Generating Chinese Classical Landscape Paintings Based on Cycle-Consistent Adversarial Networks

Abstract

Generative Adversarial Networks (GAN) has made it possible for computers to generate images autonomously. There are many approaches based on GAN, including Conditional Generative Adversarial Networks (CGAN), Deep Convolutional Generative Adversarial Networks (DCGAN) and Cycle-Consistent Generative Adversarial Networks (CycleGAN) and so on. Photos can be transformed into western oil painting styles using unpaired data based on CycleGAN. But so far, no research has been found using it to generate images with Chinese Classical Landscape Paintings' style. This paper examines the effects of different generator models and loss functions on training time and results under this framework. In order to save the training time, the Unet generator is used in CycleGAN. To get a better quality, L2 Loss is chosen. Experiments are conducted on the PyTorch platform using Python programming language. Different styles of photos are tested and satisfactory results are attained.