AI-generated Art

[Inkerrio]

The AI art is extremely useful when you need to quickly generate placeholders for UMC

[doofus-01]

The AI art is extremely useful when you need to quickly generate placeholders for UMC

Be careful with that. As far as I know, it's not settled whether AI-remix of internet images breaks copyright. If you download the program and build your own AI models, that's obviously different - or at least I think it is. I intend to try this out when I have the time to focus on it, I think it's generally a new tool like Photoshop was a few decades ago. But for those who think they can generate instant art like it was a (verbose) google search: Don't upload those images.

I mean, the third set is pretty awful, unless the woman is a stripper.

seems to be some bias in the AI for 'medieval armour' on women.

I'd say those are actually less bad than the AI-surreal metal blobs of the men. They each have some sort of personality, and if you can trust the anatomy (which you can't - yet), they can serve as models. The images aren't great if used directly, but then neither are the metal blobs or the furries.

[Inkerrio]

This is a popular misconception, the AI doesn't "remix internet images", it learns drawing patterns. In case of the most popular current technology, Stable Diffusion, the AI is taught to recognise shapes from random noise. Sort of like humans do when they scribble random pencil lines on the paper and then try to guess what it looks like.

The only legally grey area here is that it is unclear who owns the resulting image - the AI, its creators, or the human who uses the AI. But I wouldn't worry too much about this in a non-commercial project.

[gnombat]

This is a popular misconception, the AI doesn't "remix internet images", it learns drawing patterns.

Well, maybe, sometimes, the AI does that. Sometimes the AI just "memorizes" and "regurgitates" the original images it was trained with (or something very close to the original images).

[Inkerrio]

This is a popular misconception, the AI doesn't "remix internet images", it learns drawing patterns.

Well, maybe, sometimes, the AI does that. Sometimes the AI just "memorizes" and "regurgitates" the original images it was trained with (or something very close to the original images).

AI can recreate extremely well-known subjects such as the mentioned Mona Lisa but it is still a recreation from scratch, there's not a single pixel of the original image in the result. And for this to happen the subject has to be as famous as Mona Lisa. There's zero chance that a user will accidentally recreate some copyrighted game asset.

[Celtic_Minstrel]

AI can recreate extremely well-known subjects such as the mentioned Mona Lisa but it is still a recreation from scratch, there's not a single pixel of the original image in the result. And for this to happen the subject has to be as famous as Mona Lisa. There's zero chance that a user will accidentally recreate some copyrighted game asset.

You have a major misconception here. You're ascribing an "identity" to the pixels of the "original image" which simply does not exist. This is digital data. If you copy a pixel, you don't have the original pixel and a copy of it. You just have two pixels that are identical. Neither one is the "original pixel" – that would imply that there's something different about one of them that can be used to distinguish the two, but that's impossible, because they are identical. Digital systems support infinite replication without deterioration (there is no scarcity in the digital world).

You're also implying that an image's identity is based on its pixels, but that's blatantly false – the Mona Lisa for example is an analogue work with potentially thousands or even millions of different pixel representations, and even digital art can exist in multiple resolutions and file formats (for example, converting to lossy JPEG doesn't destroy its identity even though some of the pixels are changed).

If a so-called AI outputs an image that closely resembles one of its inputs, even if every single pixel is technically different, it would be difficult to argue that it's a completely different image and the resemblance is pure chance. And if the most of the pixels aren't different (the exact same shade at the exact same location), then that's essentially impossible to argue.

It may or may not be true that it's unlikely to "accidentally" recreate "some copyrighted game asset", but I don't think it's impossible. There have been sample cases of it closely recreating photos of celebrities, after all. And if the AI does happen to roughly recreate something that's copyrighted, that's not much different from going to an art gallery, taking a photo of one of the paintings, and trying to pass it off as your own. I don't think there's any way you could pass it off as an original work. It might be different pixels, but it's the same work.

Long story short, it's an interesting technology, but without total control of the inputs, the copyright is extremely nebulous, and "regurgitating" input images has been clearly shown to be something that sometimes happens even with images not as famous as the Mona Lisa.

As a side note, it also appears to be very bad at anatomy. It might be possible to generate sort of decent landscapes with the current generation of AIs, but portraits might be less reliable. Admittedly most of the ones I saw in this thread weren't that bad (probably because they all had very similar poses), but still.

I am not sure where the battle axe went in this one.
It put a guitar in and some garden tools.
I made this one 4:3

The battle axe is the guitar… although it even got that wrong – that's clearly an acoustic guitar, but "battle axe" as slang generally refers to an electric guitar.

