Research map / Video / Video generation
BLADE: Block-Sparse Attention Meets Step Distillation for Efficient Video Generation
Youping Gu, XIAOLONG LI, Yuhao Hu, Chen Minqi, Bohan Zhuang
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 |
|---|---|---|---|
| add global tokens (mean-pooled K/V) to ASA (ablation) | ASA on CogVideoX-5B · VBench-2.0 | Total score: 0.539 -> 0.569 | Improved |
| add additive ln(n) mask for global tokens (ablation) | ASA_G on CogVideoX-5B · VBench-2.0 | Total score: 0.559 -> 0.569 | Improved |
| add Gilbert space-filling-curve token rearrangement (ablation) | ASA distilled CogVideoX-5B · VBench-1.0 | quality score: 0.779 -> 0.788 | Improved |
| replace dense FA2 with ASA in 8-step model | Wan2.1-1.3B · End-to-end time | seconds: 36.11 -> 24.00 | Improved |
| replace dense attention with ASA kernel | Wan2.1-1.3B · Kernel time | ms: 73.25 -> 22.21 | Improved |
| use ASA instead of STA, training-free, sparsity ~0.75 | Wan2.1-1.3B 8-step · PSNR | PSNR: 16.72 -> 19.55 | Improved |
| use ASA instead of SVG, training-free, sparsity ~0.75 | Wan2.1-1.3B 8-step · SSIM | SSIM: 0.6390 -> 0.7433 | Improved |
| use ASA instead of Radial Attention at sparsity 0.50 | Wan2.1-1.3B 8-step · PSNR | PSNR: 22.07 -> 22.20 | Improved |
| sparsity-aware distillation (ASA_G, 8 steps) vs 50-step dense baseline | CogVideoX-5B · VBench-2.0 | Total score: 0.534 -> 0.569 | Improved |
| sparsity-aware distillation (ASA_G, 8 steps) vs 50-step dense baseline | Wan2.1-1.3B · VBench-2.0 | Total score: 0.563 -> 0.570 | No clear effect |
| use ASA_G instead of dense FA2 in 8-step model | Wan2.1-1.3B · VBench-2.0 | Total score: 0.580 -> 0.570 | Hurt |
| use STA sparse attention instead of ASA_G | Wan2.1-1.3B · VBench-2.0 | Total score: 0.570 -> 0.528 | Hurt |
Closest papers by meaning
- Mixture of Distributions Matters: Dynamic Sparse Attention for Efficient Video Diffusion Transformers (ICML 2026 · similarity 0.92)
- 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)
- Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (ICML 2025 · similarity 0.91)
- PISA: Piecewise Sparse Attention Is Wiser for Efficient Diffusion Transformers (NeurIPS 2026 · similarity 0.90)
- Parameterized Stripe Attention for Efficient 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.
