Investigating Neural Field Models For 3D-Aware GANs

3D-aware GANs generate images that are informed by 3D information and so can render these images in novel view-points with 3D-consistency. pi-GAN is one such model, which attributes its success to its use of SIREN for providing the 3D information. In this paper we compare using Gaussian activations to sine activations (used in SIREN) or ReLU activations (broadly used) in an MLP to provide the 3D information for pi-GAN. Gaussian activations have been shown to have good performance in bundle adjustment even without using a positional encoding (commonly used with ReLU but not sine activations), with easier initialization than for sine-activations. However we found that without a positional encoding using Gaussian activation functions did not produce quality images for any of the configurations we tried. With positional encoding it produced comparable results to using a sine activation or ReLU activation with positional encoding, although the method using sine activation performed the best.