Stable Diffusion 3 is now available via API, and Stability AI also plans to release the model weights for self-hosting.
The new Stable Diffusion 3 model and a Turbo version are now accessible via the Stability AI Developer Platform API thanks to a partnership with Fireworks AI. Stability AI says Fireworks AI will provide an enterprise-grade API solution guaranteeing 99.9% service availability.
Stable Diffusion 3 is said to match the performance of leading image models such as DALL-E 3 and Midjourney, but with advantages in prompt following and typography, according to Stability AI. This is achieved through a new Multimodal Diffusion Transformer (MMDiT) architecture that uses separate weight sets to represent images and text.
Besides API access, Stability AI intends to timely release the model weights for Stability AI members to self-host. The company is also inviting a small group of users to try the Stable Assistant Beta pre-release, which includes Stable Diffusion 3. In this limited release, Stable Assistant allows content creation using Stability AI's newest image and language models.
Stability AI says it has taken "reasonable steps" to prevent Stable Diffusion 3 from being misused by bad actors, starting with model training and continuing through the testing, evaluation, and deployment phases. Stability AI adds that it will continue to improve the model and security by working with researchers, experts, and the community.
Safety is especially important for Stable Diffusion as an open-source model. While Midjourney and other leading models have strong safeguards against misuse, earlier versions of Stable Diffusion, if modified, could be used to create pornography that exploits people.