Official release sample
CompVis / Stability AI
Aug 2022 · txt2img
One of the text-to-image samples shipped in the CompVis/stable-diffusion README at release.
Stable Diffusion 1.4
CompVis (LMU Munich) · Stability AI · Runway
In August 2022 Stability AI released the weights of Stable Diffusion 1.4, commercial use allowed. Built by Robin Rombach and Patrick Esser's CompVis group in Munich, it ran diffusion in a compressed latent space, so it ran on a consumer graphics card. Unlike DALL·E 2 and Midjourney, it could be downloaded, fine-tuned, and bolted onto anything.
For filmmakers the key was img2img: feed it a frame, nudge the output, feed it back, and you have animation. The Deforum notebook turned that loop into a tool, and ControlNet, AnimateDiff, and Stable Video Diffusion grew from it. Every open-weight video model since has followed the path 1.4 cleared.