Stability AI recently unveiled Stable Diffusion 3.5, a significant advancement in its open-source AI image generation technology.
This latest version includes several model variations tailored to different users, ranging from casual creators to those working on large-scale commercial projects.
The release comes after the launch of Stable Diffusion 3 Medium in June, which the company admitted fell short of expectations.
Acknowledging the feedback, Stability AI decided to take extra time to improve the technology rather than rushing to release an update. The result is a more refined and capable solution designed to meet the needs of its community.
Introducing Stable Diffusion 3.5, our most powerful models yet.
This open release includes multiple variants that are highly customizable for their size, run on consumer hardware, and are free for both commercial and non-commercial use under the permissive Stability AI Community… pic.twitter.com/KlyE8OjrxN
— Stability AI (@StabilityAI) October 22, 2024
The top-tier model, Stable Diffusion 3.5 Large, features 8 billion parameters and runs at a 1-megapixel resolution, positioning it as the most advanced model in the Stable Diffusion lineup. In addition, the Large Turbo variant provides similar image quality but speeds up the process by completing image generation in just four steps, dramatically cutting down processing time.
A Medium version, set to launch on October 29th, will include 2.5 billion parameters and offer image generation at resolutions between 0.25 and 2 megapixels. This version is designed to be more accessible, optimized for use on consumer-level hardware.
The new models integrate Query-Key Normalization within their transformer blocks, improving training stability and simplifying the fine-tuning process. However, this added flexibility introduces trade-offs, such as increased variability in output when using the same prompt with different seeds.
Stability AI has adopted a community-friendly license for these models. They are free for non-commercial use and available to businesses with annual revenues under $1 million. Companies exceeding this revenue threshold will need to obtain a separate license.
The company also underscored its commitment to ethical AI development, incorporating safety protocols from the outset. Planned future updates include advanced control features via ControlNets, which are expected to roll out after the launch of the Medium model.
These latest models can be accessed through Hugging Face and GitHub, with additional availability via the Stability AI API, Replicate, ComfyUI, and DeepInfra.
