Attributes Consistent Faces Generation Under Arbitrary Poses

Fengyi Song, Jinhui Tang, Ming Yang, Weiling Cai, Wanqi Yang

Abstract: Generating visual images automatically given semantic attributes description helps building friendly human-machine interface as criminal suspects depicting and customer-oriented product design. However, this is a hard task since there exist great semantic and structure gaps between attributes description and visual images, which further cause great inference uncertainty in transforming from attribute description to vivid visual images. We aim to reduce the posterior distribute complexity given attributes (P(I|A)) by exploring face structure knowledge and imposing semantic consistency constraint in an end-to-end learning fashion. Our contributions are three-fold: (1) we address the semantic and structure consistent problem in attribute-conditioned image generation; (2) the proposed method can generate attribute consistent face images with high quality in both detailed texture and also clear structure; (3) we provide an interface for generating attribute consistent images with diverse poses.