MeanFlow

Mean Flows for One-step Generative Modeling

Zhengyang Geng, Mingyang Deng, Xingjian Bai, J. Zico Kolter, Kaiming He · arXiv 2025

Read paper · arXiv:2505.13447

Learn an interval flow using the identity between instantaneous and average velocities for one-step generation.

What you will implement

Implement the average-flow JVP target and gradient boundaries, then run a small generative experiment.

• 16. MeanFlow: Average Velocity and JVP Targets

  Implement the total-derivative average-velocity target and its frozen-target loss, then sample in one call.

  Implement meanflow_target, meanflow_loss, one_step. toy.py provides the training and visualization scaffold.