Research map / Video / Video generation
Flow Caching for Autoregressive Video Generation
Yuexiao Ma, Xuzhe Zheng, Jing Xu, Xiwei Xu, Feng Ling, Xiawu Zheng, Huafeng Kuang, Huixia Li, XING WANG, Xuefeng Xiao, Fei Chao, Rongrong Ji
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 chunkwise caching policy instead of TeaCache uniform reuse (ablation) | FlowCache-fast · VBench MAGI-1 | VBench: 70.11% -> 77.66% | Improved |
| use chunkwise caching policy instead of TeaCache uniform reuse (ablation) | FlowCache-fast · VBench SkyReels-V2 | VBench: 80.06% -> 83.12% | Improved |
| add KV cache compression on top of chunkwise reuse (ablation) | FlowCache · VBench MAGI-1 | VBench: 77.66% -> 77.93% | No clear effect |
| add KV cache compression on top of chunkwise reuse (ablation) | FlowCache · VBench SkyReels-V2 | VBench: 83.12% -> 83.01% | No clear effect |
| add KV cache compression on top of chunkwise reuse (ablation) | FlowCache · SkyReels-V2 PSNR | PSNR: 23.74 -> 22.61 | Hurt |
| use chunkwise reuse instead of TeaCache reuse (ablation) | FlowCache · Physics-IQ MAGI-1 | Physics-IQ score: 13.69% -> 43.10% | Improved |
| add KV cache compression (importance+redundancy) (ablation) | FlowCache · Physics-IQ MAGI-1 | Physics-IQ score: 43.10% -> 39.34% | Hurt |
| keep keys at token level instead of frame level (ablation) | FlowCache · Physics-IQ MAGI-1 | Physics-IQ score: 38.62% -> 39.34% | Improved |
| use frame-level queries instead of token-level (ablation) | FlowCache · Physics-IQ MAGI-1 | Physics-IQ score: 39.34% -> 39.53% | No clear effect |
| lower lambda to weight redundancy more (ablation) | FlowCache · Physics-IQ MAGI-1 | Physics-IQ score: 37.11% (0.15) vs 39.53% (0.07) | Mixed |
| reduce KV cache budget from 8 to 5 chunks (ablation) | FlowCache · Physics-IQ MAGI-1 | Physics-IQ score: 47.60% -> 47.65% | No clear effect |
| FlowCache-fast vs TeaCache-fast | FlowCache-fast · MAGI-1 | speedup: 1.44x -> 2.38x | Improved |
Closest papers by meaning
- Ca2-VDM: Efficient Autoregressive Video Diffusion Model with Causal Generation and Cache Sharing (ICML 2025 · similarity 0.91)
- MagCache: Fast Video Generation with Magnitude-Aware Cache (NeurIPS 2025 · similarity 0.89)
- 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)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.89)
- Accelerating Rectified Flow Models via Trajectory-Aware Caching (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.
