Pose Manipulation with Identity Preservation

Abstract

This paper describes a new model which generates images in novel poses e.g. by alteringface expression and orientation, from just a few instances of a human subject. Unlike previousapproaches which require large datasets of a specific person for training, our approach may startfrom a scarce set of images, even from a single image. To this end, we introduce Character AdaptiveIdentity Normalization GAN (CainGAN) which uses spatial characteristic features extracted by anembedder and combined across source images. The identity information is propagated throughoutthe network by applying conditional normalization. After extensive adversarial training, CainGANreceives figures of faces from a certain individual and produces new ones while preserving theperson’s identity. Experimental results show that the quality of generated images scales with thesize of the input set used during inference. Furthermore, quantitative measurements indicate thatCainGAN performs better compared to other methods when training data is limited.