Definition
A diffusion model defines a forward process that gradually corrupts data with noise and learns a reverse process that removes that corruption. In a common image setup, training selects a noise level, adds a known amount of Gaussian noise to an example, and teaches a neural network to predict the noise or an equivalent denoising target.
Sampling starts from random noise and applies the learned reverse transition over a sequence of steps. The result should resemble the training distribution or a requested condition such as text, a class label, or another image. The repeated steps make ordinary diffusion sampling slower than one-pass generation, although later samplers and distillation methods can reduce the number of evaluations.
From diffusion processes to image generation
Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli presented a deep generative model based on a forward diffusion process and learned reversal in 2015. Jonathan Ho, Ajay Jain, and Pieter Abbeel's 2020 denoising diffusion probabilistic model connected the approach to denoising score matching and demonstrated high-quality image synthesis. These papers established the modern model family; neither implies that every iterative denoiser is a diffusion model.
Distinguish it from nearby terms
A generative adversarial network trains a generator against a discriminator and often produces a sample in one forward pass. A diffusion model learns reverse transitions for a specified noising process and usually samples iteratively. A denoising autoencoder may remove corruption in one pass without defining the full generative chain.
Check your understanding
During training, a clean image is corrupted at a randomly selected noise level. During generation, the system begins with noise. Which direction is fixed by design, which direction is learned, and why are multiple reverse steps needed?