Score-based diffusion models have emerged as one of the most promising frameworks for deep generative modelling. In this talk we will introduce these models, and in particular discuss their application to learning conditional probability distributions. We discuss different estimators of the conditional score and provide a theoretical justification for one of the most successful ones. Moreover, we introduce a multi-speed diffusion framework, which leads to a new estimator for the conditional score. Our theoretical findings are accompanied by experimental results for image superresolution and image inpainting.
This is joint work with Georgios Batzolis, Christian Etmann and Jan Stanzcuk.