Nerfstudio
Nerfstudio is a free, open-source Python framework for building, training and exporting neural radiance fields and Gaussian splats. It bundles a real-time web viewer, eight CLI commands and 24 documented methods, and runs entirely on your own NVIDIA GPU.
What is Nerfstudio?
Nerfstudio is an open-source Python framework covering the whole life cycle of a neural radiance field: turning photographs or video into a 3D scene, training a model on it, inspecting the result, and exporting it into something another tool can use. It started in October 2022 as a research project by Berkeley students in the KAIR lab at Berkeley AI Research, and was published at SIGGRAPH 2023 under the title Nerfstudio: A Modular Framework for Neural Radiance Field Development. It is now maintained by Berkeley students and outside contributors, with Luma AI and the BAIR commons listed as sponsors.
The governing idea is modularity. Rather than shipping one monolithic implementation, Nerfstudio splits a radiance field into interchangeable parts, namely data parsers, data managers, models, fields and pipelines, so that a researcher can swap a single component and leave the rest untouched. That structure is what lets the project document 24 methods side by side: six shipped in the box, including Nerfacto, the recommended all-round method, and Splatfacto, the in-house Gaussian splatting implementation, plus a long tail of third-party contributions such as Zip-NeRF, LERF, Instruct-NeRF2NeRF, Nerfbusters and SeaThru-NeRF.
Day to day, the framework is a command line tool. Eight commands cover the pipeline: ns-process-data prepares your footage, ns-train trains a model, ns-viewer reopens a finished one, ns-render produces video, ns-export writes meshes and point clouds, and ns-eval scores the outcome. A training run automatically launches a real-time web viewer on port 7007, where the scene can be explored while it is still converging and camera paths can be laid out by hand.
It is free software under the Apache-2.0 licence, installed with a single pip command, and it runs entirely on your own hardware, which also means an NVIDIA GPU with CUDA is not optional. What it is not is a service: no account, no hosted rendering, no pricing page and no company behind it in the legal sense. The site is a Sphinx documentation build, and that documentation is the product's only front door.
What it does
- Train a NeRF or a Gaussian splatting scene from photographs or video
- Watch training progress live in a browser-based 3D viewer
- Lay out a camera trajectory and render it out as a video
- Export a textured mesh or a point cloud from a trained scene
- Produce stereo VR output in VR180 or 360 omni-directional stereo
- Composite NeRF objects into Blender, Maya or Unreal Engine scenes
- Evaluate and benchmark methods with ns-eval and the bundled scripts
When to use Nerfstudio / When not to
A quick filter to help you decide if Nerfstudio is the right fit.
When to use Nerfstudio
- Computer-vision researchers implementing or comparing NeRF and Gaussian splatting methods
- PhD and master's students learning neural 3D reconstruction hands-on
- VFX artists and 3D animators bringing real-world captures into Blender, Maya or Unreal Engine
- Developers building a 3D capture product on a permissively licensed codebase
- Teams with an NVIDIA GPU workstation and a working comfort with the command line
When not to use Nerfstudio
- Anyone without an NVIDIA GPU: Apple Silicon Macs, AMD cards and CPU-only machines are not covered
- Users who want a click-and-go desktop application rather than a command line
- Organisations that require a contract, a DPA or a documented GDPR position, none of which exists
- Teams looking for a hosted service, since Nerfstudio ships no online product at all
- Buyers who need guaranteed commercial support or a service level agreement
How to use Nerfstudio
A typical end-to-end flow, from setup to results.
