Research map / Video / Video generation
AnyFlow: Any-Step Video Diffusion Model with On-Policy Flow Map Distillation
Yuchao Gu, Guian Fang, Yuxin Jiang, Weijia Mao, Song Han, Han Cai, Mike Zheng Shou
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 |
|---|---|---|---|
| flow map training instead of consistency ODE-init (ablation) | causal 1.3B Wan · VBench T2V overall, 4 NFE | overall: 73.97 -> 80.48 | Improved |
| flow map training instead of flow matching (few-step) (ablation) | bidirectional 1.3B Wan · VBench T2V overall, 4 NFE | overall: 74.64 -> 81.75 | Improved |
| flow map backward simulation instead of consistency backward simulation (ablation) | AnyFlow 1.3B · VBench T2V overall, 32 NFE | overall: 82.96 -> 83.96 (bidirectional); 83.49 -> 83.96 (causal) | Improved |
| add on-policy distillation with flow map backward simulation (ablation) | flow map trained model · VBench T2V causal, 4 NFE | overall: 80.48 -> 83.54 | Improved |
| consistency distillation with more sampling steps (ablation) | consistency ODE-init + consistency backward sim · VBench T2V 4 -> 32 NFE | overall: 82.96 -> 79.80 | Hurt |
| increase NFE from 4 to 32 | AnyFlow-FAR-Wan2.1-14B · VBench T2V | overall: 84.05 -> 84.41 | Improved |
| increase NFE from 4 to 32 | AnyFlow-Wan2.1-T2V-14B · VBench T2V | overall: 84.04 -> 84.10 | Improved |
| AnyFlow instead of rCM at 4 NFE | Wan2.1-T2V-14B · VBench T2V | overall: 83.73 -> 84.04 | Improved |
| interpolated timestep conditioning instead of zero-init conditioning (ablation) | flow map training · visual quality / embedding norm | embedding norm, saturation | Improved |
| time sampler Beta(2,1.5) instead of uniform weighting (ablation) | flow map training · VBench overall, 32 NFE | overall | Improved |
| flow map backward simulation instead of consistency backward simulation at 16 steps (ablation) | on-policy distillation · training cost (causal) | s/iter: 93.8 -> 53.1 | Improved |
| flow map backward simulation instead of consistency backward simulation at 4 steps (ablation) | on-policy distillation · training cost (causal) | s/iter: 45.9 -> 53.1 | Hurt |
Closest papers by meaning
- Test-time Scaling of Diffusions with Flow Maps (NeurIPS 2026 · similarity 0.89)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.88)
- Shortcutting Pre-trained Flow Matching Diffusion Models is Almost Free Lunch (NeurIPS 2025 · similarity 0.88)
- Video-Mirai: Autoregressive Video Diffusion Models Need Foresight (NeurIPS 2026 · similarity 0.87)
- AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (CVPR 2025 · similarity 0.87)
- RAVEN: Real-time Autoregressive Video Extrapolation with Consistency-model GRPO (NeurIPS 2026 · similarity 0.87)
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.
