Simple 3D Gaussian Splatting

A minimal 3DGS implementation built on gsplat.

Dependencies

• PyTorch
• gsplat
• torchvision
• omegaconf
• tqdm
• PyYAML

Installation

conda env create -f environment.yml
conda activate gsplat

Project Structure

3DRR_codebase/
├── config/             # Training configs (YAML)
├── core/
│   ├── data/           # Dataset loaders (Blender format)
│   ├── libs/           # Utilities (SSIM, config, etc.)
│   └── model/          # 3DGS model
├── train.py            # Training + validation + test rendering
├── eval.py             # Load checkpoint and render test set
└── outputs/            # Training outputs

Data Format

Expects datasets in Blender/NeRF-synthetic JSON format with calibrated intrinsics.

Training

python train.py -c config/cupcake.yaml

Outputs are saved to outputs/<experiment>/<timestamp>/:

• latest.pt — model checkpoint
• config.yaml — copy of training config
• examples/ — validation render grids (val_step*.jpg) and augmented training samples (train_aug.jpg)
• test/ — rendered test images

Evaluation (Render Only)

Load a checkpoint and render the test set (no metrics, no GT images needed):

python eval.py -w outputs/.../latest.pt

Config is automatically loaded from the checkpoint directory. Rendered images are saved to {checkpoint_dir}/test/.