- Check the hardware first: Nerfstudio requires an NVIDIA card with CUDA installed, tested against versions 11.7 and 11.8
- Create a Python environment, with the documentation recommending conda and Python 3.8 or newer
- Install PyTorch with CUDA support, then the cuda-toolkit, then the tiny-cuda-nn torch bindings
- Install the framework itself with pip install nerfstudio, or clone the repository for the development version
- Skip that whole chain if you prefer: pull the official Docker image, use pixi on Linux, or open the Google Colab demo notebook
- Fetch a sample capture to confirm the setup works, with ns-download-data nerfstudio --capture-name=poster
- Launch a first training run: ns-train nerfacto --data data/nerfstudio/poster
- Open the link printed in the terminal to reach the viewer on port 7007, forwarding that port over SSH if you train on a remote machine
- For your own footage, run ns-process-data, which converts images, video or phone captures into the Nerfstudio format once camera poses have been computed
- Finish in the RENDER tab for a video, or with ns-export for a .ply point cloud or a textured .obj mesh
Pros & Cons
Pros
- Free under Apache-2.0, with the entire source open to inspection and modification
- Runs fully locally, with no account, no upload and nothing sent to a publisher
- Unusually thorough documentation: every command, every API module and every method has its own page
- A modular architecture that lets you replace one component without rewriting the pipeline
- A real-time browser viewer included by default, with no extra setup
- Official bridges to Blender, Maya and Unreal Engine for production work
- A large, active community of 11,900 stars, 1,700 forks and a public Discord, behind a SIGGRAPH-published framework
Cons
- An NVIDIA GPU with CUDA is mandatory, which rules out Apple Silicon, AMD and CPU-only setups
- Installation is heavy, involving PyTorch, the cuda-toolkit and a tiny-cuda-nn compilation step
- The documentation itself calls Windows less tested and more fragile, and recommends Linux instead
- Camera poses must come from an external tool, usually COLMAP, whose install the docs admit is often troublesome
- No legal framework whatsoever: no terms, no privacy policy, no GDPR statement and no DPA
- No email address and no commercial support, so help means Discord or a GitHub issue
- No company is legally responsible for the software, and no maintenance is contractually promised
Pricing & Plans
Nerfstudio is free of charge. The software is released under the Apache-2.0 licence and installed with a single pip command; there is no pricing page, no paid tier, no enterprise plan and no account to open, so no lowest price point can be quoted. The real expenditure sits elsewhere: an NVIDIA graphics card capable of running CUDA, the electricity and machine time that training consumes, and any commercial third-party capture software you choose to feed it with, such as Metashape, RealityCapture or Polycam.
Data, GDPR & hosting
A consolidated view of how Nerfstudio handles your data.
GDPR overview
There is no GDPR mention on this site at all. A search across every page collected, from the documentation homepage through installation, quickstart, contributing, data, export, CLI and extension pages, plus the GitHub repository, returns zero occurrences of GDPR, data protection or any equivalent wording. There is no privacy policy, no named controller, no data protection officer, no Article 27 representative and no privacy contact address. The N recorded here therefore reflects complete silence rather than an explicit refusal, and the distinction matters: Nerfstudio processes no personal data on the publisher's side, because there is no account, no online service and no documented collection of any kind. Anyone with a regulatory obligation should note that the absence of documentation is itself the finding.
Who owns the data?
The site publishes no terms of service, no privacy policy and no data processing agreement, so no contractual text states who owns what. What the documentation does establish is that Nerfstudio is installed and executed on your own machine: input footage, camera poses, model checkpoints and exports never leave it, and there is no account to create. The Apache-2.0 licence governs the source code, not your captures. In practice you keep your data because the publisher never receives any of it, but that is an inference drawn from the architecture rather than a commitment anyone has written down. The absence of a contract remains an open question.
Reuse rights
No terms and conditions exist, so nothing in writing either restricts or authorises the reuse of your own material. The Apache-2.0 licence covers the code itself and permits commercial use, modification and redistribution provided the original notices are preserved. Everything the pipeline produces, from trained models to rendered videos, meshes and point clouds, is written to a local outputs directory and is yours to exploit as you see fit, since no third party is party to the exchange. Two exceptions deserve attention: the optional Weights & Biases and Comet loggers transmit training metrics to outside services once enabled, and the viewer's share link deliberately makes a scene publicly reachable.
