Stability AI apologizes for disappointing Stable Diffusion 3, promises 'much improved' model soonStability AI, a leader in open-source AI image generation with Stable Diffusion, has come under fire for its Stable Diffusion 3 "Medium" release.
The company has now admitted that the launch fell short of expectations in terms of quality and licensing. Users on platforms like Reddit pointed out various issues with SD 3, including peculiar problems generating images of people lying on grass.
Stability AI states that initial internal tests showed Stable Diffusion 3 Medium outperforming SDXL in prompt adherence, variety, detail, and overall quality. However, the community quickly identified weaknesses, such as issues with certain poses and words rarely present in the training data.
The company is now working to optimize the model and plans to release a "much improved version" of SD3 in the coming weeks. In addition, they are researching how best to use the new architecture and model, and intend to share their findings and those of the community to optimize its use.
Stability AI is also addressing concerns about the new license terms, which were more restrictive than previous licenses, by revising the license for individual creators and small businesses.
Individuals and organizations can install and run Stability AI models on their own devices for free for non-commercial purposes, including researchers publishing their results in scientific journals.
See also: decrypt.co/238871/stable-diffusion-3-license-revamped-amid-blowback