Stability AI, a startup that develops generative AI tools, has launched an early preview of Stable Diffusion 3, a new model that can create realistic and detailed images from text descriptions. The model is said to be the most capable text-to-image model ever released, with significant improvements over previous versions.
What is Stable Diffusion 3 and how does it work?
Stable Diffusion 3 is the latest iteration of Stability AI’s flagship product, Stable Diffusion, which allows users to generate images from text prompts. For example, if a user types “a cat wearing sunglasses and a hat”, the model will produce an image that matches the description.
Stable Diffusion 3 is based on a diffusion transformer architecture and flow matching, which are advanced techniques for modeling complex data distributions. The model has a range of sizes, from 800 million to 8 billion parameters, to suit different needs and devices. The model is also open-weights and source-available, meaning that users can run it locally and fine-tune it to their preferences.
What are the advantages of Stable Diffusion 3 over previous models?
According to Stability AI, Stable Diffusion 3 has several advantages over previous models, such as:
- Multi-subject prompts: The model can handle prompts that involve multiple subjects, such as “a dog and a cat playing chess”. Previous models struggled with such prompts and often produced blurry or distorted images.
- Image quality: The model can generate images with higher resolution, sharpness, and realism. The images also have less artifacts and noise, which are common issues in image synthesis.
- Spelling abilities: The model can correct spelling errors and typos in the prompts, such as “a hourse with wings”. Previous models would often generate images that reflected the errors, such as “a house with wings”.
How can users access Stable Diffusion 3 and what are the safety measures?
Stability AI has opened a waitlist for users who want to try Stable Diffusion 3 in early preview. The company says that this phase is crucial for gathering feedback and improving the model’s performance and safety before a public release. Users can sign up for the waitlist here .
Stability AI also says that it has taken and will continue to take reasonable steps to prevent the misuse of Stable Diffusion 3 by bad actors. The company has implemented various safeguards, such as filtering harmful or illegal prompts, monitoring the usage and outputs, and collaborating with researchers and experts. The company also adheres to its core values of openness, safety, and accessibility, and aims to empower users to unleash their creativity with generative AI.