As a self-styled digital art aficionado who works for a machine learning startup, I have been waiting for years for AI (artificial intelligence) to yield something miraculous in the field of art. But instead, all I have seen is generic style transfers that allow you to make a photo of your dog look like a poorly painted Van Gogh... more of a parlor trick than the next revolution in fine art.
But at last, our machine learning art miracle has finally come in the form of some crazy looking nude portraits generated by a recent high-school graduate from West Virginia named Robbie Barrat. The second I saw Robbie's images on his Twitter feed, I fell in love. I had one of those rare moments where you see something completely fresh and you know some very important corner has just been turned.
In aesthetics there is a concept called the uncanny valley that I believe explains much of the emotional resonance in Robbie's Nudes. According to wikipedia:
The uncanny valley is a hypothesized relationship between the degree of an object's resemblance to a human being and the emotional response to such an object. The concept of the uncanny valley suggests humanoid objects which appear almost, but not exactly, like real human beings elicit uncanny, or strangely familiar, feelings of eeriness and revulsion in observers.
Forget the valley, Robbie just tore open an "uncanny black hole."
Art nerd that I am, I immediately tried to get Robbie on the phone. We eventually nailed down a time to chat and Robbie shared with me the details of his process.
WARNING: We are about to get a little nerdy here, but don't worry, you got this! Neural networks are just a fancy name for computer programs designed to think like the human brain. Robbie is going to talk about a specific type of neural networks called GANs (generated adversarial networks).
In Robbie's own words:
A paper just came out last December called Progressive Growing of GANs. A GAN is basically two neural networks that compete. There is the generator and the discriminator. The generator tries to make images to fool the discriminator, and the discriminator's whole job is to tell the difference between generated images and real images.
So the discriminator is always comparing the images the generator sends it with pictures in the data set, and it is trying to return a value of "fake" or "real". The generator gets feedback from the discriminator on how well it is performing. It uses that feedback to adapt and to try and generate more and more realistic images that will fool the discriminator into saying "this is real".
So I fed the GAN with 10,000 nude portraits and I let it have at it, and the two networks try to fool each other. And as they start off, they are terrible and their generations might as well just be noise. But as time goes on they get better and better at imitating what is inside the data set.
So what happened with the Nudes is the generator figured out a way to fool the discriminator without actually getting good at generating nude portraits. The discriminator is stupid enough that if I feed it these blobs, it can't figure out the difference between that and people. So the generator can just do that instead of generating realistic portraits, which is a harder job. It can fall into this local-minima where it isn't the ideal solution, but it works for the generator, and discriminator doesn't know any better so it gets stuck there. And that is what is happening in the nude portraits.
Pretty amazing, right? The fact that Robbie's Nudes are so surreal can be seen as a glitch or a limitation of the discriminator. But as with traditional art-making, happy accidents are often the most important roads to creative discovery. Had the discriminator been more "discriminating," the GANs may have output just a near perfect version of a traditional nude painting. For example, Robbie's landscapes leverage the same program as the Nudes, but yields images that are more convincing. I would argue the landscapes are also much less creative than the Nudes... but is that even possible with AI art?
Is it possible that the Nudes are more creative than the landscapes? Who is the artist here, Robbie or the AI/GAN? And if the AI is the artist, is a machine even capable of creativity?
I asked Robbie to help me understand his role in the process. He compared his role to the role of artist Sol Lewitt. Lewitt is best known for writing out instructions or rule sets for creating drawings and then having others execute the rules to create his artwork. In Robbie's words:
You know how Sol Lewitt would lay down the rules for his drawings? He'd start it off with the rules and then other people would interpret those rules and then assemble the art. With traditional generative art you establish the code and the computer will perfectly execute that code. There is no room for interpretation. With AI, I think I am doing similar things to what Sol Lewitt was doing with writing a rule card and then having someone else interpret the rules. I'm laying down the rules in the data set that I feed the GAN, but it's not up to me because the GAN is not going to perfectly interpret those rules; otherwise, we'd get perfect nude portraits back. But we aren't, because the GAN has interpreted the rules laid down, the data set I provided, incorrectly. So I feel like I have less control than with traditional generative art. Now that we are dealing with something intelligent there is room for interpretation.
I pushed a little harder on Robbie's Sol Lewitt analogy, pointing out that Sol Lewitt could watch people execute his instructions and then make adjustments to the rules based on what he saw. These adjustments would then change the next execution of the rule set, and in this way he had a fair amount of control and influence on the resulting art. I asked Robbie specifically if he had any control or predictability over the outcome of his images. Was there anything he could do to enhance it in one direction or the other?
