A diffusion model trains on a forward process that gradually adds noise to data, then learns a network that predicts and removes that noise step by step at inference. Each denoising update refines a latent or pixel tensor until a coherent sample appears. The family name covers score-based, variance-preserving, and flow-matching variants that share the iterative denoising loop rather than a single architecture block recipe.
Why It Matters
Diffusion models power many image, audio, and video generators and sit beside autoregressive decoders as a major generative paradigm. Knowing the family separates denoising-based sampling from token-by-tokendecoding and connects to latent representations and conditioning signals used at each step. The architecture page places diffusion among other model families; this page is the dedicated reference for iterative denoising generators.
Simple Example
A latent diffusion image model starts from Gaussian noise in a compressed latent grid, runs forty denoising steps with a U-Net or transformer denoiser conditioned on a text embedding, then decodes the final latent through a VAE to pixels.
Common Confusions
Diffusion model is a model family and sampling paradigm, not a synonym for any single checkpoint or U-Net shape. Denoising generation describes the iterative output loop; diffusion models are the most common implementation of that loop but not the only one. Diffusion also differs from autoregressive generation: denoising updates the full spatial or channel tensor each step rather than appending one discrete token.