Stability AI press release:

Key points to remember:

• Today we present Stable Diffusion 3.5. This open version includes several model variations, including Stable Diffusion 3.5 Large and Stable Diffusion 3.5 Large Turbo. Additionally, Stable Diffusion 3.5 Medium will be released on October 29.
• These models are highly customizable based on size, run on consumer hardware, and are free for commercial and non-commercial use under the permissive Stability AI Community License.
• You can download Stable Diffusion 3.5 Large and Stable Diffusion 3.5 Large Turbo from Hugging Face and the inference code on GitHub NOW.

Today we are releasing Stable Diffusion 3.5, our most powerful models yet. This open version includes several customizable variants, running on consumer hardware and available for use under the permissive Stability AI Community License. You can download the Stable Diffusion 3.5 Large and Stable Diffusion 3.5 Large Turbo models from Hugging Faceand the inference code on GitHub NOW.

In June, we released Stable Diffusion 3 Medium, the first open release in the Stable Diffusion 3 series. This release did not fully meet our standards or the expectations of our communities. After listening to valuable community feedback, instead of a quick fix, we took the time to further develop a release that advances our mission of transforming visual media.

Stable Diffusion 3.5 reflects our commitment to giving builders and creators widely accessible, cutting-edge, and free tools for most use cases. We encourage the distribution and monetization of work across the entire pipeline, whether it’s fine-tuning, LoRA, optimizations, applications, or illustrations.

What is published

Stable Diffusion 3.5 offers a variety of models developed to meet the needs of scientific researchers, hobbyists, startups and businesses:

• Stable diffusion 3.5 large: With 8 billion parameters, with superior quality and rapid adhesion, this basic model is the most powerful in the Stable Diffusion family. This model is ideal for professional use cases with a resolution of 1 megapixel.
• Stable Broadcast 3.5 Grand Turbo: A distilled version of Stable Diffusion 3.5 Large generates high-quality images with exceptional fast adhesion in just 4 steps, making it considerably faster than Stable Diffusion 3.5 Large.
• Stable Diffusion 3.5 Medium (released October 29): With 2.5 billion parameters, with MMDiT-X architecture and improved training methods, this model is designed to run “out of the box” on consumer hardware , striking a balance between quality and ease of customization. It is capable of generating images with a resolution between 0.25 and 2 megapixels.

Development of models

When developing the templates, we prioritized customization to provide a flexible foundation to build upon. To achieve this, we integrated query key normalization into transformer blocks, thereby stabilizing the model training process and simplifying further tuning and development.

To support this level of flexibility downstream, we had to make some compromises. Greater variation in results from the same prompt with different seeds may occur, which is intentional as it helps preserve a broader knowledge base and diverse styles in the base models. However, prompts lacking specificity may result in increased uncertainty in the outcome and the aesthetic level may vary.

For the Medium model in particular, we have made several adjustments to the architecture and training protocols to improve quality, consistency, and multi-resolution generation capabilities.

Where models excel

The Stable Diffusion 3.5 version excels in the following areas, making it one of the most customizable and accessible image templates on the market, while maintaining high-level performance in terms of rapid adhesion and quality image:

• Customization: Easily adjust the template to meet your specific creative needs or create apps based on custom workflows.
• Efficient performance: Optimized to run on standard consumer hardware without heavy demands, especially the Stable Diffusion 3.5 Medium and Stable Diffusion 3.5 Large Turbo models.
• Diverse output: Creates images representative of the world, not just one type of person, with different skin tones and features, without the need for many prompts.
• Versatile Styles: Capable of generating a wide range of styles and aesthetics like 3D, photography, painting, line art, and virtually any visual style imaginable.

Furthermore, our analysis shows that Stable diffusion 3.5 large leads the market in terms of fast adhesion and rivals much larger models in terms of image quality.

Stable Broadcast 3.5 Grand Turbo delivers some of the fastest inference times for its size, while remaining highly competitive in image quality and fast adhesion, even compared to similarly sized non-distilled models

Stable Diffusion 3.5 Medium outperforms other mid-sized models, providing a balance between fast adhesion and image quality, making it a top choice for efficient, high-quality performance.

The Stability AI Community license at a glance

We are happy to release this pattern under our permissive community license. Here are the key elements of the license:

• Free for non-commercial use: Individuals and organizations can use the template free of charge for non-commercial use, including scientific research.
• Free for commercial use (up to $1 million in annual revenue): Startups, small and medium businesses, and creators can use the template for commercial use at no cost, as long as their total annual revenue or less than $1 million.
• Ownership of results: Retain ownership of the generated media without restrictive licensing implications.

For organizations with annual revenue greater than $1 million, please contact us here to inquire about a business license.

Our Commitment to Safety

We believe in safe and responsible AI practices and take deliberate steps to ensure integrity begins at the earliest stages of development. This means that we have taken and continue to take reasonable steps to prevent misuse of Stable Diffusion 3.5 by bad actors. For more information about our approach to security, please visit our Stable Security page.