Stability AI has officially released Stable Diffusion 3 Medium, its most advanced open-weights text-to-image model to date. The new model, featuring 2 billion parameters, strikes a crucial balance between high-fidelity image generation and computational efficiency, making it accessible to a wider audience of developers, artists, and enthusiasts.
Unlike its much larger, unreleased counterpart, SD3 Medium is specifically designed to run effectively on standard consumer-grade GPUs, requiring significantly less VRAM. This move democratizes access to cutting-edge AI technology, allowing users to run the model locally on their own hardware rather than relying solely on cloud-based services. The model is available with open weights for non-commercial use and can be downloaded from Hugging Face. A commercial license is also available for enterprise applications.
Key improvements in Stable Diffusion 3 Medium include a major leap in photorealism and the ability to generate images with accurate typography, a long-standing challenge for diffusion models. According to Stability AI, the model overcomes common artifacts in generating realistic hands and faces, producing images with finer detail and fewer distortions. This is achieved through a new Multimodal Diffusion Transformer (MMDiT) architecture, which uses separate sets of weights for processing image and language representations, improving the model’s conceptual understanding and final output quality.
The company collaborated with Nvidia to enhance the performance of all Stable Diffusion models using Nvidia’s TensorRT, ensuring optimized inference speeds on RTX GPUs. This release positions Stability AI to compete directly with both proprietary models like Midjourney and OpenAI’s DALL-E 3, as well as other open-source alternatives, by offering a powerful, flexible, and resource-friendly option for the global creative and developer communities.
English 
Deutsch