution checkpoint cerspense/zeroscope_v2_576w
with TextToVideoSDPipeline,
which can then be upscaled using VideoToVideoSDPipeline and cerspense/zeroscope_v2_XL
.
import torch
from diffusers import DiffusionPipeline
from diffusers.utils import export_to_video
pipe = DiffusionPipeline.from_pretrained("cerspense/zeroscope_v2_576w", torch_dtype=torch.float16)
pipe.enable_model_cpu_offload()
# memory optimization
pipe.unet.enable_forward_chunking(chunk_size=1, dim=1)
pipe.enable_vae_slicing()
prompt = "Darth Vader surfing a wave"
video_frames = pipe(prompt, num_frames=24).frames
video_path = export_to_video(video_frames)
video_path
Now the video can be upscaled:
pipe = DiffusionPipeline.from_pretrained("cerspense/zeroscope_v2_XL", torch_dtype=torch.float16)
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
pipe.enable_model_cpu_offload()
# memory optimization
pipe.unet.enable_forward_chunking(chunk_size=1, dim=1)
pipe.enable_vae_slicing()
video = [Image.fromarray(frame).resize((1024, 576)) for frame in video_frames]
video_frames = pipe(prompt, video=video, strength=0.6).frames
video_path = export_to_video(video_frames)
video_path
Here are some sample outputs:
![]() |
Text-guided video generation with ~TextToVideoSDPipeline and ~VideoToVideoSDPipeline is very memory intensive both when denoising with ~UNet3DConditionModel and when decoding with ~AutoencoderKL. It is possible though to reduce memory usage at the cost of increased runtime to achieve the exact same result. To do so, it is recommended to enable forward chunking and vae slicing:
Forward chunking via enable_forward_chunking()is explained in this blog post and
allows to significantly reduce the required memory for the unet. You can chunk the feed forward layer over the num_frames
dimension by doing:
pipe.unet.enable_forward_chunking(chunk_size=1, dim=1)
Vae slicing via enable_vae_slicing() and enable_vae_slicing() also gives significant memory savings since the two pipelines decode all image frames at once.
pipe.enable_vae_slicing()
( vae: AutoencoderKL text_encoder: CLIPTextModel tokenizer: CLIPTokenizer unet: UNet3DConditionModel scheduler: KarrasDiffusionSchedulers )
Parameters
CLIPTextModel
) —
Frozen text-encoder. Same as Stable Diffusion 2.
CLIPTokenizer
) —
Tokenizer of class
CLIPTokenizer.
unet
to denoise the encoded image latents. Can be one of
DDIMScheduler, LMSDiscreteScheduler, or PNDMScheduler.
Pipeline for text-to-video generation.
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.)
(
prompt: typing.Union[str, typing.List[str]] = None
height: typing.Optional[int] = None
width: typing.Optional[int] = None
num_frames: int = 16
num_inference_steps: int = 50
guidance_scale: float = 9.0
negative_prompt: typing.Union[str, typing.List[str], NoneType] = None
eta: float = 0.0
generator: typing.Union[torch._C.Generator, typing.List[torch._C.Generator], NoneType] = None
latents: typing.Optional[torch.FloatTensor] = None
prompt_embeds: typing.Optional[torch.FloatTensor] = None
negative_prompt_embeds: typing.Optional[torch.FloatTensor] = None
output_type: typing.Optional[str] = 'np'
return_dict: bool = True
callback: typing.Union[typing.Callable[[int, int, torch.FloatTensor], NoneType], NoneType] = None
callback_steps: int = 1
cross_attention_kwargs: typing.Union[typing.Dict[str, typing.Any], NoneType] = None
)
→
~pipelines.stable_diffusion.TextToVideoSDPipelineOutput
or tuple
Parameters
str
or List[str]
, optional) —
The prompt or prompts to guide the video generation. If not defined, one has to pass prompt_embeds
.
instead.
int
, optional, defaults to self.unet.config.sample_size * self.vae_scale_factor) —
The height in pixels of the generated video.
int
, optional, defaults to self.unet.config.sample_size * self.vae_scale_factor) —
The width in pixels of the generated video.
int
, optional, defaults to 16) —
The number of video frames that are generated. Defaults to 16 frames which at 8 frames per seconds
amounts to 2 seconds of video.
int
, optional, defaults to 50) —
The number of denoising steps. More denoising steps usually lead to a higher quality videos at the
expense of slower inference.
float
, optional, defaults to 7.5) —
Guidance scale as defined in Classifier-Free Diffusion Guidance.
guidance_scale
is defined as w
of equation 2. of Imagen
Paper. Guidance scale is enabled by setting guidance_scale > 1
. Higher guidance scale encourages to generate videos that are closely linked to the text prompt
,
usually at the expense of lower video quality.
str
or List[str]
, optional) —
The prompt or prompts not to guide the video 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
).
float
, optional, defaults to 0.0) —
Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
schedulers.DDIMScheduler, will be ignored for others.
torch.Generator
or List[torch.Generator]
, optional) —
One or a list of torch generator(s)
to make generation deterministic.
torch.FloatTensor
, optional) —
Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for video
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
. Latents should be of shape
(batch_size, num_channel, num_frames, height, width)
.
