CoPrGAN: Image-to-Image Translation via Content Preservation.

Image-to-image translation is an interesting and challenging application. At present, it has developed from single-domain to multidomain and many other aspects. But losing content structure is still an inevitable problem, which is manifested in the details of the objects or the global features of the images. We propose a new framework named CoPrGAN to alleviate the lack of content. CoPrGAN focuses on the expressive ability of content and style in different dimensions. In this way, the model uses multiple dynamic paths between the content encoder and the decoder to transform domains while preserving content structure. The content structure we are concerned with is not just posture and location, but also birthmarks, hair color, environment, etc. Experiments are arranged in animal face change that focus on local details and seasonal change that focus on global information. Both the comparative experiments with state-of-the-art and the ablation experiments demonstrate the superiority of CoPrGAN.