Christie’s just sold the first piece of AI art to be offered at a major auction house for $432,500. The piece was consigned by Obvious, a group of three friends from France with no formal art training.
There have been a lot of questions surrounding the creation of the work sold at auction, including:
Does AI artist Robbie Barrat deserve credit for this work?
Why did Obvious exaggerate the role of the algorithm?
Did Obvious really say they were considering “patenting ‘their’ algorithm”?
What role did Obvious actually have in making the Portrait of Edmond Belamy?
How does Obvious view AI and art moving forward?
In this interview, I speak with Hugo Caselles-Dupré, the tech lead from Obvious, and ask the questions above to cut through speculation and get answers straight from the source. I believe Hugo is being candid and transparent in this interview (even when it is not in his favor).
My interview with Hugo is extensive and gives his unfiltered point of view. Because I have given Hugo over 6,800 words, before we dive into the interview, I’d like to share some thoughts from other members of the AI art community to offer some balance to this story.
I asked respected AI artist Mario Klingemann who deserves the credit for the Portrait of Edmond Belamy. He shared:
“The question of who deserves the credit here is not so easy to answer. One important aspect is that this work was made in the context of art and not in that of science, so there is no obligation to cite prior work or list the tools and frameworks that you use to create your art.
“When it comes to Ian Goodfellow's GAN, I see that as a production tool, similar to Photoshop in digital art or a brush in painting. As far as I know, Ian does not identify as an artist or label his creation as an artwork, so while I would see it as good etiquette to mention that his or other researchers' work was used in one's work, I would not go so far to give him artistic credit for all the creations that get made by using it.
“Now, with Robbie's contribution, it's something else - he curated the training data set, trained the model, and put it on GitHub. So in the end, Obvious just had to fork it, have it produce a number of images based on random feature vectors, and finally make their selection. So you could say that Robbie did two thirds of the work involved in this process.”
Another highly respected artist, Tom White, has been exploring AI and art since the mid-1990s. Tom shared that he believed that Obvious used a similar codebase and the same training set.
Indeed, Tom ran Robbie’s model himself and was able to come up with images that look almost exactly like the Portrait of Edmond Belamy. It is Tom’s opinion that the Portrait of Edmond Belamy is a work of “appropriation.”
Hugo’s thoughts on this are covered in detail in the interview below. He did ask that I include this video as partial proof that Obvious did train their own GAN.
Lastly, this is the second part of my interview with Hugo. In part one we focused on how Obvious went from failing to sell their work on eBay to making $432,500 on a single print at Christie’s. It is not required that you read the first part of the interview to proceed with the second part, though I do recommend it if you have the interest and the time.
I hope you enjoy the interview.
Interview With Hugo Caselles-Dupré
Does Robbie Barrat deserve the credit for the Portrait of Edmond Belamy?
JB: Okay. So I am going to ask the hard questions, but I think this is important because I know you want to clear the air with this interview. Would you say that Robbie deserves credit for a percentage of the work you created? Or would you say he built the camera and you used the camera? How do you think about this?
HC: Yeah. I think this is a good question. We ask ourselves this question a lot, too. And we're like, 'Okay, what can we do with this?' But in the end, the fact that we did this physical piece and we signed with the formula is something that we wanted to do, and I think it has a lot of responsibility in the exposure of our project, too. So we owe him a lot, and that's what we said to him, that we wanted to be up front with him. So we owe him a lot. We wish him success. And we hope that if people like our project, they'll go to his project, too, and check out what he's doing. So we think he deserves something in this whole situation. And then, I considered even more. What can we really do?
JB: It’s just a matter of making it clear, and this interview is the exact right place to do it because we go a bit deeper. So could you have made the Belamy project without using Robbie’s code?
HC: Yeah, yeah, I really think so, because we had our eyes on many different data sets. We already knew that we wanted to do something like a classical art movement, like portraits or something like this. We already had this in mind. And so when we saw this, it was like, 'Okay, this is really convenient, so we can try working with this.' Yeah. We would have definitely found a way to do it ourselves. Before going over his code, I was already doing lots of things with GANs for my master's degree, so I already had lots of things going on with GANs. It was just a matter of collecting the data sets, and we would find a way because we have some scraping abilities. We used his code mostly for the scraper. So we used his scraper for getting all the data with the art. And in the end, the code that created the Belamy collection is the one that I -- the actual GAN permutation is the one that I talked to you about.
JB: The PyTorch DCGAN?
HC: I was already doing this with my master's degree. I was already involved in the GAN coding.
JB: You used the scraper mostly to gather the images?
