PixVid Brings Video Diffusion Back to Pixel Space
The paper “Back to Pixels: Lifting Latent Video Diffusion into Pixel Space” introduces PixVid, which retains a pretrained diffusion transformer but replaces its VAE decoder with a convolutional head that renders frames directly in pixel space. Fine-tuned from Wan2.1-14B, the authors report that PixVid became competitive with a latent baseline after roughly 1.3 million training clips, generated video at twice the resolution with the same per-frame token count, maintained alignment in depth-conditioned editing under fast motion, and responded to each action on the next frame.
PixVid provides a distinct pixel-space approach to framewise video generation and editing. Its reported results address limitations of latent-space models involving token growth at higher resolutions, grouped-frame conditions, and delayed interactive responses; the claims remain based on the authors’ paper.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2611.17336 (cs)
[Submitted on 12 Nov 2026]
Title:Back to Pixels: Lifting Latent Video Diffusion into Pixel Space
Authors:Yanke Song, Shuo Chen, Chris Rockwell, Guande He, Nan Liu, Desai Xie, Tianpei Gu, Yong-Hyun Park, Aryan V S, Dennis Yuren Menn, Ming-Hsuan Yang, Xun Huang
View a PDF of the paper titled Back to Pixels: Lifting Latent Video Diffusion into Pixel Space, by Yanke Song and 11 other authors