[WIP][LTX Video2Video] start ltx video2video.

[sayakpaul]

What does this PR do?

A bit unsure about how to deal with conditioning_mask and whether to consider adding noise to initial_latents like how it's done here:

diffusers/src/diffusers/pipelines/cogvideo/pipeline_cogvideox_video2video.py

Line 386 in 8eb73c8

latents = self.scheduler.add_noise(init_latents, noise, timestep)

Code to test

import torch
from diffusers.pipelines.ltx.pipeline_ltx_video2video import LTXVideoToVideoPipeline
from diffusers.utils import export_to_video, load_video

pipe = LTXVideoToVideoPipeline.from_pretrained("Lightricks/LTX-Video", torch_dtype=torch.bfloat16)
pipe.enable_model_cpu_offload()

input_video = load_video(
    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/hiker.mp4"
)
prompt = (
    "An astronaut stands triumphantly at the peak of a towering mountain. Panorama of rugged peaks and "
    "valleys. Very futuristic vibe and animated aesthetic. Highlights of purple and golden colors in "
    "the scene. The sky is looks like an animated/cartoonish dream of galaxies, nebulae, stars, planets, "
    "moons, but the remainder of the scene is mostly realistic."
)

video = pipe(
    video=input_video, prompt=prompt, guidance_scale=6, num_inference_steps=50
).frames[0]
export_to_video(video, "output_vid2vid.mp4", fps=24)

Input video:

hiker.mp4

Output video:

output_vid2vid.mp4

@a-r-r-o-w I am going to mention you in a couple of places where I am unsure about the implementation detail. LMK.

[sayakpaul]

@a-r-r-o-w this implementation.

[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]

Hope this answer some of your questions! Happy to help with anything else 🤗

[a-r-r-o-w]

This was only need for image-to-video because we don't want to denoise the first frame (because it is the actual encoded image latent itself)

[a-r-r-o-w]

I don't see a strength parameter here. Without strength, it is not possible to control how much effect of the original video you would like to have in the output video. Low strength -> low denoising steps -> more similar video. And high strength -> more denoising steps -> less similar video. This is more of a naive approach that we use for vid-to-vid, but with techniques like RF-inversion and FlowEdit (which we'll directly add in modular diffusers instead of pipelines), the quality and possibilities are endless!

[sayakpaul]

@a-r-r-o-w thanks for your comments. Didn't know about scale_noise() :D

I have applied the changes you proposed and I am running a sweep over params. Results: https://wandb.ai/sayakpaul/ltx_video2video/runs/pc4lfivy

Sweep

import torch
from diffusers.pipelines.ltx.pipeline_ltx_video2video import LTXVideoToVideoPipeline
from diffusers.utils import export_to_video, load_video
import wandb

pipe = LTXVideoToVideoPipeline.from_pretrained("Lightricks/LTX-Video", torch_dtype=torch.bfloat16).to("cuda")

input_video = load_video(
    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/hiker.mp4"
)
prompt = (
    "An astronaut stands triumphantly at the peak of a towering mountain. Panorama of rugged peaks and "
    "valleys. Very futuristic vibe and animated aesthetic. Highlights of purple and golden colors in "
    "the scene. The sky is looks like an animated/cartoonish dream of galaxies, nebulae, stars, planets, "
    "moons, but the remainder of the scene is mostly realistic."
)
wandb.init(project="ltx_video2video")

filenames = []
for s in [1.0, 0.8, 0.7, 0.9]:
    for steps in [50, 60, 70]:
        for cfg in [5, 6, 7, 8]:
            video_name = f"strength@{s}_steps@{steps}_cfg@{cfg}.mp4"
            video = pipe(
                video=input_video, prompt=prompt, guidance_scale=cfg, num_inference_steps=steps, strength=s
            ).frames[0]
            export_to_video(video, video_name, fps=24)
            wandb.log(
                {"videos": wandb.Video(video_name, caption=video_name.replace(".mp4", ""), fps=24)}
            )

LMK if you have any concerns on the implementation.