Yes, absolutely. First, there is what I was experimenting with over the summer when I was doing the really low-resolution portraits and landscapes. I was using data-set swapping, which I haven't really seen anybody else do. When I was training my landscape model for the first time in low res, once it was done training, I showed the discriminator these abstract paintings just very, very briefly, like, probably 400 abstract paintings vs. 14k landscape paintings. It was a very small proportion of abstract art that this thing saw, but it totally changed the outcome. I was able to generate these really cool abstract landscapes.
Also what happens with a GAN is the input for the generator is just a random high dimensional vector. In the case for my program, there is a 512 dimensional vector, which is just a fancy way of saying it is a list with 512 numbers in it. But what happens is there is this thing called latent space that emerges after you train the GAN. All the possible paintings that are possible are laid out in highly dimensional space you are feeding into the generator. But the way they are laid out isn't random, it truly makes sense. So if you want to get a similar painting to your previous painting, you can pick a point that's very close to the point for your first picture. But some of the dimensions actually mean things like color scheme. So if I had a generation that I wanted to make more colorful, I could adjust one of the dimensions. So I do have some control, but only after the fact. I can't tell the GAN to produce a specific painting, but if I find a painting I like, I can then make adjustments to it.
I'm convinced that AI art challenges us to rethink what art is and how it is made to a degree we have not experienced since Duchamp created his readymade fountain sculpture.
As the father of conceptual art, Duchamp discounted the importance of the making process and aesthetic value of art and instead emphasized the artist's concept or idea as the key element. With Fountain, he took a urinal, turned it on its side, and "created a new thought for that object," forcing us to consider if it could be art.
Artificially intelligent art flips this, not just creating a new thought for an object, but creating an object capable of doing some of the thinking and creating for us. Granted it is early and AI is mostly augmenting human creativity at this stage, I think Robbie's Nudes stand out as a watershed moment.
So what makes Robbie's work so different from previous attempts by other artists? In part he credits access to rare, expensive supercomputers and exposure to new technological breakthroughs in GANs:
I work at Nvidia, and they have absolutely insane GPU cluster supercomputers. It actually took me two weeks to train this on their supercomputers. I tried to do this with both the portraits and the landscapes a while ago, but I wasn't able to get it past 128 x 128 pixels' resolution. That's, like, horrifically small. But the fact that I have access to these supercomputers now and this paper that just came out in December on progressively growing of GANs really helps. It works by starting out with a really small GAN and it will grow in layers as the generator and the discriminator get better and better. That lets it generate super high-resolution stuff, but it takes super long to do it I don't think anybody outside of Nvidia has been able to train the model.
Access to the right tools helps, but I'm convinced it was equally important that they landed in the right hands to yield these groundbreaking results. For me, there is no question of Barrat's creative genius. At 18 he is already breaking ground in neural networks. His previous projects include a rapping AI trained on Kanye West lyrics and a house plant that creates art through the harnessing of electrical signals. If you are wondering what important artists of the next generation will look like, Robbie is the model, as far as I am concerned.
Many believe we are moving into a future where all aspects of our lives are transformed by AI. If so, I believe this series of Nudes will become increasingly important as early masterpieces of AI art. Barrat plans on making his code open source and putting the training models out so that other people can learn from them. He might wait a little bit, as he admits there is a nice little period where he is the only one generating these types of images.
I applaud his decision to go open source, but as a zealous collector of digital art, I was eager to add some of his work to my collection. Robbie graciously agreed to sell me my favorite the series of four nudes featured in this article. For those not familiar, the blockchain art market now makes it possible to by and sell digital work the same as if it were physical work.
My good friends at startup Pixura are just a few days from launching their "SuperRare" blockchain-based market for digital art. I reached out to see if Robbie and I could be their first customers and transact ahead of the launch. They were awesome and happily obliged us. Just a few clicks later and my ownership of Robbie's digital artworks is indelibly etched onto the Ethereum blockchain where they are provably rare as an artist's edition of one.
Should I want to sell the works at any time, there will also be a marketplace (actually, many marketplaces are launching) where other collectors of rare digital art can participate. And if I do sell the works, the smart contract is set up so that Robbie will receive a 10% royalty every time the work is sold moving forward (in addition to 100% percent of the funds from my initial purchase). I personally believe I now own a very important part of art history, so if you are waiting for me to sell, I wouldn't HODL your breath.