Notes by Lex Toumbourou

Diffusion Models also known as Diffusion Probabilitistic Models are a type of generative model that learn to generate data by reversing a gradual, noise-driven diffusion process, inspired by Nonequilibrium Thermodynamics.

First introduced in Deep Unsupervised Learning Using Nonequilibrium Thermodynamics, in which they describe the key steps involved in training diffusion models.

Forward Diffusion: Add Gaussian Noise to data over multiple steps. Reverse Diffusion: Learn to reverse the process. Go from noise to data.