Diffusion models generate high-quality images by learning to reverse noise addition

Diffusion probabilistic models generate data by learning to reverse a fixed forward diffusion process that gradually adds Gaussian noise to data until it becomes pure noise. This forward process is a Markov chain with known Gaussian transitions adding small noise increments.

The reverse process is a learned Markov chain starting from pure noise and progressively denoising to produce samples matching the data distribution. Both forward and reverse processes use Gaussian transitions, enabling tractable variational inference training.

Training optimizes a variational bound on negative log likelihood, decomposed into KL divergences comparing learned reverse transitions to forward process posteriors, plus a final reconstruction term. The forward process variances are fixed, simplifying training and ensuring the reverse process matches the forward noise schedule.