thisbanksydoesntexist: All of it is ART-ifcial
The Art of Training Artifice
I’ve always been fascinated by street art. As a teenager I would go to my neighborhood with a spray can and make graffiti about things that were on my mind, perhaps as an act of rebellion against the world I was living in. I wanted to be an artist, whether it was through making crappy music in my basement or tagging walls in forgotten neighborhoods with watered down spray paint — I just wanted to create something. When I was 18, I saw a documentary on Banksy that changed my perspective of art as a whole. It was called Exit Through The Gift Shop, and I recommend it to anyone reading this article.
For those of you unfamiliar with Banksy’s work, its usually a commentary on complex socio-political events with a lot of historical context. The place where the artwork is placed is also very relevant to its deeper message. Bansky is anonymous to this day, and perhaps prefers to remain so forever. His bold statements and work have shaped, broken and rebuilt much of what art is considered today.
Modern art is plagued by the necessity to sell it in galleries and validate its worth by a monetary value — sometimes just a signature on the bottom right is enough for it to be worth millions. Banksy’s “prank”, where he shredded an artwork of his that sold for $1.4 million, shocked fans and critics alike. Needless to say, his work is an interesting probe on what we consider as art, as well its deep entanglement with politics and society.
I thought it would be intriguing to “reconstruct” this deconstructed artwork and ask the question, what is the dollar value of artificially generated art?
I downloaded RunwayML, with the hopes to train a GAN (Generative Adversarial Network) in learning some of the features of Banksy’s work. I wanted to see if it could capture Banksy’s distinctive street style and combine it with other styles of graffiti.
Getting Input Data
My first try was a little less successful than the second. I combined Bansky’s work, modern graffiti and impressionist artwork to see if the model could figure out how to do a style transfer. I used a Chrome extension and extracted 7,448 images from Google Images. Some images had to be sorted Manually. We searched the keywords “graffitti” “street art” “banksy” and “impressionism”. I used a bash script to override naming conflicts, remove glyphs and icons and put everything into one directory.
Training
I used Runway ML to train it over 3000 steps. In this scenario we used one of the default Style Transfer GANs available in Runway. The whole process of training took approximately 1h 40min. If we spent more time training, its FID would be higher. One only needs to be careful of over-fitting.
The FID started at 157 and ended at 76 (The lower the FID the better)
Output
I realized that the GAN was too sensitive to diversity in the training data. Perhaps I should have focused on only specific examples of Banksy’s work and seen if the model could reconstruct the image.
Improving the model
In my second try I provided a different set of input data, this time with only 542 examples, but 3000 training steps
I tried to keep the training data thematically the same. Here are some of the results:
I thought it was interesting to see that even though it was an imitation of banksy’s previous work, the model could reach a relatively higher FID score of 53 with fewer, but more specific examples.
That still doesn’t answer the original question: is it art?
I for one strongly believe that my active decisions on what sort of training data to provide the model shaped what it perceived as art. As a human being, Banksy’s training data was a result of his education, experiences and travels around the world. A child imbues their perception of the world with their experiences. If an artificially intelligent art-creating model is to do the same, it must crawl through the internet for itself and decide what it considers as art.
References
Tsay, Jason, et al. “Runway: machine learning model experiment management tool.” (2018).
