MTCGAN: Mini CycleGAN with Tokenized MLP for Unpaired Image-to-Image Translation

Unpaired image-to-image (I2I) translation aims to learn a function that transforms images from a source domain to a target domain without paired examples. While current Generative Adversarial Networks (GANs) excel in this task, they often demand significant computational resources due to their large generator architectures. In this paper, we propose MTCGAN, a new GAN architecture optimized for efficient unpaired I2I translation. MTCGAN employs an Encoder-Decoder framework which incorporates a Token-Mixing Multi-Layer Perceptron (Tok-MLP) to serve as an efficient information bottleneck. Besides, self-supervised pre-training is introduced to improve the initialization of model parameters, and gradient penalty is utilized for regularizing the adversarial training. Experimental results show that the proposed MTCGAN achieves a favorable trade-off between computational efficiency and performance compared with other competitive GAN models, exhibiting reduced model size and inference time. Our work also provides a novel avenue in GAN research by highlighting the untapped potential of highly efficient, yet effective, generator architectures.