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.
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.
str
, optional, defaults to "np"
) —
The output format of the generate video. Choose between torch.FloatTensor
or np.array
.
bool
, optional, defaults to True
) —
Whether or not to return a ~pipelines.stable_diffusion.TextToVideoSDPipelineOutput
instead of a
plain tuple.
Callable
, optional) —
A function that will be called every callback_steps
steps during inference. The function will be
called with the following arguments: callback(step: int, timestep: int, latents: torch.FloatTensor)
.
int
, optional, defaults to 1) —
The frequency at which the callback
function will be called. If not specified, the callback will be
called at every step.
dict
, optional) —
A kwargs dictionary that if specified is passed along to the AttentionProcessor
as defined under
self.processor
in
diffusers.cross_attention.
Returns
~pipelines.stable_diffusion.TextToVideoSDPipelineOutput
or tuple
~pipelines.stable_diffusion.TextToVideoSDPipelineOutput
if return_dict
is True, otherwise a `tuple.
When returning a tuple, the first element is a list with the generated frames.
Function invoked when calling the pipeline for generation.
Examples:
>>> import torch
>>> from diffusers import TextToVideoSDPipeline
>>> from diffusers.utils import export_to_video
>>> pipe = TextToVideoSDPipeline.from_pretrained(
... "damo-vilab/text-to-video-ms-1.7b", torch_dtype=torch.float16, variant="fp16"
... )
>>> pipe.enable_model_cpu_offload()
>>> prompt = "Spiderman is surfing"
>>> video_frames = pipe(prompt).frames
>>> video_path = export_to_video(video_frames)
>>> video_path
Disable sliced VAE decoding. If enable_vae_slicing
was previously invoked, this method will go back to
computing decoding in one step.
Disable tiled VAE decoding. If enable_vae_tiling
was previously invoked, this method will go back to
computing decoding in one step.
Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
to enable_sequential_cpu_offload
, this method moves one whole model at a time to the GPU when its forward
method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
enable_sequential_cpu_offload
, but performance is much better due to the iterative execution of the unet
.
Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
text_encoder, vae have their state dicts saved to CPU and then are moved to a torch.device('meta') and loaded to GPU only when their specific submodule has its
forwardmethod called. Note that offloading happens on a submodule basis. Memory savings are higher than with
enable_model_cpu_offload`, but performance is lower.
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.
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 to save a large amount of memory and to allow the processing of larger images.
( vae: AutoencoderKL text_encoder: CLIPTextModel tokenizer: CLIPTokenizer unet: UNet3DConditionModel scheduler: KarrasDiffusionSchedulers )
Parameters
CLIPTextModel
) —
Frozen text-encoder. Same as Stable Diffusion 2.
CLIPTokenizer
) —
Tokenizer of class
CLIPTokenizer.
unet
to denoise the encoded image latents. Can be one of
DDIMScheduler, LMSDiscreteScheduler, or PNDMScheduler.
Pipeline for text-to-video generation.
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.)
(
prompt: typing.Union[str, typing.List[str]] = None
video: typing.Union[typing.List[numpy.ndarray], torch.FloatTensor] = None
strength: float = 0.6
num_inference_steps: int = 50
guidance_scale: float = 15.0
negative_prompt: typing.Union[str, typing.List[str], NoneType] = None
eta: float = 0.0
generator: typing.Union[torch._C.Generator, typing.List[torch._C.Generator], NoneType] = None
latents: typing.Optional[torch.FloatTensor] = None
prompt_embeds: typing.Optional[torch.FloatTensor] = None
negative_prompt_embeds: typing.Optional[torch.FloatTensor] = None
output_type: typing.Optional[str] = 'np'
return_dict: bool = True
callback: typing.Union[typing.Callable[[int, int, torch.FloatTensor], NoneType], NoneType] = None
callback_steps: int = 1
cross_attention_kwargs: typing.Union[typing.Dict[str, typing.Any], NoneType] = None
)
→
~pipelines.stable_diffusion.TextToVideoSDPipelineOutput
or tuple
Parameters
str
or List[str]
, optional) —
The prompt or prompts to guide the video generation. If not defined, one has to pass prompt_embeds
.
instead.
video — (List[np.ndarray]
or torch.FloatTensor
):
video
frames or tensor representing a video batch, that will be used as the starting point for the
process. Can also accpet video latents as image
, if passing latents directly, it will not be encoded
again.
