Forum - View topicINTEREST: Shinchosha Publishes AI-Drawn Manga
|
Goto page Previous 1, 2, 3, 4, 5, 6 Next Note: this is the discussion thread for this article |
| Author | Message | |||
|---|---|---|---|---|
|
AsleepBySunset
Posts: 216 |
|
|||
Most artists have been resistant to the concept of copyrighting a style because in the current... or previous climate, only other artists could copy artstyles at human speeds, and because of human limits are unable to perfectly imitate another artstyle. AI can imitate a human style to create a picture in seconds/minutes, creating an arbitrarily large number of copies which will to an untrained eye be indistinguishable from the original artists work. The idea the law will take "humans copying an artwork or artstyle" as morally identical to "machines copying an artwork or artstyle" is ridiculous. And the idea as artists we should see these two acts as morally identical is egregious too. You see, sometimes there are things which are morally valid for some actors to do, in this case, humans, which are morally invalid for other actors to do. For example, in australia it is taboo and illegal for the non aborinal people to hunt crocodiles for meat, of the top of my head. And we accept, say orcas hunting humpback whales when humans hunting humpbacks whales is considered both taboo and illegal. This ridiculous argument is asking us to hold the same moral considerations to AI as we do to humans. And I don't think that will hold up in court either. |
||||
|
||||
|
hikura
Posts: 565 |
|
|||
I dsiagree.I think there could be a market for it.It doesn't mean hand drawn manga will die off. |
||||
|
||||
Cutiebunny
 Posts: 1754 |
|
|||
|
There was a time when sampling music was new, when people thought that it was going to kill the music industry and that the artists who used bits and pieces from other well known songs were talentless hacks. Back in the 1980s (I know most of you reading this weren't alive then...), no one thought it was criminal for a group called Men at Work to use a well known Australian children's song, Kookaburra, as the basis for their famous hit. One of Run DMC's biggest songs sampled Aerosmith's "Walk This Way". Again, taking other people's works and modifying it into a new song or piece of art is not a new concept. What is new is the concept of crediting and especially, paying royalties to the original artist/copyright holders of said works. Despite the original "Land Down Under" song being released in 1983, it wasn't until 2009 when Men at Work were sued by the current rights owner, Larrikin Music, for using parts of their song.
The anime world is also no stranger to reproducing artwork from others without proper compensation. Some of the production artwork I own from a 2009 series called Kobato was taken directly from an online blog. There are screenshots from said blog with instructions for those images to be reproduced. They found pictures of Christmas trees and decorations that they liked online and then reproduced them as backgrounds. No where in that episode or elsewhere were the original blog owners mentioned or credited. It will likely take a lawsuit of similar significance to change things for artists once AI technology improves. Do I see this as the end of art? Nope. Sampling still occurs in the music world. But it may take a couple of decades for legal cases to be brought up until the original artists whose work will be featured in future works to be paid. |
||||
|
||||
PotatoGirl
 Posts: 78 |
|
|||
|
"The art is good" It can't draw facial expressions for shit and I think that's pretty important for a sequential comic featuring human beings. The tech just ain't there.
|
||||
|
||||
|
SHD
Posts: 1752 |
|
|||
As I keep saying, this situation is absolutely not the same as sampling music, being inspired by others' work, making montages of others' works, and so on. Because when these things happen a transformation always takes place, following the creator's creative vision, as well as a credit to the original creator(s). (Otherwise it's just plagiarism and/or art theft, and I think we all understand that those things are bad.) In case of AI generated images there's no such thing happening, it can't happen because the AI, having no consciousness, can't have a creative vision; and the person requesting the image has no way of influencing what the AI uses from its database, how the algorithm mixes the bits and pieces together, and how the end picture turns out. They can manipulate the end result, but even so that is not a creative input. For that matter it's also not the same as photographs, because to take a photo you must have a creative vision as well. I'm not even talking about artsy photos, but even the most basic selfies or the most mundane pet pictures. (And people seem to forget that photography didn't start out as a mainstream thing, and in the early years photos were taken incredibly seriously, with a great deal of attention paid to composition, etc., to make sure that the subject was portrayed exactly the way they wanted to be remembered.) So let's just stop with the apples and oranges comparisons, and accept that an AI generating an image using an algorithm developed by a company, from a database of artwork (or photos, etc.) made by people who never consented to their works being used for this purpose, is not the same as a person being inspired by something and transforming it in a creative way to create something new with new meaning. As such, AI-generated images (text, etc.) can't be treated the same way as sampling, etc., it doesn't overwrite the huge ethical and legal concerns. I don't mind AI, I think AI image generators can be fine to just play around, as long as the database is provided by people who are compensated for it. But once money enters into the picture it's a whole different dimension, and the ethical and legal concerns absolutely need to be sorted out before someone can make money out of it. (Never mind the absolutely immense ethical concerns about AI-generated images based on photography, that go way beyond money and copyright... Because right now I can generate a fairly realistic-looking image of a city being bombed by a particular country's airplanes, and and say "LOOK, THIS IS HAPPENING RIGHT NOW", or an image of a person of X ethnicity being tortured/murdered by persons of Y ethnicity and release it on social media...)
