IMAGE STYLE TRANSFER USING CYCLEGAN
International Research Journal of Modernization in Engineering Technology and Science · 2025
This paper explores image style transfer using Cycle-Consistent Adversarial Networks (CycleGAN), a technique that enables transformation between two image domains without requiring paired datasets.Unlike traditional style transfer methods that rely on aligned image pairs, CycleGAN leverages a cycle-consistency loss, enabling it to learn mapping functions from unpaired training data.This research focuses on applying CycleGAN to the Monet2Photo dataset, where the goal is to convert artistic images painted by Monet into realistic photographic representations and vice versa.The model architecture includes two generators and two discriminators working in tandem to ensure adversarial and cycle-consistent learning.Training was conducted utilizing PyTorch with GPU acceleration on Google Colab.The results demonstrate the model's capacity to retain semantic content while altering style, making CycleGAN a robust solution for unpaired image-to-image translation.Performance evaluation is based on visual quality, style fidelity, and convergence stability.