Neural Networks as an Alternative to Photogrammetry. Using Instant NeRF and Volumetric Rendering

This paper investigates Neural Radiance Fields (NeRF) as an emerging alternative to traditional photogrammetry for the digital acquisition of heritage objects and environments. The study reviews recent developments in machine learning and computer vision, focusing on NVIDIA Instant NeRF, volumetric rendering, and related platforms such as Luma AI and Nerfstudio. Unlike conventional photogrammetry, NeRF systems reconstruct scenes through neural networks that infer missing views and generate volumetric representations with realistic lighting, reflections, and textures. The authors test these methods on a sculptural case study, evaluating speed, geometric quality, mesh extraction, texture generation, and interoperability with external software such as Blender and Unreal Engine. Results show that NeRF workflows can reduce acquisition and processing times while performing particularly well on reflective materials and complex lighting conditions, areas where photogrammetry often struggles. Although current outputs still present limitations in mesh closure, topology control, and metric reliability, the research suggests that neural rendering may soon become a powerful tool for digital twins, immersive heritage visualization, and future survey practices.