Abstract
We propose a principled and effective framework for one-step generativemodeling. We introduce the notion of average velocity to characterize flowfields, in contrast to instantaneous velocity modeled by Flow Matching methods.A well-defined identity between average and instantaneous velocities is derivedand used to guide neural network training. Our method, termed the MeanFlowmodel, is self-contained and requires no pre-training, distillation, orcurriculum learning. MeanFlow demonstrates strong empirical performance: itachieves an FID of 3.43 with a single function evaluation (1-NFE) on ImageNet256x256 trained from scratch, significantly outperforming previousstate-of-the-art one-step diffusion/flow models. Our study substantiallynarrows the gap between one-step diffusion/flow models and their multi-steppredecessors, and we hope it will motivate future research to revisit thefoundations of these powerful models.