What is a neural radiance field?
A neural radiance field reconstructs a 3D scene from a collection of ordinary photos by training a small neural network to represent it. The network learns to output the color and density of any point in space as seen from any direction, so that once trained, it can render the scene from viewpoints the cameras never actually captured, with convincing lighting and detail.
It reframes 3D reconstruction as learning a continuous function of the scene rather than building an explicit mesh.
Key takeaways
- NeRF represents a 3D scene as a neural network learned from images.
- It renders photorealistic new viewpoints not in the original photos.
- It captures appearance and geometry implicitly and continuously.
How it works
Given many images with known camera poses, the network is trained to predict color and density along rays cast into the scene, and those samples are combined through volume rendering to reproduce each training image. Minimizing the difference between rendered and real images teaches the network the scene's structure. New views are produced by casting rays from a new camera and rendering them the same way.
Why it matters
NeRF made high-fidelity 3D capture possible from ordinary photographs, which is valuable for simulation, asset creation, and building realistic environments for physical AI. It also spurred faster follow-on methods like Gaussian splatting that trade some of its qualities for speed.
Frequently asked questions
What does a NeRF need as input?
A set of images of a scene with known camera positions, from which it learns a renderable 3D representation.
How does NeRF compare to Gaussian splatting?
Both reconstruct 3D scenes from images, but Gaussian splatting uses an explicit set of primitives that render much faster, while NeRF uses an implicit neural representation.
Related terms