Roblox's MPEG-Diffusion speeds up video generation by denoising only a self-selected high-res region

Roblox · hf · 2026-11-18

Roblox and UT Austin introduce MPEG-Diffusion, which cuts the compute cost of video diffusion transformers. Using dual-resolution attention and a mixed-resolution decoder, the model generates only a self-selected high-resolution region while keeping the background low-resolution, achieving faster denoising without quality loss. An accompanying arXiv paper shows the time-varying HR token layout.

Related event: MPEG-Diffusion speeds up video generation up to 4.8x(2 posts)→

Original post →

More from Multimodal

Multimodal channel →