Stability AI has launched Stable Diffusion 3.5, a powerful image generation model that marks a new era in open-source AI-driven creativity. Building on its prior models, Stable Diffusion 3.5 aims to provide more powerful image generation capabilities to users across various sectors, from hobbyists to enterprises. This release follows June’s Stable Diffusion 3 Medium model, which fell short of community expectations. Acknowledging the limitations of the previous version, Stability AI has invested considerable time to ensure that Stable Diffusion 3.5 meets the demand for more powerful image generation features.
The flagship model, Stable Diffusion 3.5 Large, includes 8 billion parameters with processing at 1-megapixel resolution, making it the most robust offering in the Stable Diffusion 3.5 family. Accompanying it is Stable Diffusion 3.5 Large Turbo, a version specifically tuned for speed while maintaining powerful image generation quality, completing images in just four steps for reduced processing time. Both models promise advanced, powerful image generation with customizable, efficient workflows to meet the needs of modern creators.
Stability AI is also set to launch Stable Diffusion 3.5 Medium on October 29. This model will incorporate 2.5 billion parameters and allow for powerful image generation on consumer-grade hardware, supporting image resolutions from 0.25 to 2 megapixels. The Stable Diffusion 3.5 Medium is optimized to provide consistent, powerful image generation on devices with varying processing power, ensuring that users across all levels can access its capabilities.
With the release of Stable Diffusion 3.5, Stability AI introduces enhancements like query-key Normalisation in transformer blocks. This addition increases training stability and facilitates easy fine-tuning, although it also introduces some variability in outputs based on prompts and seeds, making it a flexible tool for powerful image generation across different artistic styles and applications.
Stable Diffusion 3.5 is available under a community license that allows free use for non-commercial projects. Additionally, businesses with annual revenues below $1 million can leverage this powerful image generation model without additional licensing costs. Enterprises exceeding this revenue limit will need to secure separate licenses to access Stable Diffusion 3.5 for commercial purposes.
Responsible AI is at the forefront of Stability AI’s mission, with safety protocols integrated from the early development stages of Stable Diffusion 3.5. Along with this, Stability AI has outlined plans to introduce ControlNets in future updates, expanding powerful image generation controls for users to refine and customize outputs even further.
To facilitate access, Stable Diffusion 3.5 and Stable Diffusion 3.5 Large Turbo are hosted on popular platforms such as Hugging Face and GitHub, with additional availability through Stability AI API, Replicate, ComfyUI, and DeepInfra. Both models are part of Stability AI’s broader strategy to democratize powerful image generation technology, ensuring creators have the freedom to experiment and build using open-source tools.
diffusionSince its inception, Stable Diffusion 3.5 has focused on customization, allowing users to fine-tune the model for various purposes. Stable Diffusion 3.5 is designed with compatibility for training LoRAs and offers robust, powerful image generation features for standard consumer hardware. Stability AI’s emphasis on diverse outputs—from 3D images to painting and photography styles—makes Stable Diffusion 3.5 a versatile solution for creatives.
The release of Stable Diffusion 3.5 comes after the community’s critique of the previous model, Stable Diffusion 3 Medium, which was seen as lacking in some technical aspects, including accurate human anatomy representation. However, Stable Diffusion 3.5 builds on that feedback, refining powerful image generation to correct these issues. Stability AI underscores that this iteration is more than just a “quick fix”; it represents a thoughtful evolution in powerful image generation technology.
Stable Diffusion 3.5 maintains a similar architecture to its predecessor, with significant improvements such as QK normalization and double attention layers to ensure smoother, more flexible, and more powerful image generation. By using a permissive license model, stable diffusion 3.5 encourages community use while restricting only the creation of competing foundational models. However, customizations, including LoRAs and hypernetworks, are unrestricted, enabling further creative freedom in powerful image generation.
Later this month, the introduction of Stable Diffusion 3.5 Medium will broaden access to powerful image generation tools on consumer hardware, albeit with minor compromises in quality. The Stable Diffusion 3.5 model code is available on GitHub, while Hugging Face hosts the model itself, and platforms such as Replicate, ComfyUI, DeepInfra, and Stability AI API provide alternative access options for the community to explore powerful image generation.
