Neural Radiance Fields (NeRF) is a technique that uses a neural network to represent a 3D scene as a continuous volumetric function. Given a 3D coordinate and a viewing direction as input, the network outputs the corresponding color and volume density at that point. NeRF's key capability is photorealistic novel view synthesis and generating new camera perspectives of a scene from only a sparse set of 2D input images.
NeRF renders images through a process called volume rendering combined with ray marching. For each pixel, a ray is cast through the 3D scene and the neural network is queried at many sampled points along that ray. The resulting color and density values at each sample are composited together accumulating contributions weighted by opacity to produce the final pixel color.
Despite its impressive output quality, NeRF has notable limitations. Training a NeRF model is computationally expensive, typically requiring hours to days per scene on modern hardware. Inference is also slow, often taking seconds to render a single frame. Critically, a NeRF model is scene-specific: it encodes only the scene it was trained on and must be fully retrained to represent a different environment.
3D Gaussian Splatting has emerged as a prominent alternative, offering significantly faster rendering through an explicit point-based representation rather than NeRF’s implicit neural representation. Both techniques find application in product visualization, virtual staging for real estate, and immersive extended reality (XR) experiences, where the ability to synthesize photorealistic novel views from captured imagery provides a compelling advantage over traditional 3D modeling pipelines.