Glossary

Diffusion model

A diffusion model is a model that generates images. It doesn’t start with a blank canvas but with pure image noise, like a badly disturbed TV picture. In many small steps, it removes part of this noise. After each step, a little more structure is visible, until a finished image stands at the end. In training, the model learned the reverse route: it was given real images that were made noisier step by step and had to compute the noise back out. A text, such as “a cat on a windowsill,” steers the direction the denoising takes.

An example: Stable Diffusion is an openly available diffusion model that generates an image from a text input. To make this faster, it doesn’t work directly on the pixels but on a heavily reduced intermediate form of the image.

Not to be confused with a language model: A language model generates text piece by piece from front to back. A diffusion model works on the whole image at once and refines it step by step. Both consist of a blueprint and trained parameters, but they compute in different ways.

Where you’ll come across it: In many image generators that paint pictures from a description, in photo-editing software with AI features, and in reports about AI-generated photos and artworks.

Introduced in What an AI Model Actually Is.

Last changed on · commit c0278b1