Research map / Video / Video generation
DSA: Efficient Inference For Video Generation Models via Distributed Sparse Attention
Shenggui Li, Runyu Lu, qiaoling chen, Haiyan Yin, Yueming Lyu, Yonggang Wen, Ivor Tsang, Tianwei Zhang
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 |
|---|---|---|---|
| use DSA mixed parallelism instead of USP on 8 GPUs | USP · Wan2.1-14B 720p generation | generation time (s): 251.26 -> 175 | Improved |
| use DSA instead of USP on 8 GPUs | USP · Hunyuan-video-13B | generation time (s): 284.71 -> 189.38 | Improved |
| use DSA instead of USP on 8 GPUs | USP · Wan2.1-1.3B | generation time (s): 59.45 -> 54.11 | Improved |
| use DSA on 8 GPUs versus dense single GPU | Dense · Wan2.1-14B | speedup: 1 -> 10.79 | Improved |
| use DSA sparse distributed attention instead of SparseAttention | Sparge · Wan2.1-14B VBench | PSNR: 30.79 -> 33.19 | Improved |
| use DSA instead of SVG on quality | SVG · Wan2.1-14B VBench | PSNR: 33.03 -> 33.19 | No clear effect |
| use dynamic attention scheduling instead of naive (ablation) | DSA · Wan2.1-14B | generation time (s): 188.92 -> 180.47 | Improved |
| use spatial-only pattern for all heads (ablation) | DSA · Wan2.1-14B | generation time (s): 188.92 -> 175 | Improved |
| increase spatial sparsity from 80% to 95% (ablation) | DSA · Wan2.1-14B overall consistency | overall consistency: 0.179 -> 0.174 | Hurt |
| increase temporal sparsity to 95% (ablation) | DSA · Wan2.1-14B overall consistency | overall consistency: 0.177 -> 0.179 | No clear effect |
Closest papers by meaning
- Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.91)
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.91)
- Training-Free Efficient Video Generation via Dynamic Token Carving (NeurIPS 2025 · similarity 0.90)
- iCATS: Fast Video Generation via Interaction-Aware Sparse Attention and Timestep-Adaptive Sparsity (NeurIPS 2026 · similarity 0.90)
- Parameterized Stripe Attention for Efficient Video Generation (NeurIPS 2026 · similarity 0.89)
- MonarchRT: Efficient Attention for Real-Time Video Generation (NeurIPS 2026 · 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.
