Research map / Video / Video generation
SURF: Signature-Retained Fast Video Generation
Kaixin Ding, Xi Chen, Sihui Ji, Yuan Gao, Liang Hou, Xin Tao, Hengshuang Zhao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply SURF (two-stage preview plus refiner) to Wan 2.1 14B | Wan 2.1 14B · 720p video generation | DiT time: 3497s -> 278s (12.58x) | Improved |
| apply SURF for 1080p generation | Wan 2.1 14B · 1080p video generation | speedup: 43x vs direct Wan 2.1 | Improved |
| use SURF instead of reducing Wan 2.1 to 50% of steps | Wan 2.1 14B · Wan 2.1 720p, physical scene | Physics Commonsense: 23.64 -> 38.18 | Improved |
| use SURF instead of DMD step distillation | Wan 2.1 14B · Wan 2.1 720p, general scene | Dynamic Degree: 52.78 -> 72.22 | Improved |
| SURF vs original Wan 2.1 on physics commonsense | Wan 2.1 14B · Wan 2.1 720p, physical scene | Physics Commonsense: 45.45 -> 38.18 | Hurt |
| plug SURF into HunyuanVideo | HunyuanVideo · 720p, physical scene | Physics Commonsense: 27.27 -> 45.45, 8.7x speedup | Improved |
| plug SURF into step-distilled AccVideo | AccVideo · physical scene | speedup: 1x -> 1.3x; PC 23.64 -> 38.18 | Improved |
| switch to low resolution early (after step 5 instead of 10) (ablation) | SURF preview stage · Wan 2.1 preview stage step division | layout and motion quality: Time 201s vs 252s | Hurt |
| delay resolution switch to step 35 (35-5 vs 10-30) (ablation) | SURF preview stage · Wan 2.1 preview stage step division | final layout: time increases (610s at 40-0) | Mixed |
| switch resolution at step 10 (noise reshifting) (ablation) | SURF preview stage · Wan 2.1 preview stage step division | Time: 610s -> 252s | Improved |
| add shift-window attention in Refiner (ablation) | SURF Refiner · 1080p refinement | Refiner runtime: 107s -> 76s | Improved |
| increase refiner steps from 8 to 12 (ablation) | SURF Refiner · 720p and 1080p | quality metrics: 720p time 244.5 -> 254.2; QS 83.24 -> 83.31 | No clear effect |
Closest papers by meaning
- Training-Free Efficient Video Generation via Dynamic Token Carving (NeurIPS 2025 · similarity 0.87)
- FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention (NeurIPS 2024 · similarity 0.86)
- LinGen: Towards High-Resolution Minute-Length Text-to-Video Generation with Linear Computational Complexity (CVPR 2025 · similarity 0.86)
- SnapGen-V: Generating a Five-Second Video within Five Seconds on a Mobile Device (CVPR 2025 · similarity 0.86)
- Snap Video: Scaled Spatiotemporal Transformers for Text-to-Video Synthesis (CVPR 2024 · similarity 0.85)
- Hierarchical Patch Diffusion Models for High-Resolution Video Generation (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
