Neural Radiance Fields for Novel-View Synthesis and Neural 3D Scene Representation: Advances in Quality, Scale, and Rendering Efficiency

Abstract

Neural scene representations rapidly transformed view synthesis and computer graphics following the introduction of Neural Radiance Fields. Rather than explicitly representing geometry with conventional meshes or point clouds, NeRF models optimize continuous neural functions that map spatial positions and viewing directions to volumetric density and emitted radiance. Combined with differentiable volume rendering, this formulation enables photorealistic novel views to be synthesized from collections of posed images. This review examines neural radiance-field research through 2022, focusing on improvements in image quality, scene scale, generalization, dynamic content, anti-aliasing, and rendering efficiency. The original NeRF formulation is reviewed together with positional encoding and related neural implicit representations. Mip-NeRF introduced integrated positional encoding to reduce aliasing, while NeRF-W addressed uncontrolled photographs containing appearance variation. PixelNeRF and IBRNet investigated generalization across scenes, and D-NeRF and Nerfies extended radiance fields to dynamic and deformable environments. By 2022, Mip-NeRF 360 expanded neural rendering to unbounded scenes, Mega-NeRF addressed large-scale environments, and Instant Neural Graphics Primitives dramatically reduced training time through multiresolution hash encoding. TensorRF and Plenoxels further demonstrated that explicit or factorized representations could offer substantial efficiency improvements compared with fully implicit multilayer perceptrons. The review analyzes rendering quality, training speed, memory consumption, scene representation, generalization, and scalability. Remaining challenges included sparse-view reconstruction, dynamic scenes, editing, geometry accuracy, relighting, and deployment on resource-constrained hardware.