An introduction to Flow Matching · Cambridge MLG Blog

Flow matching is a generative modelling paradigm that combines aspects from Continuous Normalising Flows (CNFs) and Diffusion Models (DMs). It addresses key issues of these methods and allows for efficient training of CNF models. Conditional Flow Matching (CFM) is a technique used in flow matching that involves conditioning and marginalizing over latent variables to construct the probability path. By using optimal transport coupling, flow matching improves training variance and sampling speed.