float
, optional, defaults to 0.8) —
Conceptually, indicates how much to transform the reference image
. Must be between 0 and 1. image
will be used as a starting point, adding more noise to it the larger the strength
. The number of
denoising steps depends on the amount of noise initially added. When strength
is 1, added noise will
be maximum and the denoising process will run for the full number of iterations specified in
num_inference_steps
. A value of 1, therefore, essentially ignores image
.
int
, optional, defaults to 50) —
The number of denoising steps. More denoising steps usually lead to a higher quality videos at the
expense of slower inference.
float
, optional, defaults to 7.5) —
Guidance scale as defined in Classifier-Free Diffusion Guidance.
guidance_scale
is defined as w
of equation 2. of Imagen
Paper. Guidance scale is enabled by setting guidance_scale > 1
. Higher guidance scale encourages to generate videos that are closely linked to the text prompt
,
usually at the expense of lower video quality.
str
or List[str]
, optional) —
The prompt or prompts not to guide the video 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
).
float
, optional, defaults to 0.0) —
Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
schedulers.DDIMScheduler, will be ignored for others.
torch.Generator
or List[torch.Generator]
, optional) —
One or a list of torch generator(s)
to make generation deterministic.
torch.FloatTensor
, optional) —
Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for video
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
. Latents should be of shape
(batch_size, num_channel, num_frames, height, width)
.
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.
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.
str
, optional, defaults to "np"
) —
The output format of the generate video. Choose between torch.FloatTensor
or np.array
.
bool
, optional, defaults to True
) —
Whether or not to return a ~pipelines.stable_diffusion.TextToVideoSDPipelineOutput
instead of a
plain tuple.
Callable
, optional) —
A function that will be called every callback_steps
steps during inference. The function will be
called with the following arguments: callback(step: int, timestep: int, latents: torch.FloatTensor)
.
int
, optional, defaults to 1) —
The frequency at which the callback
function will be called. If not specified, the callback will be
called at every step.
dict
, optional) —
A kwargs dictionary that if specified is passed along to the AttentionProcessor
as defined under
self.processor
in
diffusers.cross_attention.
Returns
~pipelines.stable_diffusion.TextToVideoSDPipelineOutput
or tuple
~pipelines.stable_diffusion.TextToVideoSDPipelineOutput
if return_dict
is True, otherwise a `tuple.
When returning a tuple, the first element is a list with the generated frames.
Function invoked when calling the pipeline for generation.
Examples:
>>> import torch
>>> from diffusers import DiffusionPipeline, DPMSolverMultistepScheduler
>>> from diffusers.utils import export_to_video
>>> pipe = DiffusionPipeline.from_pretrained("cerspense/zeroscope_v2_576w", torch_dtype=torch.float16)
>>> pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
>>> pipe.to("cuda")
>>> prompt = "spiderman running in the desert"
>>> video_frames = pipe(prompt, num_inference_steps=40, height=320, width=576, num_frames=24).frames
>>> # safe low-res video
>>> video_path = export_to_video(video_frames, output_video_path="./video_576_spiderman.mp4")
>>> # let's offload the text-to-image model
>>> pipe.to("cpu")
>>> # and load the image-to-image model
>>> pipe = DiffusionPipeline.from_pretrained(
... "cerspense/zeroscope_v2_XL", torch_dtype=torch.float16, revision="refs/pr/15"
... )
>>> pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
>>> pipe.enable_model_cpu_offload()
>>> # The VAE consumes A LOT of memory, let's make sure we run it in sliced mode
>>> pipe.vae.enable_slicing()
>>> # now let's upscale it
>>> video = [Image.fromarray(frame).resize((1024, 576)) for frame in video_frames]
>>> # and denoise it
>>> video_frames = pipe(prompt, video=video, strength=0.6).frames
>>> video_path = export_to_video(video_frames, output_video_path="./video_1024_spiderman.mp4")
>>> video_path
Disable sliced VAE decoding. If enable_vae_slicing
was previously invoked, this method will go back to
computing decoding in one step.
Disable tiled VAE decoding. If enable_vae_tiling
was previously invoked, this method will go back to
computing decoding in one step.
Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
to enable_sequential_cpu_offload
, this method moves one whole model at a time to the GPU when its forward
method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
enable_sequential_cpu_offload
, but performance is much better due to the iterative execution of the unet
.
Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
text_encoder, vae have their state dicts saved to CPU and then are moved to a torch.device('meta') and loaded to GPU only when their specific submodule has its
forwardmethod called. Note that offloading happens on a submodule basis. Memory savings are higher than with
enable_model_cpu_offload`, but performance is lower.
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.
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 to save a large amount of memory and to allow the processing of larger images.