This one sits deep in the weeds of image and video generators. Some fast generators, called MeanFlow, make images in just a few steps by predicting an average velocity across a chunk of time instead of the instantaneous velocity at each moment. That speed is handy, but the popular way to fine-tune such models with reinforcement learning, DiffusionNFT, assumes you're working with instantaneous velocities. The two don't naturally fit.

MeanFlowNFT bridges them using a known identity relating average and instantaneous velocity: they build an implied instantaneous-velocity predictor, apply the RL objective to it, and keep sampling by the fast average-velocity route. They also prove it keeps the same policy-improvement guarantee. In experiments, four-step generation edges out much longer RL-tuned baselines on several metrics, in one case beating a 50-step method on a video benchmark.

That is the abstract's account of their own results, so the paper holds the proofs and full comparisons.