RF

In optics and computational imaging, RF stands for radiance field: a mathematical representation of how light travels through a volume of space. Rather than storing a single image, an RF model captures the intensity and direction of light rays at every point in a three-dimensional region. This allows reconstruction of photorealistic views from any camera angle, even angles where no photograph was originally taken.

The radiance field concept builds on decades of computer graphics theory, but became computationally practical around 2019 with the introduction of neural radiance fields (NeRF). A neural radiance field uses a deep neural network to learn and compress the light field data from a set of input photographs taken around an object or scene. The network learns to predict the radiance (brightness and color) along any ray passing through space, indexed by position and viewing direction. This makes RF systems particularly valuable in visual effects, 3D product visualization, and scientific imaging where dense capture geometry is infeasible.

Capture and computation

Creating an RF requires multiple input photographs of the target subject taken from known camera positions and orientations. A typical capture setup might use 20 to 100+ images depending on scene complexity and desired reconstruction quality. The input images are processed to estimate camera intrinsics, depth, surface normals, or other geometric cues, which the radiance field model uses as constraints. Rendering a novel view involves querying the learned field along rays emanating from a virtual camera, accumulating light contributions to produce a pixel value.

The primary limitation of current RF methods is memory and compute cost during training and inference. A high-resolution NeRF model can require hours of GPU training and tens of seconds per rendered frame on commodity hardware. Variants such as instant neural graphics primitives and factored radiance fields attempt to reduce this burden through compact representations and acceleration structures, but tradeoffs between speed, quality, and model size remain active research areas.

RF sits at the intersection of photogrammetry, neural rendering, and computer graphics. Unlike traditional 3D reconstruction (which produces explicit geometry) or image-based rendering (which blends input photographs), radiance fields implicitly encode both geometry and appearance in a learned function. This hybrid nature makes them less interpretable than mesh-based models but more flexible for high-fidelity novel-view synthesis in scenarios where ground-truth geometry is difficult to acquire.