Stability AI says the model will produce more photorealistic images.
(Bloomberg) — Artificial intelligence startup Stability AI is releasing a new model for generating images that it says can produce pictures that look more realistic than past efforts.
In a blog post Thursday, Stability AI, which popularized the Stable Diffusion image generator, calls the new model SDXL 0.9, short for for Stable Diffusion XL. Stability says the model can create images in response to text-based prompts that are better looking and have more compositional detail than a model called SDXL, which it introduced in April.
In an effort to illustrate the differences between versions of the model, London-based Stability AI is publishing images generated by SDXL 0.9 and those created with an earlier technology. For example, a prompt including the words “manicured hand holding up a take-out coffee” generated a more convincing-looking image of coffee with the newer AI model than the previous one, which generated fingers that didn’t look quite right.
The release comes as concerns are growing over increasingly realistic, seemingly AI-generated deep fakes. Several have gone viral on social media in recent months, including a fake image of an explosion at the Pentagon and a fake picture of Pope Francis in an ankle-length, belted white puffer coat.
Stability AI Chief Executive Officer Emad Mostaque acknowledged the concerns during an onstage interview at the Bloomberg Technology Summit in San Francisco on Thursday. The company had “photo realistic models’’ in its possession earlier this year that it decided not to release because “the time wasn’t quite right,” he said.
“You had to build in things like watermarking,” he said during the interview. “Part of what we want is to build standards around that, so you can track it and so you can use it appropriately. You want to have some way of telling what is generated and what is not.”
Read More: Altman Says World Shouldn’t ‘Overdo’ AI Laws: Tech Summit Update
The company said the improvements announced Thursday are due to a higher number of parameters — which refer to the variables an AI system learns through its training — in SDXL 0.9 as compared with the beta SDXL model.
Stability AI initially rose to prominence by popularizing Stable Diffusion, an open-source image generator that it released publicly in August 2022. Stability did not create Stable Diffusion, though the company said three of the original researchers work for the startup.
Stable Diffusion emerged as a key rival to OpenAI’s image generator Dall-E, and Stability AI has since trained and released numerous versions of the product.
The company said that SDXL 0.9 can be used via Clipdrop, which offers AI image generating and editing tools; Stability announced in March that it acquired Init ML, Clipdrop’s maker. Stability AI said its application programming interface customers will soon have access to the latest model, which will be added to its DreamStudio image-generation software. Stability AI plans to release SDXL 0.9 in mid-July as open-source software.
(Updates with comments from CEO starting in fifth paragraph.)
More stories like this are available on bloomberg.com
©2023 Bloomberg L.P.