Stability AI releases lighter version of Stable Diffusion 3

Stability AI has released a lighter version of Stable Diffision 3. The model, called Stable Diffusion 3 Medium, has two billion parameters and can run on graphics cards that consumers can have at home.

Stable Diffusion 3 Medium

The Medium version of the model, which allows users to generate images with text prompts, requires 5GB of VRAM to function, although the company recommends at least 16GB of VRAM. This makes it less demanding than the regular version of Stable Diffusion 3, which has eight billion parameters.

It works with graphics cards from Nvidia and AMD, among others, the company reports. The model can be found on Hugging Face. Stability AI already released the regular version of Stable Diffusion 3 earlier this year. Stability previously said that Stable Diffusion 3 is a combination of diffusion transformer architecture and flow matching used. It is the first time that Stable Diffusion has been given these techniques. Diffusion transformer architecture is a process of breaking down and generating a data model into small pieces, creating an image from multiple of those small pieces rather than as a whole. Flow matching is a technology that allows images to be better generated from noise without having to constantly repeat the calculation. That should make the generator a lot faster. Furthermore, Stable Diffusion 3 would contain new techniques to prevent misuse of the model.