Research map / Video / Video generation
FreeSwim: Revisiting Sliding-Window Attention Mechanisms for Training-Free Ultra-High-Resolution Video Generation
Yunfeng Wu, Jiayi Song, Zhenxiong Tan, Zihao He, Songhua Liu
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 |
|---|---|---|---|
| replace full 3D attention with inward sliding-window attention at training-scale window (ablation) | Wan2.1 1.3B · VBench 1080P | Overall Score: 65.5% -> 72.5% | Improved |
| add cross-attention override from full-attention branch (dual path) (ablation) | Wan2.1 1.3B window attention · VBench 1080P | Overall Score: 72.5% -> 72.7% | Improved |
| window attention reduces repetition only via override; without it repetitive patterns appear (ablation) | Wan2.1 1.3B · qualitative | repetition | Improved |
| omit full-attention guidance on unconditional CFG branch (ablation) | FreeSwim Wan2.1 1.3B · VBench 1080P | Overall Score: 72.7% -> 73.2% | Improved |
| add full-branch cross-attention cache and reuse (P=2) (ablation) | FreeSwim Wan2.1 1.3B · VBench 1080P | Overall Score: 73.2% -> 73.7%; time 64 min 5 s -> 54 min 40 s | Improved |
| increase cache refresh interval P from 2 to 8 (ablation) | FreeSwim Wan2.1 1.3B · VBench 1080P | Overall Score: 73.7% (P=2) -> 72.3% (P=5) -> 71.4% (P=8) | Hurt |
| set override strength lambda to 1 instead of 0.3 or 0.6 (ablation) | FreeSwim · qualitative | repetition | Improved |
| FreeSwim (training-free) vs CineScale with LoRA (training-based) | Wan2.1 1.3B · VBench 3K | Overall Score: 52.8% -> 58.7% | Improved |
| FreeSwim vs CineScale w/o LoRA on Wan2.1 14B | Wan2.1 14B · VBench 1080P | Overall Score: 73.4% -> 73.6% | No clear effect |
| FreeSwim vs Real-ESRGAN super-resolution on Wan2.1 1.3B (no cache) | Wan2.1 1.3B · VBench 1080P | Overall Score: 72.3% -> 72.7% | No clear effect |
| reduce latent frames to keep token count equal to training instead of window attention (ablation) | Wan2.1 · qualitative | visual quality | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Upscale-A-Video — Upscale-A-Video: Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution (CVPR 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.90)
- Re-ttention: Ultra Sparse Visual Generation via Attention Statistical Reshape (NeurIPS 2025 · similarity 0.90)
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.90)
- Training-Free Efficient Video Generation via Dynamic Token Carving (NeurIPS 2025 · similarity 0.89)
- Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.89)
- DiTCtrl: Exploring Attention Control in Multi-Modal Diffusion Transformer for Tuning-Free Multi-Prompt Longer Video Generation (CVPR 2025 · similarity 0.89)
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.
