News for hackers. Stuff that matters.
Re:Stable Diffusion 2.0 and the Importance of Negative Promp... by lancesells on Monday November 28, 2022 @09:11PM (#33782803)
I've always thought all the modern AI tools have an "AI illustration" style but the realistic images in that tweet are amazing and 1& uncanny valley. It's like I could be fooled until I really give a good look. I guess it's kind of the same on the illustrative stuff that look really good until you see the shadows are coming from different light sources in multiple parts of the image or there are only three fingers on the hand.

All in all I hate it because the prompts I see are things like "cyberpunk forest by Salvador Dali". You've got a tool that gives you the power of Gandalf and you prompt that?

Re:Stable Diffusion 2.0 and the Importance of Negative Promp... by ethbr0 on Monday November 28, 2022 @09:32PM (#33782970)
From what I've seen, AI generated images tend to be locally-consistent but holistically-inconsistent.

Which works, because most people on the internet tend to be detail-oblivious!

Re:Stable Diffusion 2.0 and the Importance of Negative Promp... by Filligree on Monday November 28, 2022 @10:32PM (#33783380)
I find the opposite. But perhaps we're looking at different elements? How would you describe this one here?

https://usercontent.irccloud-cdn.com/file/2csfvKjL/image.png

Re:Stable Diffusion 2.0 and the Importance of Negative Promp... by ethbr0 on Tuesday November 29, 2022 @10:29AM (#33788743)
I'd ask (1) what the zipper on her left chest is for, (2) where the necklace for the charm hanging along her centerline is (or zipper if it's a zipper pull), and (3) and how the geometry / gravity works on the patch on her left chest.

Which is about what you'd expect from a generator that understands patterns, but not meanings.

They miss on things that cannot be, because they don't understand things or rules, only patterns.

Re:Stable Diffusion 2.0 and the Importance of Negative Promp... by Terretta on Tuesday November 29, 2022 @01:31AM (#33784335)
If you mean a finger is always adjacent to a finger, locally, but holistically, the model doesn't know to stop after 4 of them, and will happily generate 8 fingers, then yes.

If you mean locally as in the size of a hand being right while holistically the person is wrong, no.

The overall images "tend" to be right (once you grasp prompting) and elements, even appear right at first, but if you focus attention on those elements, they are often not quite right.

So perhaps it's the definition of local and holistic.

Re:Stable Diffusion 2.0 and the Importance of Negative Promp... by smrtinsert on Monday November 28, 2022 @10:24PM (#33783325)
Those are average at best. Eventually you'll see enough great SD images that you'll start questioning verified photos.
Re:Stable Diffusion 2.0 and the Importance of Negative Promp... by Semaphor on Tuesday November 29, 2022 @01:06AM (#33784193)
>I could be fooled until I really give a good look

Looking at the woman on the boat [0] closely, I would still 100 % believe that’s simply a still from a movie, probably from the 90s.

[0]: https://nitter.kavin.rocks/pic/orig/media%2FFihVvliXwAEdw0y....

Re:Stable Diffusion 2.0 and the Importance of Negative Promp... by notahacker on Tuesday November 29, 2022 @06:47AM (#33786362)
> All in all I hate it because the prompts I see are things like "cyberpunk forest by Salvador Dali". You've got a tool that gives you the power of Gandalf and you prompt that?

That's one of the better prompts I've seen. Dissimilar but really strong aesthetic styles a skilled human could mesh pretty well, interesting images and shows up the strengths (some of the forests are really good, and the ones without trees are pleasantly foresty nevertheless) and weaknesses (it fails completely on 'cyberpunk' and 'Dali' once you start adding other parameters that influence the visual style) of the model.

Plus I'd be much more likely to end up with a calendar of "cyberpunk forest by Salvador Dali" images on my wall than "Mickey Mouse in a tuxedo with a cigar"

Re:Stable Diffusion 2.0 and the Importance of Negative Promp... by jonathanstrange on Tuesday November 29, 2022 @07:13AM (#33786568)
They still have an "AI illustration" style. Supposedly photorealistic images tend to look extremely photoshopped and humans in them look like they've been 3d rendered (albeit with a very high quality). They look like the heavily edited "plastic-like" images on covers of magazines.

I'm pretty sure this problem is not hard to fix in the long run, though.

Re:Stable Diffusion 2.0 and the Importance of Negative Promp... by ghaff on Tuesday November 29, 2022 @08:31AM (#33787318)
A friend of mine and I (independently) spent a day or so playing around with Stable Diffusion recently. We both came to the conclusion that, as things stand now, creating images in the style of impressionists/surrealists/cubists etc. works best because you're not really expecting realism, anatomical correctness etc.

I was able to come up with someone paddling a canoe in a Turner seascape. The only thing I couldn't get right was a proper canoe paddle and paddling motion but everything else was pretty much perfect.