import torch
from diffusers import DiffusionPipeline
from diffusers.utils import load_image, export_to_video
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("PeppX/video-pid", dtype=torch.bfloat16, device_map="cuda")
pipe.to("cuda")
prompt = "A man with short gray hair plays a red electric guitar."
image = load_image(
"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/guitar-man.png"
)
output = pipe(image=image, prompt=prompt).frames[0]
export_to_video(output, "output.mp4")Video-PiD: Pixel-Space Decoder for Wan 2.1
A small 3D pixel-space diffusion model that runs on top of Wan 2.1's VAE-decoded video frames to fix the "plastic" / waxy look of latent diffusion decoders.
Status: pre-alpha. No weights yet. This repo will hold the trained checkpoints. The source code lives at github.com/madxmoron/video-pid.
Wan 2.1 1.3B T2V Video-PiD
┌──────────┐ Wan-VAE ┌──────────┐ residual ┌──────────┐
│ text │──▶ decode ─────▶│ pixel │──▶ denoise ──▶│ pixel │──▶ video
│ latent │ (plastic) │ frames │ (4 steps) │ frames │ (sharp)
└──────────┘ └──────────┘ └──────────┘
Why
Latent diffusion decoders (Wan-VAE, SD-VAE, etc.) throw away high-frequency detail and re-introduce a "waxy" smoothness. Video-PiD is a tiny post-pass that re-denoises the decoded frames in pixel space, conditioning on the original latent, and outputs a residual that adds back the detail.
Inspired by NVIDIA's PiD (image-only). We extend it to video, in 3D, as a plug-in for Wan 2.1.
Roadmap
- Architecture spec pinned (in progress)
- Training run on Panda-70M / HD-VGGT
- First checkpoint release (v0.1.0)
- ComfyUI node
See github.com/madxmoron/video-pid/blob/main/docs/ROADMAP.md for the full plan.
License
Apache 2.0.
Citation
@software{video_pid_2026,
author = {madxmoron},
title = {Video-PiD: Pixel-Space Decoder for Wan 2.1},
year = {2026},
url = {https://github.com/madxmoron/video-pid}
}
- Downloads last month
- -