Neural Radiance Fields

A neural rendering technique that represents 3D scenes as continuous volumetric radiance functions encoded by multilayer perceptrons, mapping 5D inputs (3D spatial position plus 2D viewing direction) to colour and volume density. Novel viewpoints are synthesised by volumetric ray marching through the learned representation, enabling photorealistic view synthesis from sparse photograph sets without explicit 3D geometry. Introduced by Mildenhall et al. (ECCV 2020), NeRF has driven a generation of implicit neural scene representations spanning telepresence, virtual production, and robotics.

Semantic Classification

Content

Definition

Neural Radiance Fields (NeRF), introduced by Mildenhall et al. (ECCV 2020), revolutionised 3D scene reconstruction by representing scenes as continuous neural functions rather than discrete meshes or voxels. A NeRF encodes a scene’s geometry and appearance in the weights of a multilayer perceptron (MLP) that, given a 3D position (x, y, z) and viewing direction (θ, φ), outputs colour (RGB) and volume density (σ). Novel viewpoints are rendered by marching rays through the volume, sampling the neural function, and integrating colour/density via volumetric rendering equations, producing photorealistic images without explicit 3D geometry.

Current Landscape

NeRF has spawned 1,000+ research papers and commercial applications in telepresence TELE-053-volumetric-video-conferencing, virtual production, and VR TELE-020-virtual-reality-telepresence.

Technology Capabilities (2025):

• Training Time: 30 minutes for room-scale scenes (Instant-NGP TELE-060-instant-ngp)

• Rendering Speed: 30 FPS real-time variants (Mobile-NeRF, TensoRF)

• Quality: 32-36 dB PSNR (exceeding mesh-based methods)

  Comparison to Gaussian Splatting

  NeRF slower but more memory-efficient than TELE-051-3d-gaussian-splatting. Gaussian splatting now preferred for real-time telepresence.

  Related Concepts

• TELE-050-neural-rendering-telepresence

• TELE-051-3d-gaussian-splatting

• TELE-053-volumetric-video-conferencing

  References

  1. Mildenhall, B., et al. (2020). “NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis”. ECCV 2020.

Provenance