Semantically Decomposing the Latent Spaces of Generative Adversarial Networks

We propose a new algorithm for training generative adversarial networks that learns latent codes for both identities (e.g. individual humans) and (e.g. specific photographs). By fixing the identity portion of the codes, we can generate diverse images of the same subject, and by fixing observation portion, we can traverse the manifold of subjects while contingent aspects such as lighting and pose. Our algorithm a pairwise training scheme in which each sample from the generator of two images with a common identity code. Corresponding samples from real dataset consist of two distinct photographs of the same subject. In to fool the discriminator, the generator must produce pairs that are, distinct, and appear to depict the same individual. We augment the DCGAN and BEGAN approaches with Siamese discriminators to facilitate training. Experiments with human judges and an off-the-shelf face system demonstrate our algorithm's ability to generate convincing,-matched photographs.