Transfer photo to anime with dual discriminators GAN
Yuxin Shang, Kehua Miao, Bai-Hua Chen, Zhengrong Wen · 2022 2nd International Conference on Consumer Electronics and Computer Engineering (ICCECE) · 2022
Anime is a popular art genre. Nowadays, the way of creating anime is mainly manual creation, which is a professional and time-consuming work. In recent years, though some methods, such as CartoonGAN and AnimeGAN, are proposed to transfer photos to anime images, there are still significant flaws: 1) The converted images are dark and the colors aren't vibrant. 2) There is potential for improvement in terms of the visual effect. To address these issues, firstly, we discussed the reasons why CartoonGAN and other approaches could learn anime style and proposed a novel method. It's based on the generative adversarial networks with dual discriminators and focus on texture style in grayscale space. Then, we ran tests on different datasets and obtained high-quality results. Finally, ablation experiments were used to further explain the effectiveness of our method.