Mario Klingemann: synthesis in vision

Mario Klingemann is a Munich-based artist working with algorithms, Generative Adverserial Networks (GANs) and an artist's eye to create beautiful, unsettling and challenging artwork. His projects have included neural remakes of Beastie Boys videos; Circuit Training, exhibited at the Barbican; working with Massive Attack; and the installation Memories of Passersby I, which sold at Sotherby's for £40,000.

AI art is often derided, and in some ways is the synthesiser of the twenty first century. Critics say that art produced using GANs isn't art because there's no skill. Yet this ignores the input of the artist, the decisions made to choose what to train the GANs on, and how the output is managed.

With his fascination with glitches and the uncanny, Mario Klingemann is a perfect example of how the artist is still an important part of the process when using GANs. While some of Mario's work becomes fixed in a static image, much of it is a swirling, flexing place where different iterations emerge and fade to be replaced. I recently spoke with Mario about his work.

You've talked before about interestingness in relation to your work. What does interestingness mean to you?

MK: In the end, it's all about how we process information. We constantly want our attention captured. Interestingness is something that does that, because it comes off something you already know. You walk down the street and you constantly make predictions about what happens next. Usually, you don't have to think about it because it's happening subconsciously because it's all as expected, then something unusual happens and that is interesting.

When you talk about glitches, those are also not meant to be there, and usually that's from an external event. After a while, the glitches become expected and familiar, so you try to find the next thing. For a while, AI art was very unexpected, because it was new and different, but now, again, it already becomes normal. We're getting used to these images, so we have to ask, what's the next step?

When it comes to stories, you don't want a story to be too weird. You have certain expectations for a story, but within that you make little changes because you don't want it all to happen in the same way. That's the same as with image making. It's like a story. You have to have a lot of things that are familiar to us. You have to have these elements in there which make it different. 

This image is from Mario's BigGAN explorations entering landscapes that don't exist to bring back images of places from the AI's imagination

Why do you think the interest in AI art emerged when it did?

MK: There was this technological leap that happened with deep learning and it worked so well. Once it became possible to do visual stuff, it got more attention. Like scientists, 90% of the research nobody cares about, then once they see it in your face, especially the uncanny, it makes an impact. And, of course, we have this whole visual culture now, this attention economy. You scroll through your Twitter feed. It's all pictures and they want your eyeballs, and now you have these things that first make everybody look.

You started off, a long time ago, being interested in the idea that a bitmap contains the possibility of every possible image. Do you think the technology is approaching a point where we can find them?

MK: Yes, we do that already with GANs. The way I see it is you have the bitmap which has all possibilities, but the problem is that, from all the possibilities, 99% are noise to us. When you have a GAN, you take a slice of it, and the machine produces images that seem to make more sense to us. It cuts the noise and reduces it to the stuff that we can perceive. That makes the search more efficient. The next step is that I can produce images that look like images, like something meaningful, but amongst all these the GAN is looking for the ones that are really interesting and not just repetitive.

It sounds like the infinite monkeys and Shakespeare.

MK: The same thing with texts. You have the infinite monkeys that type something, but only a very, very small percentage is meaningful. Again, you have to find out what is meaningful, and that's where the machine can help. It tells us something about ourselves, about the way we process information and what kind of word combination we have to assemble to trigger an emotion or a reaction. This is measurable. That's the fascinating part.

You've compared the artworld to a religion, but said there isn't really another mechanism for getting this work out into the world.

MK: At the moment, I have to do art, because that's the only place where you can do these things. Currently there is something happening, maybe not exactly where I want to go. This multimedia entertainment, interactive work, like Team Lab - all these new spaces which open up and give you an interactive experience that looks artsy but is also entertaining. For me this is maybe a new genre, because for me it's not art but it's also not just pure entertainment. It's somewhere more inbetween.

Do you think games are similar?

MK: There is overlap between these things, and games are definitely a very interesting area where a lot of creative stuff happens now. If you look at it, it's getting more and more toward involving the audience... getting away from this thing where we are just consuming. People want to be the actor. They want to be in charge, so gaming is one of these art/entertainment places. In that sense, classic art is almost old school. I still like classic art. I hope it stays around for a while.

Your recent work Circuit Training sounds like it was a good way of showing people in a gallery setting what you're doing with GANs.

MK: Yes, because finding interestingness is very rare. Even if you try to automate it, it's a moving target. Once you've found it, you can't hold it for long, then you have to move to the next. In a way, the machines are still too limited. They cannot break out of a certain aesthetic.

