Neural Mesh Refinement (NMR) utilized a learned geometric prior on fine shapes to adaptively refine coarse meshes through subdivision, demonstrating robust generalization to unseen shapes, poses, and non-isometric deformation. It can also refine coarse non-organic shapes into finer ones with appropriate geometric details, even when trained on organic shapes.

NMR does not suffer from the inherent limitations of existing methods, such as volume shrinkage and over-smoothing (Loop), amplification of tessellation artifacts (Modified Butterfly), or shape damage (Neural Subdivision). Moreover, it outperforms neural subdivision in generalization across non-isometric deformations, unseen shapes, and unseen refinement levels.

This repository is a prototype implementation of NMR using Python 3.8 with PyTorch 1.12.1.

# Clone the repo.
git clone https://github.com/zhuzhiwei99/NeuralMeshRefinement.git
cd NeuralMeshRefinement

# Make a conda environment
conda create -n nmr python==3.8
conda activate nmr

Please install Pytorch first. We have tested successfully in the following environment:

• Ubuntu 20.04 with NVIDIA GeForce RTX 3090: pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3
• Windows 10 without GPU: pytorch==1.12.1

Then, install the required packages:

# Install required Packages 
pip install -r requirments.txt

For a quick demo, please use the pre-trained model and test on new shapes. To test the pre-trained model please run

python test.py -p ckpt/thingi10k_netparams.dat -t examples/coarse/sphere.obj -ns 3

# -p: path to the pre-trained model
# -t: path to the coarse mesh
# -ns: number of subdivision levels

Then, you will get refined meshes in examples/refined/thingi10k_netparams/

You can also try other pre-trained models and coarse meshes.

• ckpt/thingi10k_netparams.dat: We trained this net using the Thingi10k dataset, which can adaptively refine the coarse mesh.
• ckpt/bunny_netparams.dat: We trained this net using the examples/original/bunny.obj, which tends to smoothly refine the coarse mesh.
• ckpt/gear_netparams.dat: We trained this net using the examples/original/gear.obj, which tends to sharply refine the coarse mesh.

• examples/coarse/sphere.obj

• examples/coarse/cube.obj

You can also try other coarse manifold meshes, otherwise using fTetWild to preprocess them.

Want to see more results? Check our project page out here!

• TOSCA dataset
• Thingi10K dataset
• Stanford Bunny

If you find our work useful for your project, please consider citing the following paper.

@article{zhu2025neural,
        title   = {Neural mesh refinement},
        author  = {ZHU, Zhiwei and GAO, Xiang and YU, Lu and LIAO, Yiyi},
        journal = {Frontiers of Information Technology \& Electronic Engineering},
        year    = {2025},
        volume  = {26},
        number  = {5},
        pages   = {695--712},
        doi     = {10.1631/FITEE.2400344}
}

Zhiwei ZHU, Xiang GAO, Lu YU, Yiyi LIAO, 2025. Neural mesh refinement. Frontiers of Information Technology & Electronic Engineering, 26(5):695-712. https://doi.org/10.1631/FITEE.2400344

This project is developed by Zhiwei Zhu, if you have any questions about this project, please feel free to contact me.

We express our genuine thanks to the amazing work: Neural Subdivision, fTetWild.