Stable Diffusion 3.5 opens new doors in digital art

This spring, Stability AI introduced Stable Diffusion 3, and now, half a year later, the next model, Stable Diffusion 3.5, has been released with notable improvements and new features for AI-generated art. This release includes three model variants: Stable Diffusion 3.5 Large, Large Turbo, and Medium – all available under the Stability AI Community License. These models are customizable, optimized for consumer hardware, and open for both non-commercial and limited commercial use.

Stable Diffusion 3.5 brings significant advancements in speed, performance, and accessibility. According to Stability AI, this release reflects their commitment to democratizing high-quality, AI-driven image generation. One of the most impactful enhancements in this version is improved image quality and resolution, powered by advanced algorithms that produce sharper, more detailed images, making it easier for artists to create high-quality, realistic artwork.

What to expect:

• Stable Diffusion 3.5 Large: Featuring 8.1 billion parameters, this model delivers highly accurate image generation ideal for professional use at resolutions up to 1 megapixel.
• Stable Diffusion 3.5 Large Turbo: A streamlined version of the Large model, offering similar quality at accelerated speeds by processing prompts in just four steps.
• Stable Diffusion 3.5 Medium: Optimized for consumer hardware, this model provides high-quality image generation on standard GPUs, requiring as little as 9.9 GB of VRAM.

Stable Diffusion 3.5 also highlights Stability AI’s emphasis on flexible and diverse applications. By incorporating Query-Key Normalization in the transformer blocks, Stability AI enables smoother training and customizable adaptations. Additionally, these new models are designed to produce varied outputs with minimal prompting, accommodating a wide range of user styles and use cases.

Released under the Stability AI Community License, Stable Diffusion 3.5 is freely available for non-commercial use, including research. In addition to self-hosting via Hugging Face, Stability AI has made the model accessible through several platforms, including the Stability AI API, Replicate, DeepInfra, and ComfyUI. Committed to promoting responsible AI use, Stability AI has introduced strict safety protocols to prevent misuse.

In the near future, Stability AI plans to release ControlNets to bring advanced control options for professional users, enhancing the range of potential applications.