Willows (2020) and Lover's Hands (2020)
One of the things about AI art that’s interesting to me is that it’s not really clear where the art “happens.” There are a few stages that are relevant for GANs.
The conception of the database: what will you teach the GAN?
The construction of the database: where will the training data come from?
The output of the database: What will happen to what the GAN produces?
My first foray into GAN art was creative in that the database was novel: I trained a GAN using about 1500 images of deteriorating film photography by Romanian photographer Costică Acsinte. This was about the creative use of archives; but on the spectrum of originality, the majority of the “data” is arguably Acsinte’s, the photographer, as well as time itself, for introducing mold and degradation to the original negatives.
Of course, we ought to think more about the curation of databases as an artistic practice, because there's a lot there. With AI art, we ask about “authorship,” but authorship isn’t the only form of artistic production — curation is, too.
When I build a GAN and it starts generating new images, I slice one of these images off and explore slight evolutions a bit, like pulling a book off the shelf & writing some poetry about what I just read. But first, I had to pick what's on the shelves.
But we can also write our own books. That’s why I wanted to try to create my own series of photographs to use as a training dataset. This mean reproducing a series of images that had similarities and variation, possibly hundreds of times.
So, I borrowed a camera and got to work on taking thousands of photographs of my fiancee’s hands in various positions against a white background, a process that I called the Lover’s Hands dataset. In theory, this was going to make it possible to create images of her hands, which seemed especially poetic ahead of my upcoming departure for Australia for several months.
The result was more horrific than poetic. The GAN was able to figure out textures, but not forms: thousands of pictures of hands could tell it what fingers and wrists looked like, but the multiple positions, curling of fingers, open palms and knuckles, were too often in different positions, and the GAN was unable to make an accurate read of where these pieces should be or where the fingers should end. That, combined with a relatively solid depiction of skin tones and textures, just resulted in a horror show. This is what we get, I think, for trying to use digital means as if they can bridge a physical distance.
The next project was a little more successful.
Willows (2020)
The same weekend, I decided to photograph a collection of pussy willows and one feather in my girlfriend’s house, arranged into as many arrangements as I could muster while shooting in burst mode for about two hours on a borrowed Canon digital camera. In this case, the results were more pleasant, especially with some post-processing — the originals have color blotches which are somewhat common for GANs.
I’m happy with these results. But interestingly, it still comes back to curation: You feed a GAN a database of images, have it generate thousands of evolutions, walk through the library choosing ones you like, and designate them as the output. There’s artistry in building expressive tools. But there is *also* artistry in assembling the database, and artistry in creating the data.
We focus a lot on the first two: tool-building, and creating data for the set. But there is also an artistry to curation: good data-curators create databases for expressive results; think critically about what's included, and what the data means in aggregate; there’s potential in exploring generative results from some archives. That potential can be aesthetic, but it can also be political, and that’s an area I want to explore next.
Anyway, as a bonus, I got a viral Tweet out of the deal. As a result of my own laziness, I’m batch-editing raw photos from a digital camera to fit the size constraints of ImageGAN, which can only handle images that are 1024 x 1024 in size. When I told photoshop to resize all 1400 images, it asked me to confirm the quality settings for each one. Rather than clicking the return key to the Adobe Photoshop prompt on all 1400 images of hands during my editing phase, I looked over at the bag of tortilla chips and found a very simple solution to an annoying problem, which may be my most important contribution to AI artworks.
