Amid a series of controversies related to technical glitches and licensing changes, AI startup Stability AI has introduced its latest line of image-generation models.
The new Stable Diffusion 3.5 series is being touted as more customizable and versatile compared to Stability’s previous technology. According to the company, it is also more powerful. The series consists of three models:
Stable Diffusion 3.5 Large: This model boasts 8 billion parameters, making it the most robust in the series, capable of generating images at resolutions up to 1 megapixel. (Parameters essentially indicate a model’s problem-solving abilities, with models having more parameters generally performing better.)
Stable Diffusion 3.5 Large Turbo: A condensed version of Stable Diffusion 3.5 Large that produces images faster, although with a slight trade-off in quality.
Stable Diffusion 3.5 Medium: An optimized model for edge devices like smartphones and laptops, capable of generating images ranging from 0.25 to 2 megapixels.
While Stable Diffusion 3.5 Large and 3.5 Large Turbo are already available, 3.5 Medium is set for release on October 29. Stability claims that the new models should produce more varied outputs without needing extensive prompts.
Stay tuned as we await further updates on Stability’s approach. Unfortunately, there’s no early access impressions available at the moment.
Stability has addressed criticism from its previous flagship image generator, Stable Diffusion 3 Medium, regarding peculiar artifacts and prompt adherence. However, the company warns that Stable Diffusion 3.5 models might face similar issues due to engineering and architectural compromises. Yet, Stability assures that the models are more resilient when generating images across different styles, including 3D art.
One thing that remains unchanged with the new models is Stability’s licensing terms, catering to both non-commercial use and commercialization for small businesses. Larger organizations must secure an enterprise license.
Stable Diffusion 3.5 Large and Diffusion 3.5 Large Turbo are adaptable for self-hosting and use through Stability’s API and various third-party platforms. ControlNets for the models, allowing for fine-tuning, are expected to debut shortly.
Stability’s models are trained on public web data, which can be copyrighted or subject to restrictive licenses. While Stability relies on the fair-use doctrine to shield itself from copyright claims, data owners have initiated class action lawsuits against AI vendors.
Customers are responsible for defending against copyright claims, as Stability doesn’t offer financial assistance in such cases. On a positive note, data owners can request the removal of their data from Stability’s training datasets.
In the context of misinformation, especially surrounding upcoming elections, Stability emphasized taking appropriate measures to prevent misuse of Stable Diffusion. However, specific technical details on these measures weren’t disclosed.
Despite these challenges, Stability’s commitment to evolving its technology and addressing concerns indicates a proactive stance in the dynamic landscape of AI development.
