Back to Pixels: Lifting Latent Video Diffusion into Pixel Space
THE QUICK READ
The authors present PixVid, which lifts a pretrained latent video diffusion model to generate every frame directly in pixel space. It keeps the diffusion transformer and replaces the VAE decoder with a convolutional head. The authors report that predicting clean video rather than velocity, with a larger noise shift, is key to quality, and that finetuning from Wan2.1-14B is competitive with the latent baseline within about 1.3M training clips.
KEY POINTS
- Keeps the diffusion transformer and swaps the VAE decoder for a convolutional head that renders pixels
- Authors identify clean-video prediction instead of velocity, plus a larger noise shift, as key to quality
- Authors report 2x resolution at the same per-frame token count, better framewise conditioning, and next-frame interactive response
READING LIMITS
Based only on the title and abstract; the full paper was not read, so experimental setup, baselines and reported results are unverified.
Source abstract [EN]
Frontier video generation models work in a VAE latent space that compresses video in both space and time. This compression made training tractable, but now limits high-resolution, controllable, and interactive generation -- doubling resolution quadruples token count, conditions apply to groups of frames rather than individual ones, and each response must wait for the rest of its multi-frame block. We present PixVid, a simple yet effective method that lifts a pretrained latent video model to directly generate every frame in pixel space. PixVid keeps the diffusion transformer and replaces the VAE decoder with a convolutional head that renders its outputs into pixels. We find that predicting clean video rather than velocity, together with a larger noise shift, is key to quality. Finetuned from Wan2.1-14B, PixVid is competitive with the latent baseline within sim 1.3M training clips. Working framewise in pixels, it generates at 2times higher resolution with the same per-frame token count, keeps depth-conditioned video editing aligned under fast motion, and responds to each action at the next frame. Project page: https://chenshuo20.github.io/PixVid/
AI summaries use public titles and abstracts only. Research claims belong to the authors; consult the original paper.