Appendices A Network Architectures

Since DCGAN [1] showed astonishing image generation ability, several generator and discriminator architectures have been proposed to stabilize and enhance the generation quality. Representatively, Miyato et al. [2] have used a modified version of DCGAN [1] and ResNet-style GAN [3] architectures with spectral normalization (We abbreviate it SNDCGAN and SNResGAN, respectively). Brock et al. [4] have expanded the capacity of SNResGAN with a shared embedding and skip connections from the noise vector (BigGAN). As a result, we tested the aforementioned frameworks to validate the proposed approach. To provide details of the main experiments in our paper, we introduce the network architectures in this section.