Class-Continuous Conditional Generative Neural Radiance Field

초록

The focus of 3D-aware image synthesis lies in preserving spatial consistency while generating high-resolution images with fine details. Recently, Neural Radiance Field (NeRF) has emerged as a powerful method for synthesizing novel views with low computational cost and exceptional performance. Although existing generative NeRF approaches have achieved significant results, they are unable to handle conditional and continuous feature manipulation during the generation process. In this work, we present a novel model, called Class-Continuous Conditional Generative NeRF (C3G-NeRF), which synthesizes conditionally manipulated photorealistic 3D-consistent images by projecting conditional features onto the generator and discriminator. We evaluate the proposed C3G-NeRF on three image datasets: AFHQ, CelebA, and Cars. Our model demonstrates robust 3D-consistency, fine details, ability of 360° generation, and smooth interpolation in conditional feature manipulation. For example, C3G-NeRF achieves a Fréchet Inception Distance (FID) of 7.64 in 3D-aware face image synthesis with a 1282 resolution. Furthermore, we provide FIDs and for generated 3D-aware images of each class within the datasets, showcasing the ability of C3G-NeRF to synthesize class-conditional images.