Comparative Analysis of Generator Architectures in Cyclegan for Image Style Transfer

The study evaluated the effectiveness of alternative generator architectures within the CycleGAN framework for image stylization. Three variants were compared: a classical ResNet, a ResNet augmented with a Self-Attention mechanism, and a U-Net; the analysis additionally investigated the effects of normalization techniques (BatchNorm, InstanceNorm) and data augmentation (geometric and noise-based transformations) on training stability and output quality. Image quality was assessed using FID, SSIM, and LPIPS, complemented by qualitative visual analysis. The ResNet variant enhanced with Self-Attention and InstanceNorm, supported by geometric augmentation, yielded the most favorable outcomes. The results are discussed in the context of current state-of-the-art approaches, highlighting the need for systematic comparison with leading models to advance practical and efficient image style transfer. The findings demonstrate that even minor modifications to architecture and training procedures materially influence CycleGAN performance in image style transfer tasks.