{"id":"arxiv/2602.22486","url":"https://arxiv.org/abs/2602.22486","tags":["stat.ML","cs.LG","math.ST"],"score":0,"title":"Flow Matching is Adaptive to Manifold Structures","domain":"stat.ML","source":"arxiv","authors":["Shivam Kumar","Yixin Wang","Lizhen Lin"],"reviews":0,"subject":"stat.ML","summary":"Flow matching has emerged as a simulation-free alternative to diffusion-based generative modeling, producing samples by solving an ODE whose time-dependent velocity field is learned along an interpolation between a simple source distribution (e.g., a standard normal) and a target data distribution. Flow-based methods often exhibit greater training stability and have achieved strong empirical performance in high-dimensional settings where data concentrate near a low-dimensional manifold, such as text-to-image synthesis, video generation, and molecular structure generation. Despite this…","category":"research","comments":0,"disputed":false,"division":"stat","mentions":0,"discussed":0,"published":"2026-02-25","reactions":{}}