pi-GAN: Periodic Implicit Generative Adversarial Networks for 3D-Aware Image Synthesis

We have witnessed rapid progress on 3D-aware image synthesis, leveragingrecent advances in generative visual models and neural rendering. Existingapproaches however fall short in two ways: first, they may lack an underlying3D representation or rely on view-inconsistent rendering, hence synthesizingimages that are not multi-view consistent; second, they often depend uponrepresentation network architectures that are not expressive enough, and theirresults thus lack in image quality. We propose a novel generative model, namedPeriodic Implicit Generative Adversarial Networks ($\pi$-GAN or pi-GAN), forhigh-quality 3D-aware image synthesis. $\pi$-GAN leverages neuralrepresentations with periodic activation functions and volumetric rendering torepresent scenes as view-consistent 3D representations with fine detail. Theproposed approach obtains state-of-the-art results for 3D-aware image synthesiswith multiple real and synthetic datasets.