CartoonGAN: Generative Adversarial Networks for Photo Cartoonization

CartoonGAN: Generative Adversarial Networks for Photo Cartoonization

Ayeesha Siddiqha1, Charmitha Jain K J 2, Chinmayi 3, Chinmayi N S4, Aisiri H M5

Dept. of Computer Science and Engineering, Malnad College of Engineering, Hassan, India

1ayeeshasiddiqha4@gmail.com, 2charmithajainkj@gmail.com, 3cchinmayi.1608@gmail.com, 4chinmayi87@gmail.com, 5aisirihm@gmail.com

Abstract—Cartoonization, the transformation of real-world images into stylized cartoon-like representations, has become increasingly significant in digital art, animation, augmented re- ality, and entertainment. Traditional methods, reliant on manual techniques or predefined filters, often fall short in efficiency, scalability, and capturing nuanced artistic styles. Recent ad- vancements in deep learning, particularly Generative Adversarial Networks (GANs), offer promising solutions but face challenges such as preserving semantic content, replicating diverse cartoon styles, and avoiding visual artifacts. Additionally, many existing GAN-based approaches require paired datasets, limiting their applicability. To address these challenges, we propose Cartoon- GAN, a novel deep learning framework designed for automated cartoonization using unpaired datasets. CartoonGAN employs specialized loss functions, including content loss to maintain structural integrity and style loss to emulate cartoon aesthetics, alongside edge-smoothing techniques to minimize artifacts. By integrating Adaptive Instance Normalization (AdaIN), Cartoon- GAN enables dynamic adaptation to various artistic styles, enhancing its versatility.