What is NeRF (Neural Radiance Fields)?

Introduced in a 2020 research paper, NeRFs made photorealistic novel-view synthesis possible from ordinary photos. Rather than building a mesh, a NeRF learns a function: give it a 3D position and a viewing direction, and it returns a colour and a density. Rendering an image means sampling that function along a ray for every pixel.

Strengths

• Strikingly realistic results, including reflections and soft details that meshes struggle with
• Handles fine structures like foliage and hair better than photogrammetry

Limits

• The original method was slow to train and render, though later work sped both up a lot
• It isn't a mesh, so it can't be dropped into a game engine or 3D-printed without converting it
• Editing the scene is hard

Much of the field's attention has since moved to Gaussian splatting, which renders in real time. Ideas from both feed into today's image-to-3D generators, which ultimately output standard meshes like GLB.