MaGNET: Uniform Sampling from Deep Generative Network Manifolds Without Retraining

Deep Generative Networks (DGNs) are extensively employed in Generative\nAdversarial Networks (GANs), Variational Autoencoders (VAEs), and their\nvariants to approximate the data manifold and distribution. However, training\nsamples are often distributed in a non-uniform fashion on the manifold, due to\ncosts or convenience of collection. For example, the CelebA dataset contains a\nlarge fraction of smiling faces. These inconsistencies will be reproduced when\nsampling from the trained DGN, which is not always preferred, e.g., for\nfairness or data augmentation. In response, we develop MaGNET, a novel and\ntheoretically motivated latent space sampler for any pre-trained DGN, that\nproduces samples uniformly distributed on the learned manifold. We perform a\nrange of experiments on various datasets and DGNs, e.g., for the\nstate-of-the-art StyleGAN2 trained on FFHQ dataset, uniform sampling via MaGNET\nincreases distribution precision and recall by 4.1\\% \\& 3.0\\% and decreases\ngender bias by 41.2\\%, without requiring labels or retraining. As uniform\ndistribution does not imply uniform semantic distribution, we also explore\nseparately how semantic attributes of generated samples vary under MaGNET\nsampling.\n