AnyFlow: Any-Step Video Diffusion with On-Policy Flow Map Distillation
MIT / NVIDIA
AnyFlow addresses the quality-vs-speed tradeoff in video diffusion by targeting full ODE trajectory using flow-map transitions across arbitrary time intervals, rather than fixed-step consistency mapping. A 'Flow Map Backward Simulation' decomposes Euler rollouts into shortcut transitions for efficient training. The method scales from 1.3B to 14B parameter models and matches or surpasses consistency-based approaches while supporting flexible step budgets.
Why it matters
78 upvotes on HuggingFace Daily Papers for May 14. Tackles the core efficiency challenge in video diffusion from first principles, with results across a wide range of model scales demonstrating practical utility for production video generation pipelines.
Importance: 2/5
HF Daily Papers May 14 (78 upvotes); principled video diffusion distillation with flexible step budget