Addendum: there are a fair number of artists who weren't "good" from a conventional craft POV and their "artistic value" comes entirely from the interesting way in which they were uniquely fucked up. But AI slop isn't unique or interesting, which is sort of the whole deal.
I saw an interesting observation that part of why AI images are so immediately recognisable is that they combine a high level of technical detail with a complete lack of compositional understanding. By the time a human artist could paint like that they would have learnt to compose a picture.
With AI "art" I come back over and over to an observation from Roger Ebert: "[bad movie director] knows that better directors tilt their cameras, but he does not know why."
there was actually a really interesting (IMO) thing with the guy who made the sword waifu back on Twitter. he showed off this AI image of a schoolgirl beheading a dragon, some real artists banged out the same concept as a dunk, all very funny the really interesting one was an MSPaint stick figure
because even though it was just barely more than stick figures, you could really see that sense of motion and composition in that stick figure image, while Shad's thing was just a bunch of detailed components in each other's general vicinity
📌
From a technical perspective, I naively would have thought that one of the first things you could teach AI is proper composition rules. If you train it on a lot of [illegally acquired] art with proper composition, why hasn't it pick that up?
Composition happens at a higher level of abstractions and that's the kind of thing LLMs don't do well. Knowing what pixels are statistically likely to show up next to one another does not tell it how to lead the eye or use sightlines and illumination to show what is important about an image.
Also it has been trained on a lot of images with descriptions added by cheap labour in the developing world. Those descriptions probably didn't include any compositional detail.
It doesn't know what composition is because it doesn't know what an image is, all it can do is blobs in relation to other blobs. It's why it struggles so much with the right number of fingers.
If you give two A1s the same training, will their output be the same from the same prompt?
Nope. What they generate is non-determinative. If you give a single model the same prompt twice the output will differ. This is why consistent characters are so hard w AI, & part of why AI video is so computationally expensive.
Shit, even the same model won't output the same image from the same prompt
Gen AI models have a degree of randomness generally that determines the odds that it chooses a less optimal outcome. If you set this value in the settings to zero (or use the same seed for the RNG in two runs) you will actually get the same result (unless the model is hiding it somehow)
Consumer facing models do not generally expose this because it breaks the image of "magic machine" and because always choosing the optimal next action actually breaks how these models operate (like how if you choose the next suggested word on your phone keyboard you eventually get it to repeat)
And they start from random noise each time they can't generate the same image twice.
The funny thing is that right wingers, as a rule, do not understand this. Since the days of Hitler's "degenerate art" exhibition the emphasis has been entirely on photorealistic technique. Real computational artists (people doing actual algoritmic art, not just prompting) hate genAI dorks too.
Oh man i had a conversation with a coworker about a year ago how art that isn’t photorealistic is arguably more important because with the advent of cameras we dont need paintings to capture whats “actually” there but to capture what the artist sees/feels and thats way more expressive
Also this
And in my view, if you want photorealistic art, take a photo.
I... Did the Al prompter actually compare himself to... Steven Spielberg?!? 😳
He did -- and, in his estimation, came out ahead!
I've said it elsewhere before, it's perversely aspirational to them, comparing themselves to Spielberg, because they want to be seen as Spielberg. But NOT Spielberg the auteur, but Spielberg the brand. The name alone that commands status, respect, and millions of dollars confidently attached.
They think you can just bypass all "unnecessary years" of: - honing your craft - networking and forming connections - proving people you can be relied upon to deliver - being someone people would WANT to work with AI is seen as an express elevator in a world full of retractable ladders to them.
That's a great point - it's not just that it takes multiple people to do the work, but even more people to do the work around the work. "If you build it, they will come" is already a key tech bro belief, and AI of course exacerbates that instinct.
They like genAI tech because they want to be artists without having to work for it. Not out of a drive to create; rather they want the praise and prestige of being an artist. Without having to work for it. They think it's unfair. Same energy as wanting straight pride parades and white history month.
They don’t understand that the process is part of the art. You don’t get Jaws by feeding AI a prompt, you get it by dealing with locals and a shark-bot that keeps breaking and actors who hate each other and pivoting when things don’t work!
Gromit was initially going to be a cat, and speak, but it was too hard to do that in clay; a dog was chunkier (so the clay model was more robust) and animating just the eyebrows was much easier. The limitations of the medium act as an enabling constraint.
They did make a cat later in Shawn the Sheep though which was quite well imo 😊
Favorably, even
Thus displaying for all to see that they've no clue what a film director does.
To be fair...i.dont actually know what a director -or a producer!- does either. But i know films dont get made without directors and producers running the project. So like... even with that level of knowledge
Indeed.
This is par for the course for the AI promoter. They think they’re some kind of amazing voice in the field of media analysis, but they’re… it’s funny actually to watch. The sense of righteousness while being a total fucking ghoul with a smile.
not only the ways in which they were fucked up but bc it came from human hands which is why generative art is interesting if the code was crafted by the artist, and not if it isn’t if you code a generative art (not “AI”) tool and you use it, that’s art. if i copy it, & run it again, it’s not.
people don’t consider greeting cards made from commercially bought stamps art but if you carve your own stamp from a block, that’s another story
(missing background: making elaborate greeting cards with rubber stamps, embossing die, cut outs, etc is a whole thing.)
There's so much art to it, too! You collect and curate a library of stamps, choose what goes together, choose composition, match colours, apply techniques to create the effect that ties it together. It's far more than the imagined printing an image on a folded bit of cardstock.
I've taught a beginner session of cardmaking, just the absolute basics of "here are the materials and tools, here's how to put things together, have a play with it, I'm here to troubleshoot or advise" and people are stunned to find out how much actually goes into it.
i agree there’s a lot to it! but it’s not art, it’s craft
Eh, the lines between art and crafts are blurry and based significantly in snobbery.
You bet your sweet rubber brayer it’s a whole thing.
“sweet rubber brayer” is a brand new sentence and i am here for it
The Ramones, "four great guys, three great chords."
Case in point: Anthony Kiedis. To a lesser extent: Will Ferrell. Coming in with that spice/
They can’t explain why the AI chose the style that it did or the color selection it made which a real artist would have no trouble with.
The only variants of generative visual ML that I found interesting when all this was new was the stuff that was essentially interactive, requiring the user to steer the process. Stuff like controlnet, tools that reorient / blend objects in images, style transfer, etc
Text prompts alone isn't more creative than commissioning an artist. It's their creativity, not yours.
Exactly. Prompting an AI isn't making art, it's commissioning art. It's like claiming Pope Julius II did the ceiling of the Sistine Chapel instead of Michaelangelo. Except in generative AI's case, "Michaelangelo" is also doing plagiarism.
This is why nothing AI can make will ever hold a candle to The Room or Troll 2.