4 ms·
I'm an AI researcher working on generative modeling, so I want to note my bias upfront. But I'm also a musician/artist and so I find some of these conversation
by godelski 4y ago
I'm an AI researcher working on generative modeling, so I want to note my bias upfront.
But I'm also a musician/artist and so I find some of these conversations odd. The problem with them I see is that they are oversimplified. To get better at drawing I often copy other works. Or I'll play a piece exactly as intended. Then I get more advanced and learn a style of someone I admire and appreciate. Then after that comes my own flair.
So I ask, what is different between me doing it and a machine? The specific images being shown in this article shows that Hollie is doing the same process as me and the machine. Their work is in fact derivative. There's absolutely nothing wrong with that either. They say a good artist copies but a great artist steals. I don't think these generators are great artists, but they sure are good ones.
I learn by looking at a lot of art and styles. I learn by copying. I am able to be more efficient than the machine because I can understand nuance and theory, but the machine is able to do the former much better than I can. It can practice its drawings a hundred or thousand times an hour.
Now there are highly unethical stuff that is actually going on and I don't want the conversation getting distracted from. Just today a twitter account posted a demo of their work and actively demonstrated that one can remove watermarks with their tool[0]. This is bold and borderline illegal (promoting theft). There are also people presenting AI generated work as human digital paintings (we need to be honest about the tools we use). People presenting work in ways that it was not actually created is unethical. But there are other generative ethical concerns.
Now there are concerns about photographer's/artist's rights. If I take someone else's work and post it as my own, that is straight up theft. Even celebrities can't post photos of themselves that were taken by others[1]. This gets muddled if I make minor changes but it's been held up in court that the intention matters and it needs to be clear that the changes were new in an artistic manner and not a means of circumventing ownership rights. These are some of the bigger issues we're running into.
A problem with these generative models is interpretation. How do you know if the image you produced actually exists in the wild or if it is new and unique? There's been papers that show that there are privacy concerns[2] and that you can pull ground truth images out of the generator. I'd argue that this question is easier to answer the more explicit the density your model calculates. Meaning that this is very hard for implicit density models (such as GANs), moderately difficult for approximate density models (such as diffusion and VAEs), but not too bad when pulling from explicit density models (such as Autoregressive or Flow based models).
This is a concern that is implicit by articles such as this, but fail to actually quantify the problem here: "How do we meaningfully differentiate generated images from those made by real people?" I'm a strong advocate for the study and research for explicit density models, but a good chunk of the community is against be (they aren't currently anywhere near as powerful, but there's nothing theoretically in the way. I'd argue it is that few people are researching and understanding these models. There is a higher barrier to entry). So I'd argue that the training methods aren't the major concern, but what is actually produced. While the generators learn in a similar fashion to me it is clear that I'd get in trouble if I was passing off a fake Picasso as a legitimate one. But it is also fine for me to paint something in that same style as long as I'm honest about it.
The nuance here really matters and I think we need to not lose sight of that. This is a complex topic and I would like to hear other views. But I'm not interested in mic drops or unmovable positions. I don't think anyone has the right answer here and to solve it we must get a lot of different view points. So do you agree or disagree with me? I especially want to hear from the latter.
[0] https://twitter.com/ai_fast_track/status/1587475575479959559 https://twitter.com/ai_fast_track/status/1587475575479959559
[1] https://collenip.com/taylor-swift-entitled-say-photographers-rights-belong/ https://collenip.com/taylor-swift-entitled-say-photographers...
