SURF: Signature-Retained Fast Video Generation

CVPR 2026

Kaixin Ding^{1 †}, Xi Chen¹, Sihui Ji¹, Yuan Gao², Liang Hou², Xin Tao², Pengfei Wan², Hengshuang Zhao^{1 ✉
1}The University of Hong Kong
²Kling Team, Kuaishou Technology

†: Intern at Kling Team, KuaishouTechnology, ✉: Corresponding Author

📋 News

• [2026.03] Release arXiv paper.
• [2026] Accepted to CVPR 2026.

📖 Introduction

High-resolution video generation is slow: for example, Wan 2.1 takes over 50 minutes to generate a 720p video. Existing acceleration methods often compromise model priors (layout, semantics, motion). We propose SURF, a two-stage framework: first, a fast low-resolution preview using a pretrained model; second, a Refiner to upscale while preserving priors. Key techniques include noise reshifting to reduce prior loss and shifting windows with careful training design. SURF is simple, efficient, and compatible with various base models, achieving 12.5× speedup for generating 5-second, 16fps, 720p Wan 2.1 videos and 8.7× speedup for generating 5-second, 24fps, 720p HunyuanVideo.

🏆 Quantitative Results

Comparison on Wan 2.1 14B

Method	QS↑	AQ↑	DD↑	MS↑	OC↑	SA↑	PC↑	Time↓	Speed↑	PFLOPs↓
Wan 2.1	83.31	66.9	63.89	97.65	27.08	41.82	45.45	3497 (58min)	1×	658.46
30% Step	77.92	58.43	56.94	96.95	24.56	18.18	16.36	1049	3.34×	197.54
50% Step	81.51	63.52	66.67	96.99	25.90	25.45	23.64	1748	2×	329.23
SVG	83.36	65.6	68.06	97.69	27.32	25.45	20.00	2712	1.29×	429.86
DMD	83.31	66.11	52.78	98.96	26.77	34.55	30.91	282	12.40×	39.51
Ours	83.26	66.86	72.22	97.95	27.38	41.82	38.18	278	12.58×	34.3

⚙️ Gallery

720p_compress.mp4

Comparison

720p Video

720p_compress.mp4

1080p Video

1080p_compress.mp4

Citation

@inproceedings{ding2026surf,
  title={SURF: Signature-Retained Fast Video Generation},
  author={Ding, Kaixin and Chen, Xi and Ji, Sihui and Gao, Yuan and Hou, Liang and Tao, Xin and Zhao, Hengshuang},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={9171--9181},
  year={2026}
}