Stability AI’s Stable Diffusion 3.5: A New Benchmark in Open-Source AI Image Generation

Stability AI has made waves again with the release of Stable Diffusion 3.5, a significant step forward in open-source AI image generation. This latest release showcases Stability AI's dedication to innovation and refinement in the generative AI space, addressing various user needs, from casual creators to enterprise-level applications.

Moving Beyond Stable Diffusion 3.0

This announcement comes after the June release of Stable Diffusion 3.0 Medium, which Stability AI admits did not fully meet expectations. Acknowledging the community's feedback, Stability AI opted to take a more strategic approach, focusing on a robust, quality-driven update rather than a rushed fix. The result is a set of models that set new standards in image generation.

Stable Diffusion 3.5 Large: Power and Speed Combined

The flagship Stable Diffusion 3.5 Large model boasts impressive specifications:

- 8 billion parameters

- 1-megapixel resolution

This makes it the most powerful model in the Stable Diffusion family to date. Alongside the Large model, Stability AI has introduced a Large Turbo variant that offers similar quality but reduces processing time by generating images in just four steps. This enhanced speed optimizes workflows without compromising quality.

The Medium Variant: Optimised for Consumer Hardware

Scheduled for release on October 29th, the Stable Diffusion 3.5 Medium model is tailored to consumer hardware. With 2.5 billion parameters and support for image generation resolutions ranging from 0.25 to 2 megapixels, this model brings high-quality image generation within reach for users with more modest computing resources.

Enhanced Training Stability with Query-Key Normalisation

One of the key technical advancements in this release is the use of Query-Key Normalisation in transformer blocks. This innovation improves training stability and streamlines the fine-tuning process, allowing for greater flexibility. However, this enhanced adaptability also introduces more output variation when identical prompts are used with different seeds—a notable consideration for users seeking precise consistency in results.

A Licence with Flexibility in Mind

Stability AI has implemented a permissive community licence for this release. The models are available for free for non-commercial use and for businesses with annual revenues under $1 million. For larger enterprises, Stability AI offers custom licensing options, providing a flexible approach to commercial use and supporting smaller businesses and developers.

Commitment to Responsible AI Development

Throughout development, Stability AI has remained committed to responsible AI practices. Safety measures were integrated from the beginning, and Stability AI plans to release additional features—such as ControlNets for advanced control options—following the Medium model's launch.

Accessing the Models

Stable Diffusion 3.5 models are accessible on platforms such as Hugging Face and GitHub, as well as via Stability AI API, Replicate, ComfyUI, and DeepInfra. This wide accessibility underscores Stability AI's mission to democratize access to powerful AI tools.