“The text-to-image models in this release can generate images with default resolutions of 512x512 pixels and 768x768 pixels.”
Version 2 swaps in a new OpenCLIP text encoder and filters adult content out of the LAION-5B training set. Users found right away that it got worse at copying famous artists, which was the point. Stability is trying to look clean for investors and regulators. The open weights for version 1 are already on everyone’s hard drive.