If that is legit and not cherry picked after 300 reiterations, I'm fucking blown away. Goddamn, that is accurate.
First try i got a similar image just dog/cat was reversed, but it was closer than i expected. Using sdxl
Well, Copilot tried I guess. Same prompt, though I had to add “generate a” to get it to make an image and for some reason it cropped out “photo of” in the final result:
Let's see how SDXL does @aihorde@lemmy.dbzer0.com draw for me a red sphere on top of a blue cube. Behind them is a green triangle, on the right is a dog, on the left is a cat
Here are some images matching your request
Prompt: a red sphere on top of a blue cube. Behind them is a green triangle, on the right is a dog, on the left is a cat
Style: fustercluck
It's ok SDXL, you tried
What a fustercluck
Yeah... SD up till now has been just really good at people but terrible at multiple concepts. I've been pretty impressed with Dall-E 3, hoping SD 3 catches up or surpasses it.
Given the infinite pockets of OpenAI, I doubt this is possible, But if they get close enough, the FOSS community is having weekly breakthroughs and can take it much further. Just look at how good the SD 1.5 finetunes and customization is by now
SD 1.5 needs something like controlnet and inpaint to get close to Dall-E 3. I'm just amazed how Dall-E can do all that without any extra work.
But yeah, really hoping 3 has the community friendly tunability with at least some of that power that Dall-E has.
Heh, that third picture with the blue cat face. Funny, the other cat has the colors of the dog it wanted, but turned it into a cat.
I see they trained their AI on Word clipart.
That’s not a cube tho
Could be a prism or a more complicated shape, but it could be a cube.
All faces of a cube are square. The face visible in the picture is definitely not square. Thus, no matter what the non-visible parts of the blue shape look like, it's for sure not a cube.
It's one atom thick.
Have they stated what license they'll be using for the model?
Probably the same as stable cascade
Is stable diffusion 3 released?
https://stability.ai/stablediffusion3
Wait list opened up today if you want to apply
Nah, I have Bing if I need closed source internet dependant image generators, I want this running on my computer
Oh, so it's not open source?
They did the same with SDXL, it was closed during beta, then they open sourced it after fine tuning.
They said it will be "open"
How many watt-hours did that cost?
EA cover
Discuss matters related to our favourite AI Art generation technology
