Abstract:
Image-to-image translation models, such as StarGAN v2, have enabled the translation of diverse images over multiple domains in a single framework. However, StarGAN v2 is computationally expensive making it challenging to execute on resource-constrained environments. To reduce the computation requirement of StarGAN v2 while maintaining accuracy, we propose a novel cross distillation method that is specially designed for knowledge distillation (KD) of multiple networks in a single framework. By leveraging this new KD method, the knowledge of a multi-network large teacher StarGAN v2 can be effectively transferred to a small student TinyStarGAN v2 framework. Without losing the quality and diversity of generated images, we reduce the size of the original framework by more than 20× and the computation requirement by more than 5×. Experiments on CelebA-HQ and AFHQ datasets show the effectiveness of the proposed method.
