NeRF — Neural Radiance Fields — is the 3D scene representation introduced in a 2020 Berkeley paper that learns to encode an entire 3D scene inside the weights of a small neural network. Train the network on a few dozen photos of a scene from different angles, and you can then render that scene from any new viewpoint with photorealistic quality. NeRF was the first technique to show that you could capture and re-render real-world 3D scenes at near-photographic fidelity from ordinary photo input.
How NeRF works conceptually:
- The neural network is a function from (x, y, z position, viewing direction) → (colour, density).
- During training, you show it 50–200 photos of a scene with known camera positions.
- The network learns a continuous field representing the entire 3D volume.
- At render time, for each pixel of a new viewpoint, you cast a ray through the field, sample many points along it, and accumulate colour and density to produce the pixel.
NeRF's strengths:
- Photorealistic novel views — including reflections, refractions and complex lighting that mesh-based methods struggle with.
- Compact — the entire scene fits in a few MB of network weights.
- High fidelity from sparse input — works well with 50–200 photos where traditional photogrammetry might need thousands.
The historical limitations:
- Slow to train — original NeRF took hours to days per scene.
- Slow to render — original NeRF was multiple seconds per frame.
- Static scenes only — original NeRF could not handle moving objects, time-varying lighting or dynamic content.
The 2026 NeRF landscape — the field has moved fast:
- Instant NGP (NVIDIA) — train a NeRF in minutes, render in real time on consumer hardware.
- Mip-NeRF, Zip-NeRF — anti-aliasing improvements for higher quality at distance and detail.
- Dynamic NeRFs — handle moving subjects (talking heads, animated scenes).
- Generative NeRFs — text-to-3D-scene via NeRF representations.
- Gaussian Splatting (separate but related technique) — has largely supplanted NeRF for many real-time applications because it renders faster.
Where NeRF (and its descendants) earn their keep in 2026:
- VR and AR scene capture — film a real environment with a phone, view it in VR.
- Real estate — walkable photorealistic virtual property tours from a phone scan.
- Cultural heritage — digital preservation of monuments and historical sites.
- Visual effects — environment capture for film and TV; replacement for traditional set scanning.
- E-commerce 3D — product photography that can be rotated and inspected from any angle.
- Robotics and simulation — capturing real environments for training and testing autonomous systems.
For a US team in 2026, NeRF is the right tool when you need photorealistic novel-view synthesis from photo or video input. For interactive use cases, Gaussian Splatting has often replaced NeRF on quality-per-millisecond grounds. For asset pipelines that need traditional meshes, NeRF is more typically a capture step that gets converted to meshes downstream. The category remains research-heavy but has matured into shippable products in vertical applications.