Data retention & training
Hosting summary
There is no publisher-side hosting to describe. Nerfstudio is a Python package installed on your own machine; the training run, the viewer and the export steps all execute locally, and the outputs are written to a directory you control. No country and no hosting region is mentioned anywhere on the site, because no user data ever reaches a server operated by the project. Two things should not be confused with the software itself. The documentation is served by GitHub Pages, and the nerf.studio domain resolves to 34.94.31.43, an address geolocated to Los Angeles in the United States on Google infrastructure; that concerns the website, not your captures. Separately, the optional experiment trackers the framework can log to, namely Weights & Biases and Comet, are third-party services with their own hosting arrangements, and enabling them is a deliberate choice on your part. In short, your data sits where you put it, under whatever jurisdiction your own hardware sits in. No data processing agreement, no subprocessor list and no jurisdiction statement exists, because the architecture never creates the need for one.
Things to keep in mind
Risks and trade-offs to weigh before adopting Nerfstudio.
- No terms of service, no privacy policy and no legal notice are published on either nerf.studio or docs.nerf.studio
- The site never mentions the GDPR, so nothing here can be relied upon in a compliance review
- No legal entity stands behind the software: the copyright line simply reads nerfstudio Team, with no address and no email
- The Apache-2.0 licence supplies the software as is, without warranty of any kind
- The viewer's share option creates a publicly reachable link, which deserves care with sensitive captures
- Input data usually passes through third-party tools such as COLMAP, Polycam or Record3D, each with its own policies
- Nothing commits anyone to maintaining the project: no roadmap obligation, no SLA and no support contract
Setup & Integrations
Technical difficulty
High for a first install, low afterwards. The full route runs conda, then PyTorch with CUDA, then the cuda-toolkit, then a tiny-cuda-nn compilation, then pip install nerfstudio, which is comfortable only if you already manage Python environments. Windows additionally requires Git, Visual Studio 2022 and its C++ build tools, activated before CUDA. Two shortcuts exist: the official Docker image removes the CUDA setup entirely, and pixi installs COLMAP, tinycudann and hloc in one command on Linux. For a first look, the Colab notebook needs nothing at all.
Deployment
Integrations
Behind Nerfstudio
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
Does Nerfstudio cost anything?
What hardware do I need to run it?
Can I run it on a Mac or without a GPU?
Is there a way to try it without installing anything?
What can I export from a trained scene?
Does it handle Gaussian splatting as well as NeRFs?
Is any of my data sent anywhere?
How do I get support?
Can I use it alongside Blender or a game engine?
Who actually develops Nerfstudio?
Should you pick Nerfstudio?
Nerfstudio has become the default entry point into neural radiance fields, and the reasons are shown rather than claimed. Twenty-four methods are documented one by one, eight command line tools are specified down to their flags, the Python API reference is generated from the code itself, and twelve capture pipelines are described end to end. Behind that sit a SIGGRAPH 2023 paper, 11,900 GitHub stars, 1,700 forks and six named maintainers who put their institutional affiliation next to their handle. Very few free tools show their work this completely.
The price of entry is technical rather than financial. You need an NVIDIA card with CUDA, patience with a PyTorch and tiny-cuda-nn installation that the project itself admits is fragile on Windows, and a working relationship with the command line. COLMAP, on which most custom captures depend, adds friction of its own. The Docker image and the Colab notebook soften the first hour considerably, and are the sensible way to find out whether the tool fits your work.
What deserves a clear-eyed look is everything outside the code. There is no company, no postal address, no email, no terms of service, no privacy policy and no mention of the GDPR anywhere on the site. For a research framework that never sees your data, that is coherent rather than alarming, but it means Nerfstudio should be assessed the way a library is assessed, not the way a vendor is. Nobody is contractually obliged to maintain it, to answer you, or to keep the documentation online.
On those terms it is an excellent choice for researchers, students, VFX artists and developers who want a permissive, inspectable, locally run foundation for 3D reconstruction. Organisations that need contracts, guarantees or compliance paperwork will have to look at a commercial product instead.
- Choosing a selection results in a full page refresh.
- Opens in a new window.