Stable Diffusion 3 to debut on June 12
- Reference: 1717389014
- News link: https://www.theregister.co.uk/2024/06/03/stable_diffusion_3_release_schedule/
- Source link:
The date was revealed at the Computex conference in Taiwan today, where Stability AI CTO and co-interim CEO Christian Laforte revealed the news as he joined AMD CEO Lisa Su on stage.
Laforte told the event "A lot of thought went into this" – especially considerations of AI safety – so his firm feels the revised model is ready for wider testing.
[1]
He framed the release as a change in the way AI models are developed.
[2]
[3]
"It used to be that the frontier of research led to these models," he noted. "Now it seems like a natural evolution. By releasing models openly, we let people unlock new use cases."
Laforte didn't offer technical details of Stable Diffusion 3 (SD3) but used black and white images of a human hand hovering over an ornate acoustic guitar's strings to illustrate the advances in this edition – one created with SD XL 1.0 in 2023, and the other made with SD3.
[4]
"Anything that involves hands and repeating patterns is really hard," he explained, praising the image created with SD3 as offering "more realistic details like the shape of a guitar and hands."
[5]Hugging Face to make $10M worth of old Nvidia GPUs freely available to AI devs
[6]Stability AI decimates staff just weeks after CEO's exit
[7]Stability AI reportedly ran out of cash to pay its bills for rented cloudy GPUs
[8]What is Model Collapse and how to avoid it
The Register was able to squeeze off a quick snap of the images Laforte showed.
[9]
Stable Diffusion XL 1.0 vs SD3 demo – Click to enlarge
The co-CEO rated the above as containing "a few imperfections here and there," but "a big change" made possible by a multimodal diffusion transformer that allows it to better understand prompts.
Laforte also showed off images that he explained were generated by SD3 in less time than it took to enter the prompt.
He attributed the model's improved prowess to hardware offering more high-bandwidth memory. "Ninety-two gigabytes of HBM is a game changer," he beamed, before expressing hunger for even more powerful hardware.
"We need more memory and compute," he declared – an observation that saw AMD's Lisa Su observe "I think you need some GPUs." Su heartily approved Laforte's news that SD3 will run just fine on the Ryzen 300 AI CPUs Su [10]announced in the same keynote.
[11]
Stability AI has created a "weights list" for would-be users. Interestingly, the signup form includes a field for a Discord ID – perhaps reflecting SD's popularity among game developers. ®
Get our [12]Tech Resources
[1] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2Zl2URkdRZdOUwTsrbFuCUwAAAEM&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[2] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Zl2URkdRZdOUwTsrbFuCUwAAAEM&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Zl2URkdRZdOUwTsrbFuCUwAAAEM&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Zl2URkdRZdOUwTsrbFuCUwAAAEM&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[5] https://www.theregister.com/2024/05/17/hugging_face_nvidia_zerogpu/
[6] https://www.theregister.com/2024/04/18/stability_ai_layoff_staff/
[7] https://www.theregister.com/2024/04/03/stability_ai_bills/
[8] https://www.theregister.com/2024/01/26/what_is_model_collapse/
[9] https://regmedia.co.uk/2024/06/03/simon_sharwood_stable_diffusion_3_demo.jpg
[10] https://www.theregister.com/2024/06/03/amd_ryzen_300_ai/
[11] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Zl2URkdRZdOUwTsrbFuCUwAAAEM&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[12] https://whitepapers.theregister.com/
Re: Perfectly ordinary 7-string guitar
7 strings aren't common but do exist.
At least the fingers look more normal, the ones in the first pic look like they've missed a payment and had a run in with a loan shark's DIY tools
Re: Perfectly ordinary 7-string guitar
Yes, they exist. And SD 1 XL at least got string spacing reasonably correct. SD 3 does not. :)
Re: Perfectly ordinary 7-string guitar
They are bad but the image as a whole is a great improvement.
I suspect a little prompt tuning would fix the strings but even so, as a piece of creative work, it's far better as a whole.
(And my bad, I read your title as not knowing 7 strong guitars existed)
Re: Perfectly ordinary 7-string guitar
> Is this the best they can do?
Yeah, you'd hope they could string something better together...
Re: Perfectly ordinary 7-string guitar
Indeed, they should have spent more time tuning it.
Re: Perfectly ordinary 7-string guitar
You see a lot of porn generated using Stable Diffusion, they've managed to get faces correct although often too perfect looking, but as for the rest of the body, there are all sorts of weird arm, legs and fingers in places they shouldn't be - it's enough to put you off your stroke!
HW reqs?
1. if it needs 92 GB of VRAM to run, then realistically you can't run it locally: I'm not interested
2. no, the fact it can run on the CPU AI accelerators is not good enough: 7900XTX runs circles around M2 and M3 macs that have more than double the memory, and the Ryzen 300 CPUs won't have memory that's even half as fast as the stuff in Apple's SoCs
Perfectly ordinary 7-string guitar
And I can't help noticing that they are proud of going from slightly weirdly contorted dark hands to slightly weirdly muscular lighter hands.
Is this the best they can do?