Roblox's PixVid Drops the VAE to Hit 960p at Nearly 480p Cost

Video generators including Sora, Veo, Wan, and Cosmos typically compress clips with a variational autoencoder (VAE) before processing them with a diffusion model. Researchers at Roblox and UC Merced have introduced PixVid, which removes the VAE from the generation path and converts a pretrained latent video model into a pixel-space generator. The change targets three limits of compressed video: resolution costs, coarse temporal control, and motion artifacts during editing.

Starting from Wan2.1-14B, PixVid reached quality comparable to the original model after 20,000 fine-tuning iterations. It generates 960p video with the same transformer token count used for 480p, preserves fast motion during controlled editing, and can reflect user input in the next rendered frame.

Compression now gets in the way

A video VAE compresses pixels into a smaller grid of latent values, often combining four consecutive frames into one temporal block. A diffusion transformer, or DiT, then denoises that compact grid. The compression made large-scale video training practical because the transformer processes far fewer tokens.

Spatial and temporal compression also constrain later stages of the pipeline. Doubling both output dimensions produces four times as many latent tokens at a fixed compression ratio. Frame-level conditions such as depth, pose, or user input must pass through shared temporal blocks. At 30 frames per second, a four-frame block spans 133 milliseconds before inference and system overhead are counted. Fast objects can occupy several positions inside that interval, producing blur or distorted motion.