Stability AI has introduced Stable Diffusion 3.5, their latest version of the open-source AI image generation models. This release brings multiple variants tailored for diverse user needs, ranging from casual hobbyists to large-scale enterprises. The new models emphasize improved performance and versatility, aiming to address previous limitations and enhance user experiences.
Stable Diffusion 3.5 builds upon earlier versions by significantly increasing parameter counts and optimizing processing speeds. Earlier iterations, such as the June release of Stable Diffusion 3 Medium, faced criticism for not meeting user expectations, prompting Stability AI to refine their approach for this new launch.
“This release didn’t fully meet our standards or our communities’ expectations,” Stability AI stated.
What Are the New Features of Stable Diffusion 3.5?
The flagship model, Stable Diffusion 3.5 Large, features 8 billion parameters and operates at a 1-megapixel resolution, making it the most powerful in the Stable Diffusion lineup. Additionally, the Large Turbo variant maintains similar quality but completes image generation in only four steps, effectively halving processing times.
How Do the Variants Cater to Different Users?
Alongside the Large models, a Medium version is set for release on October 29th, with 2.5 billion parameters and support for 0.25 to 2 megapixel resolutions. This variant is optimized for consumer hardware, providing a balance between performance and accessibility for everyday users.
What Is the Licensing Structure?
Stability AI has adopted a permissive community license for Stable Diffusion 3.5, allowing free use for non-commercial purposes and businesses generating under $1 million annually. Enterprises exceeding this revenue must arrange separate licensing agreements, ensuring wide accessibility while maintaining commercial control.
The company has incorporated Query-Key Normalisation in transformer blocks, enhancing training stability and simplifying fine-tuning processes. However, this flexibility introduces greater variation in outputs from identical prompts with different seeds. Stability AI continues to prioritize responsible AI development, implementing safety measures from the early stages. Future updates will include ControlNets for advanced control features following the Medium model’s launch.
Users can access the new models through various platforms such as Hugging Face, GitHub, and the Stability AI API. Additional access options include Replicate, ComfyUI, and DeepInfra, facilitating widespread adoption and integration into diverse applications.
The release of Stable Diffusion 3.5 by Stability AI represents a focused effort to enhance AI image generation capabilities. By addressing the shortcomings of previous versions and introducing adaptable models, the company caters to a broader audience while maintaining flexibility through its licensing policies. Users can access these models through various platforms, facilitating widespread adoption and integration into diverse applications.
