Definition
Plain language
A way to train a generative model by teaching it, from any point, which direction points toward real data.
As stated in the literature
A diffusion-style training objective that learns a velocity field along straight paths between noise and data samples; generation integrates the learned field over a fixed number of steps. Standard in modern image and video diffusion stacks.
Why it matters: It gives generative models a stable training target and lets them produce samples in a fixed, predictable number of steps.
For example, from any noisy starting point the model learns which way to nudge the numbers so they move steadily toward something that looks like real data.
Heard on the show
“… those latents — that's the Diffusion Transformer, the DiT — you train it with something called flow matching, and you steer generation with classifier-free guidance. …”Episode 127 — What Diffusion Language Models Were Missing: A Map, Not an Algorithm