Improved Generative Convolution Method for Image Generation

In our recent study, we introduced a novel convolution method, called GConv, which improves the performance of generative adversarial networks (GAN) by modulating the convolution kernels following the given latent vector. In this paper, we analyze the limitations of GConv, and propose an improved GConv to address those problems. While GConv modulates the convolution kernel equally at all pixels, the proposed method produces pixel-wise different kernels following not only the given latent vector but also the feature in each pixel. Even though the proposed method is a simple modification of GConv, it shows better performance compared to the standard convolution as well as GConv. To show the superiority of the proposed method, this paper provides experimental results on the CIFAR-10 and CIFAR-100 datasets. Quantitative evaluations reveal that the proposed method improves both GAN and conditional GAN (cGAN) performance in terms of Frechet inception distance (FID).