MeanFlowNFT: 4-Step Video RL Tops 50-Step Diffusion on VBench

Reinforcement learning has become the default polish step for image and video generators, but the recipe until now assumed the underlying model was a diffusion or flow-matching sampler that runs in tens of steps. A new arxiv paper from Yushi Huang and collaborators introduces a working RL alignment method for MeanFlow generators, the family that produces samples in as few as four steps by predicting an average velocity instead of the usual instantaneous one.

The bridge they build is called MeanFlowNFT. It takes DiffusionNFT, a forward-process RL objective that comes with a strict policy-improvement guarantee, and adapts it to MeanFlow by constructing what the authors call an induced instantaneous-velocity predictor via the MeanFlow identity. The DiffusionNFT objective is then applied to that predictor, so the reward optimization runs on familiar instantaneous-velocity math while sampling still happens in MeanFlow's fast average-velocity mode.

The headline number is on video. A 4-step Wan 2.1 tuned with MeanFlowNFT reportedly hits 84.33 on VBench, which the authors say surpasses a 50-step LongCat-Video RL baseline at 82.57. On the image side with SD3.5-M, they claim wins on 6 of 8 metrics against prior RL-tuned few-step generators. If those numbers replicate, it is a real efficiency shift, because reward-tuned quality would stop paying the roughly ten-times sampling tax that came with the diffusion-RL default.

The honest caveat is that this is a single-team result on two specific base models against two specific baselines, and an automated benchmark win does not always translate to gains a human rater would notice. The paper as summarized doesn't lay out training-cost numbers or blind human preference studies, so take the 84.33 vs 82.57 gap as reported, not settled.

The direction of travel is the part worth watching. If MeanFlow-family generators can be RL-aligned with the same theoretical footing as diffusion, the choice between cheap-to-serve inference and reward-tuned quality stops being a tradeoff, and the teams that already bet on few-step samplers for cost reasons get to keep that advantage without falling behind on the alignment leaderboards.