Finally also with grass: Stability AI makes Stable Diffusion 3.5 available
Stability AI has made the image generator Stable Diffusion 3.5 available in several versions – Large, Large Turbo and Medium.
Stable Diffusion 3.5 is available in three versions that are adapted to different needs. According to Stability AI, the image generators can run on consumer hardware and are easy to adapt to individual requirements. They can be used free of charge for most applications – under the Stability AI Community License, even for commercial purposes. All models are available via Hugging Face and Github.
The Large model has eight billion parameters, making it the largest of the model family. According to the blog post, its resolution of up to one megapixel makes it suitable for professional use. The Turbo version is particularly fast and creates images in just four steps, according to the blog post. It is apparently the result of so-called model distillation, i.e. the knowledge of a larger model has been transferred to a smaller one. This method makes training easier.
Stable Diffusion 3.5 Medium is a small model that was only released a few days after the other versions. It is based on 2.5 billion parameters and an improved architecture, which should make 3.5 Medium particularly interesting for end users. It even achieves a resolution of between 0.25 and 2 megapixels.
Stable Diffusion 3.5 should be competitive again
Stability AI points out that compromises were made during development. Particular attention was paid to ensuring that the models are easy to adapt and can be used for as many purposes as possible. In return, the output also varies more – in terms of aesthetics or continuity, for example.
As bullet points, Stability AI emphasizes that the models are customizable, do not require any special hardware to run, and that the people show a wide variety of types. Different styles are also possible. The quality of the images can keep pace with that of much larger models.
The previous model, Stable Diffusion 3, was not particularly well received by users. Shortly after its release, images emerged showing, among other things, human-like figures lying in the grass. Apparently, the image generator was not able to display healthy people in the grass – This is said to have been fixed, as the lead image of the current blog post shows. There had also been changes to the licenses. Stability AI admitted that it had not met expectations: "We are aware that our latest release, SD3 Medium, did not meet the high expectations of the community." They promised to do better and changed the licenses again.
(emw)