Last Updated: 5 November 2024
Stability AI, a leader in open-source image generation, has launched its latest offering: the Stable Diffusion 3.5 models. This new series aims to overcome past limitations by offering more varied outputs and improved prompt responsiveness, while addressing criticisms regarding image quality.
The Stable Diffusion 3.5 series includes three distinct models:
Stable Diffusion 3.5 Medium: Designed for consumer devices like laptops and smartphones, this model uses 2.5 billion parameters and produces images ranging from 0.25 to 2 megapixels. Its release is expected on October 29.
Stable Diffusion 3.5 Large: The flagship model with 8 billion parameters, capable of generating high-quality images up to 1 megapixel. It excels in prompt adherence, making it ideal for professional applications.
Stable Diffusion 3.5 Large Turbo: A streamlined version focusing on efficiency, balancing speed and output quality to deliver high-quality images rapidly.
Stability AI has ensured that these models depict individuals with different skin tones and features, thereby promoting greater representation. A unique training strategy involving multiple prompt versions helps the models overcome issues experienced by earlier versions, ensuring more culturally sensitive and accurate outputs.
Stability AI aims to address issues with earlier releases, such as poor prompt adherence and visual inconsistencies. The previous flagship, Stable Diffusion 3 Medium, faced criticism for creating bizarre artifacts and not meeting user expectations. The 3.5 models focus on improving prompt responsiveness and consistency across different artistic styles, including 3D art.
However, Stability AI acknowledges that some prompting issues may persist due to engineering trade-offs. While the goal is to maintain a broad knowledge base, inconsistent results may occur with less specific prompts, which is an intentional feature aimed at encouraging diversity in outputs.
The licensing terms for Stable Diffusion 3.5 remain consistent: the models are free for non-commercial use, and small businesses can use them commercially without cost. Larger enterprises require an enterprise license. Stability AI has also made adjustments to their licensing terms to allow more flexible commercial usage, enabling creators to freely distribute and monetize their work.
Like other AI image generators, the models are trained on publicly available data, which may include copyrighted material. Stability AI cites fair-use doctrine as its defense, though data owners have pursued legal actions. Stability allows data owners to request removal of their content from training datasets, with 80 million images removed to date.
Measures to prevent misuse of these models, especially in light of the upcoming U.S. elections, have been put in place, although details remain undisclosed. Stability AI prohibits misleading content but does not categorically ban content related to public figures or elections.
Stability AI views the Stable Diffusion 3.5 models as a significant advancement in the generative AI landscape, highlighting diversity, prompt adherence, and user ownership as core principles. Whether these improvements will meet community expectations and restore user trust remains to be seen.
