I’ve come to learn that if a person can run neural networks and has deep interest in art, they are probably a pretty creative and interesting person. Australian artist Liam Ellul is no exception.
Ellul recently shared a new portrait series he working on called Just Tell Me Who To Be on Twitter. His portraits are simultaneously of nobody in particular and yet also everybody in his life. The series explores identity through four 12”x12” acrylic-on-cotton paintings. Each painting was painted directly on printouts of images Ellul created by training a GAN (generative adversarial network) on 10k photographs from his Facebook account.
Without going into a full description of how GANs work (you can find that here), the process involves a neural network inventing new images based on a set of training images provided by the artist. So in Ellul’s case, he is essentially asking the GAN, “If I give you photos of all the important people and moments in my life, can you go and invent me some new people and moments?” After training, the GAN outputs a large number of potential images and stores them in “latent space,” which you can see in the animated GIF below.
Ellul then short listed a dozen of faces seen in the GIF above, printed them out, and laid them around his apartment for a few days. According to Ellul:
After extracting all the faces from my archives — the data preparation was somewhat manual — I found myself looking at thumbnails most of the time. Pre-processing in this context reminded me of mixing colors on a pallet, but instead of colors, I was mixing forms. It became pretty clear which ones had the strongest hold on me — then I painted them. The common thread was that the selected outputs I chose gave me an impression of something I identified with in a really deep way. Like, out of the latent space, it touched on something that I couldn’t have represented unless I saw it first.
Like Dropping Your Family Photo Album Into a Blender
Though some AI artists make their own training image sets — notably, Anna Riddler, with her painstaking photographic collection of tulips, and Helena Sarin, who trains GANs on her drawings and paintings — it is rare. For practical reasons (scale and availability), most AI artists select large public data sets to train GANs. However, because these public data sets are widely available, as are the GANs used to process them, there are signs that the results are becoming increasingly homogenous.
Ellul bucks this trend by not only using his own materials, but by using the most personal materials possible: photographs from his own life’s relationships, experiences, and memories, which are no doubt loaded with personal meaning and associations. He owns the material, in the truest sense of the word, as he has quite literally lived it. From Ellul:
It was a surprising realization just how much data I have created over my life and how effectively it can be harnessed in the creative process. Some look like me physically, but the face and expression I would never pull in a photo — it’s this surreal look that captures a feeling and encourages me to express it. Others look like a blend of me and a friend with similar surreal expressions.
Once I was happy with the outputs of the model, I spent a long while just watching the waves of eerily familiar faces that it produced. Often, I’d recognize a face as my own or fused with a close friend – despite never being captured with that expression – certain frames would perfectly resonate with a part of me when I saw them.
Fascinated by Ellul’s use of GANs as a departure point or inspiration for creating physical paintings, I asked him about his both his artistic and technical background.
Ellul shared that he has been creating portraits as a sort of visual journal since his grandfather first taught him to draw with charcoal when he was 10 years old (though he later switched to painting in acrylic). He initially went to school for law but realized “it wasn’t something I wanted to do professionally,” and he eventually shifted his focus to a rapidly growing interest in analytics. This led to Ellul and a friend launching “a small company focused on agricultural crop analysis and research.” It was there that Ellul learned about neural networks while testing predictive models for plant growth. Again from Ellul:
The first time I saw a GAN was 2017 in Alex Radford’s GitHub repo where he showed the generation of bedrooms, faces, and album art. My brain broke. Then mid-last year I saw the incredible high resolution faces you could get with GANs — something clicked in my brain and I felt compelled to do this portrait series.
Ellul now works in strategy and product development at Microsoft and creates his artwork on the side. I asked Ellul if he has any upcoming projects and if so what was next:
Yes! I love the adventurous nature of this area and the experience of running through a personal gauntlet to get these the paintings out! In terms of what’s next, I have two ideas bubbling away that are very much still coming together. Network design and exploring ways they can be linked together is something I will put more time into as I develop my approach. I am going also see if I can make the switch from acrylics to oils!
While the purist in me loves seeing work created digitally staying digital, I suspect we will increasingly see artworks executed in a variety of media as GANs come into their own as a tool for augmenting creativity (imagine what a GAN-inspired sculpture might look like). I think this is an interesting direction, and I’m encouraged by the exploration and work of artist/technologists like Ellul and his recent portraits.
As always, feel free to reach out to me at firstname.lastname@example.org with any questions or suggestions. You can also hit me up on Twitter, my social media of choice, at @artnome.com.