Flow Matching for Generative Modeling

Flow Matching asks how to train a continuous generative flow when the distribution-level velocity target is unavailable. Read it to see how per-example conditional paths create tractable regression labels, why the conditional and marginal objectives have identical gradients, how path geometry changes solver work, and why straight conditional paths do not prove globally straight transport.

Reading focus: How a neural velocity field transports source noise into the data distribution through an ODE. Why conditional velocity regression has the same parameter gradient as the inaccessible marginal Flow Matching objective. How to read the likelihood, FID, and solver-evaluation results without turning conditional straightness into a global optimal-transport claim.

ICLR 2023. Yaron Lipman et al.. 29 min read, medium difficulty.