Discriminating Against Unrealistic Interpolations in Generative Adversarial Networks

Interpolations in the latent space of deep generative models is one of the\nstandard tools to synthesize semantically meaningful mixtures of generated\nsamples. As the generator function is non-linear, commonly used linear\ninterpolations in the latent space do not yield the shortest paths in the\nsample space, resulting in non-smooth interpolations. Recent work has therefore\nequipped the latent space with a suitable metric to enforce shortest paths on\nthe manifold of generated samples. These are often, however, susceptible of\nveering away from the manifold of real samples, resulting in smooth but\nunrealistic generation that requires an additional method to assess the sample\nquality along paths. Generative Adversarial Networks (GANs), by construction,\nmeasure the sample quality using its discriminator network. In this paper, we\nestablish that the discriminator can be used effectively to avoid regions of\nlow sample quality along shortest paths. By reusing the discriminator network\nto modify the metric on the latent space, we propose a lightweight solution for\nimproved interpolations in pre-trained GANs.\n