Glossary Entry

Diffusion Model

A generative model that learns to reverse a gradual noising process, creating new samples by iteratively denoising pure noise.

Generative AI Models

Also called: diffusion models, denoising diffusion, DDPM

Seed source: Ho et al., Denoising Diffusion Probabilistic Models

A diffusion model is trained on a deceptively simple task: given a data point with a known amount of Gaussian noise mixed in, predict the noise. Because the noising (forward) process is fixed and has a closed form, every training example comes with an exact supervised target, which makes training far more stable than adversarial approaches like GANs.

At generation time, the model starts from pure noise and applies its denoiser repeatedly, walking backwards through the noise levels until a clean sample emerges. This iterative refinement is the engine behind modern image generators (Stable Diffusion, FLUX), video models (Sora, Veo), and diffusion modules in scientific models like AlphaFold 3.