Denoising Diffusion Probabilistic Models

Denoising Diffusion Probabilistic Models turns generation into a sequence of small denoising steps. Read it to connect variational learning, Gaussian noise, and score matching to the practical recipe that became the foundation of modern diffusion image models.

Reading focus: How a fixed forward Markov process gradually turns data into Gaussian noise. How a learned reverse process generates samples by predicting and removing noise one step at a time. Why the variational bound and its connection to denoising score matching lead to a simple noise-prediction training objective.

NeurIPS 2020. Ho, Jain, and Abbeel. 50 min read, medium difficulty.