MeanFlowNFT: Bringing Forward-Process RL to Average-Velocity Generators
MeanFlowNFT adapts forward-process RL to MeanFlow by optimizing an induced instantaneous-velocity predictor while keeping few-step sampling.
The paper says this bridges DiffusionNFT’s instantaneous-velocity objective with MeanFlow’s average-velocity generator. The authors claim the method preserves MeanFlow’s fast sampling and inherits DiffusionNFT’s strict policy-improvement guarantee. In image and video tests, it improves baselines and beats prior RL-tuned few-step generators on most reported SD3.5-M metrics. On Wan 2.1, the paper reports a 4-step MeanFlowNFT VBench score of 84.33, above 50-step LongCat-Video RL at 82.57. ArXiv · AI/CL/LG's note
The paper says this bridges DiffusionNFT’s instantaneous-velocity objective with MeanFlow’s average-velocity generator. The authors claim the method preserves MeanFlow’s fast sampling and inherits DiffusionNFT’s strict policy-improvement guarantee. In image and video tests, it improves baselines and beats prior RL-tuned few-step generators on most reported SD3.5-M metrics. On Wan 2.1, the paper reports a 4-step MeanFlowNFT VBench score of 84.33, above 50-step LongCat-Video RL at 82.57. ArXiv · AI/CL/LG's note
score 4