Denoising Diffusion Probabilistic Models

DDPM made diffusion models practical for high-quality image synthesis: corrupt data with a fixed Gaussian chain, train one time-conditioned U-Net to predict the added noise, then reverse the chain from pure noise. Follow the closed-form training shortcut, the simplified objective, and the evidence behind the paper's sample-quality claims.

Reading focus: How a closed-form forward marginal lets training jump directly to any noise level instead of simulating every earlier corruption step. Why predicting injected noise turns the reverse-process objective into a simple denoising regression problem across timesteps. What the CIFAR-10 ablations establish about the simplified objective, and why 1,000-step sampling remained the central cost.

NeurIPS 2020. Jonathan Ho, Ajay Jain, Pieter Abbeel. 13 min read, hard difficulty.