Neural Radiance Field (NeRF)

NeRF represents a scene as a function F(x, y, z, θ, φ) → (r, g, b, σ), implemented as a small multilayer perceptron. Training optimises this network to reproduce the input photographs when rendered via differentiable volume rendering (ray marching through the field and accumulating colour weighted by density).

The original NeRF required minutes of rendering per frame. Subsequent work — Instant-NGP, TensoRF, Zip-NeRF — reduced training from hours to seconds and rendering to near-real-time by using hash-grid encodings, tensor decompositions, and multi-resolution representations.

While Gaussian splatting has largely overtaken NeRF for real-time rendering, NeRFs remain influential as the conceptual foundation for many 3D generative models. Several text-to-3D and image-to-3D systems optimise NeRF-like representations under diffusion priors.

Datameister has tracked NeRF research from its ECCV origins and applies the underlying principles in scan-to-3D and scene reconstruction pipelines.