Art Machines: International Symposium on Computational Media Art, City University of Hong Kong, 2019, pp. 170–171.
The Electronic Curator examines whether a computer can not only generate art, but also evaluate its quality. The work uses a Generative Adversarial Network (GAN), which constitutes a dialog between two competing neural networks. Here one represents a painter, who turns a human face into a vegetable portrait. The other represents a curator, who evaluates whether the portrait indeed looks like vegetable faces and encourages the painter to improve. The dialog between the competing networks represents the artistic process.
Training is unsupervised based on the cycle-consistent generative adversarial networks (CycleGAN). Thus we require only a set of face images and an unpaired and unrelated small set of vegetable-faces collected from a Google search on the Internet. In order to avoid mode collapse and get diverse and interesting results, we use a modified loss function inspired by DistanceGAN.
In exhibition mode, the painter observes the spectator's face and turns it in real time into a vegetable-face. The curator then grades the outcome. If the outcome is good enough to confuse the curator, a curatic text is generated based on the vegetables and fruits found in the portrait by object detection. In a world of AI art and creative machines, will the art of curation remain reserved for humans?
In the talk, we will review the techniques that helped in training and in inference, as well as those which did not help. Namely, we will discuss data collection and training strategies, modifications to the loss, and inference time normalization.
Gruss, Eyal, and Eran Hadas. 2019. “The Electronic Curator or How to Ride Your CycleGAN.” Art Machines: International Symposium on Computational Media Art, 170–171.