Flow Matching revisited: the simulation-free CNF training method behind today's generative models

burny_tech · x · 2026-09-21

A thread revisits the 2023 Meta paper "Flow Matching for Generative Modeling" (Lipman, Chen et al., arXiv:2210.02747), which generalizes score-based generative models via continuous normalizing flows. Key points: Flow Matching trains CNFs simulation-free by regressing vector fields of fixed conditional probability paths; diffusion paths are just a special case of its Gaussian path family, and FM makes training more stable; optimal-transport interpolation paths yield faster training/sampling and better generalization; on ImageNet, FM-trained CNFs beat diffusion baselines on both likelihood and sample quality. It underpins today's rectified-flow generative models.

Original post →

More from Multimodal

Multimodal channel →