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

TL;DR

AnyFlow distills a video flow-matching teacher into a student that can sample at any step budget, not just a fixed few-step grid. The framing rejects consistency distillation’s endpoint mapping (zₜ → z₀) in favour of flow-map transition learning (zₜ → zᵣ) over arbitrary [t,r] intervals, which preserves the teacher’s test-time scaling behaviour. A Flow Map Backward Simulation procedure decomposes a full Euler rollout into shortcut flow-map transitions for efficient on-policy distillation. Trained on bidirectional and causal/AR Wan2.1 backbones from 1.3B to 14B, AnyFlow matches or beats rCM, Self-Forcing, Krea-Realtime, LightX2V-CausVid, and FastVideo-CausalWan at 4 NFEs while continuing to improve with more steps — a property consistency-distilled models lose.

Key claims

Consistency-distilled video models degrade as test-time NFE increases, because consistency distillation replaces the teacher’s PF-ODE trajectory with a consistency-sampling trajectory and loses the desirable test-time scaling of ODE sampling [Abstract, §Method Overview].
Shifting the distillation target from endpoint consistency mapping (zₜ → z₀) to flow-map transition learning (zₜ → zᵣ) over arbitrary [t,r] intervals preserves a fine-grained instantaneous flow field, so the distilled model retains continued-training and any-step sampling [§Method Overview, Fig. 1].
Flow Map Backward Simulation preserves the teacher’s original Euler sampling trajectory but decomposes the long trajectory into shortcut segments, reducing on-policy rollout cost [§Method Overview, Fig. 1].
On 1.3B bidirectional Wan2.1, AnyFlow uniformly lifts performance over its flow-matching teacher at every step budget (4 / 16 / 32 NFEs), with the largest gains in the few-step regime [Fig. 2; Fig. 3 — “AnyFlow vs. Flow-Matching Teacher” row].
AnyFlow matches or surpasses rCM (consistency-distilled) on the same 1.3B Wan2.1 backbone at 4 NFEs while extending the test-time scaling curve [Fig. 2; Fig. 3 — “AnyFlow vs. Consistency Distillation (rCM)” row].
On 14B causal/AR Wan2.1, AnyFlow-FAR at 4 NFEs achieves better dynamic motion and quality than community-trained consistency counterparts: Krea-Realtime-Wan2.1-14B (4 NFEs), LightX2V-Wan2.1-14B-CausVid (9 NFEs), and FastVideo-CausalWan2.2-A14B-Preview (8 NFEs) [§Causal Video Diffusion Model (14B) §1.1].
The same 14B AnyFlow-FAR model serves I2V at 4 NFEs at quality comparable to Wan2.1-I2V-14B running 50×2 NFEs, and additionally supports V2V at 4 NFEs without retraining [§1.2, §1.3].
Because flow-map distillation preserves the instantaneous flow field, the distilled model can be continued-trained on a downstream dataset while keeping few-step sampling — domain fine-tuning on AnyFlow-FAR-Wan2.1-1.3B improves identity preservation (robot-arm type) and trajectory accuracy (moving pedestrians) at 4 NFEs [§3, Fig. 4].

Method

Standard consistency distillation parameterises the student as an endpoint map f(zₜ, t) ≈ z₀, then trains it so that consecutive timesteps along the teacher’s PF-ODE map to the same endpoint. AnyFlow replaces this with a two-time transition map f(zₜ, t, r) that predicts the teacher’s ODE state at an arbitrary later time r ∈ [0, t]. Forward training initialises the student on supervised transition pairs sampled from the teacher; on-policy refinement is then done via Flow Map Backward Simulation, which preserves the teacher’s full Euler trajectory but rolls it out in shortcut segments — i.e. several composed flow-map transitions whose composition equals one long Euler step — so each on-policy rollout reduces wall-clock cost without abandoning the teacher’s trajectory geometry. The same recipe is applied to a bidirectional Wan2.1 teacher and to a causal Wan2.1 teacher (AnyFlow-FAR), where the AR variant is benchmarked against Self-Forcing-class consistency distillations. Continued training on downstream data is run by initialising from the AnyFlow distilled checkpoint and resuming flow-map transition training on the new corpus while preserving few-step sampling.

Results

Bidirectional 1.3B Wan2.1: AnyFlow > flow-matching teacher at every of 4/16/32 NFEs; AnyFlow ≥ rCM at 4 NFEs and continues improving where rCM plateaus or degrades [Fig. 2, Fig. 3].
Causal AR 1.3B: AnyFlow-FAR at 4/16/32 NFEs beats Self-Forcing at the same step counts [Fig. 3, “AnyFlow-FAR vs. Self-Forcing”].
Causal AR 14B (Wan2.1): AnyFlow-FAR-Wan2.1-14B at 4 NFEs beats Krea-Realtime-Wan2.1-14B (4 NFEs), LightX2V-Wan2.1-14B-CausVid (9 NFEs), and FastVideo-CausalWan2.2-A14B-Preview (8 NFEs) on T2V dynamic/quality (qualitative comparisons) [§1.1].
I2V: AnyFlow-FAR-Wan2.1-14B at 4 NFEs reaches similar quality to Wan2.1-I2V-14B at 50×2 = 100 NFEs [§1.2].
Continued training: domain-finetuned AnyFlow-FAR-Wan2.1-1.3B improves identity and trajectory metrics on a specialised downstream dataset while preserving 4-NFE inference [§3.2].

No quantitative VBench/HPS/AeS numbers are reported on the project page — all comparisons are visual / step-budget-anchored.

Why it’s interesting

AnyFlow attacks the exact failure mode that Diffusion Distillation has been circling: consistency distillation buys few-step inference but throws away the teacher’s test-time scaling. That trade-off is at the core of why production stacks ship 4-NFE distilled models that you cannot improve at inference. By distilling the flow map itself (zₜ → zᵣ for any r) rather than the endpoint (zₜ → z₀), AnyFlow recovers monotone scaling over NFEs and matches few-step quality. This makes it the natural any-step counterpart to Streaming Autoregressive Video Generation via Diagonal Distillation‘s per-chunk-asymmetric schedule and Cross-Resolution Distribution Matching for Diffusion Distillation‘s resolution cascade — all three are “more compute where it matters” cascades on different axes (NFE, AR chunks, resolution). On the AR side, AnyFlow-FAR’s win over Self-Forcing and the LightX2V/Krea/FastVideo CausVid descendants is a direct test of the Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation thesis that consistency-style AR distillation is structurally lossy: AnyFlow keeps the teacher’s full trajectory and outperforms them. Conceptually it sits next to Learning the integral of a diffusion model (flow maps)‘s flow-map framing — Dieleman’s piece argues flow maps are the right primitive; AnyFlow ships a concrete on-policy distillation recipe for them, on production-scale video.