Yes, and? When that happens and is found out it's usually a big deal, because that is called plagiarism and Japan really doesn't like that. |
||||
|
||||
NeverConvex
 Subscriber Posts: 2341 |
|
|||
This seems like it's taking a fairly strong and controversial position on what constitutes creativity. I don't think we really know how or to what extent human creation differs from a neural net nonlinearly transforming a few million image inputs. The field's long since moved past just directly copying human biology, but the structure of these neural nets is at least broadly inspired by human brains..
They can change the input prompt in order to alter how the AI uses its database. Not typically with pinpoint accuracy, though various researchers in the privacy community have developed techniques for trying to induce a neural network to cough up specific input elements. There is clearly some skill involved in this, though not so much as to make "dude who figures out what prompts to give to an ANN to make art" an enviable job description. I think it's likely more economically efficient -- and, culturally healthier, since the alternative seems likely to significantly substitute cheaply, tediously generated ANN art for hand-created work -- for creators of commercial-use neural networks to compensate the people who created them than for them to be able to re-use inputs freely, but it is interesting how difficult it is, in principle, to separate these marginal uses of any single input from a typical entirely human act of artistic creation drawing inspiration from others' work. The economic and practical implications seem much clearer, to me. |
||||
|
||||
dm
Subscriber Posts: 1388 |
|
|||
This is not how this stuff works. What is stored in the model is not "a database of images", what is stored in the model is a set of probabilities. "A pixel like that will have a pixel like this next to it with a certain probability under such-and-such conditions". You can't ask these systems to reproduce a specific artist's specific work, because they aren't storing the artists' works in the model. You can't ask them to reproduce Monet's lily pads at Giverny, you can only ask them to create an image of lily pads in the "style of Monet". Where "style" is a matter of "probably shaped like this", "probably using pixels colored in this way", "probably using transitions and juxtapositions of color like this". These things only do transformations. It's true that music-sampling is a bad analogy because these things aren't doing collage. But "being tutored by others' works" is how these things work (I've replaced "inspired" with "tutored" because it's wrong to suggest these things are "inspired" by anything). I keep comparing these things as a medium to photography for a couple of reasons. The first reason is that, as you say, a good photograph --- a photograph as art --- is not just a matter of randomly snapping pictures. It's a matter of framing, composition, the use of contrast. The second reason is that most of what I'm hearing about this echoes exactly what was said when Photoshop was first introduced, and, to a degree, about photography when it was first introduced. I keep saying this is a new medium because an AI image is good only by accident. To produce decent images with these things takes a lot of coaching skill and usually a lot of post-processing with Photoshop and other tools. It might be possible (at great expense) to train these things in such a way that they'll learn composition and framing, and they'll learn how to produce images of aesthetic quality. It will be interesting to see if the model builders can find a way to train that sort of thing that is remotely economical. Anyway, I'll probably take a look at the first tankubon that gets released. If the story is decent, I can overlook some flaws in the art. If the art is decent (samples I've seen suggest the artist should be making more use of Photoshop to correct things), it will go partway to compensating for a mediocre story. Last edited by dm on Mon Jan 09, 2023 10:54 am; edited 2 times in total |
||||
|
||||
Gem-Bug
 Posts: 1231 |
|
|||
It's pretty simple, honestly: Did a human being make this artwork? Y/N If you selected "Yes", then yes, it is art. If you selected "No", it is not.
