Any-to-one Face Reenactment Based on Conditional Generative Adversarial Network

Tianxiang Ma, Bo Peng, Wei Wang, Jing Dong

Abstract: Face reenactment refers to the process of transferring the expressions and postures of a given face to the target face. We present a novel Any-to-one Face Reenactment Model based on Conditional Generative Adversarial Network, which has a simple dual converter structure: Any-to-one Face Landmarks Map Converter(AFLC) and Landmark-to-face Converter based on Conditional Generative Adversarial Network(LFC). The former transfers any source face into the landmarks map of the target face, and the map has the expression and posture attributes of the source face. The latter has a generator that transfers the landmarks map of the target face into the realistic and identity-preserving target facial image. The whole model is purely learning-based without any 3D model, and can generate high quality transferred face comparable to the state-of-the-art. What's more the model is highly robust to wild faces, including various faces of different complexions, ages, and genders. We performed an ablation study on our proposed AFLC to verify its importance for face reenactment of any object. AFLC helps the overall model to achieve an effective facial reenactment.