nerfstudio-project/nerfstudio

Nerfstudio is a Python framework for working with NeRFs, which stands for Neural Radiance Fields. A NeRF is a technique that takes a set of photos of a real-world object or scene, trains a neural network on them, and produces a model that can then generate new views of that scene from any camera angle, including angles that were never photographed. The result is a kind of interactive 3D reconstruction made entirely from ordinary photos. This library, started by students at Berkeley AI Research in 2022, aims to make that process more approachable. It breaks the NeRF pipeline into modular, documented pieces so that researchers and developers can swap out individual components, try different approaches, and understand what each part does. The goal is to lower the barrier for people learning about NeRF techniques, while also giving more experienced researchers a flexible base for experiments. To use nerfstudio you need a computer with an NVIDIA graphics card and CUDA installed, since training a NeRF requires GPU processing. After installing via pip or conda (or running a Docker image), you can start training on a provided test dataset with a single command. The default model is called nerfacto, which the project recommends for real-world scenes. Training produces a model you can explore through a browser-based 3D viewer that lets you move a virtual camera through the reconstructed scene. The library supports multiple different NeRF methods beyond the default, and the project encourages contributors to add new models, datasets, and improvements. There is a Discord community for questions, a feedback form for suggestions, and documentation hosted separately. The project is sponsored by Luma AI and the Berkeley AI Research Commons and is available under the Apache 2.0 license.