MeanFlow

A generative method that learns the 'average velocity' over a time interval, letting it produce a sample from noise in a single step.

MeanFlow was proposed in May 2025 by Zhengyang Geng, J. Zico Kolter, Kaiming He, and colleagues as a one-step generation method. Flow matching trains a network to predict the 'instantaneous velocity' at a given moment, and generation requires integrating along that velocity field over many small steps, which is slow at inference. MeanFlow instead predicts the 'average velocity' over a time interval — the total displacement across that interval divided by its length — and derives an identity relating average velocity to instantaneous velocity to use as the training objective, so it needs no pretrained teacher model or distillation and can sample in one step even trained from scratch. The paper reports an FID (a metric for how close generated images are to real ones, lower is better) of 3.43 for one-step generation on ImageNet 256×256. In robotics, work such as MP1 has already used it as an action-generation head to cut a policy's inference latency.

ExampleMP1 uses MeanFlow on a point-cloud-based robot-arm policy, generating a whole action chunk in a single network forward pass; the paper reports about 6.8 ms of inference time, roughly 19x faster than the 3D diffusion policy DP3, with a 10.2-point higher average success rate.

Also called

Mean Flows

Related

Flow Matching · Velocity Field · One-step Generation · Consistency Model · Rectified Flow · Denoising Steps

Sources

Mean Flows for One-step Generative Modeling (arXiv:2505.13447)
MP1: MeanFlow Tames Policy Learning in 1-step for Robotic Manipulation (arXiv:2507.10543)

As of

2025-07