Core principles of a Neural Radiance Field

WHY THIS MATTERS NeRF reframes novel-view synthesis as learning a continuous volumetric function of a scene rather than reconstructing explicit geometry, achieving photorealistic view interpolation from a set of posed images.

WHAT THIS TESTS Whether you understand the implicit MLP representation, the role of view direction and density, and the differentiable volume-rendering step.

A GOOD ANSWER COVERS A NeRF represents a single scene implicitly in the weights of a small multilayer perceptron. The MLP takes a five-dimensional input: a 3D spatial location and a 2D viewing direction. It outputs an emitted RGB color and a scalar volume density at that point. Density depends only on position, ensuring geometric consistency, while color depends on both position and direction so the model can reproduce view-dependent effects like specular highlights. Positions and directions are passed through a positional encoding of sinusoids so the MLP can represent high-frequency detail. To synthesize a novel view, for each pixel cast a camera ray into the scene, sample a series of 3D points along that ray, query the MLP at each sample to get color and density, then composite them with the volume rendering integral: each sample's contribution is its color weighted by its density and by the accumulated transmittance, the probability light reached it unobstructed. Summing along the ray gives the pixel color. Because every step is differentiable, NeRF is trained by rendering rays from known training cameras and minimizing the squared error against the real pixels.

COMMON WRONG ANSWERS Saying NeRF stores an explicit mesh, voxel grid, or point cloud; it is a continuous implicit function. Forgetting view dependence in color. Omitting volume rendering and just sampling one point.

LIKELY FOLLOW-UPS Why positional encoding? Why hierarchical coarse-to-fine sampling? Why is vanilla NeRF slow and how do Instant-NGP or Gaussian Splatting speed it up? How does it differ from photogrammetry?

ONE CONCRETE EXAMPLE Given 50 photos of a statue with known poses, train the MLP so its rendered rays match the photos; at inference place a virtual camera at an unseen angle, march rays through the implicit field, volume-integrate, and obtain a photorealistic new image of the statue, including correct highlights.