Add CogVideoX text-to-video generation model

[zRzRzRzRzRzRzR]

What does this PR do?

This PR converts the CogVideoX model into a diffuser-supported inference model, including a complete pipeline and corresponding modules. The paper is still in the process of being written, which may result in temporary omissions regarding the paper in the documentation.

Who can review?

@yiyixuxu
@stevhliu and @sayakpaul

[sayakpaul]

We can also include a tip section like we have in Flux:
https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux

This way users are aware of the optimizations that are possible.

[sayakpaul]

We should probably also mention that users can benefit from context-parallel caching.

[sayakpaul]

Curiosity: why is this fake?

[zRzRzRzRzRzRzR]

Because this is executed serially on a single GPU.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[a-r-r-o-w]

@zRzRzRzRzRzRzR Just to notify, I have removed the clear_fake_cp_cache parameter and switched to something I thought was more clean and understandable. It should be consistent with the old implementation AFAICT. Now trying to debug why we use double the memory than the SAT implementation.

[a-r-r-o-w]

@zRzRzRzRzRzRzR Is there any reason why we're moving the latents to CPU here but then moving them back to inputs.device in fake_cp_pass_from_previous_rank? Since this implementation is tailored to single GPU, and since I don't think the latents will reside on multiple devices for our implementation, we can get rid of the .contiguous().detach().clone().cpu() part and just keep them on same device. cc @yiyixuxu

Edit: Tried it and it looks like .contiguous().clone() will be required but don't think we would need to move to CPU here for single GPU case

[yiyixuxu]

let's merge this soon!

[yiyixuxu]

@a-r-r-o-w we normally just adjust the weights (ok to keep this! I don't think we need to update the weights now just FYI ) https://github.com/huggingface/diffusers/blob/main/scripts/convert_sd3_to_diffusers.py#L37

[stevhliu]

Great work, thanks!

[stevhliu]

Do we need to include this code block to demonstrate torch.compile, or is it to show inference time without torch.compile? If it's not necessary, I'm more in favor of just showing the below to keep it simpler.

# create pipeline
pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-2b", torch_dtype=torch.bfloat16).to("cuda")

# set to channels_last
pipeline.transformer.to(memory_format=torch.channels_last)
pipeline.vae.to(memory_format=torch.channels_last)

# compile
pipeline.transformer = torch.compile(pipeline.transformer)
pipeline.vae.decode = torch.compile(pipeline.vae.decode)

# inference
prompt = "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical atmosphere of this unique musical performance."
video = pipeline(prompt=prompt, guidance_scale=6, num_inference_steps=50).frames[0]
export_to_video(video, "output.mp4", fps=8)

[yiyixuxu]

i made a PR to update scheduler config on the hub, https://huggingface.co/THUDM/CogVideoX-2b/discussions/2 we can merge that after this PR is merged here

you can test this pr with revision="refs/pr/2, here is the script I used to run tests on both scehdulers with and without dynamic scheduler

from diffusers.utils import export_to_video
import torch
import numpy as np
import PIL

import tempfile
import imageio

def export_to_video_imageio(video_frames, output_video_path: str = None, fps: int = 8):
    """
    Export the video frames to a video file using imageio lib to Avoid "green screen" issue (for example CogVideoX)
    """
    if output_video_path is None:
        output_video_path = tempfile.NamedTemporaryFile(suffix=".mp4").name

    if isinstance(video_frames[0], PIL.Image.Image):
        video_frames = [np.array(frame) for frame in video_frames]

    with imageio.get_writer(output_video_path, fps=fps) as writer:
        for frame in video_frames:
            writer.append_data(frame)