HC: Yeah. Gather the data. So it's a Python script that when you run it, it collects all the data, like for the portrait class, for example. So you can collect all the data of the portraits.
JB: Where did you get the images?
HC: In the wikiart.com.
Why did you say, “Creativity is not only for humans,” implying that AI was autonomously making the work, even when you knew that was a false statement?
JB: What about your narrative that “creativity isn’t only for humans”? Were you playing up the machines and now saying that is not what you meant?
HC: Yeah. Exactly. I think that's what happens when you're doing something and nobody cares, then you’re just goofing around and doing really clumsy stuff. And then when everybody has this view, then they go back to what you did before and then you have to justify it. We kept justifying, because we still think that this part of the GAN operator that creates the images is really interesting and there is some form of creativity there … and we just thought it was cool to just do it like this. For us, it was just a funny way to talk about it.
JB: You didn't know you were going to be under the microscope.
HC: If we knew we were going to have to 400 press articles on what we do, we most definitely would have done that. But at [that] moment we were like, 'Yeah, it’s silly, okay, whatever, let's put this.' But retrospectively, when we see that, we are like, 'That's a big mistake.'
JB: All you can do is admit the mistake. What creative behavior do GANs exhibit? Many feel they don’t exhibit creative behavior.
HC: For me, the fact that you give it a certain number of examples and then you can continue to see results in the latent space, for me, the gap has to be [bridged]. So necessarily, there's some kind of, like, inventing something. So I guess there is some kind of creativity for me… because creativity is a really broad term, so it can be misunderstood, because creativity is something really related to humans. But at the basic, low level, it was given a set of images, it can create images that does not belong to the training set. So that's something that is transformed by the model, and there's some kind of creativity. So it's just a way you interpret the word "creativity." Maybe from certain perspectives you can say it's creativity.
JB: So it sounds like you believe it is dependent upon your personal definition of creativity? Some people say GANs are just are approximate distributions and that is not really creative - but it sounds like you think it is creative?
HC: Yeah. It's like, whatever you think creativity is, if we fit on the same definition, we are obliged to agree on something. So if we go to the same definition that creativity is something like, let's say, this ‘Concept A,’ then GANs will fit this concept. Or not? It's just a point of view thing, I guess -- and I understand that people can argue that [it’s] not great, we understand that, but it's just a point of view.
Did you claim you were going to “patent” the algorithm even though it was not yours?
JB: So there was an article where you are quoted saying you decided not to patent your algorithm. You mentioned to me that you never actually said that. But the formula on the front of your painting is by Ian Goodfellow, but you don’t credit him there.
HC: Yeah. Yeah. “Belamy” is translated to “Goodfellow” in French. So I think this argument is really not good, because we said many, many, many times that “Belamy” is the French translation for “Goodfellow,” because we admire Goodfellow and that he created GANs, and so we put the formula there. So it's a mathematical expression -- it's not ours, it's not his. It does not belong to anybody. So it's exactly like GANs, but we have the respect to pay to Goodfellow because he created this paper, but it's open source. So we never thought something about copyrighting the GAN algorithm. It doesn't make sense. Because for me, as a researcher in machine learning, it's really ridiculous to think that, because, like, you cannot put a patent on a theorem or an algorithm because it's part of the general knowledge of humankind, and anybody could call it in their own and use it. It's part of general knowledge. So yeah. There's more and more phrases in articles [that] we never said.
What contributions did Obvious make in creating Portrait of Edmond Belamy?
JB: So somebody else wrote the GAN code you used, correct? Did you use DCGAN for the Belamy painting?
HC: Yeah, yeah. Exactly. So we used DCGAN. It turns out this is the implementation available in PyTorch, it is the Soumith Chintala repository, so we used that because we tried both variants, because in my research I already had the code for many different types of GANs, so I already had code. And in the end, when I did the full search, DCGAN was fine. It was not [about] technological performance; we were just like, 'Hey, it's just this new way of using GANs.' I guess something like Big GANs, it's interesting for AI art, but it's also research. Like, there's an actual technological innovation with GANs, and we didn't claim to do something like this. We just wanted to have a regular GAN that worked well and allowed us to do what we wanted to do.
Because right now we are working on a project with 3D GANs, and I guess this time the technological innovation is a bit better, I guess. We are in contact with some researcher at the Max Planck Institute to use one of their models in order to create and train a GAN. And in this project, I think we are getting more involved in how it works. But since it was our first project … everybody's got to start somewhere, so you start with this. It seems like a reasonable idea. This project seems good. So let’s roll with it. So yes, we used this GANs, which we did, and we curated the assets in order to have the best result that we got. We tried many super resolution algorithms, and so we tried one with GANs, we tried others that don't really use machine learning techniques, more traditional techniques. And in the end, we found an enhancer, and so that worked really, really great and that gave really beautiful results, so we were like, 'We think it's really cool, and we are just going to stick with this.'
