Skip to content
All apps

Neural Rendering (NeRF)

About this app

A scene as a function instead of a mesh: a tiny network returns colour and density for every point in space. A ray is marched through the volume and the samples are composited into one pixel. Novel viewpoints are computed — including view-dependent shine.

What a NeRF is

A neural radiance field (NeRF) stores a scene neither as photos nor as a 3D mesh. It is a function of space, usually a small neural network, that returns a colour and a density for a point and a viewing direction. To render a pixel, a ray is sent through the scene, the function is sampled at many points, and the colours are blended, weighted by density and the remaining transmittance (volume rendering). Because colour depends on the viewing direction, a NeRF can also show highlights that move with the viewer.

What the app shows

The app shows the rendering process for a single ray. You choose the number of sample points and step along the ray. At each point it shows density, colour and the remaining transmittance, and the pixel colour builds up. Two sliders change the viewing angle horizontally and vertically, and the image is recomputed. An option emphasises the view-dependent shine. The scene function in the app is an analytic field standing in for the trained network.

Try it yourself

Increase the number of samples and watch the pixel colour settle. Step slowly through the march: once the ray hits dense material, transmittance drops and later points contribute almost nothing. Turn on the shine and change the view: the highlight moves across the surface. A plain collection of photos cannot do this, because it only contains the views that were captured.

Frequently asked questions

What is NeRF in simple terms?
NeRF stands for neural radiance field. A neural network learns from photos of a scene what colour and density each point in space has from each viewing direction. New views that were never photographed can then be computed.
How does a NeRF render an image?
For each pixel a ray is cast into the scene and sampled at many points. The colours of those points are summed, weighted by density and transmittance. This volume rendering is repeated for every pixel, which is computationally expensive.
What is the difference between NeRF and photogrammetry?
Photogrammetry reconstructs explicit geometry from triangulated feature points. NeRF learns a continuous function for colour and density and can represent view-dependent effects such as shine. In return, rendering a NeRF is much slower.
Does the app use a real trained NeRF?
No. The app uses an analytically defined field for colour and density in place of the network. The rendering process with ray, sampling and volume rendering follows the NeRF principle.

Subject: Computer Vision

More from Computer Vision

A VisuApp by heyprof: interactive, free in your browser, no sign-up. What is a VisuApp? · All apps