Dynamics-Aware Sparse Attention for Efficient Autoregressive Video Diffusion
In this paper, we propose Dynamics-Aware Sparse Attention (DASA), a unified framework that shifts from dense computation to sparse, motion-driven generation to eliminate both storage and computational redundancies.
ProofPaper ↗
Key points
- Autoregressive Diffusion Transformers have emerged as a powerful paradigm for long video generation.
- This inefficiency arises from dense computation mechanisms that allocate uniform processing resources across all spatiotemporal regions, failing to leverage the inherent temporal redundancy of video where motion is typically confined to sparse areas.
- Inspired by video codec principles, our approach explicitly decouples video content into dynamic and static regions.
- To mitigate memory bottlenecks, we selectively compress the historical KV cache by retaining only dynamic features within the generated chunk.
Sources (1)
- [1]Dynamics-Aware Sparse Attention for Efficient Autoregressive Video DiffusionarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 12:19 PM
In this paper, we propose Dynamics-Aware Sparse Attention (DASA), a unified framework that shifts from dense computation to sparse, motion-driven generation to eliminate both storage and computational redundancies.
Autoregressive Diffusion Transformers have emerged as a powerful paradigm for long video generation.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 19, 2026TriTD: Tri-Partite Trajectory-Distribution Distillation for Real-Time Autoregressive Video Generation
- Oct 19, 2026TamedFlow: Trajectory-Aware Distillation\\for Teacher-Assisted Diffusion Sampling
- Oct 19, 2026SpecKCast: Spectral Kalman Forecasting for Diffusion Acceleration
- Oct 19, 2026Any 3D Scene is Worth 1K Tokens: 3D-Grounded Representation for Scene Generation at Scale
- Oct 18, 2026GPT-2 KV cache from scratch: correctness, latency, and memory measurements [P]
- Oct 17, 2026Nemotron 3.5 Bolt, Muse Spark Open, and a Fast LFM2.5 24B