MPEG-Diffusion: Mixed-Resolution Video Generation
The authors report 4.8× faster denoising on Wan2.1 14B while processing just 30% of full-resolution tokens.
MPEG-Diffusion automatically selects visually important regions to generate at high resolution, leaving the rest at low resolution. It uses resolution-specific attention and blends the two sets of latents before decoding to reduce distortion. The reported denoising speedup is 3.8× on Wan2.1 1.3B. In a user study, raters preferred its output over prior mixed-resolution methods in more than 90% of pairs and rated its quality close to full-resolution generation.
HF Daily Papers' note
MPEG-Diffusion automatically selects visually important regions to generate at high resolution, leaving the rest at low resolution. It uses resolution-specific attention and blends the two sets of latents before decoding to reduce distortion. The reported denoising speedup is 3.8× on Wan2.1 1.3B. In a user study, raters preferred its output over prior mixed-resolution methods in more than 90% of pairs and rated its quality close to full-resolution generation.
HF Daily Papers' note
score 5