MeanFlow

MeanFlow studies how to model the average change between states over a time interval, rather than only the instantaneous velocity used by conventional flow matching. Its methods learn finite-interval transport directly, sometimes distilling multi-step diffusion or flow models, and extend the idea to discrete transition kernels, control policies, and physical dynamics. The goal is to cut sampling or action-generation steps in settings such as image and video generation, restoration, reinforcement learning, and PDE solving. A central challenge is preserving fidelity and training stability at very low step counts, since average-velocity objectives can introduce bias or struggle with complex trajectories.