Itô maps enable fast, differentiable sampling from stochastic differential equations at inference time, making it practical to do posterior sampling and stochastic control without expensive iterative procedures.
This paper introduces Itô maps, a new way to learn stochastic flow maps that can predict future states from any intermediate point in a stochastic process. Unlike recent one-step generative models that work with deterministic flows, Itô maps handle randomness (Brownian motion) directly, enabling efficient sampling and control in a single forward pass.