Taming the GAN
Or, how to use artificial intelligence as a tool in personal artistic work.
I am particularly interested in how to realise the human artistic intent with the serendipitous potential of neural media. My art has to reflect my personal experiences, I am not interested in autonomous art machines. I also like to keep everything small: small datasets, not too large prints, and a single GPU setup to keep carbon footprint under reasonable limits.
During the past fifteen months I have struggled with Generative Adversarial Networks (GANs) for image creation, writing my own code, experimenting, and finally, hopefully, taming this technology into something that I can comfortably use as an artistic tool. At least for the time being.
How do I get the generative GAN to make images that feel like mine? The most important factor is the selection of images into a training set, i.e. a set of images used to train the network with. Those images will to a large degree set the content and, to some extent, the style of the resulting images. But while the dataset usually consists of photographs, my style is not photographic. My aesthetic target is rather roughly something like art printmaking. In fact, I seldom consider a work ready before it is printed on fine art paper.
So where does the visual flavour of my GAN-based work come from? Maybe as a result of imperfections, both unintentional and deliberate, in the architecture and the training process. Balancing between overfitting and failing to train, the network is not able to reach perfect similarity, and somewhere within this middle ground a style is created, and elements from different images are seamlessly recombined. It is there I now look and find my own works.
Let us have a look at some examples.
November (2019). A series based on photos taken last November, close to where I live. A relatively focused, homogenous training set has produced images which match the late autumn mood, and combine elements from different photos in a surprising but meaningful way.
What you see is within (2019). A series based on photos of various small household objects against a blank background. The objects have been transformed beyond recognition, what we see in these pictures comes from our own mind, but the setting with a blank background with shadows is preserved.
Expressions of Kuhmo (2019). This July I spent two weeks in a countryside setting yet close to the unique Kuhmo Chamber Music Festival. The training photos combine the wild nature with closeups like wooden walls with the paint peeling off, with the created images reflecting both the surroundings and the compactness of chamber music.
Cityscapes from a dream (2019). My life has been stretched between the urban environment of my everyday life and the surrounding Finnish nature. Likewise, my works contain elements from both. Urban views but with an added emptiness and silence, perhaps.
Abstract. I feel as if my artistic work is suspended between abstract and representation. In fact, I originally got interested in neural media when I was looking for methods to transform photographs into abstract images.