D‘ARC, J. - A variety of organic and fair trade garments

Feeding the GAN —
(Sub)conscious, abstract expressionism & Francis Bacon

When I came across GANs recently, some of the first things that popped into my head were: (1) this is weird, (2) this is scary, and (3) this is mesmerizing. It reminded me of idea generation and the process of designing. Hundreds, if not thousands of impressions a day, each day, some perceived more consciously than others, yet all of them informing the final outcome. How could I utilize this network?

In abstract expressionism, especially action painting, a fundamental element is giving up control – thus embracing chance. Jackson Pollock could determine certain parameters, like the color of the paint, the size of the hole he poked into the bottom of the paint can, or how and where he would move next, but chance also played a crucial role in determining which splashes and drips ended up on the canvas, and in what form (although he claimed otherwise). Controlled chance. Uncontrolled consciousness.

At some point, I made the decision to apply this concept to my design process by using a GAN. Setting parameters the two neural networks have to work with. A conscious selection of “inspiration” constituted the dataset: faved, saved, and pinned posts from Tumblr, Instagram, and Pinterest, plus long-forgotten screenshots and images from the darkest shoals of my harddrive.

To receive an output one would consider “realistic”, GANs need a dataset of at least 500 similar images (cats, cars, people). Small and/or incoherent datasets are likely to result in either shapeless blobs or nightmare fuel. But sources of inspiration are incoherent. One is not influenced by only one thing, but a myriad of impressions from different fields. Ultimately, I ended up feeding the machine a chaotic mixture: photos of faces, flowers, sculptures, and butterflies, book covers, film posters, paintings, drawings, fonts… And to confuse the machine even more, the selection did not consist of the recommended 500, but a mere 191 images in total. The GAN would function as a substitute for my subconscious. My role would be shifting from designer to selector: selector of the dataset, and selector of one of the infinite images the trained model would eventually disgorge. Training a model takes hours or days (depending on factors like the size of the dataset and the number of training steps). I decided on the default setting of 3000 steps. The model training would take four hours.