news

NVIDIA's AnyFlow video checkpoints are on Hugging Face, where eight times the sampling steps buys 0.06 VBench points

Four any-step models distilled from Wan 2.1 ship in Diffusers format, and the test-time scaling claim behind them rests on VBench gaps of under half a point, scored by the authors.

By Daniel Ochoa · 19 October 2026

NVIDIA's AnyFlow checkpoints are now listed on Hugging Face in Diffusers format, four text-to-video models distilled from Wan 2.1 at 1.3B and 14B, with the four listings showing download counts between 28 and 41. The claim attached to them is test-time scaling: unlike distilled models locked to a fixed step budget, one AnyFlow model is supposed to run at any number of steps and keep improving as you add more. The paper's own VBench table is where that claim can be priced, and for the 14B bidirectional model it comes to 84.04 overall at four function evaluations against 84.10 at 32.

The gaps are fractions of a point, including the ones over the teacher

At 1.3B, the Wan2.1-T2V teacher scores 83.24 overall running 50 steps with classifier-free guidance, and the AnyFlow student scores 83.48 at four function evaluations and 83.96 at 32. At 14B, the teacher is 83.74 and the student 84.04 at four steps. The causal variant, AnyFlow-FAR-Wan2.1-14B, reports 84.05 at four and 84.41 at 32, which the paper compares against 83.25 for Krea's Realtime Wan2.1 14B at the same four steps. On image to video, AnyFlow-FAR reports a VBench-I2V score of 87.87 at four function evaluations against 87.71 for Wan2.1-I2V-14B at 50 steps with guidance.

Those are real step reductions, and the headline comparison against a 50 step teacher is the useful one. The any-step story is weaker. Across the two 14B configurations, moving from four steps to 32 is worth 0.06 and 0.36 points respectively, on a benchmark that aggregates 16 dimensions into a single total and offers no confidence interval to say whether a 0.06 point move means anything at all.

What the tables leave out

There is no matched inference latency comparison, so the benchmark says what the steps buy in score and nothing about what they cost in seconds or dollars on a given card. The 32 step result also has no matched 32 step Krea Realtime number next to it, which leaves the only head to head at the step count that favours the few-step regime. Every figure is produced by the authors, who re-evaluate the baselines themselves under their own unified protocol, a setup that is normal for distillation papers and still means the comparison and the claim come from the same desk. The work comes out of NVIDIA with Show Lab at the National University of Singapore and MIT, was presented as an ECCV 2026 poster, and the code sits under NVlabs on GitHub.

The field is also moving faster than the checkpoints. Distillation papers posted this month already list AnyFlow as a baseline to beat on the same Wan 2.1 backbones, which is a reminder that a VBench total is a snapshot of one protocol on one day rather than a durable ranking. We have seen the same pattern in upscaling, where FlashVSR+ trained against a black-box reward model and reported its wins against it.

The shipped version of this trade is the turbo tier

Few-step distillation is not an academic curiosity for anyone buying generation by the clip. It is exactly what the fast tiers in commercial products are, a student model sold alongside the full-step original at lower latency and lower price. In our benchmark, Kling 3.0 and Kling 3.0 Turbo are scored separately on the same prompts, which is the closest public read on what a distilled sibling actually gives up.

Sources: NVIDIA AnyFlow project page, arXiv, Hugging Face, GitHub.

Daniel Ochoa: Covers model launches, shutdowns and pricing changes as they happen. Reads deprecation notices for a living so you do not have to.