Impact of Hyperparameters on the Quality of Image Translation Using CycleGAN

Abstract

The study delves into the intricate relationship between hyperparameters and image translation quality in Cycle-Consistent Generative Adversarial Network (CycleGAN), which is one of the most widespread tools for unsupervised image-to-image translation. The hyperparameters investigated include learning rate, identity loss, and cycle consistency loss, which are aimed at enhancing the model's quality of producing realistic and good images. Furthermore, by employing horse2zebra dataset author was able to execute various experiments in order to understand what influence hyperparameters in CycleGAN have to its performance in style transfer problems. The analysis shows that changing the learning rate and identity loss weight have a drastic effect on whether the translation done yields detailed or correct images, tallied with the low-learning rate producing more explicit and accurate image translation. Loss weight transform manipulation of the identity is a pivotal element in the restoration of the original and the transformed features from the images. The study provides more insights into the understanding of how CycleGAN behaves in different training environments, and it stresses the significance of adequate parameters fine-tuning for the optimal image translation.