MeanFlowNFT: Bringing Forward-Process RL to Average-Velocity Generators

MeanFlowNFT adapts DiffusionNFT’s forward-process reinforcement-learning objective to MeanFlow generators, which predict average velocities over time intervals. The method constructs an induced instantaneous-velocity predictor using the MeanFlow identity, applies reward optimization to that predictor, and keeps sampling based on average velocities. The paper reports consistent gains on image and video generation. On SD3.5-M, it reportedly leads on 6 of 8 metrics among few-step RL-tuned generators. On Wan 2.1, 4-step MeanFlowNFT reaches a VBench score of 84.33, compared with 82.57 for 50-step LongCat-Video RL.