Finally there’s an official release date for Stable Diffusion 3, and it has been set to june 12th 2024.
Watch the full announcement on the Stable Diffusion 3 Medium weight release at @computex_taipei by our Co-CEO, @chrlaf: https://t.co/lcVl0tEr8M 🎉 https://t.co/IoHXwZPYhE
— Stability AI (@StabilityAI) June 3, 2024
WHAT IS STABLE DIFFUSION 3?
A few months ago I did an early review of Stable Diffusion 3, and it’s important to keep in mind that the version that will be released next week is not the same version I reviewed. The version I reviewed was still being fine-tuned and could only be run through API, and was heavily censored.
For those who wants to read all the technical information and science behind SD 3 can find their research paper here:
Scaling Rectified Flow Transformers for High-Resolution Image Synthesis
I’m not going to pretend that I understand all the technical data myself, because I most certainly don’t. But I will write the dumbed-down information that I have gathered from here and there.
Stable Diffusion 3 will be released in 4 different sizes, starting with the medium one. The sizes are as follows.
- Small (1 billion parameters)
- Medium (2 billion parameters)
- Large (4 billion parameters)
- Huge (8 billion parameters)
On June 12th the medium (2B parameters) will be released and available for download at Huggingface and CivitAI.
IS 2 BILLION PARAMETERS A LOT?
it depends. If you look at the parameters alone, then Stable Diffusion XL which was released almost 1 year ago today (June 27th 2023), and it has 2.6 billion parameters. But it’s not all about how many parameters there are, it’s how they are being used.
Something that one of the staff members from Stability AI pointed out.
I can’t say exactly what all this means, but among other things Stable Diffusion is better at photorealism and won’t mess up faces and hands as much as previous models. It’s also capable to embed text in images with a lot higher accuracy than SD 1.5 and SD XL. Not to forget that all Stable Diffusion models, from SD 1 to SD 3 is completely free for personal use!
WHO CAN USE IT, AND HOW?
The medium checkpoint (2B parameters) that’s to be released first should be able to run on a graphics card with at least 6GB VRAM. That’s pretty amazing as my old computer with 6GB VRAM struggled with SDXL checkpoints, so that says alot about how much they have managed to effectivize running generative AI locally.
If you are a ComfyUI user, you will be able to run SD3 from day 1.
As for those of us using Forge or Automatic1111 we will just have to wait and see, since it seems neither of these are being regularly updated anymore.
The below image have been generated with the version of SD 3 that will be released first.