You're not just producing one type of art. The projects you're more known for concern latent space, but you've also got the visual matching to Beastie Boys and True Detective videos, and you're starting to work with text too.

MK: And I do physical stuff. At the moment, I'm still fascinated by AI. Text interests me too, but it's a lot harder to sell. To get people to read from a tweet generator is OK, but to get them to read a story means they have to invest more time. That's maybe why movies are much easier to consume. Visuals are a very quick way to get people interested in your work. At the same time, I find text and storytelling much more interesting. It's much harder.
 

Is it a different process working with text than images? When I've been reading you talking about latent space and the image based stuff, it feels like it's got a geographic feel. You talk about going in and bringing out images.

MK: That's the fascinating thing: the magic about latent spaces. For the machine, everything is data, so it doesn't know if this is an image or if it's a text or if it's music. There are models custom built for a certain type of data. A model that processes text works slightly differently than one that processes images, but it's still possible to combine the two. In the end, you can have a latent space for text. In that space, similar concepts are close to each other.  You have this space where you can go in and extract text. There are certain rules - it's more sequential - whereas images are more two dimensional, but that’s the fascinating thing. In the end, that's what I'm looking for; latent story space. Right now, we have the latent word space. Traversing a latent story space is something that will be very interesting at some point. All these spaces get connected and then latent space forms a universal translation space, and it becomes like our brain where this happens already.

Where do you see yourself? As part of the continuity of art or technology, or inbetween?

MK: I think art also works by having these stepping stones. You get something that is not yet part of art that people get interested in. They explore it and offer it up. Art is this weird organism, decides this might be something worth incorporating. You cannot really be too far outside. You always have to expand the surface of art iteratively.

I see myself somewhere on the outside. It seems like what I have been doing has been accepted by some parts of the art world. Who defines that? The art world is not homogeneous, and consists of multiple different players and sub-genres. In the end I'm just making an offering.

In some ways it might seem like it has a similar relationship to comics, or to your background where you did posts for raves, which have traditionally not been seen as ‘gallery art‘, for want of a better term.

MK: That's the thing. Also, some people might not even want to be seen as part of the art world, because it's commercial. Of course, people look back and say “this work was art“. What you do now might not be recognised as art yet. In the end I'm having fun, but if you want to exhibit in a certain context, you have to be making work that is perceived as part of that. It comes back to this idea that you have to do ninety percent that's kind of familiar, and have ten percent that inflates the balloon a little bit.

The other thing I notice about the aesthetic of your work is that textures come out really strongly, like the hair in the Arcana pieces. Had you fed the GAN a particular data set to get it to produce those textures?

MK: First of all, I love textures. I have always found texture lacking in AI art, so I did lots of research into how I can get textures into the work, because they are in this intermediate space between form and noise. There is one model that I have trained on faces, hair, and skin colour, because it's the way of bringing us into this weird uncanny area. I used that model, but then I broke the model with a method called neural glitch. I run the model over some input data, but before I do that I break parts inside the model, like you unplug connections and plug them back in the wrong way. The machine has its own knowledge about what it's supposed to do - in this case generate human parts - but it applies them to the wrong place or the wrong way. There's some part of that texture that makes it through, but it gets semantically applied the wrong way. That's where it gets interesting. It's glitching again, but on a semantic level.

A parallel it reminds me of is teratoma tumours that produce hair, bone and teeth.

MK: Yes, that's exactly how it works. The model has a DNA, but then you splice it and put it back in the wrong way. As long as it's not totally broken, it still functions and produces these things. Us looking at it, the way our visual systems work... we still recognise this as texture, but it's in the wrong place. Yes, it's very much like a tumour. My art's like a tumour!

While it might seem an unfair comparison, Mario's work shares another characteristic with misfiring cells in the way it is constantly changing and mutating, rarely staying still to allow it to become a stereotype or conform to people's expectations of AI art. In fact this is probably the main character of emerging AI art; it's unwillingness to stay still long enough to be characterised. 

Whether it's Refik Anadol transforming architectural spaces with interactive installations, Jake Elves taking a queer approach to AI art, or Mario's latest piece Appropriate Response which has such an analogue feel, AI art is as diverse and varied as anything produced with a paintbrush or chisel.

Steve Toase is a Munich-based writer, originally from North Yorkshire, England. He writes regularly for Fortean Times and Folklore Thursday. He also likes old motorbikes and vintage cocktails. There‘s more on Steve‘s website.

Images: Hanging Man, a piece for the Arcana exhibition, inspired by Tarot cards and shows the organic textures within Mario's work. You can follow Mario's work on his website and Instagram.