Definition
Flow Matching is a training objective for continuous generative models that learns a vector field defining straight-line “flows” between a noise distribution and the data distribution, replacing the traditional score-matching loss used in earlier diffusion models. It yields faster training and cleaner probability paths, and has become a standard objective in state-of-the-art image generators like Stable Diffusion 3 — a recipe that transfers directly to text latent diffusion without modification.