Stochastic Interpolants

In this work, we approach generative modelling from the standpoint of stochastic interpolants, a dynamical approach to measure transport. We discuss its theoretical construction, giving formal definitions and demonstrating favorable properties. We parametrize the corresponding vector fields associated with the conditional expectation of the stochastic interpolant using deep neural networks. We show how the learned vector field can be used in a probability flow ODE or a family of SDEs, allowing us to sample deterministically or stochastically. We end giving examples of transported distributions with maps learned by stochastic interpolant methods, comparing deterministic to stochastic sampling.

You can find a detailed report, results, and code here.