Yushi Huang

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

Yushi Huang, Xiangxin Zhou, Jun Zhang, Liefeng Bo, Tianyu Pang

MeanFlow generators achieve fast few-step sampling by predicting average velocities over time intervals, making them attractive for efficient generation. Reinforcement learning (RL) has become a powerful way to align diffusion and flow models with human preferences and task-speci…Read more