So yeah, we just tried a bunch of things, and when we thought the result was correct enough for our first project, we said, 'Okay, now let's try to show it to the world' and maybe use it to finance our further research and see where we can go with it. Because the actual first idea was, like, 'Okay, let's try this.' If we manage to have a little bit of expression and people are interested and we start new projects, then we'll continue with that. If nobody cares, we're just going to stop working, and then my two friends were planning on getting back their job, and we would stop and we would continue with our lives -- I have my Ph.D., they have jobs -- and we go on with our lives. And the fact that it blew up really changed everything.
I guess, yeah, a really big misconception [about it is] that it's just our first project, so we wanted to do this.
JB: Some of the engineers I have worked with would call this using off-the-shelf technology. There is not a lot of technical innovation going on here on your part. If it’s not technical, then where is the innovation in the Belamy project?
HC: So for this project, we guessed that the innovation with it is … we presented it in such an easy, not subtle way. Since it's really easy to comprehend, I think that's what the innovation is. But since it has resonated with so many people, is that there must be something here that is different to what was done before. But at first, we wanted to do something original, something unique. But you can't really control what people think of what you are doing. So yeah, maybe the fact that it was really accessible was the key. But we don't really know. So for further projects, we have lots of ideas.
But also one thing that must be really considered here is that we don't have any money. We don't have any computational power, so we spend lots of money on just trying this first project, and when we feel that we've got enough results, we stop there, because it was costing us money and costing us time, so we couldn't really afford to do something really innovative, because if you don't have the computational power, you just can't. So of course, I knew about progressive GAN from the day that they posted it on Reddit, and I wanted to try it the day after, but I just couldn't. So it's exactly the same thing with the big GAN papers, it's like, 'Okay, it requires like 512 GPU cores,’ something that we don't have, we don't have the budget for this. So for now, if you wanted to train this, you just can't. So yeah, we want to do this innovative stuff, but we've got to start somewhere to get some financing and continue working, having some credibility, having opportunities to get to access to more computational power. It was a way to have the means of doing something really innovative. At the time we created the Belamy Family, we didn't have the means to do something really creative -- or, I can’t really say that. It was really hard for us to try something really innovative, because when you try something really innovative -- and I see it in my research, too -- you need to try and fail a lot. So if you fail, you are going to train that model for nothing, and then you have to pay for it. So we couldn't really afford that.
Why didn’t you open source your code?
JB: A few more quick questions. I let some folks know I was interviewing you and asked if they had any questions. Someone is asking if you really want to make this understandable to the public, why don’t you open source the code?
HC: Yeah, okay, why not? But it's already -- we can do it, but it's already open source. It's like, what I told you, we used DCGAN, so yeah, it's already online. But we can do it if some people are interested. But mainly we just want to point to different things that we used and that's what we did also in our Medium blog post. So there is this DCGAN PyTorch repository. We also were inspired by the art-DCGAN repository of Robbie Barrat. …We can release the version of the data set that is curated, but yeah. It's not really interesting. We just removed all the paintings that have double faces or that were really like a real portrait. But we are not against open sourcing. For me, open sourcing this would be like taking someone's code and open sourcing it, like, it’s not yours, just point to where you get this code, and I got this code there at Robbie Barrat’s GitHub and Sumith’s GitHub. You can just use it. These tools are already available.
JB: So open sourcing it would imply that you are taking more credit than you want to take because you did not actually write any code?
HC: Yeah, yeah, yeah. I think you're making a good point. If I was in the shoes of Robbie Barrat, I wouldn’t like to see my code on something that Obvious released because, yeah, it's my code. And we already with lots of journalists talked about him, saying we were inspired by what he did, and on our main blog post, the link to his GitHub is available. It's one of the first things we talk about. And we are also really up front with him with this -- like, when we did this, we were like, ‘Okay, we need to send him a message and ask him if he's okay with that,’ and so he told us he was okay, and he told us, “Okay, I thought you were using just the code but not training models.” And then we said, “No, no, we trained our models. We tweaked the hyper-parameters to make it work.” We really had fun with it, and in the end, he said it's totally okay. And he asked us to make references to his code, and that's what we did on our website. We wanted really not to steal his ideas or to steal his code, we wanted to be really honest with him.
Where are GANs going from here?
JB: Given that you have some passion around art, how do you look at AI and art in general? What is AI adding to art, who is making interesting work, how do you position it relative to the larger sphere of art and your own passion for it?
