Neural Radiance Fields (NeRFs)
A technique that reconstructs full 3D scenes from a set of 2D photographs - letting AI synthesise realistic views of a scene from any angle, including angles never photographed.
Added May 18, 2026 · 3 min read
NeRFs represent a fundamental shift in 3D scene representation. Traditional 3D graphics requires artists to manually model geometry, materials, and lighting. NeRF derives all of this automatically from photographs. This has profound implications for product visualisation, architecture, film production, robotics scene understanding, and any field where accurate 3D understanding of real-world scenes is valuable.
Given a collection of photos of an object or scene taken from different angles, how do you reconstruct a complete 3D representation? Traditional methods used explicit 3D geometry - point clouds, meshes, depth maps. Neural Radiance Fields, introduced in 2020, took a fundamentally different approach: represent the scene as a continuous function that, given a 3D position and viewing direction, predicts what colour and density of material occupies that point.
This function is parameterised by a neural network - specifically, a multilayer perceptron (MLP) that takes as input a 3D coordinate (x, y, z) and a viewing direction (theta, phi), and outputs a colour (RGB) and density (sigma). The density represents how opaque the space is at that point - empty space has near-zero density, solid surfaces have high density. To render an image from a given viewpoint, rays are cast through the scene, the network is queried at many points along each ray, and the resulting colours and densities are composited using classical volume rendering equations.
Training a NeRF requires only the original 2D photographs and their associated camera positions. The network is trained to minimise the difference between its rendered views and the actual photographs. With enough photos from enough angles, the network learns a consistent 3D representation that explains all the observed images simultaneously.
The results are striking. Novel view synthesis from NeRF - rendering the scene from a camera position that was never photographed - produces photorealistic images with correct geometry, lighting, and occlusion. Objects can be viewed from any angle, behind occluders can be inferred, and lighting effects like specular highlights are correctly modelled.
NeRFs have been extended significantly since the original paper. Instant-NGP (Neural Graphics Primitives) dramatically accelerated training and inference using hash grids. Dynamic NeRFs handle moving scenes. Deformable NeRFs support articulated objects like humans in motion. Blocknerf scales to city-scale scenes from drone footage.',
3D Gaussian Splatting, released in 2023, emerged as a faster alternative that represents the scene as a collection of 3D Gaussians rather than an implicit neural function, achieving real-time rendering at comparable or better quality.
Analogy
A forensic artist who, given photographs of a crime scene from a few different angles, mentally reconstructs the full 3D layout and can then draw what the scene would look like from any other angle. The artist understands the 3D geometry from the 2D observations. NeRF teaches a neural network to do the same - infer the full 3D structure from multiple 2D views.
Real-world example
The set visit scenes in some film productions now use NeRF-like techniques to enable virtual camera moves in post-production: multiple cameras capture the scene simultaneously, a NeRF is trained on the footage, and the director can then choose any camera path through the reconstructed 3D scene - including camera positions that were never physically present on set.
Why it matters
NeRFs represent a fundamental shift in 3D scene representation. Traditional 3D graphics requires artists to manually model geometry, materials, and lighting. NeRF derives all of this automatically from photographs. This has profound implications for product visualisation, architecture, film production, robotics scene understanding, and any field where accurate 3D understanding of real-world scenes is valuable.
In the news
No recent coverage - search for Neural Radiance Fields (NeRFs).
Related concepts
3D Gaussian Splatting
A technique that represents 3D scenes as millions of tiny coloured ellipsoids - enabling real-time photorealistic rendering of scenes reconstructed from photographs, much faster than NeRF.
Diffusion Models
The generative AI technique behind Stable Diffusion and DALL-E 3 - which creates images by learning to reverse a process of gradually adding noise, turning pure static back into coherent pictures.