one year on
Stability AI announces Stable Diffusion 3 with scaled parameter options from 800M to 8B
The next generation text-to-image model promises better multi-subject handling, image quality, and spelling, available through waitlist for early preview.
Stability AI today unveiled Stable Diffusion 3, its most capable text-to-image model yet, offering parameter configurations from 800 million to 8 billion. The model is available for early preview through a waitlist, with a promise of eventual open release.
Stable Diffusion 3 adopts a diffusion transformer architecture combined with flow matching, similar to the underlying approach in OpenAI’s Sora video model. Stability AI CEO Emad Mostaque said the model can accept multimodal inputs and scale further, and that a video version is possible with sufficient GPUs. The company says the preview phase is intended to gather feedback and improve safety before a full open release.
The parameter range spans from 800M — suitable for mobile and edge devices — to 8B, aimed at high-end GPUs. By comparison, Stable Diffusion 1.5 is 983M params and SDXL is 3.5B. The HN thread saw users debating whether the 8B model’s hardware demands will push hobbyists toward the API, while others noted that quantization will likely keep local inference affordable. One commenter recalled that SD2.1’s adoption suffered from licensing and CLIP replacement issues, asking whether SD3 will follow a similarly restrictive path.
Noted that SD3 uses a new diffusion transformer similar to Sora, combined with flow matching, and will be released open source after preview. Emphasized the model's ability to scale across parameter sizes and potential for video generation with sufficient GPUs.
Discussed the implications of parameter range and hardware requirements, noting that the 8B model may require more VRAM but quantization and smaller variants could preserve accessibility. Some expressed cautious optimism about safety and licensing after SD2.1's adoption issues.
One year later — open only if you can handle spoilers
Stable Diffusion 3 eventually released open source in June 2024 to mixed reception, with strong image quality but lingering concerns over prompt adherence and safety filters. The model's transformer-based architecture influenced later open-weights image generators, though Stability AI's financial struggles overshadowed the launch.
The Weekly Replay · free by email
This week, one year ago — every Sunday.
One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.
Free · double opt-in · unsubscribe anytime · privacy