NeRF (Neural Radiance Fields) is a method for reconstructing a three-dimensional scene from a set of two-dimensional photographs with known camera poses. The key idea is to store the scene not as a point cloud or a polygon mesh, but as a continuous function represented by the weights of a small neural network: for any point in space and any viewing direction, the network returns the density of the medium and the emitted colour. An image from a new camera is then obtained by volume rendering – a series of points is sampled along each ray and their colours are composited into the resulting pixel with weights given by density. Because this whole process is differentiable, the network can be trained by ordinary gradient descent so that the rendered images match the input photographs; the scene is thus learned not from annotations but from itself. The result is photorealistic synthesis of novel views, including reflections and transparency. The original formulation suffered from very long training and slow rendering, which grid-based and hashed representations and especially the later Gaussian splatting resolved. Applications include virtual reality, heritage digitisation, film post-production, robotics and real-estate presentation.
You walk around a monument and photograph it from thirty angles. From those photos NeRF does not assemble a model out of triangles, as classical software does; instead it learns to answer a single question: “If I stood here and looked in this direction, what would I see?” It knows the answer for every place and every direction, including those you never shot from. So in practice you do not acquire a statue, but the ability to paint any view of it. And because that ability includes how light bounces off polished stone at different angles, the result looks like a photograph rather than computer graphics.