New Beckmann Transport Models Enable 1-Step Generation for Diffusion

kastnerkyle · x · 2026-08-11

This thread explains a new paper on Beckmann Transport Models, introducing a novel approach to streamline generative model inference.

Background: Current flow matching and diffusion models use time-dependent velocity fields to turn noise into data. Attempts to simplify this into time-independent flows (like Equilibrium Matching) lacked solid theoretical guarantees.
Theoretical Breakthrough: The authors prove that as long as target data lies on a lower-dimensional manifold (true for almost all real-world data like images), a time-independent flow naturally converges onto it.
Method: By tying this back to the flux constraints in Beckmann’s optimal transport problem, they derive a direct 1-step generative map.
Result: This eliminates the need for multi-step ODE solvers at inference, drastically simplifying the generation process.

More from Research