Research map / Architectures and efficiency / Attention and transformers
DFSAttn: Dynamic Fine-grained Sparse Attention for Efficient Video Generation
Jie Hu, Zixiang Gao, Yutong He, Kun Yuan
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 |
|---|---|---|---|
| 3D Hilbert token reordering instead of raster order (ablation) | DFSAttn · HunyuanVideo-T2V-13B | PSNR: 27.794 -> 29.378 | Improved |
| Hilbert3D instead of Hilbert2D reordering (ablation) | DFSAttn · HunyuanVideo-T2V-13B | PSNR: 29.265 -> 29.378 | Improved |
| Hilbert3D instead of Block3D reordering (ablation) | DFSAttn · HunyuanVideo-T2V-13B | PSNR: 29.156 -> 29.378 | Improved |
| Hilbert reordering reduces intra-block variance | DFSAttn · query/key variance | Var(q): 1.22 -> 0.98 | Improved |
| add hierarchical sub-block scoring (size 16 vs none) (ablation) | DFSAttn · HunyuanVideo-T2V-13B | PSNR: 28.565 -> 29.378 | Improved |
| add hierarchical sub-block scoring (ablation) | DFSAttn · HunyuanVideo-T2V-13B | latency: 827.3s -> 830.4s | No clear effect |
| adaptive decreasing sparsity schedule instead of fixed (ablation) | DFSAttn · Wan/Hunyuan, 60-90% sparsity | PSNR | Improved |
| sparse mask caching, update every 25% of steps (ablation) | DFSAttn · video DiT | latency | Improved |
| DFSAttn vs SVG2 at higher sparsity | DFSAttn · Wan2.1-T2V-14B | PSNR: 18.034 -> 22.370 | Improved |
| DFSAttn vs SVG2 speed | DFSAttn · HunyuanVideo-T2V-13B | speedup: 2.20x -> 2.10x | Hurt |
| DFSAttn vs SVG2 imaging quality | DFSAttn · HunyuanVideo-T2V-13B | VBench IQ: 63.13 -> 66.06 | Improved |
Closest papers by meaning
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.92)
- Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (ICML 2025 · similarity 0.92)
- PISA: Piecewise Sparse Attention Is Wiser for Efficient Diffusion Transformers (NeurIPS 2026 · similarity 0.92)
- Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.91)
- iCATS: Fast Video Generation via Interaction-Aware Sparse Attention and Timestep-Adaptive Sparsity (NeurIPS 2026 · similarity 0.91)
- Parameterized Stripe Attention for Efficient Video Generation (NeurIPS 2026 · similarity 0.91)
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.
