Unpaired Faces to Cartoons: Improving XGAN

Stev H. Ramos, Joel Cabrera, Daniel Ibáñez, Alejandro B. Jiménez-Panta, César Beltrán-Castaño, Edwin Villanueva; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2022, pp. 1518-1527


Domain Adaptation is a task that aims to translate an image from a source domain to a desired target domain. Current methods in domain adaptation use adversarial training based on Generative Adversarial Networks (GAN). In the present work, we focus on the task of domain adaptation from real faces to cartoon face images. We start from a baseline architecture called XGAN and introduce some improvements to it. Our proposed model is called W-XDGAN, which uses a form of GAN called Wasserstein-GAN, learns to approximate the Wasserstein Distance, and adds a denoiser to smooth the output cartoons. Whereas the original XGAN paper only presented a qualitative analysis, the advantages of this solution are demonstrated both quantitatively and qualitatively by comparing the results with models such as UNIT and original XGAN. Our code and models are publicly available at https://github.com/IAmigos/avatar-image-generator.

Related Material

@InProceedings{Ramos_2022_CVPR, author = {Ramos, Stev H. and Cabrera, Joel and Ib\'a\~nez, Daniel and Jim\'enez-Panta, Alejandro B. and Beltr\'an-Casta\~no, C\'esar and Villanueva, Edwin}, title = {Unpaired Faces to Cartoons: Improving XGAN}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops}, month = {June}, year = {2022}, pages = {1518-1527} }