Flow matching is a generative modeling method that defines a continuous path between an easy-to-sample distribution and the target data distribution. A neural network learns a time-dependent vector field that describes how samples should move along that path.
At generation time, a numerical solver follows the learned field from noise or another simple starting point toward a sample. The method can generate continuous representations such as audio latents without first turning them into many discrete token streams.
The choice of path, solver and number of steps affects quality, speed and computational cost. Flow matching provides a framework for continuous generation, but it does not remove the need for evaluation of stability, fidelity and controllability.
ELI5
Imagine starting with a cloud of random dots and learning a smooth set of directions that moves the dots into the shape of real data. Flow matching learns those directions.
For example, a speech system can start from noise and follow the learned path toward a voice representation. More careful steps can improve quality, while fewer steps can make generation faster.
