Generative artificial intelligence startup Stability AI Ltd. today announced the release of what the company called its “most advanced” text-to-image open-source AI model within its Stable Diffusion 3 series, called Stable Diffusion 3 Medium.

The new AI model is built on 2 billion parameters. It supports features such as photorealistic image production that overcomes common artifacts in hands and faces, adheres to complex user text prompts and styles, can understand and render text without spelling errors, and is highly resource-efficient, according to the company.

With the Stability Diffusion 3 family of models, Stability focused especially on the models’ ability to accurately generate words and spell text correctly. One thing that text-to-image generators have struggled with is creating clear words and sentences from user prompts without spelling errors or producing pure gibberish. The company claims that SD3 Medium achieves much better results and attributes that to its Diffusion Transformer Architecture.

For fine-tuning, users can quickly and simply adjust the model using small datasets to customize its outputs. This makes it particularly nimble and easily focused for projects that need rapid turnaround even when there aren’t a lot of examples of a particular image to work with to get the model to train on a specific theme or picture.

With the smaller parameter size, SD3 Medium is condensed compared to heavier models, which weigh between 800 million and 8 billion parameters. This means it could be optimized to run on personal computers with consumer or gaming graphics processing units without performance degradation from its smaller VRAM footprint.

To tighten up its resource use, Stability said, the company collaborated with Nvidia Corp. to enhance the performance of all Stability Diffusion models, including SD3 Medium, but taking advantage of Nvidia RTX GPUs and TensorRT. Nvidia cards with TensorRT cores can provide a 50% increase in performance.

The company also collaborated with Advanced Micro Devices Inc. to optimize inference for SD3 Medium on the company’s devices including the company’s accelerated processing units and consumer GPUs.

For developers, Stable Diffusion 3 is available via the company’s application programming interface and the model weights are available open source to the community.

Image: Stability AI