Nerfstudio
Berkeley-born Python framework for training, viewing, and exporting NeRF and Gaussian splat scenes.
About
Started in 2022 by students in Angjoo Kanazawa's lab at Berkeley, Nerfstudio turns photos or video of a real place into a trainable neural radiance field that you can fly through live in a browser viewer. Methods are broken into swappable parts, so one pipeline can train its own nerfacto model, Instant-NGP, Mip-NeRF, TensoRF, NeuS, or splatfacto, its Gaussian splatting option. Each stage has a command: ns-process-data runs COLMAP or ingests captures from Polycam, Record3D, Metashape, and others; ns-train fits a model; ns-render makes videos along camera paths; ns-export writes point clouds, meshes, or splat files. Progress can stream to the viewer, TensorBoard, Weights and Biases, or Comet. An NVIDIA GPU with CUDA is required, and tiny-cuda-nn must be compiled during setup unless you use the Docker image. A SIGGRAPH 2023 paper documents the design. Nothing has been committed since July 2025 and the newest tagged release, v1.1.5, dates to November 2024, yet the code remains a common base for radiance field research. Apache 2.0 allows commercial use. Suited to researchers and 3D artists comfortable with Python and CUDA.
Should you use Nerfstudio?
Pick it when
Use Nerfstudio when you want a modular Apache-2.0 framework to process captures, train radiance fields, and inspect them in an interactive web viewer, or to prototype your own NeRF variant from reusable parts.
Look elsewhere when
Skip it if you need a textured mesh for games, where Meshroom fits better, or only need a splat rasterizer inside your own code, where gsplat is leaner. For paper baselines use the original Gaussian Splatting code.
Alternatives to Nerfstudio
- gsplat
Just the CUDA Gaussian rasterizer as a pip library, leaner if you are writing your own training loop rather than using a framework.
- Gaussian Splatting
The original 3DGS reference with real-time rendering, but research-use only and without Nerfstudio's viewer and data tooling.
- Meshroom
Photogrammetry to textured meshes through a node GUI, better when you need editable geometry rather than radiance fields.
- VGGT
Recovers cameras and point maps from many images in seconds with COLMAP export, a fast pose step before radiance-field training.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Category
- 3D Model Generation
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Intermediate (3/5)
- License
- Apache-2.0
- Minimum VRAM
- 8 GB
- Added
- Apr 3, 2026