Research map / Video / Video generation
Mixture of Distributions Matters: Dynamic Sparse Attention for Efficient Video Diffusion Transformers
Yuxi Liu, Yipeng Hu, Zekun Zhang, Kunze Jiang, 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 |
|---|---|---|---|
| use MOD-DiT sparse attention instead of full attention | HunyuanVideo-13B · VBench prompts | speedup: 1.00× -> 2.05× | Improved |
| use MOD-DiT sparse attention instead of full attention | Wan2.1-14B · VBench prompts | speedup: 1.00× -> 1.75× | Improved |
| use MOD-DiT instead of full attention | HunyuanVideo-13B · VBench Subject Consistency | Subject Consistency: 0.9582 -> 0.9398 | Hurt |
| use MOD-DiT instead of SVG | Wan2.1-14B · VBench Subject Consistency | Subject Consistency: 0.7986 -> 0.9427 | Improved |
| use MOD-DiT instead of Radial attention | HunyuanVideo-13B · VBench prompts | PSNR: 26.72 -> 27.73 | Improved |
| use MOD-DiT instead of Radial attention | HunyuanVideo-13B · VBench prompts | SSIM: 0.885 -> 0.879 | Hurt |
| use static one-time pattern prediction instead of dynamic piecewise-linear tracking (ablation) | MOD-DiT · VBench | generation quality | Hurt |
| reduce full-attention warm-up steps m to 8 (ablation) | MOD-DiT · VBench | speed and quality | Mixed |
| add full-attention warm-up before sparse attention (ablation) | MOD-DiT · early denoising steps | quality | Improved |
| add MOD-DiT predictor overhead (sparsity conversion and solve kernel) (ablation) | MOD-DiT · HunyuanVideo 512x832x81 | TFLOPS vs full attention: 0.0297 and 0.0002 vs 7.6127 TFLOPS | No clear effect |
| use custom CUDA least-squares kernel instead of torch.lstsq (ablation) | MOD-DiT · solver time | solve time: 100x faster | Improved |
| mask 50% of blocks with MOD-DiT | HunyuanVideo · HunyuanVideo 512x832x81 | attention TFLOPS: 7.6127 -> 5.7096 | Improved |
Closest papers by meaning
- Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (ICML 2025 · similarity 0.92)
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.92)
- Parameterized Stripe Attention for Efficient Video Generation (NeurIPS 2026 · similarity 0.91)
- Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.91)
- PISA: Piecewise Sparse Attention Is Wiser for Efficient Diffusion Transformers (NeurIPS 2026 · similarity 0.91)
- Training-Free Efficient Video Generation via Dynamic Token Carving (NeurIPS 2025 · 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.
