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

Forum topic · 小凯 · 2026-07-18

Summary

MeanFlowNFT (arXiv:2607.15273) is a new reinforcement learning framework that applies forward-process RL tuning to MeanFlow generators. MeanFlow achieves fast few-step sampling by predicting average velocities over time intervals, but existing RL methods like DiffusionNFT optimize instantaneous velocities, leaving a gap between RL and MeanFlow. The authors bridge this gap by constructing an induced instantaneous-velocity predictor based on the MeanFlow identity, applying the DiffusionNFT objective to it, while sampling remains based on average velocities to retain few-step efficiency. MeanFlowNFT inherits DiffusionNFT's strict policy improvement guarantees. Experiments on image and video generation show consistent improvements, outperforming prior few-step RL-tuned generators on 6 of 8 metrics with SD3.5-M. On Wan 2.1, a 4-step MeanFlowNFT reaches a VBench score of 84.33, surpassing the 50-step LongCat-Video RL (82.57), and even exceeds multi-step RL-tuned diffusion models using only a few sampling steps.

Paper Overview

Field: Computer Vision (CV) Authors: Yushi Huang, Xiangxin Zhou, Jun Zhang, Liefeng Bo, Tianyu Pang Published: 2026-07-16 arXiv: 2607.15273

Abstract (translated)

MeanFlow generators achieve fast few-step sampling by predicting average velocities over time intervals, making them attractive candidates for efficient generation. Reinforcement learning (RL) has become a powerful way to align diffusion and flow models with human preferences and task-specific objectives. In particular, DiffusionNFT offers an efficient forward-process RL framework that requires neither reverse-process trajectories nor likelihood estimation. However, applying such RL methods to MeanFlow remains underexplored: DiffusionNFT optimizes instantaneous velocities, whereas MeanFlow samples with average velocities.

To bridge this gap, the authors introduce MeanFlowNFT. Inspired by the MeanFlow identity — which bridges average and instantaneous velocities — they construct an induced instantaneous-velocity predictor and apply the DiffusionNFT objective to it, making reward optimization well-defined for MeanFlow. Sampling still relies on average velocities, preserving MeanFlow's fast few-step generation advantages. The paper further proves that MeanFlowNFT inherits DiffusionNFT's strict policy improvement guarantees.

Key Results

Experiments on image and video generation show MeanFlowNFT consistently improves over baselines.
It outperforms prior state-of-the-art few-step RL-tuned generators on 6 of 8 metrics with SD3.5-M.
It can surpass multi-step RL-tuned diffusion models using only a few sampling steps.
Notable example: on Wan 2.1, 4-step MeanFlowNFT achieves a VBench score of 84.33, exceeding the 50-step LongCat-Video RL (82.57).