Unusual Transformation: A Deep Learning Approach to Create Art

Abstract

In this research, we proposed the concept of “unusual transformation,” which realizes transformation between two very different image sets as an extension of CycleGAN. CycleGAN is a new deep-learning-based AI technology that can realize transformation between two image sets. Although conventional CycleGAN researchers have tried transformation between two similar image sets, we applied CycleGAN to the transformation of two very different image sets such as between portraits photos and Ikebana or Shan-Shui paintings. Then to obtain a better result, we improved CycleGAN by adding a new loss function and developed “UTGAN (Unusual Transformation GAN).” We found that by using UTGAN, portrait photos and animal photos are transformed into Ikabana-like and Shan-Shui-like images. Then we carried out an analysis of the obtained result and made a hypothesis that the unusual transformation works well because both Ikebana and Shan-Shui are fundamental and abstracted expressions of nature. Also, we carried out various considerations to justify the hypothesis.