Instant Neural Graphics Primitives is an NVIDIA research project that implements four neural graphics primitives: neural radiance fields (NeRF), signed distance functions (SDFs), neural images, and neural volumes. The core innovation is the use of multiresolution hash input encoding combined with multilayer perceptrons trained via the tiny-cuda-nn framework, enabling dramatically faster training and rendering compared to previous approaches. The project was published in ACM Transactions on Graphics at SIGGRAPH in July 2022 by Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller.

The repository is written in CUDA and provides both a standalone executable application with an interactive graphical interface and Python bindings for developers. The application demonstrates the capability to train NeRF models in under five seconds, as exemplified by the included fox dataset. Users can load NeRF-compatible datasets from various sources including the original NeRF dataset, SILVR dataset, and DroneDeploy dataset, or create their own using provided tutorials.

The interactive GUI includes comprehensive features for exploring neural graphics primitives. Users can navigate scenes using WASD controls, adjust camera velocity and exposure, toggle training on and off, and switch between different visualization and render modes. The interface supports saving and loading snapshots for sharing trained models, includes a camera path editor for video creation, and provides NeRF-to-Mesh and SDF-to-Mesh conversion capabilities. Additional features include camera pose and lens optimization, multi-view visualization of neural model layers, and accumulated error map visualization.

The application supports virtual reality viewing through OpenXR-compatible runtimes, including OculusVR and SteamVR. VR users can navigate scenes with stick controls, manipulate the neural graphics primitive with grab gestures, and erase portions of the NeRF by pressing the stick near the hand.

Windows users can download precompiled releases optimized for specific GPU architectures including RTX 5000 series Blackwell cards, RTX 3000 and 4000 series Ampere and Ada cards, RTX 2000 series Turing cards, and GTX 1000 series Pascal cards. Linux users and those requiring Python bindings must build the project from source using CMake, requiring a C++14 capable compiler, CUDA 10.2 or higher, and optionally OptiX 7.6 or higher for faster mesh SDF training.

According to GitGenius activity tracking, the repository has a median issue and pull request response latency of approximately 9073.6 hours with a mean of 10024.7 hours across 81 tracked items. The most active contributors tracked include Tom94 with 5 events, yoponchik with 4 events, and FlyCole with 3 events. The repository shares overlapping contributors with microsoft/vscode, microsoft/typescript, and rust-lang/rust, indicating cross-project collaboration within the broader software development community.