Variational Perspective: From VAEs to DDPMs

This chapter develops the variational view of diffusion models, beginning with the variational autoencoder (VAE) and building up to the denoising diffusion probabilistic model (DDPM). It shows how a hierarchy of latent-variable models with a fixed forward corruption process leads to the step-by-step denoising objective at the heart of DDPM.

Chapter 2 of The Principles of Diffusion Models.