MeanFlowNFT: Forward-Process RL for Average-Velocity Generators

Forum topic · 小凯 · 2026-07-18

Summary

MeanFlowNFT is a reinforcement learning framework that adapts the forward-process RL method DiffusionNFT to MeanFlow generators, which sample fast in few steps by predicting average velocities over time intervals. The key challenge is that DiffusionNFT optimizes instantaneous velocities while MeanFlow samples with average velocities. To bridge this gap, the authors construct an induced instantaneous-velocity predictor inspired by the MeanFlow identity, apply the DiffusionNFT objective to it, and retain average-velocity-based sampling for fast few-step generation. MeanFlowNFT inherits strict policy improvement guarantees from DiffusionNFT. Experiments on image and video generation show consistent gains over baselines, outperforming prior state-of-the-art few-step RL-tuned generators on 6 of 8 metrics for SD3.5-M, and even surpassing multi-step RL-tuned diffusion models: a 4-step MeanFlowNFT on Wan 2.1 achieves a VBench score of 84.33, beating 50-step LongCat-Video RL (82.57). Paper: arXiv 2607.15273.

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

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 does not require reverse-process trajectories or 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. The DiffusionNFT objective is applied to this predictor, making reward optimization for MeanFlow well-defined. Sampling still relies on average velocities, preserving MeanFlow's fast few-step generation advantage.

Key Contributions

An induced instantaneous-velocity predictor that makes forward-process RL (DiffusionNFT-style) applicable to MeanFlow generators
Sampling remains based on average velocity, retaining few-step generation efficiency
A proof that MeanFlowNFT inherits the strict policy improvement guarantees of DiffusionNFT

Experimental Results

MeanFlowNFT consistently improves baselines on both image and video generation
It outperforms prior state-of-the-art few-step RL-tuned generators on 6 of 8 metrics for SD3.5-M
It even surpasses multi-step RL-tuned diffusion models with only a few sampling steps
Notably, on Wan 2.1, a 4-step MeanFlowNFT achieves a VBench score of 84.33, exceeding the 50-step LongCat-Video RL score of 82.57