Paper 2605.13724
AnyFlow: Any-Step Video Diffusion Model with On-Policy Flow Map Distillation
01 In brief
Summary
AnyFlow is a video diffusion distillation framework that enables any-step generation by learning flow-map transitions between arbitrary time pairs, unlike consistency models that degrade with more sampling steps.
It uses a two-stage pipeline: forward flow map training (with interpolated timestep conditioning, guidance-fused training, and adaptive loss reweighting) and on-policy distillation via flow map backward simulation, which decomposes Euler rollouts into shortcut transitions to reduce discretization error and exposure bias.
Experiments on bidirectional and causal architectures (1.3B–14B parameters) show AnyFlow matches or outperforms consistency-based baselines in few-step regimes and improves with more steps.
For causal text-to-video, AnyFlow-FAR reaches 84.05 VBench at 4 NFEs and 84.41 at 32 NFEs, surpassing Krea-Realtime-14B (83.25 at 4 NFEs).
For image-to-video, it achieves 87.87 VBench-I2V at 4 NFEs, comparable to Wan2.1-I2V-14B at 50×2 NFEs (87.71).
In bidirectional 14B text-to-video, AnyFlow scores 84.04 at 4 NFEs, outperforming rCM-14B (83.73).
The method also supports continued fine-tuning on downstream datasets.
Training costs are practical, with flow map backward simulation being 43–47% cheaper than consistency backward simulation at 16 steps.
02 From the paper
Abstract
Few-step video generation has been significantly advanced by consistency distillation. However, the performance of consistency-distilled models often degrades as more sampling steps are allocated at test time, limiting their effectiveness for any-step video diffusion. This limitation arises because consistency distillation replaces the original probability-flow ODE trajectory with a consistency-sampling trajectory, weakening the desirable test-time scaling behavior of ODE sampling. To address this limitation, we introduce AnyFlow, the first any-step video diffusion distillation framework based on flow maps. Instead of distilling a model for only a few fixed sampling steps, AnyFlow optimizes the full ODE sampling trajectory. To this end, we shift the distillation target from endpoint consistency mapping $(z_{t}\rightarrow z_{0})$ to flow-map transition learning $(z_{t}\rightarrow z_{r})$ over arbitrary time intervals. We further propose Flow Map Backward Simulation, which decomposes a full Euler rollout into shortcut flow-map transitions, enabling efficient on-policy distillation that reduces test-time errors (i.e., discretization error in few-step sampling and exposure bias in causal generation). Extensive experiments across both bidirectional and causal architectures, at scales ranging from 1.3B to 14B parameters, demonstrate that AnyFlow achieves performance matches or surpasses consistency-based counterparts in the few-step regime, while scaling with sampling step budgets.