A 3D-Consistent Super-Resolution Framework for Efficient and Enhanced 3D-Aware Image Synthesis

Neural volume rendering techniques, such as neural radiance fields (NeRFs), have revolutionized three-dimensional (3D)-aware image synthesis by enabling the generation of images of a single scene or object from various camera poses. However, NeRF is computationally expensive for high-resolution (HR) synthesis. Most existing methods use two-dimensional (2D) super-resolution at the cost of 3D consistency, whereas radiance manifolds and two-stage generation are limited to specific synthesis tasks. To tackle these challenges, we propose SuperNeRF-GAN, a general framework for 3D-consistent super-resolution in NeRF-based 3D-aware image synthesis. Given a pretrained generator, we obtain an HR NeRF representation through a NeRF super-resolution module. Depth aggregation and normal-guided depth super-resolution produce a boundary-correct multi-depth map that guides efficient HR rendering. Experimental results demonstrate the superior efficiency, 3D consistency, and quality of SuperNeRF-GAN.