Nerfstudio

Overview

Launched in 2022 by Berkeley students, Nerfstudio is now developed by students and community contributors. Its modular API and command-line tools cover data processing, model training and evaluation, video rendering, viewing, and export. That breadth makes it a capable foundation for building a 3D capture workflow, but it is a developer-oriented project rather than a turnkey consumer scanning app.

The method lineup includes Nerfacto, Instant-NGP, NeRF, Mip-NeRF, TensoRF, and Splatfacto, its Gaussian Splatting implementation. This gives users several approaches to work with in one toolkit, rather than committing to a single representation. The tradeoff is that getting useful results still depends on a compatible GPU and comfort with a technical workflow.

Key features

Training and visualization

The command line handles processing, training, evaluation, rendering, and export. In the web viewer, users can watch training progress in real time, inspect scenes interactively, and render custom camera trajectories. That combination is valuable for iterating on a model and presenting a scene without separating training from visualization.

Share URLs let others view training progress, which can help with remote review. Those links are publicly accessible and introduce latency, so they are less suitable when access should be private or a responsive local view is essential. The viewer is a web application that connects to a server over a websocket.

Capture pipeline and development tools

Nerfstudio documents a capture-to-render workflow using COLMAP, Polycam, or Record3D to turn phone video into a 3D render. Tensorboard and Wandb logging, code profiling, and built-in debugging tools support more systematic experiments. Textured mesh output and exports in PLY, OBJ, PNG, and JPEG formats make the project useful beyond viewing a trained scene; it does not include measurement tools.

Pricing

Nerfstudio is free, with a free plan and no paid tier described. Its Apache-2.0 license suits users who want an open-source project they can incorporate into their work. The main cost is practical rather than financial: users need an NVIDIA video card with CUDA installed, and the repository quickstart says the library was tested with CUDA 11.8.

Installation is available through pip, from source, or with Docker. Docker requires NVIDIA GPU drivers capable of CUDA 11.8. The project directs questions to its Discord community and also provides tutorials, documentation, and a feedback form, giving new users several routes into a toolset that is more technical than a consumer app.

Platforms

Nerfstudio supports Linux, Windows, web, API access, and self-hosting. Linux is the safer installation choice: Windows installation is less tested and more fragile, and the project recommends Linux when available. The web viewer adds browser-based scene viewing, but it is connected to a server rather than replacing the GPU-backed workflow.

Capture input and mobile platforms are both supported. The documented COLMAP, Polycam, and Record3D pipeline is particularly relevant to users starting with phone video and moving toward model training on a computer.

Who it's for

Choose Nerfstudio if you want an open-source, GPU-backed environment for experimenting with NeRF and Gaussian Splatting methods, or need a documented route from capture to training, visualization, and export. Its logging and debugging support also suits users refining models as part of a development workflow.

Look elsewhere if you need a polished phone-first scanner, Windows reliability, or measurement tools. Nerfstudio’s CUDA requirement and command-line-centered workflow make it a poor fit for casual capture without compatible hardware or technical setup.

Pros and cons

• Pro: Multiple methods, including Splatfacto, let users explore NeRF and Gaussian Splatting workflows within one project.
• Pro: Training, evaluation, video rendering, interactive viewing, and export are covered in one command-line and web-viewer workflow.
• Pro: COLMAP, Polycam, and Record3D workflows connect phone-video capture to 3D rendering.
• Con: An NVIDIA CUDA GPU is required, narrowing the audience and making hardware a prerequisite rather than an optional upgrade.
• Con: Windows is less tested and more fragile than Linux, so Windows users face a less dependable installation path.
• Con: Public share URLs add latency, and there are no measurement tools for users who need to take measurements from scenes.

Alternatives

For a broader comparison of tools in this category, see AI 3D Scanning Software.

• WebODM is worth considering if you want a free, locally installed tool that can run offline, or need macOS support; its Starter plan is 24.00 USD per month billed yearly and includes 12 maps per 30 days.
• COLMAP is a free option for users who want a separate tool in the same capture workflow and support for macOS.
• Polycam is a better fit for a mobile-first capture app on iOS or Android; its free plan has limited space, object, and AI captures, with GLTF exports and public link sharing.
• RealityScan is an alternative for users seeking a freemium option with Android and iOS support.
• KIRI Engine may suit users who want a mobile and web scanning app with a free plan allowing up to 150 photos per scan, a 2 GB scan limit, unlimited scans, and free exports.
• OpenDroneMap is another free-software option for users who need macOS support or a self-hosted workflow.
• 3DF Zephyr is a Windows alternative with a free personal-use edition capped at 50 photos or video frames, one GPU, and limited exports and editing.
• Luma AI Video Generator is an alternative for users considering an API-based tool with paid monthly plans.

Verdict

Nerfstudio is a strong choice for developers and 3D creators who want a free, open-source toolkit spanning model training, scene viewing, capture workflows, and export. Its GPU requirement is the central hurdle, while Windows users also face a less dependable installation path. Choose it for a technical, extensible workflow on compatible hardware; look elsewhere for a simpler mobile scanner or a more reliable Windows-first experience.