Trying Out 3D Gaussian Splatting Part 1 (Environment Setup, WSL2 Ubuntu 20.04)

2026年2月19日 2026年2月19日

管理人

table of contents

Info

This article is translated from Japanese to English.

https://404background.com/program/3dgs-1/

Introduction

In this post, I set up the environment to use 3D Gaussian Splatting.
I have been creating 3D models using Scaniverse, a smartphone app, and 3D Gaussian Splatting is the technology used behind it.
▼3D Gaussian Splatting was also featured in NVIDIA's Physical AI announcement.

Since I cannot use Scaniverse on my usual iPhone 8, I’ve been wanting to run it on my PC.
▼Previous articles are here:

Building the Environment

Changing the CUDA Version

▼I am using a gaming laptop purchased for around 100,000 yen, running Windows 11.

【Amazon.co.jp限定】 ASUS ゲーミングノートPC TUF Gaming A15 FA506NCR 15.6インチ RTX 3050 AMD Ryzen 7 7435HS メモリ 16GB SSD 512GB リフレッシュレート144Hz RGB イルミネートキーボード Windows 11 動画編集グラファイトブラック FA506NCR-R7R3050A

ASUS

Amazon

ポチップ

https://amzn.to/4aaSMlT

Building this environment involves dependencies like CUDA, and I set it up after several trials within a WSL2 Ubuntu environment.
▼The Ubuntu 20.04 environment for WSL2 was downloaded in the following article:

The README mentioned that it was verified with CUDA 11.8, so I switched to CUDA 11.8.
▼The GitHub repository is here:

https://github.com/graphdeco-inria/gaussian-splatting

▼Checking with "nvidia-smi" beforehand showed CUDA 12.6 was installed.

Note that "nvidia-smi" shows the CUDA version on the Windows side, while the CUDA version in the WSL2 environment can be checked with "nvcc --version."
▼Sometimes different versions are installed.

While using CUDA's "update-alternatives" seemed like a good method, I didn't try it this time. Since I’ll likely need to switch CUDA versions again in the future, I want to verify that method thoroughly another time.
▼The following articles seemed helpful:

https://note.com/kmykpurin/n/n2bf91dc63f16

I executed the commands while consulting with ChatGPT.
First, I removed the existing installations:

sudo apt-get --purge remove "cuda*"
sudo apt-get --purge remove "nvidia*"
sudo apt autoremove -y

Next, I installed the packages for Ubuntu 20.04 and CUDA 11.8:

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/cuda-ubuntu2004.pin
sudo mv cuda-ubuntu2004.pin /etc/apt/preferences.d/cuda-repository-pin-600
wget https://developer.download.nvidia.com/compute/cuda/11.8.0/local_installers/cuda-repo-ubuntu2004-11-8-local_11.8.0-520.61.05-1_amd64.deb
sudo dpkg -i cuda-repo-ubuntu2004-11-8-local_11.8.0-520.61.05-1_amd64.deb
sudo cp /var/cuda-repo-ubuntu2004-11-8-local/cuda-*-keyring.gpg /usr/share/keyrings/
sudo apt-get update
sudo apt-get -y install cuda

Finally, I added the environment variables to ".bashrc" so they are reflected whenever the terminal starts:

echo 'export PATH=/usr/local/cuda-11.8/bin:$PATH' >> ~/.bashrc
echo 'export LD_LIBRARY_PATH=/usr/local/cuda-11.8/lib64:$LD_LIBRARY_PATH' >> ~/.bashrc
source ~/.bashrc

I verified the CUDA version with "nvcc -V" or "nvcc --version."
▼It was successfully changed to 11.8.

Installing Miniconda

Since I needed to run "conda" commands, I installed Miniconda.
▼The installation method is described on the following page:

https://www.anaconda.com/docs/getting-started/miniconda/install#macos-linux-installation

I used the commands for Linux x86:

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash ~/Miniconda3-latest-Linux-x86_64.sh
source .bashrc

▼The conda environment is now activated.

Installing Gaussian Splatting

Download the Gaussian Splatting repository, then create and activate the conda environment:

git clone https://github.com/graphdeco-inria/gaussian-splatting --recursive
cd gaussian-splatting
conda env create --file environment.yml
conda activate gaussian_splatting

▼When you reopen the terminal later, running the last command will activate it again.

Running a Sample

I tried reconstructing a 3D model using the provided COLMAP data.
I downloaded the "T&T+DB COLMAP (650MB)" linked at the beginning of the README.
▼The link is located at the top of the image.

https://github.com/graphdeco-inria/gaussian-splatting

I added the downloaded folder to the execution directory and ran the command:

python train.py -s tandt_db/tandt/train/

▼The process started. It appears there are iterations at 7,000 and 30,000.

After the process finished, the results were saved in the "output" folder.
▼The .ply files are the 3D models.

I used SuperSplat, which allows you to edit 3D Gaussian Splatting .ply files in a browser, to check the results.
▼SuperSplat is here:

https://superspl.at/editor

▼Result of iteration_7000:

▼Result of iteration_30000:

From a glance, it was hard to tell the difference between 7,000 and 30,000 iterations. My main impression was that the surrounding "mist" (artifacts) was quite noticeable.

Finally

Now that I can run 3D Gaussian Splatting on my PC, I'm curious to see if I can reconstruct 3D models for much larger environments than I could with Scaniverse. Scaniverse was convenient and easy for trimming parts I wasn't focused on, so I wonder if there are tools for PC that improve on this aspect.
Since I’ve successfully run the sample, I plan to try 3D model reconstruction from my own images and videos next.