On Open-Set, High-Fidelity and Identity-Specific Face Transformation

Abstract

In this paper, a Generative Adversarial Networks-based framework has been proposed for identity-specific face transformation with high fidelity in open domains. Specifically, for any face, the proposed framework can transform its identity to the target identity, while preserving attributes and details (e.g., pose, gender, age, facial expression, skin tone, illumination and background). To this end, an autoencoder network is adopted to learn the transformation mapping, which encodes the source image into the latent representation, and reconstruct it with the target identity. In addition, the face parsing pyramid is introduced to help the decoder restore the attributes. Moreover, a novel perceptual constraint is applied to the transformed images to guarantee the correct change of the desired identity and to help retrieve the lost details during face identity transformation. Extensive experiments and comparisons to several open-source approaches demonstrate the efficacy of the proposed framework: it can achieve more realistic identity transformation while better preserving attributes and details.