Multi-Attribute Balanced Sampling for Disentangled GAN Controls

Various controls over the generated data can be extracted from the latent\nspace of a pre-trained GAN, as it implicitly encodes the semantics of the\ntraining data. The discovered controls allow to vary semantic attributes in the\ngenerated images but usually lead to entangled edits that affect multiple\nattributes at the same time. Supervised approaches typically sample and\nannotate a collection of latent codes, then train classifiers in the latent\nspace to identify the controls. Since the data generated by GANs reflects the\nbiases of the original dataset, so do the resulting semantic controls. We\npropose to address disentanglement by subsampling the generated data to remove\nover-represented co-occuring attributes thus balancing the semantics of the\ndataset before training the classifiers. We demonstrate the effectiveness of\nthis approach by extracting disentangled linear directions for face\nmanipulation on two popular GAN architectures, PGGAN and StyleGAN, and two\ndatasets, CelebAHQ and FFHQ. We show that this approach outperforms\nstate-of-the-art classifier-based methods while avoiding the need for\ndisentanglement-enforcing post-processing.\n