instant-ngp

Why we included this project

For anyone doing 3D reconstruction, this is the reference implementation for training neural radiance fields (NeRF) in seconds instead of hours. It bundles four neural graphics primitives, NeRF, signed distance functions, neural images, and neural volumes, on a multiresolution hash encoding that keeps a small MLP fast enough for real-time rendering. Photogrammetry and computer-vision teams will find it the quickest practical path from a set of photos to an interactive 3D scene, and the bundled viewer lets you fly around the result right away. The code is CUDA-heavy and expects a modern NVIDIA GPU, so it fits researchers and engineers who want a solid baseline for novel-view synthesis rather than a plug-and-play tool. It is also a useful reference for anyone studying how hash-based input encodings speed up neural field training.