La Fonda – Dieb #92 (2018)
This body of work is based on Holly’s interest in altering the output of a CycleGAN with values of art composition such as texture, contrast, and color. For training the CycleGAN, Holly used her original art, the first being a set of watercolor life drawings and the second a set of paintings. By using her art, she finds that complex compositions are preserved, as opposed to the “allover” compositions that are often generated by GANs.
She started this project by labeling 500 images from WikiArt with eight art attributes. She then trained on these images, fine-tuning a ResNet50 pre-trained on ImageNet. The resulting Art Attributes Composition Network (ACAN) is then used in CycleGAN training. In addition to the standard losses of a CycleGAN (Adversarial, Cycle-Consistency, Identity) she has added eight loss terms from the ACAN network.
Further details on this project can be found on this blog post.