Stability AI has unveiled Stable Diffusion 3.5, a major advancement in open-source AI image generation, pushing the boundaries of what’s possible in this space.
The new models come in various versions, catering to a wide range of users—from hobbyists to large-scale enterprise applications. This release follows the earlier launch of Stable Diffusion 3 Medium in June, which didn’t fully meet user expectations.
Stability AI acknowledged this, stating: “That release didn’t meet our standards or the community’s expectations.” Rather than rushing a quick fix, the company focused on building a more robust solution. The flagship model, Stable Diffusion 3.5 Large, features an impressive 8 billion parameters and supports image resolutions up to 1 megapixel, making it the most powerful model in the lineup. Its Large Turbo variant delivers similar quality but accelerates the process, generating images in just four steps, significantly cutting down on processing time.
Scheduled for release on October 29, the Medium version will feature 2.5 billion parameters and generate images at resolutions between 0.25 to 2 megapixels. It’s optimized for consumer-grade hardware. These models also integrate Query Key Normalization in their transformer blocks, which improves training stability and simplifies fine-tuning. However, this added flexibility introduces some variability in outputs, particularly when using identical prompts with different seeds.
Stability AI is offering a community license for this release, making the models free for non-commercial use and for businesses with annual revenues under $1 million. Larger enterprises will need to negotiate licensing agreements.
The company reaffirmed its commitment to responsible AI development, embedding safety measures right from the start. Additional features, such as ControlNets for advanced control capabilities, will be rolled out after the Medium model’s release.
The latest Stable Diffusion models are available now on Hugging Face and GitHub, with further access through platforms like Stability AI API, Replicate, ComfyUI, and DeepInfra.
