Skip to content

nerfstudio — How-To (Training, Export, and Visualization)

This document describes the workflow used for the nerfstudio (splatfacto) implementation, including:

  • Environment setup (Windows)
  • Dataset preparation
  • Training
  • Export of the Gaussian Splatting reconstruction

References

Important Note

During the original experiments, nerfstudio was installed inside an environment where a working gsplat installation was already present.

Because of this:

  • CUDA extensions were not compiled from scratch
  • The installation procedure is not fully reproducible
  • Some official installation steps were not required

The installation guide described below follows the official documentation for reproducibility.

1. Overview

nerfstudio is a neural rendering framework providing tools for data preparation, training and visualization.

The Gaussian Splatting pipeline (splatfacto) requires known camera poses and intrinsics.

nerfstudio provides the command ns-process-data to process the data into the nerfstudio format.

The workflow therefore consists of two stages:

  1. Dataset preparation (ns-process-data or existing data with ns-download-data)
  2. Gaussian Splatting optimization (ns-train)

In this work, ns-process-data was used.

2. Prerequisites

  • NVIDIA GPU with CUDA support
  • PyTorch with CUDA support (tested runtime with CUDA 11.8)
  • Python 3.8+ (Conda/Miniconda recommended) (tested with Python 3.10.19)
  • Git

3. Create the Environment

Using Conda:

conda create -n nerfstudio python=3.8 -y
conda activate nerfstudio
python -m pip install --upgrade pip

Install PyTorch compatible with CUDA 11.8:

pip install torch==2.1.2+cu118 torchvision==0.16.2+cu118 --extra-index-url https://download.pytorch.org/whl/cu118
Install the cuda-toolkit:

conda install -c "nvidia/label/cuda-11.8.0" cuda-toolkit

Install the torch bindings for tiny-cuda-nn:

pip install git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch

Install nerfstudio from pip:

pip install nerfstudio

Or install nerfstudio from source:

git clone https://github.com/nerfstudio-project/nerfstudio.git
cd nerfstudio
pip install --upgrade pip setuptools
pip install -e .

4. Dataset Processing

In order to run the preprocessing, the command ns-process-data is used. The data can be a video or images.

ns-process-data {images, video} --data {DATA_PATH} --output-dir {PROCESSED_DATA_DIR}

Example (images):

ns-process-data images --data C:\Users\ernes\nerfstudio\data\video-interno\input --output-dir C:\Users\ernes\nerfstudio\data\video-interno-processed

5. Training

The pipeline used in the benchmark is splatfacto.

ns-train splatfacto --data {DATA_PATH} --output-dir {OUTPUT_DIR}

Example:

ns-train splatfacto --data C:\Users\ernes\nerfstudio\data\video-interno-processed --output-dir C:\Users\ernes\nerfstudio\results\video-interno --pipeline.model.stop-split-at 20000 --max-num-iterations 30000

Parameter description

  • splatfacto → Gaussian Splatting training pipeline implemented in nerfstudio
  • --data → path to the processed dataset
  • --output-dir → directory where the output is stored
  • --pipeline.model.stop-split-at 20000 → stops Gaussian densification after 20k iterations
  • --max-num-iterations 30000 → total number of optimization iterations

At the end of the terminal is present a link that loads the webviewer.

After training, the folder contains:

<results>
└── splatfacto
    └── <experiment_name>
        ├── config.yml
        └── ...

6. Export

The command ns-export is used for exporting .ply from the config.yml file.

ns-export gaussian-splat --load-config {config.yml} --output-dir {OUTPUT_DIR}

Example:

ns-export gaussian-splat --load-config C:\Users\ernes\nerfstudio\results\video-interno\splatfacto\2026-01-20_131913\config.yml --output-dir C:\Users\ernes\nerfstudio\results\video-interno\exported

Parameter description

  • gaussian-splat → export format for 3D Gaussian Splatting reconstruction
  • --load-config → path to the config.yml generated after training (contains dataset path, model parameters and checkpoint reference)
  • --output-dir → directory where the exported reconstruction will be saved