The new involution operator was proposed recently, but it was only implemented in the discriminative models. We first associate the involution with generative adversarial networks(GAN) to explore its effectiveness in image generation problems. We use the human face dataset CelebA. The experiments show the superior performance of involution GAN over deep convolution GAN and self-attention GAN. The generated images of involution has rather saturated and smooth colors. Besides, the hyperparameters (e.g.kernel size, group, and reduction ratio) are discussed to tuning the involution layer, where reduction size is found to play an essential role. Finally, the stability issues introduced by new involution in GAN are presented.
