Kim Jiwook ; Lee Minhyeok

Abstract

The 3D-aware image synthesis focuses on conserving spatial consistencybesides generating high-resolution images with fine details. Recently, NeuralRadiance Field (NeRF) has been introduced for synthesizing novel views with lowcomputational cost and superior performance. While several works investigate agenerative NeRF and show remarkable achievement, they cannot handle conditionaland continuous feature manipulation in the generation procedure. In this work,we introduce a novel model, called Class-Continuous Conditional Generative NeRF($\text{C}^{3}$G-NeRF), which can synthesize conditionally manipulatedphotorealistic 3D-consistent images by projecting conditional features to thegenerator and the discriminator. The proposed $\text{C}^{3}$G-NeRF is evaluatedwith three image datasets, AFHQ, CelebA, and Cars. As a result, our model showsstrong 3D-consistency with fine details and smooth interpolation in conditionalfeature manipulation. For instance, $\text{C}^{3}$G-NeRF exhibits a Fr\'echetInception Distance (FID) of 7.64 in 3D-aware face image synthesis with a$\text{128}^{2}$ resolution. Additionally, we provide FIDs of generated3D-aware images of each class of the datasets as it is possible to synthesizeclass-conditional images with $\text{C}^{3}$G-NeRF.