Abstract—An Image-to-Image (I2I) translation technique is a method that transforms an image from one domain to another by mapping one domain onto another. This technique involves two generators and two discriminators. Each generator can only translate one domain to another. This paper proposes a new approach called Knowledge Distillation Generative Adversarial Network (KD-GAN). The KD-GAN uses an image generated from Cycle-Consistent Generative Adversarial Networks (CycleGAN) as part of the target in training for a new generator. Our experiment involved translating between males and females in the CelebA dataset. We compared our model’s results with the state-of-the-art using Fréchet Inception Distance (FID) and Kernel Inception Distance (KID). The experiment showed that while KD-GAN is not the best regarding FID and KID, the output image can better keep the skin tone and hairstyle from the input image than other methods.
 
Keywords—Generative Adversarial Network (GAN), unpaired Image-to-Image (I2I) translation, Knowledge Distillation (KD), deep learning, cycle-consistency loss


Copyright © 2024 by the authors. This is an open access article distributed under the Creative Commons Attribution License (CC BY-NC-ND 4.0), which permits use, distribution and reproduction in any medium, provided that the article is properly cited, the use is non-commercial and no modifications or adaptations are made.