TLDR: SYNBUILD-3D

This repository contains code and instructions for working with the SYNBUILD-3D dataset, a large-scale dataset of more than 6.2 million semantically enriched 3D building models at Level of Detail 4, including corresponding floor plan images and LiDAR-like roof point clouds.

• Paper: https://arxiv.org/abs/2508.21169
• Dataset: https://purl.stanford.edu/kz908vb7844

💻 Generation Pipeline

Code to replicate and extend the dataset will be added soon. Paper under review.

📦 Dataset

Download the dataset from Stanford Research Data (SDR).

It contains:

• prod_run_X_2000.tar.gz

  • The full dataset, split across 8 .tar files

• sample_100.zip

  • A random sample of 100 3D buildings with their floor plans and segmentation masks — useful for quick visualization and inspection.

After downloading, place the files as follows:

Synbuild-3D-NeurIPS-Release/
├── data/
│   └── prod_run_X_2000/  # Extracted from prod_run_X_2000.tar.gz
│   └── sample_100/       # Extracted from sample_100.zip

🧪 Visualization Setup

Set up your environment with:

conda create -n synbuild-3d python=3.11
conda activate synbuild-3d
python -m pip install -r requirements.txt

➤ Visualize the dataset:

cd notebooks
jupyter notebook

Then open and run:

• dataset_visualization.ipynb — for 3D visualizations

Citation

If you use this dataset, please cite our work as:

@dataset{synbuild3d, 
    title = {SYNBUILD-3D: A large, multi-modal, and semantically rich synthetic dataset of 3D building models at Level of Detail 4}, 
    author = {Mayer, Kevin and Vesel, Alex and Zhao, Xinyi and Fischer, Martin}, 
    year = {2025}, 
    publisher = {Stanford Research Data}, 
    doi = {https://doi.org/10.25740/kz908vb7844}, 
    url = {https://purl.stanford.edu/kz908vb7844} 
}