Diffusion models

A high level mathematical overview

Abstract

Diffusion models are tools of deep learning which underpin AI image generators such as Stable Diffusion and DALL-E. While some of these ideas existed since at least 2015 and theoretical foundations were laid in the 1980s, they became popularized in the 2020 article [SSDK+20]. Here, we summarize some of the ideas of ‘score-based’ diffusion models and their theoretical underpinning.