instant-ngp

Instant Neural Graphics Primitives is a CUDA-based research project for training and rendering neural radiance fields, signed distance functions, neural images, and neural volumes. Each primitive uses an MLP with multiresolution hash input encoding through NVIDIA's tiny-cuda-nn framework. The downloadable Windows application provides an interactive GUI for loading NeRF datasets, OBJ SDFs, EXR/images, and NanoVDB volumes. It supports real-time scene exploration, training controls, DLSS-related rendering options, crop controls, snapshot save/load, camera-pose and lens optimization, NeRF/SDF-to-mesh conversion, and camera-path keyframing for MP4 or JSON export. An OpenXR-compatible VR mode can connect to Oculus, SteamVR, or other supported runtimes. For automation and experimentation, the project includes Python bindings and a scripts/run.py interface, plus headless operation and a Colab notebook. Building from source is supported on Windows and Linux with an NVIDIA GPU, C++14 compiler, CUDA, CMake, and optionally Python, OptiX, and the Vulkan SDK. A Dev Container/Docker setup is also provided. The repository includes dataset guidance, example scenes, keyboard/VR controls, troubleshooting notes, and citation information. The code is released under the Nvidia Source Code License-NC, so commercial use is not granted by that license.