Effective shortcut technique for generative adversarial networks

The generator in the generative adversarial network (GAN) learns image generation a coarse-to-fine manner which earlier layers learn an overall structure of and latter ones refine details. To propagate coarse information well, recent works usually build their generators by stacking up multiple residual blocks. Although block can produce high-quality as well be trained stably, it often impedes flow network. alleviate this problem, brief introduces novel architecture that produces combining features obtained through two different branches: main auxiliary branches. goal branch is to passing blocks, whereas convey layer later one. combine branches successfully, we also propose gated feature fusion module controls those prove superiority proposed method, provides extensive experiments using various standard datasets including CIFAR-10, CIFAR-100, LSUN, CelebA-HQ, AFHQ, tiny- ImageNet. Furthermore, conducted ablation studies demonstrate generalization ability method. Quantitative evaluations method exhibits impressive GAN performance terms Inception score (IS) Frechet inception distance (FID). For instance, boosts FID IS scores on tiny-ImageNet dataset from 35.13 25.00 20.23 25.57, respectively.