“Flow Matching for Generative Modeling,” posted to arXiv on October 6, 2022 by Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le of Meta, introduced a simple and scalable way to train a class of generative models called continuous normalizing flows. These models describe generation as a smooth, continuous transformation from noise to data, but they had historically been expensive and awkward to train.
Flow matching makes the training simulation-free. Instead of running the full generative process during training, the method directly regresses a neural network onto the velocity field of a chosen probability path connecting noise to data. The authors showed that picking paths based on optimal transport produces especially efficient and stable training, and that the resulting models match or exceed diffusion models in sample quality while being faster to sample from. The framework also generalizes diffusion as one special case among many possible paths.
Flow matching, together with the closely related rectified flow, has become one of the dominant training paradigms behind the newest generation of high-quality image and video generators. For a general reader, it represents a maturation of generative modeling: a unifying, mathematically clean recipe that explains diffusion as a particular choice and gives engineers a flexible set of options for balancing quality and speed.