Overview
Mean Flow learns the average velocity or transition over a finite time interval, rather than repeatedly estimating instantaneous velocity and numerically integrating it. Methods build on flow matching and consistency learning, with variants using semigroup or transition-kernel objectives, distillation, and geometry-aware formulations for spaces such as SO(3), SE(3), and action manifolds. The goal is low-latency generation for image and audio tasks, molecular and protein modeling, and robot control, with evaluations spanning perceptual datasets, geometric tasks, and control benchmarks. The main challenge is preserving quality, diversity, and stable training when compressing many refinement steps into one or a few evaluations; variance, distributional bias, and manifold constraints remain active trade-offs.
Latest papers
25 papers1 paper in the last four weeks, down 67% on the four weeks before. 0.0% of all new papers.