Flow Matching — analysis and distillation into a one-step image generator

🔧 Flow Matching and distillation into a one-step generator — a breakdown

Flow Matching is one of the main approaches to image generation. Stable Diffusion 3.5 and FLUX.2 are built on it.

Without the jargon: Flow Matching doesn't draw the picture right away — it moves noise along a vector field toward the finished image. The vector fields are the key here: they set the direction of movement at each step.

The mechanism. Step-by-step movement means that for a single image the model has to be run dozens, and sometimes hundreds, of times. Hence the slow generation — that's how multi-step processing gets in the way.

The solution is distillation. A trained model is collapsed into a fast one-step generator: all the steps are packed into one ⚡ The breakdown describes what the two-model game has to do with it.

Model training is covered separately: Flow Matching plus distillation into a one-step generator. There's a lot of material — intuition, math, and practical training details. It's all in the article on Habr