Research map / Video / Video generation
Dynamics-Aware Sparse Attention for Efficient Autoregressive Video Diffusion
Yuanyu He, Zhuokun Chen, Yefei He, Zhiwei Tang, Jiasheng Tang, Yinghao Yu, Jianfei Cai, Bohan Zhuang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a dynamics-aware sparse attention method for autoregressive video diffusion transformers to reduce computation and memory when generating long videos.
Problem
Autoregressive Diffusion Transformers for long video generation suffer from prohibitive computational overhead and large memory footprints because dense computation allocates uniform resources across all spatiotemporal regions, ignoring temporal redundancy where motion is sparse.
Approach
Uses video dynamics (motion) to allocate sparse attention so that computation concentrates on dynamic regions instead of uniformly over all spatiotemporal positions; details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use layer-wise memory retention (sigma 0.8 to 0.2) instead of uniform 0.6 (ablation) | DASA on Self Forcing · VBench 5s | throughput / total score: 24.11 -> 26.21 FPS; 83.30 -> 83.81 total | Improved |
| uniform retention sigma 0.8 instead of layer-wise (ablation) | DASA on Self Forcing · VBench 5s | throughput (FPS): 26.21 -> 22.34 | Hurt |
| add motion-aware early exit with local alignment (ablation) | Self Forcing · VBench 5s | throughput (FPS): 18.42 -> 26.21 | Improved |
| add motion-aware early exit with local alignment (ablation) | Self Forcing · VBench 5s | total score: 84.19 -> 83.81 | Hurt |
| add dynamics-guided adaptation on top of alignment (ablation) | DASA on Self Forcing · VBench 5s | throughput and semantic score: 26.21 -> 27.58 FPS; semantic 80.77 -> 81.32 | Improved |
| increase max age threshold tau for static token refresh to 6 or more (ablation) | DASA · VBench 5s | visual quality | Hurt |
| use early-exit threshold gamma=0.92 (ablation) | DASA · VBench 5s | speedup / total score: 1.42x speedup, total 83.98 | Improved |
| apply DASA instead of Dummy Forcing | Self Forcing · VBench 1080P | throughput and total score: 2.64 -> 2.81 FPS; 82.46 -> 82.79 | Improved |
| apply DASA to LongLive at 720P | LongLive · VBench 720P | total score: 83.31 -> 83.01 | Hurt |
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.
- LongLive — LongLive: Real-time Interactive Long Video Generation (ICLR 2026) · better on 2 of 2 shared results
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
- Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.90)
- VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.90)
- Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (ICML 2025 · similarity 0.90)
- Generative Pre-trained Autoregressive Diffusion Transformer (NeurIPS 2025 · similarity 0.90)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.89)
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 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.
