ContraFeat: Contrasting Deep Features for Semantic Discovery

StyleGAN has shown strong potential for disentangled semantic control, thanks to its special design of multi-layer intermediate latent variables. However, existing discovery methods on rely manual selection modified layers obtain satisfactory manipulation results, which is tedious and demanding. In this paper, we propose a model that automates process achieves state-of-the-art performance. The consists an attention-equipped navigator module losses contrasting deep-feature changes. We two variants, with one samples in binary manner, another learned prototype variation patterns. proposed are computed pretrained deep features, based our assumption the features implicitly possess desired structure including consistency orthogonality. Additionally, metrics quantitatively evaluate performance FFHQ dataset, also show representations can be derived via simple training process. Experimentally, models achieve results without relying layer-wise selection, these discovered semantics used manipulate real-world images.