Neural radiance field (NeRF)
Also known as: NeRF, NeRFs, neural radiance fields
A neural radiance field (NeRF) is an AI technique that learns a 3D scene from a set of photos and can then generate realistic images from viewpoints that were never photographed.
Instead of storing a 3D model, it trains a neural network that predicts colour and density at any point in space. While NeRFs amazed researchers in 2020, slow rendering limited practical use, and faster Gaussian splatting has since taken over many applications.
Neural radiance field (NeRF), explained at three levels
At home
Imagine a friend who has studied photos of your kitchen from every angle so well they can sketch it from a spot where no photo was taken. A neural radiance field is that friend, made of AI. You feed it photos of a scene, and it learns what colour and how solid every point in space is. Then it can paint convincing new views from any angle.
At work
A neural radiance field (NeRF) is an AI method that learns a 3D scene from a set of photos taken from known positions and then generates realistic new views. It captures reflections and fine detail that traditional 3D scanning misses. Businesses have used it for virtual tours, visual effects, and product imagery, though rendering is slow and faster Gaussian splatting has replaced it in many tools.
At home
A NeRF stores a scene inside a small neural network instead of a 3D model. Ask it about any point in space and a viewing direction, and it answers with a colour and a density, meaning how solid that point is. To draw a pixel, the computer sends a ray from the camera, samples dozens of points along it, asks the network about each, and blends the answers into one colour.
At work
Training a NeRF starts with camera positions, usually estimated by structure-from-motion tools such as COLMAP. The network then renders rays through the scene, compares the predicted pixel colours with the real photos, and adjusts its weights until they match. Because the colour depends on view direction, NeRFs capture reflections and translucency well. Instant NeRF sped training with hash grids, and variants handle large outdoor scenes, changing lighting, and moving objects.
At home
A NeRF has no objects inside it, just a mathematical function stored in network weights. Rendering one pixel means querying that function dozens or hundreds of times along a ray, which is why NeRFs are slow. Editing is harder still: to move a chair, you'd have to retrain how the whole function behaves in that region. In my view, that 'black box' storage is why NeRFs never became a mainstream content format.
At work
Compared with Gaussian splatting, NeRFs are compact, since a network can be just megabytes, but slow to render without heavy acceleration. Their implicit form still suits research on scene understanding, relighting, and generative 3D, and tools such as Nerfstudio keep them accessible. Capture needs accurate camera poses and static scenes. Consider privacy and consent when capturing people and private spaces, since realistic reconstructions can reveal sensitive details.
Real-world examples
At home
Jerome used a phone app to capture a NeRF of his vintage guitar. Rotating the 3D view, he could see the light glinting off the finish from angles he'd never photographed.
At work
Aisha works in visual effects in Vancouver. Her team used NeRF captures of a film location to plan camera moves before the shoot, saving time on set.
How it compares
| feature | Neural radiance field (NeRF) | Gaussian splatting |
|---|---|---|
| Storage | Neural network weights | Millions of explicit blobs |
| Rendering | Slower | Real time |
| File size | Compact | Larger |
| Current momentum | Mostly research | Rapid adoption in apps |
NeRF pioneered AI 3D capture. Splatting made it fast enough for everyday use.
For analysts
For practical 3D capture today, I'd evaluate Gaussian splatting tools first and consider NeRF-based tools for research or specialized visual effects work.
History & origin
Ben Mildenhall and colleagues from UC Berkeley, Google, and UC San Diego introduced NeRF at the European Conference on Computer Vision in 2020. Their results, showing reflections and fine detail from new angles, launched thousands of follow-up papers. Nvidia's Instant NeRF, presented in 2022, cut training time from hours to seconds. Google used neural rendering in Maps' Immersive View, announced in 2022, and apps such as Luma AI let users create NeRFs from phone videos. Gaussian splatting, published in 2023, offered real-time rendering, and research attention shifted. In my opinion, NeRF's legacy is less about the method itself and more about proving that AI could learn 3D scenes from ordinary photos.