Stability AI

Contribution to AI

Text-to-image generation existed before August 2022, but only as a service: you sent a prompt somewhere and waited for a picture to come back. Publishing the Stable Diffusion weights — a latent diffusion architecture from the CompVis group at LMU Munich, developed with Runway, trained on LAION's public image-text corpus on compute Stability paid for — moved the model onto other people's hard drives. Because latent diffusion denoises in a compressed representation rather than in pixels, the released checkpoint ran on a consumer graphics card with roughly 8GB of memory, and that engineering fact determined everything that followed. Within months there were interfaces nobody at the company had written, fine-tuning methods that let one person teach the model a face or a style on a single GPU, and conditioning techniques such as ControlNet that added pose and depth control the original model never had. None of that required permission, and none of it was possible against an endpoint. The cost side is equally instructive. Weights that cannot be recalled put the training-data question in front of judges rather than policy panels — Getty Images sued in London and in Delaware in January 2023, artists filed in California the same month — and made one company the defendant of record for a practice the whole field relied on. Giving away the flagship also left it hunting for revenue it had already surrendered.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).