NEW Details Announced - Stable Diffusion 3 Will DOMINATE Generative AI!

Stability AI has released the research paper for Stable Diffusion 3, their latest text-to-image model that pushes the boundaries of open-source AI. With improved image generation capabilities and a focus on safety and responsibility, SD3 aims to democratize AI art creation. The model introduces advanced techniques like conditioning on images, inpainting, and outpainting for greater creative control. Stability AI emphasizes the importance of AI ethics, aiming to develop models that are safe, unbiased, and transparent. Get ready to explore the cutting-edge of AI art with Stable Diffusion 3!

Tell us what you think in the comments below!

[Adreitz]

Correction -- SDXL does not use T5. You might be thinking of DeepFloyd IF. SDXL uses CLIP-L and OpenCLIP-G, which SD3 also uses in addition to T5. Also note that, from my reading of the paper, they are not "completely removing" T5 from SD3, but showing that SD3 can run either with or without it, and that removing T5 does not affect image aesthetics, though it somewhat negatively impacts text quality and prompt-following for complex prompts (notice that the ferret is not inside of the jar in any of the images generated without T5).

[jibcot]

Great video, I love the more technical depth you go into than most other YouTubers. Yes I will be using this model upon release, hopefully it will gets TensorRT support quickly and it gives the same 60% speed up as on SDXL.

[voxyloids]

UPBELIVIBLE researching speed. They developed this thing soo much fast OMG

[greypsyche]

Focusing on text and prompt adherence are important but what about the overall quality of the generations? Are they drastically better, almost the same, or worse? I'm especially interested in the quality of hands/fingers. These models struggle a lot with that.

[fimbulvntr]

Yep, it will be better. IMO SDXL tunes already outperform everything else out of the box, and once you start adding workflows into the mix it's not even a competition.

SD3 will take the crown instantly, though it might be a few weeks or months before it's in A1111 or Comfy with all the controlnets and the other bells and whistles, and a few months after THAT before we start seeing the really great finetunes show up.

At that point, they'll probably come out with a video model, taking the crown from OpenAI+Sora too.

StabilityAI is such a godsend ❤

[lindsay]

I think, when they talk about language models and understanding text better, they are not talking about rendering text.

[HolidayAtHome]

so are loras compatible between the 8B and 800M model of SD3?

[viddarkking]

The SD3 base model no.
But after it's fine tuned.
Yes I very much will.

[DoctorMandible]

Wake me up when it can be used commercially

[Earthball_Productions]

5 months later, it failed, Flux is the new open source winner!

[aa-xnhc]

this video did not age well + too much click-bait title