Stable Diffusion is published, and the weights are anyone's
Stability AI released Stable Diffusion to the public, publishing the weights themselves. A single consumer graphics card with around 6.9GB of VRAM could produce a 512×512 image in seconds. It was the first capable text-to-image model whose weights anyone could download. Built with the CompVis group at LMU Munich, Runway, LAION, and EleutherAI, it shipped under a Creative ML OpenRAIL-M licence permitting both commercial and non-commercial use. Unlike DALL·E 2, which was invitation-only, or Midjourney, which lived in Discord, it ran on your own machine and could be modified, and fine-tunes and derivatives multiplied immediately. The fights over the rights to the works in its training data start here too.