SURF: Signature-Retained Fast Video Generation
SURF is a two-stage acceleration framework for high-resolution video generation that tries to preserve the base model’s distinctive behavior instead of trading it away for speed. The first stage generates a low-res preview using the pretrained model, but adds “noise reshifting” so the model can start denoising at the original resolution before switching down later. The second stage uses a Refiner to map that preview back to the target high-res output with fewer denoising steps.
For developers, the appeal is obvious: faster iteration on video generation without the usual drift in layout, motion, or semantics that can make accelerated methods feel like a different model entirely. The paper claims 12.5x speedup for 5-second, 16fps, 720p Wan 2.1 videos and 8.7x for 5-second, 24fps, 720p HunyuanVideo. It was submitted on 25 Nov 2025 and revised on 18 May 2026, and the authors position SURF as compatible with different base models and acceleration methods.
“most of them compromise the distinctive signatures”
- what
- SURF is a two-stage framework for fast high-resolution video generation that aims to retain the base model’s layout, semantic, and motion signatures.
- who
- Authors are Kaixin Ding, Xi Chen, Sihui Ji, Yuan Gao, Liang Hou, Xin Tao, and Hengshuang Zhao.
- when
- Submitted on 25 Nov 2025 and last revised on 18 May 2026.
- impact
- Claims 12.5x speedup on 5-second, 16fps, 720p Wan 2.1 videos and 8.7x on 5-second, 24fps, 720p HunyuanVideo.
Meaningful speedups while preserving output quality
Follow video-generation updates
See relevant stories in your personalized news feed.
Discussion