Nope, Duchamp was not an idiot. Also, he was making art. Anti-art is still art.
Do we agree that photographs can be art? If so, the ‘art’ of a photograh is not in the technical ability of the artist. It’s more about the ‘eye’ of the artist, and using photography to capture what the artist sees. The physiccal act of photography involves pushing a button.
Even in painting, and especially in modern art, the vision of the artist is what’s important. Mark Rothko could probably have supervised someone else to paint his works - they aren’t products of technical genius or exsquisite skill. The art come from the choice of medium, choice of colors, choice of size, etc.
I’m not sure i see much difference between that and using an AI as a ‘camera’ to produce a vision the artist has for something. I do think it’s important to disclose that AI aspect so the art is judged fairly for what it is.
All the buttons and settings and controls and bags of expensive lenses and bodies and filters and accessories and post-processing suggest otherwise, beyond the photographer’s pure imagination.
Even the ready-to-use AI generative art scripts we are talking about in this thread have several commands for various tasks as well as a few parameters to tweak, not to mention the problem of understanding how to control the input to get what you want as output.
That may or may not be difficult, but it’s not part of the ‘art’. There have been artistic shots made with iphones or brownie cameras. And you could have assistants to set up the lighting and such. You certainly need knowledge to get the most out of the equipment, but you also need knowledge to pull off really good AI art, if you are trying to actually match a vision in your head.
I think it’s cool that the AIs understand photography. You can have a prompt that says, “Portrait of X, 35mm lens, f2.8, lighting from right side at 5500K”, and it will faithfully recreate that. The act of actually turning an aperture ring on a physical camera doesn’t really add anything.
Well, there’s technical ability in photography, too. But certainly, art is more than just technical ability.
(I’m just using this quote as a reference point, not a direct response)
I haven’t played with Dall-E but the developers of MidJourney (used to create the image in the OP) said that at least one of their models had a tendency to come really close to copying art if asked. If you asked for the Mona Lisa, it would just throw up the Mona Lisa give or take some pixels because, hey, that one was easy. They took that build down and must have messed with it since now asking for the Mona Lisa directly gives you something like a blown out overexposed version or it hanging on a wall and seen at an angle, etc. You can still ask for the Mona Lisa wearing a hat or with Steve Buscemi eyes but not a straight up depiction of the original. So the concern about the AI plagiarizing content isn’t entirely without merit but it is addressable.
I think it also only really applies to a fair small set of art pieces. The Mona Lisa is easy for the same reason that getting a picture of Taylor Swift eating cake is easy – both the Mona Lisa and Taylor Swift are overrepresented in the model by sheer weight of fame so the AI makes very accurate depictions of them but has to get more creative with lesser known works of art or celebrities. Rothko’s art may be well known but obviously not Mona Lisa famous or having the same representation in the training model.
I asked Dall-e just now for the Mona Lisa and it produced paintings that were obviously her. None were nearly as good as the original - like what you’d get if you asked a bunch of amateur artists to recreate it. Still a much more faithful reproduction than the Rothko attempt.
There is no scale parameter to play with? I did not see one on Dall-E, but E.g. I went to the Stable Diffusion demo and asked for “Mona Lisa, oil painting” and if you drop the “scale” down enough, you can get all sorts of bizarre Monas
There is but it’s a manual tag. “Picture of Mona Lisa --s 4000” or whatever. Someone just playing with the site because they heard about it on Facebook wouldn’t be likely to know it and even a lot of experienced users don’t mess with it much unless they’re in the 2nd or 3rd stage of trying to make something.
Also, part of the concern was that the AI could (and would) make these accurate reproductions, not that you were locked into them. After all, you could always just not ask for the Mona Lisa.
Yes you can. https://arxiv.org/pdf/2208.12242.pdf
I only skimmed the paper, but it appears to be using multiple real-life images of the same subject “(~3-5)” as its initial input, and then using existing models to generate more photorealistic digital images of that subject.
It’s remarkable work, but it’s not the same as a generator being able to look at a single generated character and generating more examples of that same character in different poses and settings.
Your point was about lack of consistency, which has been demonstrated to be false. 3 images vs 1 is an engineering problem. If you think that demonstrates a fundamental inability of these models we’ll have to just agree to disagree for the next couple months until it’s improved.
Right- the paintbrush/paint/canvas, chisel/stone/wood, etc… are equivalent with the AI prompt/guidance by my way of thinking. It’s the tool used to make the art, no more, no less.
And there should absolutely be a different category for AI art, just like there are different categories for watercolors, oils, acrylic, fresco, encaustic, etc… Maybe even different categories for different types/styles of AI art even.
You, uh, you saw the word currently in the bit that you quoted? You’re inventing a position that I don’t hold.
That paper shows that the models can take multiple real-world images and make new ones. That’s awesome, but is also very different than taking a single AI-generated image and making new ones. Two completely different problems, only one of which is currently solved.