TopoLiDM: Topology-Aware LiDAR Diffusion Models for Interpretable and Realistic LiDAR Point Cloud Generation

Jiuming Liu^†1, Zheng Huang^†1, Mengmeng Liu², Tianchen Deng¹, Francesco Nex², Hao Cheng², Hesheng Wang^*1

¹Shanghai Jiao Tong University  |  ²University of Twente
^†Equal contribution  |  ^*Corresponding author

Installation

Note: Run the following commands one by one rather than as a single script, as some steps (e.g. conda activate) require shell re-sourcing.

Step 1 — Install Rust (required by some packages)

# Optional: set Tsinghua mirror for faster downloads in China
export RUSTUP_DIST_SERVER=https://mirrors.tuna.tsinghua.edu.cn/rustup
export RUSTUP_UPDATE_ROOT=https://mirrors.tuna.tsinghua.edu.cn/rustup/rustup

curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y
export PATH="$HOME/.cargo/bin:$PATH"

Step 2 — Create the conda environment

conda create -n ld python=3.10.11 -y
conda activate ld

Step 3 — Install PyTorch (CUDA 11.8)

pip install --upgrade pip
pip install setuptools==69.5.1
pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 \
    --index-url https://download.pytorch.org/whl/cu118

Step 4 — Install core dependencies

pip install \
    torchmetrics==0.5.0 pytorch-lightning==1.4.2 omegaconf==2.1.1 \
    einops==0.3.0 transformers==4.36.2 imageio==2.9.0 imageio-ffmpeg==0.4.2 \
    opencv-python kornia==0.7.0 wandb more_itertools gdown \
    numpy==1.26 easydict joblib scipy

pip install -e git+https://github.com/CompVis/taming-transformers.git@master#egg=taming-transformers
pip install -e git+https://github.com/openai/CLIP.git@main#egg=clip

Step 5 — Install TopologyLayer and Dionysus (required for TopoLiDM)

pip install -e git+https://github.com/bruel-gabrielsson/TopologyLayer.git --no-build-isolation

apt-get update && apt-get install -y libboost-all-dev
conda install -c conda-forge dionysus

After installing TopologyLayer, patch a NumPy API deprecation inside the library:

# Replace deprecated np.int with int throughout TopologyLayer source
cd path/to/TopologyLayer   # typically inside your pip src/ folder
grep -rl "np.int" . | xargs sed -i 's/np.int/int/g'

Step 6 — Install torchsparse (optional, required for evaluation)

apt-get install -y libsparsehash-dev

# Clear LD_LIBRARY_PATH to avoid library conflicts (see Troubleshooting below)
LD_LIBRARY_PATH="" pip install git+https://github.com/mit-han-lab/torchsparse.git@v1.4.0 \
    --no-build-isolation

Step 7 — Prepare dataset directory and pretrained weights

mkdir -p dataset/        # symlink or place KITTI-360 data here
mkdir -p pretrained_weights/   # place pretrained model checkpoints here

See lidm/eval/README.md for details on preparing evaluation checkpoints. If a checkpoint is a .zip file, remove the .zip suffix before use.

Pretrained Weights

We release pretrained weights on Hugging Face: hhzz0505/TopoLiDM

Model	Description	File
Topological Autoencoder	Trained on KITTI-360	autoencoder/model.ckpt
Diffusion Model	Unconditional LiDAR generation on KITTI-360	diffusion/model.ckpt

Download with Python:

from huggingface_hub import hf_hub_download

# Autoencoder
hf_hub_download(repo_id="hhzz0505/TopoLiDM", filename="autoencoder/model.ckpt", local_dir="pretrained_weights/")

# Diffusion model
hf_hub_download(repo_id="hhzz0505/TopoLiDM", filename="diffusion/model.ckpt", local_dir="pretrained_weights/")

Or via CLI:

huggingface-cli download hhzz0505/TopoLiDM --local-dir pretrained_weights/

Quick Start

Training the Topological Autoencoder

# Train on KITTI with topological regularization (multi-GPU)
python main.py -b configs/autoencoder/kitti/autoencoder_c2_p4_topo.yaml -t --gpus 0,1,2,3

# Resume from a checkpoint  (note: -n and -r are mutually exclusive)
python main.py -b configs/autoencoder/kitti/autoencoder_c2_p4_topo.yaml -t --gpus 0 \
    -r path/to/checkpoint

