AnyFlow: Any-Step Video Diffusion Model with On-Policy Flow Map Distillation

Researchers introduced AnyFlow, an any-step video diffusion distillation framework for learning flow-map transitions over arbitrary intervals.

Instead of fixed-step consistency learning, it uses flow maps and an on-policy backward simulation scheme along ODE trajectories.

The method reports strong results on both bidirectional and causal video models ranging from 1.3B to 14B parameters.

AnyFlow addresses a key weakness of consistency-distilled video models by preserving better test-time scaling for few-step and multi-step generation.