TNT-GS: Truncated and Tailored Gaussian Splatting.

You can explore our visual demonstrations at the link below:

TNT-GS Webpage

More examples will be added soon—stay tuned!

This repository includes submodules. To clone it properly, please use one of the following commands:

# SSH
git clone git@github.com:GoogolplexGoodenough/TNT-GS.git --recursive

or

# HTTPS
git clone https://github.com/GoogolplexGoodenough/TNT-GS --recursive

And install the torch and torchvision packages from the official Pytorch website. Make sure that the CUDA version matches your local NVCC version to ensure proper GPU support, especially for compiling CUDA-based submodules. Once PyTorch is installed, use the following command to install the remaining dependencies:

pip install -r requirements.txt
pip install submodules/simple-knn
pip install submodules/diff_TNT_rasterization

To launch the optimizer, use the following command:

python train.py -s <path to COLMAP or NeRF Synthetic dataset> -m <output path> --output_size <desired model size>

The --output_size argument specifies the target model size (MB). (Default: 50). Generally we set it to be ~50% of SOTA model size for compactness in the paper.

The command-line arguments for train.py are largely consistent with those of 2D Gaussian Splatting. 2DGS.

We provide a demo script for the DTU dataset. To run it, use:

python auto_run_DTU.py

Please modify the dataset path before running the script. The DTU dataset we use can be downloaded from this Google Drive Link, which is shared by the 2DGS repository.

We provide the CUDA code based on GSSurfels. And the code based on 2DGS with more results ans scripts is coming soon.