Our training pipeline leverages PyTorch and optimized CUDA extensions to efficiently produce high-quality trained models in record time.
- GPU: CUDA-ready GPU with Compute Capability 7.0+
- Memory: 24 GB VRAM (for paper-quality results; we recommend NVIDIA RTX4090)
- Conda (recommended for streamlined setup)
- C++ Compiler compatible with PyTorch extensions
- CUDA SDK 11 (or compatible version)
⚠️ Important: Ensure C++ Compiler and CUDA SDK versions are compatible
Our testing environment uses the following CUDA configuration:
| Component | Version |
|---|---|
| Conda environment CUDA version | 11.6 |
Ubuntu system nvidia-smi CUDA |
12.2 |
nvcc -V compiler version |
11.8 (v11.8.89) |
Note: The Conda CUDA and system CUDA versions may differ. The compiler version (
nvcc) is what matters for PyTorch extensions compilation (diff-gaussian-rasterization_fastgs).
git clone --branch fast-pgsr https://github.com/fastgs/FastGS.git
cd FastGSWe provide a streamlined setup using Conda:
# Windows only
SET DISTUTILS_USE_SDK=1
# Create and activate environment
conda env create --file environment.yaml
conda activate fast-pgsrPlease download the Tanks & Temples Dataset from the official website.
The data folder should like this:
data
├── tnt_dataset
│ ├── tnt
│ │ ├── Ignatius
│ │ │ ├── images_raw
│ │ │ ├── Ignatius_COLMAP_SfM.log
│ │ │ ├── Ignatius_trans.txt
│ │ │ ├── Ignatius.json
│ │ │ ├── Ignatius_mapping_reference.txt
│ │ │ └── Ignatius.ply
│ │ └── ...Then run the scripts to preprocess Tanks and Temples dataset:
# Install COLMAP
Refer to https://colmap.github.io/install.html
# Tanks and Temples dataset
python scripts/preprocess/convert_tnt.py --tnt_path your_tnt_path# Tanks and Temples dataset
python scripts/run_tnt.py📋 Advanced: Command Line Arguments for train.py
Threshold of the loss map; a lower value generally results in more Gaussians being retained.
Absolute gradient (same as Abs-GS) threshold for split.
Gradient(same as vanilla 3DGS) threshold for clone.
Learning rate for high-order SH coefficients (features_rest).
Learning rate for low-order SH coefficients (features_dc).
Percentage of scene extent (0--1) a point must exceed to be forcibly densified.
Multiplier for the compact box to control the tile number of each splat
Path to the source directory containing a COLMAP or Synthetic NeRF data set.
Path where the trained model should be stored (output/<random> by default).
Alternative subdirectory for COLMAP images (images by default).
Add this flag to use a MipNeRF360-style training/test split for evaluation.
Specifies resolution of the loaded images before training. If provided 1, 2, 4 or 8, uses original, 1/2, 1/4 or 1/8 resolution, respectively. For all other values, rescales the width to the given number while maintaining image aspect. If not set and input image width exceeds 1.6K pixels, inputs are automatically rescaled to this target.
Specifies where to put the source image data, cuda by default, recommended to use cpu if training on large/high-resolution dataset, will reduce VRAM consumption, but slightly slow down training. Thanks to HrsPythonix.
Add this flag to use white background instead of black (default), e.g., for evaluation of NeRF Synthetic dataset.
Order of spherical harmonics to be used (no larger than 3). 3 by default.
Flag to make pipeline compute forward and backward of SHs with PyTorch instead of ours.
Flag to make pipeline compute forward and backward of the 3D covariance with PyTorch instead of ours.
Enables debug mode if you experience erros. If the rasterizer fails, a dump file is created that you may forward to us in an issue so we can take a look.
Debugging is slow. You may specify an iteration (starting from 0) after which the above debugging becomes active.
Number of total iterations to train for, 30_000 by default.
IP to start GUI server on, 127.0.0.1 by default.
Port to use for GUI server, 6009 by default.
Space-separated iterations at which the training script computes L1 and PSNR over test set, 7000 30000 by default.
Space-separated iterations at which the training script saves the Gaussian model, 7000 30000 <iterations> by default.
Space-separated iterations at which to store a checkpoint for continuing later, saved in the model directory.
Path to a saved checkpoint to continue training from.
Flag to omit any text written to standard out pipe.
Spherical harmonics features learning rate, 0.0025 by default.
Opacity learning rate, 0.05 by default.
Scaling learning rate, 0.005 by default.
Rotation learning rate, 0.001 by default.
Number of steps (from 0) where position learning rate goes from initial to final. 30_000 by default.
Initial 3D position learning rate, 0.00016 by default.
Final 3D position learning rate, 0.0000016 by default.
Position learning rate multiplier (cf. Plenoxels), 0.01 by default.
Iteration where densification starts, 500 by default.
Iteration where densification stops, 15_000 by default.
Limit that decides if points should be densified based on 2D position gradient, 0.0002 by default.
How frequently to densify, 100 (every 100 iterations) by default.
How frequently to reset opacity, 3_000 by default.
Influence of SSIM on total loss from 0 to 1, 0.2 by default.
Percentage of scene extent (0--1) a point must exceed to be forcibly densified, 0.01 by default.
This project is built upon PGSR, Taming-3DGS, Speedy-Splat, and Abs-GS. We extend our gratitude to all the authors for their outstanding contributions and excellent repositories!
License: Please adhere to the licenses of Deformable-3D-Gaussians, 4DGaussians, Taming-3DGS, Speedy-Splat, and Abs-GS.
Special thanks to the authors of DashGaussian for their generous support!
If you find this repo useful, please cite:
@article{ren2025fastgs,
title={FastGS: Training 3D Gaussian Splatting in 100 Seconds},
author={Ren, Shiwei and Wen, Tianci and Fang, Yongchun and Lu, Biao},
journal={arXiv preprint arXiv:2511.04283},
year={2025}
}
⭐ If FastGS helps your research, please consider starring this repository!
FastGS: Training 3D Gaussian Splatting in 100 Seconds