Stability AI has announced the release of Stable Diffusion 3.5, marking a leap forward in open-source AI image generation models.
The latest models from Stability AI include multiple variants designed to cater to different user needs, from hobbyists to enterprise-level applications.
The announcement follows June’s Stable Diffusion 3 Medium release, which the company acknowledges didn’t meet expectations.
“This release didn’t fully meet our standards or our communities’ expectations,” Stability AI stated.
Rather than rushing a quick fix, Stability AI says it invested time in developing a more robust solution.
The flagship model, Stable Diffusion 3.5 Large, boasts 8 billion parameters and operates at 1 megapixel resolution—making it the most powerful in the Stable Diffusion family. Alongside it, the Large Turbo variant offers comparable quality but generates images in just four steps, significantly reducing processing time.
A Medium version, scheduled for release on 29th October, will feature 2.5 billion parameters and support image generation between 0.25 and 2 megapixel resolution. This variant is specifically optimised for consumer hardware.
The models incorporate Query-Key Normalisation in transformer blocks, enhancing training stability and simplifying fine-tuning processes. However, this flexibility comes with trade-offs, including greater variation in outputs from identical prompts with different seeds.
Stability AI has implemented a notably permissive community licence for the release. The models are free for non-commercial use and available to businesses with annual revenues under $1 million. Enterprises exceeding this threshold must secure separate licensing arrangements.
The company emphasised its commitment to responsible AI development, implementing safety measures from the early stages. Additional features, including ControlNets for advanced control features, are planned for release following the Medium model’s launch.
Stability AI’s latest image generation models are currently available via Hugging Face and GitHub, with additional access through platforms including the Stability AI API, Replicate, ComfyUI, and DeepInfra.
