What Are Diffusion Models?

Presentation overview

Chen Cheng will present an introduction to diffusion models based on Lilian Weng’s tutorial. Diffusion models generate data by learning to reverse a process that gradually adds noise to observed samples. The forward process corrupts the data through a sequence of Gaussian transitions; the learned reverse process starts from noise and progressively reconstructs a sample.

The tutorial develops the probabilistic formulation of denoising diffusion models and the training objective used to learn the reverse transitions. It also connects denoising to score-based generative modeling. The presentation will introduce these foundations and discuss how conditioning and guidance steer generation, along with approaches that reduce the cost of sampling.