How Neural Radiance Fields and Gaussian Splatting Reconstruct 3D Scenes from Photos

Neural Radiance Fields (NeRF), introduced by Mildenhall and colleagues in 2020, encode a single specific scene into a small neural network of roughly 600,000 parameters trained on 50–100 photographs, with no mesh or texture atlas involved. The technique casts rays through each pixel and queries the network at multiple points along each ray, making rendering computationally expensive — around 92 million network evaluations per frame. NVIDIA's Instant-NGP (2022) addressed this bottleneck by replacing the bulk of the network with a multiresolution hash table of feature vectors, cutting training time from hours to minutes and enabling interactive rendering at the cost of higher memory usage. 3D Gaussian Splatting, presented at SIGGRAPH 2023 by Kerbl and colleagues, eliminated the neural network from scene representation entirely, instead modeling scenes as explicit 3D Gaussian blobs with per-element parameters. This shift to a fully explicit representation is why Gaussian Splatting can render scenes in milliseconds rather than seconds.