Stable Diffusion 3

A new diffusion transformer and better prompt understanding

Better prompts

Stable Diffusion 3 is the newest model from Stability AI. It converts text into images and does a better job than ever before. It’s especially good when dealing with prompts that have multiple subjects. Plus, the images it creates are of higher quality, and it’s even better at spelling!

The best balance

This new version includes 800 million to 8 billion parameters. It offers a range of choices to users, allowing them to pick the best balance between scalability and quality for their creative projects. Stable Diffusion 3 uses a special structure called a diffusion transformer and a technique known as flow matching.

Available now

Stable Diffusion 3 Medium has been public since June 2024, and Stable Diffusion 3.5 (Large, Large Turbo and Medium) since October 2024, under the Stability AI Community Licence: free for personal use and for businesses under one million dollars of yearly revenue.
You can try Stable Diffusion 3 Medium right here, without an account, or download the weights to run it locally with ComfyUI.

Try Stable Diffusion 3 Medium online

Stable Diffusion 3 is the third generation of Stability AI's text-to-image model, built on a diffusion transformer instead of a U-Net. Released as SD 3 Medium in June 2024 and improved as SD 3.5 in October 2024, it follows sentences, handles several subjects and renders short text. You can try it free on this site.

How to use Stable Diffusion 3

Open the Stable Diffusion 3 Medium demo on this site, no account needed.
Write a full sentence describing the scene, with text to render between quotes if needed.
Compare the result with SDXL on the same prompt to see the difference in prompt understanding.
To run it locally, download SD 3.5 Medium or Large and use ComfyUI, see the installation guide.

Limits and questions

Is Stable Diffusion 3 free?

Yes for personal use and for businesses under one million dollars of yearly revenue, under the Stability AI Community Licence. The weights are downloadable from Hugging Face.

How much VRAM does it need?

About 10 GB for SD 3 Medium in fp16, 16 GB or more for SD 3.5 Large. SD 3.5 Medium is the balanced choice for 12 GB cards.

Does it still need a waitlist?

No. The early-access waitlist of February 2024 is over: SD 3 Medium has been public since June 2024 and SD 3.5 since October 2024.

Read the guide: SDXL vs Stable Diffusion 3 vs SD 1.5