A neural radiance field, or NeRF, is a method for representing a three-dimensional scene inside a neural network. Given a few dozen photographs of a scene with known camera positions, a NeRF learns a function that maps any 3D coordinate and viewing direction to color and density; rendering along camera rays then produces photorealistic images of the scene from viewpoints never captured. The technique, introduced by Ben Mildenhall and colleagues at UC Berkeley and Google in 2020, reframed 3D reconstruction as a deep learning problem and started a large research subfield.
How it works
The core model is a small multilayer perceptron trained per scene with gradient descent: for each pixel of each training photo, points are sampled along the camera ray, the network predicts color and density at each point, and volumetric rendering composites them into a pixel color that is compared with the ground truth. The scene itself becomes the network's weights, a compact and continuous representation of geometry and appearance stored as a learned representation rather than meshes or voxels.
Evolution and successors
Early NeRFs took hours to train and seconds to render a single frame. Follow-up work such as NVIDIA's Instant-NGP (2022) cut training to seconds using hash-grid encodings and GPU optimization, and hundreds of variants addressed dynamic scenes, reflections and large-scale environments. In 2023, 3D Gaussian splatting displaced NeRF for many applications by offering comparable quality with real-time rendering, though NeRF-style implicit representations remain influential in research on world models and robotics simulation.
Applications
NeRF techniques appear in photogrammetry, visual effects, e-commerce product capture and mapping, including Google's Immersive View. In text-to-3D generation they served as the underlying 3D representation optimized by early systems like DreamFusion (2022). They also inform embodied AI and computer vision research, where photoreal digital twins of real spaces are used to train robots in simulation.