Generative residual block for image generation

摘要

Generative adversarial network (GAN) consisting of the generator and discriminator is widely studied to synthesize photorealistic images. Recently, researchers usually build their generator and discriminator using multiple residual blocks (RBs) which enables both networks to ease the adversarial learning. Following this trend, most prior studies have focused on developing the normalization or attention techniques that are compatible with the RB without modifying the block architecture. This paper proposes a novel architectural unit, called generative residual block (GRB), which is effective to produce high-quality images. GRB contains an additional residual path which effectively emphasizes the informative feature while suppressing the less useful one. We provide comprehensive empirical evidence proving that the proposed method brings significant improvements in GAN performance with slight additional computational costs. Furthermore, we reveal the generalization ability of GRB by conducting extensive experiments across various datasets including CIFAR-10, CIFAR-100, LSUN, and tiny-ImageNet. Quantitative evaluations show that the proposed method significantly improves the performance of GAN and conditional GAN in terms of Frechet inception distance (FID). For instance, GRB boosts FID scores on the tiny-ImageNet dataset from 32.78 to 26.57.