Literally anyone can pick up a pen/pencil/paintbrush/clay/whathaveyou and create art; it doesn't matter how "good" or "bad" it may be. The only way to not be creative is to never try ever. Someone typing in prompts that then pull from stolen artwork and recompiles it into something else is creating something, sure, but it is not art. No matter how many good, lengthy explanations SHD and others give in this thread, there is a steady stream of bad faith actors, contrarians, crypto-shills, sock puppet accounts and folks who obviously don't create anything flooding in with the same non-arguments without reading any of the previous posts. 
|
||||
|
||||
|
NeverConvex
Subscriber Posts: 2341 |
|
|||
Technically, a human being derived the theorems and wrote the code that represent the neural network, not unlike a human might write Python code to automate image processing in GIMP or something. But, yeah, if the narrow distinction desired is "Was an ANN used near-exclusively in the process or was it principally the work of a human using traditional tools?", it is simple to separate in practice. Just not a very principled distinction, which is more what I meant.
Not sure if this is in reference to me, but I don't even disagree with SHD's final policy conclusion, just with the way they reasoned their way there. I'd like to see contributors to training corpuses financially compensated, and I think I was even the first person in this thread to advocate for that. EDIT: I do also create in my spare time, FWIW; code and 3D art, mostly. Nothing I'd consider good enough to bother selling, but for fun. Last edited by NeverConvex on Mon Jan 09, 2023 12:43 pm; edited 2 times in total |
||||
|
||||
|
dm
Subscriber Posts: 1388 |
|
|||
This is not what's happening. No artwork is "stolen". No artwork is even "quoted". Art work has been looked at, and the visual aspect of techniques learned. But what is produced only looks good (literally) by accident, as the result of chance. So, for this manga to look good, someone has to generate a lot of images and choose the ones that evoke the appropriate response. For those images to cross the uncanny valley, the compiler has to do a lot of work with tools like Photoshop (the samples I've seen don't indicate that a lot of this is happening in this case). Somewhere, maybe, it stops being a form of "collage" (for want of a better term --- no actual images are being reproduced, maybe I should say "pastiche") and becomes something more, I don't know. |
||||
|
||||
|
NeverConvex
Subscriber Posts: 2341 |
|
|||
Some input images are almost certainly memorized, and sometimes exactly or near-exactly reproduced -- out of concern for privacy, there is a research community that has examined methods for prompting artificial neural networks to behave like this (example). This is unintended behavior/not the primary way in which they're used, though, and there are algorithmic ways of provably limiting the extent to which an ANN could do this, so it's more of an implementation detail than a fundamental issue, I think. Your "and becomes something more, I don't know." is the same kind of hesitation I feel about this in trying to draw a principled distinction between art generated by ANNs (or, in part by ANNs, possibly in combination with other pipelines; and always with a human present somewhere in small or large ways) and art generated by humans. The separation is .. fuzzy. |
||||
|
||||
|
Gem-Bug
Posts: 1231 |
|
|||
No, the last bit wasn't specifically aimed at you, even though the rest of my post was in reply to yours. You seem like you're trying to have a well-thought-out discussion on the subject compared to like 50% of the replies in this thread. |
||||
|
||||
|
mrsticky005
Posts: 122 |
|
|||
|
There is no such thing as AI art. At the very least not yet.
AI cannot decide that it wants to draw something. So called "AI art" is simply humans using computers to generate an image. CGI. It may be more efficient now but it's still a human making the call. As far as this manga goes I say it depends on how the images are generated. There's no rule that says manga has to be drawn. It could be made with photos and photoshop or a million other ways. The issue is when those photos belong to someone else. How much editing before it stops being a derivative work and therefore violates intellectual property rights and it being something unique and creative and it's own piece of art. and I say a big part of that answer is effort. Unless your AI machine is North No 2 (press F to pay respects) then it's probably not making art. |
||||
|
||||
|
dm
Subscriber Posts: 1388 |
|
|||
That was an interesting presentation, but I'm going to suggest that if you want to be able to retrieve a specific work of art from one of these models you have to present that work to the model several thousand times. So, I'm doubtful about that "almost certainly". Well, it might memorize corporate logos. I bet it's easy to get it to put a Nike swoosh onto something, but hard to get the swoosh in the right place with the right orientation. |
||||
|
||||
|
NeverConvex
Subscriber Posts: 2341 |
|
|||
Oh, fair enough! I agree a lot of the replies are clearly not made in good faith. That seems pretty standard, unfortunately.
I think it may often be the opposite that's true; rare and unusual images may be more likely to be memorized, because this is compatible with training incentives (can't get high scores on these training images easily by representing them as functions of other images), as suggested here. I am not sure, though; I don't specialize in prompting ANNs to regurgitate memorized inputs -- I just know it is a well-known phenomenon, and at this point an expected property of ANNs as traditionally fit. |
||||
|
||||
 |
All times are GMT - 5 Hours |
|
|
|
Powered by phpBB © 2001, 2005 phpBB Group