[2] https://arxiv.org/abs/2107.06018 https://arxiv.org/abs/2107.06018
- aqsalose 4y ago>But I'm also a musician/artist and so I find some of these conversations odd. The problem with them I see is that they are oversimplified. To get better at drawing I often copy other works. Or I'll play a piece exactly as intended. Then I get more advanced and learn a style of someone I admire and appreciate. Then after that comes my own flair. >So I ask, what is different between me doing it and a machine? You are a human. If you practice art as a hobby you can feel pleasure doing it, or you can get informal value out of the practice (there is social value in showing and sharing hobbies and works with friends). One could try to formalize that value and make a profession out of it, get livelihood selling it. When all that "machinery" to (learn to) produce artistic works was sitting inside human skulls and difficult to train, the benefits befell on the humans. When it is a machine that can easily and cheaply automate ... the benefits are due to the owner of machine. Now, I don't personally know if the genie can be put back into bottle with any legal framework that wouldn't be monstrous in some other way. However, ethically it is quite clear to me there is a possibility the artists / illustrators are going to get a very bad deal out of this, which could be a moral wrong. This would be a reason to think up the legal and conceptual framework that tries to make it not ... as wrong as it could be. It could be that we end up with human art as a prestige good (which it already is). That wouldn't be nice, because of power law dynamics of popularity already benefit very few prestige artists. So it could get worse. But could we end up with a Wall-E world where there are no reason for anyone to learn to draw any well? When a kid asks "draw me a rabbit", they won't ask any of the humans around, they ask the machine. The machine can produce a much more prettier rabbit, immediately and tailored to their taste.
- godelski 4y ago> One could try to formalize that value and make a profession out of it, get livelihood selling it. I really think it is bad to frame this about profits. I mean in my case I am doing it purely for the pursuit of pleasure. I'd argue that these models allow more people to do so as it lowers the barrier to create quality work. It can also be used as a great tool for practice, as you can generate ideas and then copy and/or edit them. It is also a great tool to make quick iterations as you explore what certain ideas and concepts might look at. But I do not believe that it is anywhere near ready to be a replacement for humans. Especially since it is highly limited in its creativity (something not discussed). Also, I want to add that these methods allow for new types of art that didn't exist before. There are artists working and exploring this path. Questioning how these tools can be used to modify things or create things to be modified. > When it is a machine that can easily and cheaply automate ... the benefits are due to the owner of machine. In what way? If you are not paying for the system and it is freely handed over, why is it not "you" who is benefiting? I would understand this comment if the benefit was behind a paywall (e.g. Dall-E) but it isn't (e.g. Stable Diffusion). > Now, I don't personally know if the genie can be put back into bottle with any legal framework that wouldn't be monstrous in some other way. However, ethically it is quite clear to me there is a possibility the artists / illustrators are going to get a very bad deal out of this, which could be a moral wrong. This would be a reason to think up the legal and conceptual framework that tries to make it not ... as wrong as it could be. I guess part of the issue I have with this is that it sounds a lot like the arguments made when digital art itself was beginning. How do we differentiate the "I hate it because it's new" from the "I hate it because it is unethical?" This is not so obvious to me to be honest, because one can think the former and say the latter. I am not going to shy away from the fact that transitionary periods can be rough, but I'm not convinced it is going to kill artists' livelihoods. Especially since there is a lot of effort that is still needed to produce high quality images. I think this might be a point where people working on these machines (like me) vs the people that aren't (maybe you? idk) have different biases. All day I see a ton of crap come out of these. But if you just payed attention to articles like this or Twitter you'd think they are far more powerful than they are. These selected images are being created by expert artists too, that deeply understand aesthetics and the prompt engineering required to make high quality work. Maybe we'll get there, and that makes the point moot, but I'd argue that we're still pretty far away from that. I don't think this is going to kill off professional artists by any measure (especially because this exclusively affects the digital media domain and no other form) but may make the barrier to entry slightly higher (but it also might help artists become more creative as you can quickly explore ideas). > The machine can produce a much more prettier rabbit, immediately and tailored to their taste. Actually I'd argue the opposite. While this may be true for your average person making a drawing I still have significant doubts that the machines will be able to create better results than professional artists within the next 5-10 years (plenty would bet against me though, so that's fair). I also think there's issues with the diversity of these images and that it can't be resolved simply by adding more data (the "scaling" paradigm). I think they will rather reinforce that certain things look a specific way. Especially since these models do not understand a lot of basic concepts that we humans take for granted (a fundamental problem in AI: causal understanding). I don't think these issues are insurmountable, but they are a lot harder than many give credit for. But I do want to make it clear that while we disagree I respect your position and still do think you bring up some good points. But I do think we also have fundamentally different vantage points. Which probably makes it a good thing that we're actually discussing this together and not in our respective bubbles.