    return output_video_path



prompts = [
    "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical atmosphere of this unique musical performance.",
    "A detailed wooden toy ship with intricately carved masts and sails is seen gliding smoothly over a plush, blue carpet that mimics the waves of the sea. The ship's hull is painted a rich brown, with tiny windows. The carpet, soft and textured, provides a perfect backdrop, resembling an oceanic expanse. Surrounding the ship are various other toys and children's items, hinting at a playful environment. The scene captures the innocence and imagination of childhood, with the toy ship's journey symbolizing endless adventures in a whimsical, indoor setting.",
    "The camera follows behind a white vintage SUV with a black roof rack as it speeds up a steep dirt road surrounded by pine trees on a steep mountain slope, dust kicks up from it’s tires, the sunlight shines on the SUV as it speeds along the dirt road, casting a warm glow over the scene. The dirt road curves gently into the distance, with no other cars or vehicles in sight. The trees on either side of the road are redwoods, with patches of greenery scattered throughout. The car is seen from the rear following the curve with ease, making it seem as if it is on a rugged drive through the rugged terrain. The dirt road itself is surrounded by steep hills and mountains, with a clear blue sky above with wispy clouds.",
    "A street artist, clad in a worn-out denim jacket and a colorful bandana, stands before a vast concrete wall in the heart, holding a can of spray paint, spray-painting a colorful bird on a mottled wall.",
    "In the haunting backdrop of a war-torn city, where ruins and crumbled walls tell a story of devastation, a poignant close-up frames a young girl. Her face is smudged with ash, a silent testament to the chaos around her. Her eyes glistening with a mix of sorrow and resilience, capturing the raw emotion of a world that has lost its innocence to the ravages of conflict."
    ]


pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-2b", torch_dtype=torch.float16, revision="refs/pr/2")
pipe.enable_model_cpu_offload()

for prompt in prompts:
        for seed in [3]:
                # test ddim
                pipe.scheduler = CogVideoXDDIMScheduler.from_config(pipe.scheduler.config)
                assert pipe.scheduler.config._class_name == "CogVideoXDDIMScheduler" and pipe.scheduler.config.timestep_spacing=="trailing"

                generator= torch.Generator(device="cpu").manual_seed(seed)
                video = pipe(prompt, guidance_scale=6, num_inference_steps=50, generator=generator).frames[0]
                export_to_video_imageio(video, f"{prompt[:10]}_{seed}_ddim.mp4", fps=8)


                assert pipe.scheduler.config._class_name == "CogVideoXDDIMScheduler" and pipe.scheduler.config.timestep_spacing=="trailing"
                generator= torch.Generator(device="cpu").manual_seed(seed)
                video = pipe(prompt, guidance_scale=6, num_inference_steps=50, generator=generator, use_dynamic_cfg=True).frames[0]
                export_to_video_imageio(video, f"{prompt[:10]}_{seed}_ddim_dynamic_cfg.mp4", fps=8)

                # test dpm
                pipe.scheduler = CogVideoXDPMScheduler.from_config(pipe.scheduler.config)
                assert pipe.scheduler.config._class_name == "CogVideoXDPMScheduler" and pipe.scheduler.config.timestep_spacing=="trailing"

                generator= torch.Generator(device="cpu").manual_seed(seed)
                video = pipe(prompt, guidance_scale=6, num_inference_steps=50, generator=generator).frames[0]
                export_to_video_imageio(video, f"{prompt[:10]}_{seed}_dpm.mp4", fps=8)
                
                
                assert pipe.scheduler.config._class_name == "CogVideoXDPMScheduler" and pipe.scheduler.config.timestep_spacing=="trailing"
                generator= torch.Generator(device="cpu").manual_seed(seed)
                video = pipe(prompt, guidance_scale=6, num_inference_steps=50, generator=generator, use_dynamic_cfg=True).frames[0]
                export_to_video_imageio(video, f"{prompt[:10]}_{seed}_dpm_dynamic_cfg.mp4", fps=8)

[sayakpaul]

This should have ideally gone to check_inputs().

[a-r-r-o-w]

Ah cool, then we should do it here too:

diffusers/src/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3.py

Line 454 in 2dad462

if prompt is not None and type(prompt) is not type(negative_prompt):

as it was copied over from there

[sayakpaul]

Should have been raised as a ValueError.

[sayakpaul]

It wouldn't hurt to have a comment explaining why this addition is required.