---
license: mit
language:
- en
tags:
- 3d
- 3d-reconstruction
- sketch-based-modeling
- surface-reconstruction
- pytorch
- siggraph-2026
---

# S2V-Net for NeuralSketch2Surf

This repository hosts the pretrained S2V-Net weights used by **NeuralSketch2Surf: Fast Neural Surfacing of Unoriented 3D Sketches**.

S2V-Net reconstructs closed surfaces from sparse, unoriented 3D sketch curves. The model predicts a `112^3` volumetric occupancy field from voxelized sketch strokes; the final mesh is extracted with Marching Cubes and can be refined with the smoothing tool from the project repository.

![Architecture](./Architecture.png)
![Results](./UpdatedResults.png)
![Cross-dataset results](./UpdatedDataset.png)

## Model

- **Input:** voxelized 3D sketch strokes on a `112^3` grid.
- **Output:** binary surface occupancy probabilities.
- **Backbone:** SwinUNETR-style 3D transformer for global shape inference.
- **Refinement:** lightweight 3D residual module for local boundary correction.
- **Use case:** interactive sketch-based surface reconstruction from raw 3D strokes.

## Notes

- The model is trained for closed-surface reconstruction.
- Very thin structures may be limited by the `112^3` voxel resolution.

## Citation

If you use these weights, please cite:

```bibtex
@article{ye2026neuralsketch2surf,
  title   = {NeuralSketch2Surf: Fast Neural Surfacing of Unoriented 3D Sketches},
  author  = {Ye, Hongsheng and Sureshkumar, Anandhu and Wang, Zhonghan and Cani, Marie-Paule and Hahmann, Stefanie and Bonneau, Georges-Pierre and Parakkat, Amal Dev},
  journal = {ACM Transactions on Graphics (Proceedings of SIGGRAPH)},
  year    = {2026}
}
```