NeRF
A neural network that represents a 3D scene by learning to predict color and density for any ray direction through space.
Think of it like
Instead of storing a 3D mesh or voxel grid, you have a function: "Give me a ray, Ill tell you what color and density it sees."
Example
Input: multiple photos of a sculpture from different angles. NeRF learns to synthesize any novel viewpoint, rendering high-quality new photos.
How it actually works
Encodes 3D space as a function: f(x, y, z, theta, phi) -> (r, g, b, sigma). Positional encoding (Fourier features) maps coordinates into high-dimensional space for better learning. Volume rendering integrates along rays (standard differentiable raycasting). Training: minimize photometric loss on input views. Advantages: compact, high quality, no mesh topology. Disadvantages: slow rendering (1-100s per frame), tricky optimization, requires many input views.
For product teams
Novel view synthesis, 3D reconstruction, VR/AR content creation.
For engineers
Input: multi-view images + camera poses. Output: implicit 3D representation rendering any viewpoint. Metric: PSNR, LPIPS, SSIM.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome