ARTFEED — Contemporary Art Intelligence

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

ai-technology · 2026-05-14

Researchers introduce AnyFlow, a novel video diffusion distillation framework that enables any-step sampling without performance degradation. Traditional consistency-distilled models lose quality when more sampling steps are used at test time because they replace the original probability-flow ODE trajectory with a consistency-sampling trajectory. AnyFlow addresses this by optimizing the full ODE sampling trajectory through flow-map transition learning (z_t → z_r) over arbitrary time intervals, rather than endpoint consistency mapping (z_t → z_0). The framework is the first to support any-step video diffusion with on-policy flow map distillation. The paper is published on arXiv under ID 2605.13724.

Key facts

  • AnyFlow is the first any-step video diffusion distillation framework based on flow maps.
  • It optimizes the full ODE sampling trajectory instead of only a few fixed steps.
  • Distillation target shifts from endpoint consistency mapping (z_t → z_0) to flow-map transition learning (z_t → z_r).
  • Traditional consistency-distilled models degrade with more sampling steps at test time.
  • AnyFlow addresses the limitation of consistency distillation for any-step video diffusion.
  • The paper is available on arXiv with ID 2605.13724.
  • The framework uses on-policy flow map distillation.
  • It improves test-time scaling behavior of ODE sampling.

Entities

Institutions

  • arXiv

Sources