Spinning this off from the image-sharing thread to avoid a hijack, but also because it’s interesting.
Plenty of folks are pointing out that photography was also heralded as the the beginning of the end for artistic endeavors, and it was years before it was widely considered anything even approaching art.
From my perspective, there are two distinct areas of concern at play here.
The general feeling that using tools to generate art cheapens the process. In my anecdotal experience, artists - particularly those who have put lots of sweat equity into developing their skills - are more likely to be against the democratization of art. I think this argument is exactly as specious as all the other “art isn’t art if it isn’t good enough or difficult enough” arguments.
The concern that since neural networks are trained from existing images on the internet, their work is inherently derivative and dishonest. I don’t think this one holds up either, since most visual artists use other works as references and tools when developing their own skills. Art isn’t produced in a vacuum. Worries about attribution and credit are more relevant, but the “things on the internet can be used by anybody for anything” genie came out of the bottle many many years ago.
As someone who can word good but can’t art good, I’ve been getting a real kick out of the image generators. Getting the programs to make what you want them to make requires a firm understanding of language, semantic relationships, stylistic influences, and a dozen other things. It’s an exciting time to be alive!
It’s worth noting that this win was in the “Digital Arts/Digitally Manipulated Photography” category, which kind of makes the whole thing a nothingburger imho. Unless they had specific rules about what tools you were and weren’t allowed to use to “digitally manipulate” the photography, then using machine learning to do it seems perfectly reasonable.
If the fair doesn’t like it, they can add a “Deep Learning Art” category, but if Photoshop doesn’t already use Deep Learning algorithms to do some ordinary functions it will soon, and that distinction will become harder to maintain.
Also, there was a person in this loop. Somebody created the prompt, ran it many times, selected the images to submit, etc. These are all artistic endeavors. It may be a different ‘kind’ of art, but I believe it is in fact art.
There’s a lot of philosophizing about art that I’m not qualified to do, but I think there’s a couple of other considerations to add to this. One is that the AI is incapable of doing something truly new. At some point in the history of all art, humans had to have done stuff that hadn’t been done before. Otherwise all art would be derivative of cave paintings or whatever. DALL-E can only combine existing images in new ways, which isn’t a criticism of the images it creates, as humans can also limit themselves in the same way and still produce novel and good art. But it is a criticism of (current) AI generated art in general. Let DALL-E try for a thousand years and it’s not going to be able to grow as an artist.
The other thing to consider is that DALL-E could, based on the inputs, output something that already exists. And it wouldn’t even know it. It’d be up to a human to weed out outputs that are pure plagiarism.
Outright plagiarism is a potential problem, though I don’t know how likely or unlikely it is. From my extremely limited understanding, the likelihood would be based on the size of the learning set and the number of steps the program uses to generate images.
One of the crazy things about these programs is that even their creators don’t know all they’re capable of producing. They can influence output by curating the learning sets and by blacklisting keywords, but that’s it.
Maybe one of these programs will spit out an innovation as revolutionary and (relatively) simple as linear perspective. Maybe it already has, but the person using the program didn’t have any way to recognize it.
Plagiarism is taking something (ideas or expression) from a source without acknowledging the source. It is not a legal matter.
Copyright infringement is using a creative and original work of expression without permission of the copyright owner. Copyright law acknowledges the possibility of two creators independently coming up with the same expression coincidentally.
The two can overlap, but they have different conceptual bases, different sources of authority, different standards for behavior, and different methods of enforcement.
I am not going to try to define “deep learning” precisely, but, for instance, the highly useful plug-in where you mask off an object (e.g., a person) in your photograph and— poof!— it’s gone, with the correct background seamlessly reconstructed behind it perhaps falls into that category. The newer inpainting tools now reverse this, so you can turn a bare sofa to one with a tortoiseshell cat sleeping on it with a few keystrokes in a quite different manner to cutting and pasting. And of course you can generate the entire image from scratch, either guided by a text prompt, or by image(s) you supply— there are so many possibilities, we should not start trying to enumerate them all here.
The only problem I can see is that many people do not understand exactly how these tools work, but a real artist who is also not afraid to dive into some technical aspects might, and that is where we are going to see some original and astonishing creations. Not limited to prints on canvas, either.
Absolutely. For example, one thing the machines currently lack is consistency. If my prompt draws a character I really like, I can’t tell it, “okay cool, now draw that guy from another angle.” Dall-E will never again create that same character.
But it can add to existing art. So if I can draw my own characters, I could have it create backgrounds and environments and throwaway characters he’s interacting with. It’s even capable of mimicking styles as it expands on preexisting works. Pairing the consistency of a visual artist with the unlimited flexibility of the neural networks is really exciting territory.
