Stability AI released Stable Diffusion XL 1.0 on 26 July, making its flagship image generator’s weights available alongside source code. The launch follows the research-only SDXL 0.9 release and targets native 1024 × 1024 images.
Illustration: Sulat. Logo belongs to Stability AI. Independent editorial use; no endorsement.
The company says users preferred its output to earlier open image models in preference testing. It also promotes improvements in hands, text and spatial arrangements. These are Stability AI’s claims, not Sulat test results. Further image-control features remain forthcoming.
SDXL can generate an image with its base model alone. An optional refiner handles the final denoising steps in a two-model pipeline. Developers therefore do not have to run both models for every image.
The release-day model card is more cautious about quality than the announcement: it says legible text remains beyond the model, difficult spatial compositions can fail, and faces may be generated incorrectly. It also warns about social biases and describes intended use as research. Those limitations matter when judging the launch’s polished examples.
The weights use the CreativeML Open RAIL++-M licence dated 26 July 2023. It allows use and redistribution subject to restrictions, including bans on certain harmful uses, and requires downstream users to retain those restrictions. It does not license the training data or grant logo rights.
The accompanying repository’s code has an MIT licence. That permissive code licence does not replace the separate terms governing model weights. Developers can inspect and adapt the implementation while still having to respect the model’s use restrictions.
Sources: Launch announcement, Release documentation, Model licence, Code licence, Stability AI logo.
Discover more from Sulat AI & Tech News
Subscribe to get the latest posts sent to your email.
