A Lagrangian View of Flow Matching: Why only one step is needed

docmilanfar · x · 2026-08-31

This is a pedagogical post offering an intuitive, bottom-up perspective on Flow Matching. It explores why some models require hundreds of iterative steps to generate an image while Flow Matching needs only a few, or even one.

The underlying mechanics of image generation often rely on extremely slow, iterative solvers. Standard diffusion models require many sequential steps, whereas newer frameworks like Flow Matching and Rectified Flow achieve the same in just a few. While literature often explains this via Eulerian frameworks (Optimal Transport, continuity equations), this post adopts a Lagrangian (particle-centric) perspective to explain the mechanics.

Original post →

More from Research

Research channel →