I am not an expert on the current tools (nor do I have the bank of A100’s necessary to be one), but some of these limitations may be with the user interface and/or scripts built around the model rather than with the model itself. (Without implying that the models do not suffer from severe limitations, of course.) For example, some reproducibility is achieved by saving the values of random seeds used. And as for drawing a character, I have not tried this but you could take a picture of that character and use that in combination with the text encoding, so that you fine-tune your model with the concept of [*] and then ask it for things like [*] in three-quarter profile, or [*] depicted as embroidery.
Naturally anything like this should also work much better with more sophisticated future models.
This just isn’t true. It is genuinely synthesizing new imagery. It is trained on existing images, but–what artist hasn’t?
It’s probably not sophisticated enough yet to come up with an entirely new art movement, like Impressionism or whatever. But only a few artists ever have. It is sophisticated enough to blend styles, and even when you specify an existing style, it has its own take on it. Just as existing artists working within a certain genre do.
That’s true in the limited sense that it is a fixed set of weights, and has deterministic output. But the weights will be retrained over time, and eventually I expect that (human) artists using these tools will train their own versions in a way that they do grow individually as artists. Is it the human or the AI that’s growing? Impossible to say.
It’s not really possible to be pure plagiarism. The dataset is too small. The one running on my computer is a handful of gigabytes. Not enough to store a tiny fraction of the imagery it comes up with. Even if it comes up with something similar to what’s out there, it did so very nearly from scratch.
This is like the claim that you could suffocate in a room full of air because the random motion of the air molecules could happen to just move them to places away from you. Like, mathematically that is true in some sense, but in reality, it will never ever ever happen.
There are way too many pixels for DALL-E to ever output an existing piece of art before the heat death of the universe.
I don’t have a cite handy because I was learning about this on youtube, so take my understanding for what it’s worth (not much). But the way it was explained, if the inputs don’t lead it to a varied result set, it’s likely that it’s going to produce something resembling to one of its training images. For example, if you ask it to produce a “Rothko Red and Yellow in the style of Mark Rothko,” it’s not going to get creative. It will, however, introduce randomness at several stages, so the result isn’t going to be a pixel perfect recreation, but it may be something that is easily recognizable by humans as, well, call it a “reproduction” if you’d like.
Actually if anyone has access and could run that, I’d be interested in seeing what it spits out.
Ok, I put in “Rothko Red and Yellow in the style of Mark Rothko”
And these are the 4 results I got:
Those are… honestly not even that close to the Rothko paintings I see when I search his name. Like, they are vaguely rectangular chunks of color that appear to be painted, but the Rothko paintings all have the rectangles painted against a field of another color. Certainly far from plagiarism. If I told a human painter “give me a Rothko-style orange and yellow and red painting” they could/would almost certainly produce something more similar to his actual paintings. I mean, I haven’t painted anything since art class in 7th grade and I bet I could produce a closer facsimile. While the discussion of artistic merit is obviously larger than this thread, I kind of struggle to understand where line would be drawn here. Surely Rothko doesn’t own the concept of painted rectangles in a few colors, even if he is a famous artist who painted in that style.
I went a step further and used the following prompt:
Rothko’s “Orange, Red, Yellow” in the style of Mark Rothko
Only one of them (the last one) even looks like a painting, and it would be very much out of place in a lineup of Rothko’s other color field paintings.
I suspect that, with enough fiddling, I could get it to output something that more closely resembles his works, but that would require actively attempting to ape his style.
Perhaps that ability will come when someone trains an AI to produce 3D models, which are then sent through a rendering engine to get the nice 2D art images. Save the model the AI produced, stick it in Poser or whatever and the artist can make it pose the way they want.
A lot (not all, but a lot) of what DALL-E produces is, effectively, just collages composed from its training database. Some of the images that get a lot of attention from it can’t be anything but, because it would be impossible to construct those images without a world-model far more sophisticated than it has.
But DALL-E isn’t the only art AI out there, and they don’t all work the same way. And some of them have, I think, created art. Like, when I asked Nightcafe Coherent for an oil painting of “Life, but not as we know it”, I had no idea it was going to give me a poignant scene of a hobo resting against a shipping container…
but there we go.
Well, it seems to me that Duchamp answered “what is art” pretty conclusively when he selected the urinal for “Fountain”. If you select it and present it to the world as art, it gets to be art. You selecting it is the important part, not how it is made. After that, it’s out in the world, and the rest of the world gets to decide if its good or not. They really don’t get to decide whether it’s art or not.
Even artists concentrating on the process of making their art will discard or withhold works they see as unsuccessful or unfinished. The act of selecting is the last step, and what makes it actually art.
So, in my mind, if you decide it’s art, and you present it as art, it gets to be art. It doesn’t matter if you ground your own pigments and wove your own canvas to make a painting, recorded wind chimes for hours and selected and edited the best 15 minutes of it together, fed some terms into a generative art program, or selected an industrial appliance and signed it. Your process is just your process for making art. That process can advise the viewer or the listener about whether they like it or think it is good (two different things), and the manufacturing process can be interesting in its own right. However, it seems to have little bearing on determining what is and isn’t art. It’s just how you used your tools.
