Sagar Panwar^*, Nissim Maruani^*, Céline Loscos, Mathieu Desbrun,Pierre Alliez

SuperSDF is a learning-based method for signed distance field super-resolution that reconstructs high-fidelity meshes from coarse inputs, without mesh supervision or auxiliary surface representations. Using a sparse voxel network near the surface, our approach learns how to directly refine the input SDF, outperforming prior methods in quality, efficiency, and scalability. More details in our SIGGRAPH 2026 paper.

1. Start by cloning the repository and fVDB submodule:

git clone --recursive https://github.com/Sagar160/SSU.git

2. Create the ssu conda environment (tested with CUDA 12.1):

conda env create -f dev_env.yml
conda activate ssu

3. Our code requires building fVDB, which can take a while (please refer to the original README for more details). Run:

cd openvdb/fvdb
export MAX_JOBS=$(free -g | awk "/^Mem:/{jobs=int($4/2.5); if(jobs<1) jobs=1; print jobs}")
pip install .
cd ../..

For a hands-on experience, please refer to the demo notebook located at demo/demo.ipynb. You can load the trained model directly to test results and run inference on your own data.

Please consider this config file: config_test.yaml

conda activate ssu
cd ssu/run
python main.py --config config_test.yaml

Now you can play with the parameters and run different experiements.

if you want to change model, can be done in main.py file.

mkdir data 
cd data
pip install gdown
gdown 1YyYOgn8uxGH6Nz_gGk8OR7IuLKUF89Ze

untar

pip install py7zr
py7zr x groundtruth.7z 

Remove unecessary file

rm groundtruth.7z 

Our benchmarking methodology is directly inspired by PoNQ. We employ a nearly identical evaluation framework to ensure consistency and comparability in our results.

cd ssu/benchmarking
python get_prediction.py --config config_eval.yaml

• logging: Enable or disable logging to Weights & Biases (WandB).

• dataset_grids: Specific grid data to be loaded.
• mask_threshold: The threshold value used for masking.
• sdf_scaling_value: The value applied for SDF scaling.
• unique_random_direction: Boolean; determines if a random direction is assigned to each voxel.

• use_pre_train_model: Toggle to use a pretrained model.
• pre_train_model_name: The name or path of the pretrained model to load.

• only_eval: Set to true if you only want to run evaluation (e.g., if training completed but evaluation failed).
• run_eval: Determines whether to run evaluation at the end of a session.
• normalize: The specific normalization method used for evaluation.

We experimented with a U-Net-inspired architecture for hierarchical feature extraction and multi-scale reconstruction.

We would like to express our gratitude to the following project teams and organizations for their invaluable support and contributions to this work:

• TITANE Project-Team: For their expertise and guidance in geometric modeling and 3D vision.
• Inria Centre at Université Côte d'Azur: For providing the research environment and resources necessary for this project.