Stability AI partnered with Fireworks AI to launch the Stable Diffusion 3 API

Stability AI recently announced the availability of Stable Diffusion 3 (SD3) and Stable Diffusion 3 Turbo (SD3-Turbo) on the Stability AI Developer Platform API. Additionally, the models can now be accessed as part of the Fireworks AI distributed inference service to obtain an enterprise-grade API solution with a guaranteed 99.9% service availability. Stable Diffusion 3 and Stable Diffusion 3 Turbo will be available for self-hosting as part of the Stability AI Membership when the remaining details are finalized.

Stable Diffusion 3 is a state-of-the-art matching or outperforming other text-to-image models, including DALL·E 3, Midjourney v6, and Ideogram v1, in typography and prompt adherence according to the results from extensive human preference evaluations. The model's groundbreaking performance is due to its novel Multimodal Diffusion Transformer (MMDiT) architecture, which uses separate sets of weights for image and language representations, thus improving text understanding and spelling capabilities compared to its predecessors. In parallel with the API launch, Stability AI also announced the Stable Assistant Beta early preview, where a limited number of users have been invited to test the Stable Assistant capabilities that simplify content creation using Stability AI's language and image models.

Fireworks AI has revealed the benchmarking results for SD3 and SD3-Turbo, placing the former at 3.8 seconds per image at a resolution of 1024x1024 (although enterprise deployments can be optimized up to 1.8 seconds per image), and the latter at 0.37 seconds per image at the same resolution. Fireworks has also published a tutorial to get started with SD3 and SD3-Turbo using their distributed inference service. The company has yet to update its pricing table, so it is unclear whether the prices currently listed for image models will be applicable for SD3 and SD3-Turbo in pay-per-usage settings.