Asymmetric Two-Stage CycleGAN for Generation of Faces With Shadow Puppet Style

In this paper, we study the problem of cross-domain image-to-image transformation from real faces to faces with shadow puppet style. Our aim is to make target images retain the salient features of real faces and capture the artistic style of shadow puppet faces. This task involves dual changes of contents and geometric structures to the source image, and the target distribution is a transitional distribution, that cannot be solved by existing solutions. We propose a new CycleGAN-based scheme that extends the transformation of two image domains to that of three domains. We adopt a two-stage generation strategy in the forward transformation and urge the generator of the first stage to learn the target distribution. To this end, we set up a discriminator for the generator to guide its generation. We also present a multiple contrastive training strategy to address the problem of difficulty in training the discriminator, because there is no way to obtain real samples from the target domain. The experimental results show that our scheme is effective and that the target generator can produce intermediate images that meet the requirements.