Artificial intelligence company Stability AI will release its image generation model Stable Diffusion 3 on Wednesday, June 12. The company announced this during AMD’s keynote. The new model was previously available as a closed beta.

To work on a wide range of devices, Stable Diffusion 3 can handle different numbers of parameters: from 800 million to several billion. In particular, the maximum of 8 billion is much higher than previous versions of the image generation model. Stable Diffusion XL 1.0, released last year, still has a maximum number of parameters of 3.5 billion.

During the keynote, AMD and Stability AI demonstrated the new model on the Instict MI300X accelerator card. Compared to the Stable Diffusion XL, the result was clearly better, with performance also increased.

Stable previously said that Stable Diffusion 3 is a combination of Structure of diffusion transformers And Flow matching user. This is the first time these techniques have been given for stable propagation. Diffusion transformer architecture is the process of dividing and creating a data model into small pieces, and creating an image from several of those small pieces rather than as a whole. Flow matching is a technique that allows images to be generated better than noise without having to constantly repeat the calculation. This would make the generator much faster. Furthermore, Stable Diffusion 3 will contain new technologies to prevent model abuse.