🧊 DCx

Dual Contouring over Expanded Cubes (DCx) for Zero-Level Set Extraction from Neural Unsigned Distance Functions
(ACM SIGGRAPH 2026)

Extraction of complex non-manifold structures using DCx from Unsigned Distance Functions.

This repository contains the official implementation of the SIGGRAPH 2026 paper: "Dual Contouring over Expanded Cubes (DCx) for Zero-Level Set Extraction from Neural Unsigned Distance Functions".

DCx provides a robust framework to extract high-quality, non-manifold zero-level sets from both Ground Truth and Neural Unsigned Distance Functions (UDFs).

Overview of the DCx pipeline: from UDF sampling to the final mesh extraction via Expanded Cubes.

Visual comparison of extracted meshes using our proposed DCx algorithm.

We recommend using Conda to manage your environment. Follow the steps below to set up the dependencies:

# 1. Create and activate a new conda environment
conda create -n dcx python=3.9 -y
conda activate dcx

# 2. Install PyTorch with CUDA 11.8
conda install pytorch==2.2.1 torchvision==0.17.1 torchaudio==2.2.1 pytorch-cuda=11.8 -c pytorch -c nvidia

# 3. Install CUDA toolkit and related packages for CAPUDF
conda install -c nvidia cuda-toolkit=11.8 cuda-nvcc=11.8 cuda-cccl=11.8 -y

# 4. Install other Python dependencies
pip install open3d scikit-image tqdm pyhocon==0.3.57 trimesh PyMCubes scipy point_cloud_utils==0.29.7 "numpy<2"

# 5. Install DCx
cd DCX
pip install .

CAPUDF: We use CAPUDF to compute Neural Unsigned Distance Functions (NUDF). You need to compile the Chamfer Distance extension first:

cd CAPUDF/extensions/chamfer_dist
python setup.py install

cubvh: We use cubvh to compute Ground Truth Unsigned Distance Functions (GTUDF). Install it directly via git:

pip install git+https://github.com/ashawkey/cubvh --no-build-isolation

DCx/
├── evaluate_finetune.py     # Main entry point for mesh extraction
├── ckpt/                    # Directory for CAPUDF trained checkpoints
├── confs/                   # Configuration files for experiments 
│   └── ...                  
├── DCX/                     # Core DCx algorithm implementation
│   └── ...                  
├── CAPUDF/                  # NUDF computation module based on CAP-UDF 
│   ├── extensions/          # Custom C++/CUDA extensions 
│   │   └── chamfer_dist/    # Chamfer distance extension 
│   └── ...                  
├── data/                    # Put your datasets here (update dataset_dir in conf)
│   └── dataset/             
├── assets/                  # Images for README (Pipeline, results, etc.)
├── LICENSE                  # MIT License
└── README.md                # Project documentation and setup instructions

The main entry point for mesh extraction and evaluation is evaluate_finetune.py. You can extract zero-level sets from either Ground Truth UDF (GTUDF) or Neural UDF (NUDF).

By default, the script extracts the mesh from a GTUDF (--udf_type 1). This method reads the raw .ply mesh and computes the UDF using cubvh.If we have enough sampling points, there's no need for supplementary sampling.

python evaluate_finetune.py \
    --conf ./confs/test.conf \
    --datadir <your_data_dir> \
    --dataname <your_shape_name> \
    --udf_type 1 \
    --num_points 200000000 \
    --gpu 0 \
    --thinning \
    --postprocessing \
    --finetune

To extract from a learned Neural UDF using CAP-UDF (--udf_type 0), ensure you have the pre-trained checkpoints placed in the ./ckpt/ directory as specified in your configuration file.

python evaluate_finetune.py \
    --conf ./confs/test.conf \
    --datadir <your_data_dir> \
    --dataname <your_shape_name> \
    --udf_type 0 \
    --gpu 0 \
    --thinning \
    --supplementary_sampling \
    --postprocessing \
    --finetune

Here is the full list of arguments you can pass to the script:

💡 Note on outputs: The extracted .ply meshes will be saved into the results directory specified in your .conf file (e.g., dir_path.result_dir).

If you find our work useful in your research, please consider citing:

@article{10.1145/3811388,
author = {Bao, Qingchao and Chen, Xuhui and Yin, Jingpeng and Hou, Fei and Wang, Wencheng and Qin, Hong and He, Ying},
title = {Dual Contouring over Expanded Cubes (DCx) for Zero-Level Set Extraction from Neural Unsigned Distance Functions},
year = {2026},
issue_date = {July 2026},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
volume = {45},
number = {4},
issn = {0730-0301},
url = {https://doi.org/10.1145/3811388},
doi = {10.1145/3811388},
journal = {ACM Trans. Graph.},
month = jul,
articleno = {56},
numpages = {20},
keywords = {dual contouring, zero-level set extraction, unsigned distance functions, neural representations, non-manifold structures}
}