# Single-GPU debug mode
python main.py -b configs/autoencoder/kitti/autoencoder_c2_p4_topo.yaml -t --gpus 0 -d

Sampling and Evaluation

CUDA_VISIBLE_DEVICES=0 python scripts/sample.py \
    -d kitti -r models/lidm/kitti/uncond/model.ckpt -n 2000 --eval

Project Structure

topolidm/
├── main.py                                    # Main training entry point
├── configs/
│   └── autoencoder/kitti/
│       └── autoencoder_c2_p4_topo.yaml        # Topological autoencoder config
├── lidm/
│   ├── models/
│   │   ├── autoencoder.py                     # Standard LiDM autoencoder
│   │   └── autoencoder_topo.py                # Topological autoencoder
│   └── modules/diffusion/
│       ├── model_lidm.py                      # LiDM UNet components
│       └── model_topolidm.py                  # Graph encoder + decoder

Troubleshooting

torchsparse Installation Issues

Two errors are commonly encountered when installing torchsparse in a conda environment.

---

Error 1: undefined symbol: ucs_config_doc_nop

ImportError: /opt/hpcx/ucc/lib/libucc.so.1: undefined symbol: ucs_config_doc_nop

Cause: The system-level LD_LIBRARY_PATH includes /opt/hpcx/ucc/lib/, whose libucc.so.1 conflicts with the library expected by PyTorch. Python loads the system library instead of the conda environment's library at startup.

Fix:

# Clear LD_LIBRARY_PATH before running pip
LD_LIBRARY_PATH="" pip install git+https://github.com/mit-han-lab/torchsparse.git@v1.4.0 \
    --no-build-isolation

# Alternative: use env -i to fully isolate the subprocess environment
env -i PATH="$PATH" HOME="$HOME" pip install \
    git+https://github.com/mit-han-lab/torchsparse.git@v1.4.0 --no-build-isolation

---

Error 2: CUDA version mismatch

RuntimeError: The detected CUDA version (12.4) mismatches the version that was used
to compile PyTorch (11.8).

Cause: The nvcc visible to pip comes from a different CUDA toolkit than the one used to build the installed PyTorch. PyTorch's cpp_extension.py checks for this mismatch.

Fix: Reinstall PyTorch pinned to your system's CUDA version, then reinstall torchsparse:

pip uninstall -y torch
pip install --no-cache-dir torch==2.0.1+cu118 --index-url https://download.pytorch.org/whl/cu118

LD_LIBRARY_PATH="" pip install git+https://github.com/mit-han-lab/torchsparse.git@v1.4.0 \
    --no-build-isolation

---

Root Cause: LD_LIBRARY_PATH priority

The underlying problem is that conda's environment variables (set via conda env config vars set) only take effect when conda explicitly activates the environment in the current shell. A LD_LIBRARY_PATH set in ~/.bashrc or by a system init script takes precedence over conda's settings in any shell that inherits it without re-activating conda.

Diagnosis commands:

# Check which library paths are currently active
echo $LD_LIBRARY_PATH

# Check that PyTorch is loading from the right place
unset LD_LIBRARY_PATH && python -c "import torch; print(torch.__version__, torch.version.cuda)"

# Verify torchsparse after a clean install
unset LD_LIBRARY_PATH && python -c "import torchsparse; print(torchsparse.__version__)"

---

Common Errors Quick Reference

Error	Cause	Fix
undefined symbol: ucs_config_doc_nop	LD_LIBRARY_PATH points to wrong libucc	Run with LD_LIBRARY_PATH=""
CUDA version mismatch	nvcc version differs from PyTorch's CUDA	Reinstall matching PyTorch version
Unknown CUDA arch (8.7)	GPU arch not supported by this build	Set TORCH_CUDA_ARCH_LIST="8.0;8.6"
No module named 'pkg_resources'	New setuptools removed pkg_resources	pip install setuptools==69.5.1
libtorch_cpu.so: undefined symbol: iJIT_NotifyEvent	PyTorch version conflicts with system libs	Reinstall CUDA-matched PyTorch

Citation

If you find this work useful, please cite our paper:

@inproceedings{liu2025topolidm,
  title={TopoLiDM: Topology-Aware LiDAR Diffusion Models for Interpretable and Realistic LiDAR Point Cloud Generation},
  author={Liu, Jiuming and Huang, Zheng and Liu, Mengmeng and Deng, Tianchen and Nex, Francesco and Cheng, Hao and Wang, Hesheng},
  booktitle={2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  pages={8180--8186},
  year={2025},
  organization={IEEE}
}

Acknowledgements

• The diffusion model framework builds heavily on Latent Diffusion
• Topological regularization concepts from GLiDR
• LiDAR data processing pipeline from LiDM