Image-to-pixel-art Translation Based on CycleGAN
Yifei Hang · Theoretical and Natural Science · 2025
Image style transfer has gained significant attention in the computer vision fields in recent years, especially with the emergence of generative models. While numerous style transfer tasks have been handled by various models, image-to-pixel-art translations were not extensively explored, which is seemingly trivial yet requires delicacy in practice. To this end, this paper introduces the Pixel-Landscape-CycleGAN (PL-CycleGAN), which is a CycleGAN model that addresses the translation from, but not limited to, real-world landscape images to pixel art. The model is quantitatively evaluated using Frechet Inception Distance (FID) and Kernel Inception Distance (KID) scores, and got 85.472 for FID and 0.0366 for KID. These scores achieved a 52.46% and 64.98% reduction in comparison to the scores from interpolation. Further visual analysis also proved the model’s efficacy at translating to pixel art while maintaining the initial content of landscape images, as well as its constant performance when encountering non-scenery objects within images.