The Nuts and Bolts of Adopting Transformer in GANs

Transformer becomes prevalent in computer vision, especially for high-level vision tasks. However, adopting Transformer in the generative adversarial network (GAN) framework is still an open yet challenging problem. In this paper, we conduct a comprehensive empirical study to investigate the properties of Transformer in GAN for highfidelity image synthesis. Our analysis highlights and reaffirms the importance of feature locality in image generation, although the merits of the locality are well known in the classification task. Perhaps more interestingly, we find the residual connections in self-attention layers harmful for learning Transformer-based discriminators and conditional generators. We carefully examine the influence and propose effective ways to mitigate the negative impacts. Our study leads to a new alternative design of Transformers in GAN, a convolutional neural network (CNN)-free generator termed as STrans-G, which achieves competitive results in both unconditional and conditional image generations. The Transformer-based discriminator, STrans-D, also significantly reduces its gap against the CNN-based discriminators. Models and code will be publicly available.1