[hermestrismi]

I didn't know that such thread existed.
Anyway, I am WITH the idea of using ai generated images if the developers of the ai claims that they are open to non-industrial usage.
AI doesn't create anything but 'copy, manipulate and modify'? ok, but what artists do? what humans do in general? Do I create a new language right now or I just 'copy, manipulate and modify' something I already memorise? In fact, none can create something without imagine it first, and none can imagine something completely out of nothing.
As for the rights of the generated product and since the ai himself doesn't recognise his rights (if he has any), the rights goes to the developer who created the ai. If I created a microscope and you discovered a medicament using my tool (along with others tools), I cant claim that I created that medicament.
Sorry for the poor English and the illness area of the example. I couldn't imagine another example right now hhhh

[Celtic_Minstrel]

The difference is that a human remixing other people's art can somewhat judge what level of similarity is "okay". An AI, or at least the current generation of AIs, cannot do that. So it might by chance churn out a 90% identical copy of someone else's art, which is clearly not "okay".

[hermestrismi]

The difference is that a human remixing other people's art can somewhat judge what level of similarity is "okay". An AI, or at least the current generation of AIs, cannot do that. So it might by chance churn out a 90% identical copy of someone else's art, which is clearly not "okay".

But you can't judge all artists as 'copycat' just because there is a tiny chance that an artist is copying others arts.
We need to take the AI as child who is trying to grow up while we have to prepare ourselves to all goods and wrongs that he can and surely will make. Am I wrong?

[Pentarctagon]

I would say that falls into the category of "anthropomorphizing the AI too much". These AIs are not children - they're huge, complex statistical models being built and refined by any companies who are able to pay the millions upon millions in cloud compute costs that it takes to create them.

[hermestrismi]

I would say that falls into the category of "anthropomorphizing the AI too much". These AIs are not children - they're huge, complex statistical models being built and refined by any companies who are able to pay the millions upon millions in cloud compute costs that it takes to create them.

I see. I agree with the second part. I mean the part about the ai being the product and the tool of the multinational conglomerates but I let myself disagree about the first part and precisely in two parts:
First, I don't agree that the Ai 'thinking' structure is far different from the Sapiens mind. We both had been programmed and we are the result of what our environments. In fact, the ai is not able to evolve to a certain point just because we didn't teach him (we didn't allow it) the basic of humanity: learning by mistakes, surviving at any cost... . If we are looking to ai as mimicking humanity, it is because we programmed it to do so.
Second, the purpose of the ai is to explore profits for companies . Still, it can be used to help humanity including helping us to be more resistant to the modern slavery, too. The computer and the ink I use at every set-in are both sold from big companies that they will complete their growing faster if the resistance choice to hide far from the modern world and its new technology

[Celtic_Minstrel]

The current AIs do not "think" or even "learn". They perform statistical analysis to produce output that's similar to the training set. I think it's misleading to call them AIs at all – they are not intelligent by any means. All they know is the training set that has been fed to them, and they don't understand that training set – it's just a collection of bits to them. In basic form they are no different from Wesnoth's random name generator.

[dwarftough]

The difference is that a human remixing other people's art can somewhat judge what level of similarity is "okay". An AI, or at least the current generation of AIs, cannot do that. So it might by chance churn out a 90% identical copy of someone else's art, which is clearly not "okay".

The user of AI can handpick ones looking most original, why not?

[Celtic_Minstrel]

The difference is that a human remixing other people's art can somewhat judge what level of similarity is "okay". An AI, or at least the current generation of AIs, cannot do that. So it might by chance churn out a 90% identical copy of someone else's art, which is clearly not "okay".

The user of AI can handpick ones looking most original, why not?

If you were given 100 images, of which 10 are by famous living artists, 10 are by obscure but "published" living artists, 10 are by famous historical artists, and the rest are AI art, are you confident that you could pick out the 20 living artists' works with 100% accuracy?

If you answered no, or frankly if anyone answers no, then that's your answer.

Human judgement can't be trusted to identify whether a work is original or not. Usually we rely on a chain of sources to determine this. Current AI art tools obfuscate or remove that chain of sources, making it even harder to tell if it's original.

[dwarftough]

If you were given 100 images, of which 10 are by famous living artists, 10 are by obscure but "published" living artists, 10 are by famous historical artists, and the rest are AI art, are you confident that you could pick out the 20 living artists' works with 100% accuracy?

If you answered no, or frankly if anyone answers no, then that's your answer.

I don't understand how telling AI art from human art relates to determining of originality. If a picture is unoriginal, it looks similar to some other picture. In the other case, it's original.

Human judgement can't be trusted to identify whether a work is original or not. Usually we rely on a chain of sources to determine this. Current AI art tools obfuscate or remove that chain of sources, making it even harder to tell if it's original.

Human judgement is not trusted, but by this logic we can't be sure in originality of any art. How can we be sure that this or that non-AI art is original? We don't use any tools generally (although one can try google the image to check), it's all about human judgement in the end. With all of its shortcomings, we have no alternatives anyway