README.md

---
license: cc-by-nc-4.0
tags:
  - diffusion-model
  - 3d-generation
  - dental
  - medical-imaging
  - signed-distance-function
  - pytorch
library_name: pytorch
---

# ToothCraft
---
license: cc-by-nc-4.0
tags:
  - diffusion-model
  - 3d-generation
  - dental
  - medical-imaging
  - signed-distance-function
  - pytorch
library_name: pytorch
---

# ToothCraft

**ToothCraft** is a diffusion model for **patient-specific dental crown completion** — it reconstructs a missing or damaged tooth crown as a 3D shape, conditioned on the surrounding (local) anatomical context of the dental arch. It is the official pretrained-weights release accompanying the paper *"From Synthetic Data to Real Restorations: Diffusion Model for Patient-specific Dental Crown Completion"* (VISAPP 2026).

- 📄 Paper: [arXiv:2603.26588](https://arxiv.org/abs/2603.26588)
- 💻 Code: [github.com/ikarus1211/VISAPP_ToothCraft](https://github.com/ikarus1211/VISAPP_ToothCraft)
- ✍️ Authors: Dávid Pukanec, Tibor Kubík, Michal Španěl

## Model description

ToothCraft learns to complete incomplete tooth geometry by running a diffusion process over **Signed Distance Functions (SDFs)**, conditioned on a local context cutout that includes the neighboring crowns in the arch. The goal is to generate a crown that fits both the local morphology of the patient's dentition and the occlusion with the opposing (antagonist) teeth, reducing the risk of occlusal interference in real restorations.

Because paired real damaged/complete crown data is scarce, the model was trained primarily on **synthetically damaged** teeth: an augmentation pipeline carves incomplete geometry out of complete dental arches from the public **3DS** and **ODD** datasets, which the model then learns to complete. The paper also reports that the resulting model transfers directly to real-world restoration cases.

### Checkpoints in this repository

| File | Variant | Trained on | Notes |
|---|---|---|---|
| `ToothCraft_normal_control_branch.pth` | Normal | ODD | Control-branch weights |
| `ToothCraft_normal_diff_branch.pth` | Normal | ODD | Diffusion-branch weights |
| `ToothCraft_classfree_control_branch.pth` | Classifier-free | ODD | Control-branch weights |
| `ToothCraft_classfree_diff_branch.pth` | Classifier-free | ODD | Diffusion-branch weights |
| `ToothCraft_antag_control_branch.pth` | Antagonist-aware | ODD | Control-branch weights |
| `ToothCraft_antag_diff_branch.pth` | Antagonist-aware | ODD | Diffusion-branch weights |
| `ToothCraft_gingiva-mod_control_branch.pth` | Gingiva-aware ⭐ | ODD | Control-branch weights |
| `ToothCraft_gingiva-mod_diff_branch.pth` | Gingiva-aware ⭐ | ODD | Diffusion-branch weights |

> **Note:** each of the variants ships a *control branch* and a *diffusion branch* checkpoint, which are loaded together at inference time.


> ⭐ **Recommended for clinical cases:** the gingiva-aware (`gingiva-mod`) variant reports improved results and is the best-suited checkpoint for real/clinical restorations, where the tooth does not properly connect to gingiva. 
## Usage

Inference and training code, configuration files, and the data-augmentation pipeline are provided in the companion GitHub repository: [ikarus1211/VISAPP_ToothCraft](https://github.com/ikarus1211/VISAPP_ToothCraft). Download the checkpoint(s) for the variant you need from this repository and point the relevant config (e.g. `configs/test/test_normal.yaml`) at the local file paths, then run:

```bash
python test.py --config-path "configs/test" --config-name "test_normal.yaml"
```

See the GitHub README for environment setup (Python 3.10, PyTorch 2.9, Hydra-managed configs).

## Training data

- **3DS** and **ODD** public dental arch datasets, processed with the repository's `AugmentPipeline/` to synthesize incomplete crowns from complete arches for training and evaluation.

## License

This model (code is licensed separately — see the GitHub repository) is released under **[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/)**: you may share and adapt it with attribution, for **non-commercial purposes only**. Contact the authors for commercial licensing.

***In addition to the rights granted under the CC BY-NC 4.0 license, we permit commercial entities to internally test and evaluate the model
for a period of 90 days from the date the model is first obtained, provided that the model is not deployed in production and its outputs are
not provided to third parties. For any other commercial use, please get in touch with the authors.***
## Citation

```bibtex
@inproceedings{pukanec2026toothcraft,
  title     = {From Synthetic Data to Real Restorations: Diffusion Model for Patient-specific Dental Crown Completion},
  author    = {Pukanec, D\'avid and Kub\'ik, Tibor and \v{S}pan\v{e}l, Michal},
  booktitle = {Proceedings of VISAPP 2026},
  year      = {2026},
  doi       = {10.5220/0014646500004084}
}
```