Adopted from Gaussian Splatting By Kerbl et. al

This is a final project code submission for Brown University's CS1430 With Professor Srinath Sridhar

• Setup
• Running the Code

The code has been tested in the following setup

• Ubuntu 22.04 and WSL/Ubuntu24.04
• Python 3.12
• PyTorch 2.5
• CUDA 12.4-12.6

The provided environment.yml serves as a suggestion:

$ conda env create -f environment.yml
$ conda activate g-splat

You need to compile the diff-gaussian-rasterizer and fused-ssim manually:

NOTE: You might run into a compilation issue in diff-gaussian-rasterizer. Easily fixable by adding #include <cstdint> in the file that complains.

$ pip install 3rdparty/diff-gaussian-rasterization
$ pip install 3rdparty/fused-ssim

We have 3 modes:

1. Single-image: overfits on a single image
2. Blender: Runs on our synthetic Blender-generated datasets
3. Colmap: More in-line with the upstream gaussian splatting data format

See examples below

python main.py --mode blender --data data/yosemite1.jpg --viz_interval 100

python main.py --mode blender --data data/monkey --viz_interval 5

python main.py --mode colmap --data data/db/drjohnson --viz_interval 5

The code is pretty bare-bones. Please adjust they hyperparameters stright in the code. Specifically see main.py and scene.py.

Not for commercial use :)