AnimeFaceGAN: Deep Learning-Based Stylized Migration of Anime Faces

The AnimeFaceGAN proposed in this paper generates an anime face with real face features by using unpaired training data between real and anime faces. A primary challenge in this task arises from the disparate structural domains and appearances of real and anime faces, posing difficulty in producing high-quality anime faces. Additionally, preserving facial individuality poses a significant challenge. Existing methods often fail to retain the unique features of real faces, leading to anime faces that deviate from reality. To address these issues, this study introduces an attention mechanism to guide the model towards crucial regions, distinguishing between source and target domains. Additionally, an auxiliary classifier is incorporated to enhance the generator's understanding of areas needing improvement via CAM loss, facilitating model convergence. To enhance fidelity to facial characteristics, this paper introduces a novel sketch loss function to augment the generative adversarial network's capability in generating anime faces. Additionally, a three-stage training strategy is proposed in this study to prevent network instability through fine-tuning. Experimental results demonstrate that the proposed method effectively transforms real face photographs into high-quality anime face images, surpassing current methodologies.