Identity-preserving Conditional Generative Adversarial Network

Zhonghua Zhai, Jian Zhai

Abstract: This article proposes an identity-preserving conditional generative adversarial network (IPcGAN) for image-to-image translation. The proposed framework can learn the translation function without any corresponding image in two domains, and map a real image into a latent space and a conditional representation. Contrary to other works, we introduce a fine-tuning process and a novel joint-loss to preserve the persons original identity while maintaining realism in the synthetic faces. We evaluate the new model with feature-wise errors and demonstrate that it produces much better results than existing models.