neural radiance field (NeRF)
/ NUR-ul RAY-dee-uns feeld (nurf) /
A neural radiance field, or NeRF, builds a 3D model of a scene from a set of ordinary photos taken from different angles — and then lets you fly a virtual camera to any new viewpoint you never actually photographed. Snap a statue from twenty positions, and a NeRF can render how it would look from in between any two of them, with correct perspective, lighting, and even reflections. It is a way to capture a real place or object as something you can revisit in 3D.
The clever twist is what it stores. Instead of a traditional 3D mesh of triangles, a NeRF is a small neural network that has memorized the whole scene as a continuous function: ask it "at this 3D point, looking from this direction, what color is there and how solid is it?" and it answers. To make an image, the computer shoots imaginary rays through each pixel, asks the network many points along each ray, and adds up the colors — a process called volume rendering. Train the network until its rendered views match your real photos, and it has implicitly learned the scene's geometry.
It is a striking result, but keep the limits clear. The original NeRF was slow to train and slow to render, captures one specific scene (it is not a general world model), and assumes the scene held still — moving people or cars confuse it. It interpolates beautifully between your photos but cannot truthfully invent the back of an object you never showed it. The field moved fast: faster variants and a related technique called Gaussian splatting now render in real time. Still, a NeRF reconstructs an appearance you can re-view; it is not a measurement-grade 3D scan, and the unseen parts are guesses.
You walk around a fountain taking 40 photos. A NeRF trains on them, then lets you glide the camera to a low angle you never stood at — and the water, the splashing reflections, and the depth all look right. But aim the virtual camera at the fountain's far side, which your photos never captured, and the result smears into vague guesswork.
Free movement between the photos you took — but the parts you never showed it remain a guess.
A NeRF reconstructs one specific, mostly static scene as a re-viewable appearance — not a general 3D world model and not a measurement-grade scan. It interpolates between your photos but cannot truthfully fill in surfaces it never saw. The original was slow; newer methods (and Gaussian splatting) render far faster.