Stability AI publishes the Stable Diffusion weights
Stability AI, working with LMU Munich researchers, RunwayML, EleutherAI and LAION, published the full weights of Stable Diffusion on Hugging Face on 22 August 2022 under the CreativeML OpenRAIL-M license. The model ran on a single consumer graphics card.
Why it mattered Anyone could now run, alter, or sell work from a capable image generator on their own machine, and the arguments over training-data consent and copyright grew louder as they did.
Text-to-image systems in the middle of 2022 were reached on someone else’s terms. OpenAI announced DALL-E 2 in April and admitted users from a waitlist. Midjourney opened a public beta in July that ran inside a Discord server. In both cases the model sat on company hardware, and the company decided what it would draw and who could ask.
What Stability AI published on 22 August was the model itself. The weights of Stable Diffusion, a latent diffusion system built with researchers at LMU Munich and with RunwayML, EleutherAI and LAION, went onto Hugging Face under the CreativeML OpenRAIL-M license, which permitted commercial use. The file could be downloaded and run on a single consumer graphics card. It was the first capable text-to-image model of which that was true.
That moved the controls. Someone with a gaming computer could now generate images without an account, a queue, or a content filter, and could retrain the model, wire it into other software, or sell what came out. Tools built on the open weights followed within weeks. The developer Simon Willison published an assessment on 29 August under the title “Stable Diffusion is a really big deal”.
The objections arrived on the same schedule as the tools. Artists asked what had been in the training data and whether their work had been used without consent. A week after the release, an image made with Midjourney won a prize at the Colorado State Fair. In September, Getty Images barred AI-generated pictures from its marketplace, citing unresolved questions about copyright and training data. The lawsuits over both came the following January.
The difference between the three image generators of that summer was not mainly one of quality. DALL-E 2 and Midjourney could be changed, filtered, priced, or withdrawn by the companies that ran them. A downloaded copy of Stable Diffusion could not.