Diffusion Models Using a Single Equation

Anonymous

Keywords: generative modeling, denoising diffusion, DDIM, DDPM

Abstract: In this work I present a novel viewpoint and a simplistic implementation and explanation of denoising diffusion models, and also the intuition that we force these models to sample from the data distribution by misleading them. The aim of this blogpost is to lower the barrier of entry to the field of diffusion models, by providing an explanation of their inner workings that is mathematically very light (only uses a single equation, with carefully meaningfully named variables), and try to understand them from a different point of view compared to previous work. I also discuss new intuitions about these models, and provide an end-to-end, simple to use implementation for diffusion models, that is easy to customize and is intended be a good starting point for future projects on this subject.

ICLR Paper: https://arxiv.org/abs/2010.02502