Research map / Video / Video generation
Motion-Aware Caching for Efficient Autoregressive Video Generation
Jing Xu, Yuexiao Ma, Xuzhe Zheng, XING WANG, Shiwei Liu, Chenqian Yan, Xiawu Zheng, Rongrong Ji, Fei Chao, Songwei Liu
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 motion-aware token-wise caching instead of chunk-level FlowCache | SkyReels-V2 caching · SkyReels-V2 slow config | PSNR: 21.83 -> 23.46 | Improved |
| use motion-aware token-wise caching instead of chunk-level FlowCache | MAGI-1 caching · MAGI-1 slow config | PSNR: 18.16 -> 19.71 | Improved |
| use MotionCache instead of TeaCache-fast | MAGI-1 caching · MAGI-1 fast config | VBench: 68.81% -> 74.59% | Improved |
| increase soft-mapping floor alpha from 0.0 to 0.6 (ablation) | MotionCache · SkyReels-V2 | PSNR: 20.22 -> 23.46 | Improved |
| increase alpha from 0.6 to 1.0 (ablation) | MotionCache · SkyReels-V2 | PSNR: 23.46 -> 23.44 | No clear effect |
| increase alpha (more updates of static tokens) (ablation) | MotionCache · SkyReels-V2 | latency | Hurt |
| add Phase 1 chunk-wise warm-up (K=0 -> K=6) (ablation) | MotionCache · SkyReels-V2 | PSNR: 20.79 -> 23.46 | Improved |
| increase K from 6 to 9, 12 or 15 (ablation) | MotionCache · SkyReels-V2 | PSNR: 23.46 -> 23.39 / 23.41 / 23.43 | No clear effect |
| use frame difference as proxy for token importance | MotionCache · SkyReels-V2 | NDCG: avg 0.9687 | Improved |
Closest papers by meaning
- Flow Caching for Autoregressive Video Generation (ICLR 2026 · similarity 0.94)
- MagCache: Fast Video Generation with Magnitude-Aware Cache (NeurIPS 2025 · similarity 0.90)
- AdaState: Self-Evolving Anchors for Streaming Video Generation (NeurIPS 2026 · similarity 0.89)
- LeMiCa: Lexicographic Minimax Path Caching for Efficient Diffusion-Based Video Generation (NeurIPS 2025 · similarity 0.89)
- Accelerating Rectified Flow Models via Trajectory-Aware Caching (NeurIPS 2026 · similarity 0.89)
- FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality (ICLR 2025 · similarity 0.88)
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.
