Neural Rendering Techniques

Abstract

Neural rendering has transformed 3D scene representation and novel view synthesis by replacing explicit geometric primitives with implicit neural representations that encode scene appearance and geometry directly in neural network weights. This study presents a systematic evaluation of five neural rendering approaches -- Neural Radiance Fields (NeRF), Instant-NGP (hash-encoded radiance fields), 3D Gaussian Splatting (3DGS), Mip-NeRF 360 (unbounded scenes), and NeuS (surface reconstruction) -- across four rendering tasks: bounded object synthesis (NeRF Synthetic / Blender dataset, 8 objects), unbounded forward-facing scenes (LLFF, 8 scenes), unbounded 360-degree scenes (Mip-NeRF 360 dataset, 9 scenes), and surface reconstruction (DTU dataset, 15 objects). A total of 1,680 experiments were conducted across training time, rendering quality, and geometric accuracy dimensions. 3D Gaussian Splatting achieved real-time rendering at 68.4 +- 4.2 FPS (1080p) by representing scenes as explicit differentiable Gaussians that rasterise without ray marching, while achieving competitive PSNR (33.2 +- 0.4 dB on Blender vs. NeRF's 31.0 +- 0.6 dB). Instant-NGP reduced NeRF training from hours to 2.4 +- 0.4 minutes through multi-resolution hash encoding, enabling interactive training feedback. Mip-NeRF 360 achieved the highest unbounded scene quality (PSNR = 29.4 +- 0.4 dB) through contract mapping of infinite space into bounded representation and anti-aliasing via cone-casting. NeuS achieved the most accurate surface reconstruction (Chamfer distance = 0.62 +- 0.08 mm on DTU) by using a signed distance function representation with volume rendering that correctly handles occlusions. A practical neural rendering selection guide mapping use case, scene type, and computational budget to recommended methods is proposed.