HC: I think one of the reasons we got so much exposure is that AI art is something that is revealing what people think about AI, and revealing the fear and the misconception about AI. And that's why it also gets so much attention. So in the art spectrum, I would say this is really interesting because this is really showing something about this society, and so in this way, I guess art is a great way reveal the mood of society and what people are thinking right now. So this is really representative of the current atmosphere around AI and around all the misconceptions. So I think that's one of the reasons AI art is also interesting, because it goes to show something about the humans of today. I see that GANs were created in 2014, [and] the first results were not that great. Now, every six months we get a really big leap in technology — so we got GANs, then we got DCGANs, [then] we got Progressive GANs, then we got Big GANs that were the first result with the faces that were realistic. Then Progressive GANs was a big leap. Then Big GANs really, I think, is a huge leap, too, because that diversity of images is really big. So from the technology point of view, since the researchers were being really fast, I think that there will be more and more possibilities around this technology.
And it's not only GANs, because GANs is a huge example and something really interesting right now, but we won't be interested in AI research, there will be more breakthroughs in the future because people are putting lots of effort into doing AI research. So I think there will be more and more tools that are created, and that should create new artists and new art and new artistic approaches using these tools. Because we really do think that this tool is something incredible. When we talk about photography, we really meant it that when photography first appeared, it was just like a technology for highly qualified engineers, and so we do have the same thing with AI tools right now. So maybe AI could be something like photography that sparked a whole new art movement. So we hope it's that way, but we can't really know for sure. I don't know if it will last as long as and will be as important as photography, but I think there is a good chance.
JB: I look at something like Google Deep Dream, and once the code became open source, anyone could add photos. When I added photos, I was at first amazed, thinking, “This is better than Salvador Dali,” but after about 10 of them it loses its novelty. So nobody gets excited about Deep Dream images anymore because they realize even their grandmother can do it with the click of a button. How will GANs be different from Deep Dream?
HC: I see what you're getting at. I think here again we can compare it to photography. Anyone can take their cell phone and take anything. And so what makes photography really incredible or really interesting? And so I think as in photography, you have to add craftsmanship which will be [amplified] with your tool, and the message that is conveyed. When you see photographs of Weiwei with his middle finger and things like this, you see, anyone could have done this photography, but the way he did it was really relevant and was attached to a strong message. So that's why it's really important. And so I think that as time goes on, we will see that artists with the best ideas, I would say, or the most creative way to use the tools, will eventually be recognized for this and not just using something really new or something like that. And so that's exactly something that could be said, too, is like, ‘Okay, you are just using Deep Dream for GANs.’ We totally agree with that. In order for people to start getting cameras and take photos — and so for the great artists of tomorrow to rise, then you need to present technology.
Also, I think that what we do may spark people to get to know this tool and maybe be more creative than us. And we don't care. We're not in a competition. Since we want to make this technology shine, and so that more and more people know about AI, know about machine learning -- I'm passionate about machine learning, so I want people to know about it. I want them to know how great it is and how interesting it is. And I think in the end, you cannot fool people -- artists will eventually get sorted out and the best will naturally rise. This process has been seen for a million years, and it's always the same thing. We hope and believe that the best artists will get what they deserve and get the exposure that they really deserve.
So I think it's the craftsmanship that will be the most determinant thing in why AI art is interesting in the future. I don't think it will be a succession of technological innovation and stuff like this. I think already what we have right now, you can explore it in so many ways that there are things to be created that are potentially masterpieces. And you need that work and you need that dedication to find a way to find these masterpieces.
JB: When you project forward from your description it sounds like humans are going to become more and more important in what differentiates good art in GANs, not less and less. The public has this dystopian vision that AI is going to replace artists. But what you just described is the opposite.
HC: I totally agree with that.
JB: What are some examples of humans making good GAN art? What is good GAN art?
HC: Yeah, what is an example of good GAN art? To reproduce something that was done before is something that has been done a lot through the history of art. I think that's a bit of what we do. Like, portraits have been something really important in art for a long time, and reinventing portraits and seeing it from a different perspective is something that is interesting. So we were inspired by that and we thought that it was a really striking way to show how it's really interesting.
What you are saying is that I don't think that machine will replace artists and things like that. In the end, art is made for humans, so it needs to have this human part. And I guess that's also one thing that is in our work, is that it is really goofy and human in a way, so I guess that's why it resonated for people, too. Like, if you make something really striking for the machine, we couldn't comprehend it, so it would stop being interesting for us. And in the end, it's people that enjoy art, not machines, so it doesn't make any sense that the human part is totally removed from the art process.