Stable Diffusion is released

Robin Rombach and Björn Ommer's group at LMU Munich had shown in December 2021 that diffusion could be run not on pixels but in the compressed latent space of an autoencoder, which made it roughly fifty times cheaper. Trained on LAION-5B, a public dataset of five billion image–text pairs scraped from the web, with compute paid for by Stability AI, the resulting model fitted in the memory of a consumer graphics card. On 22 August 2022 it was released with its weights, under a licence that allowed almost any use.

Within days it was running on laptops, in browsers and, crudely, on phones. Within weeks it had been fine-tuned to draw particular people, styles and products, extended with ControlNet to follow sketches and poses, and wrapped in a hundred interfaces. It was also used to make sexual images of real people, to imitate living artists by name, and to flood art sites with output. Getty Images and a group of artists sued in January 2023.

Stable Diffusion is on this timeline at the top level because it settled a question. DALL·E 2 had shown what the technology could do and had kept it behind a waiting list and a filter. Stable Diffusion showed that once a model existed it would be free, and everything since, in images and then in language, has had to assume that.