DCTrellis2 — dental fine-tuned TRELLIS.2 denoisers
Two fine-tuned denoiser checkpoints that adapt TRELLIS.2 to single-photo reconstruction of dental arch casts. They are the weights used by JeremyHo1123/DCTrellis2 — a pipeline that turns one photo of a dental model into a 3D mesh and then into a uniform-thickness clear-aligner shell ready for printing.
[!WARNING] Research use only. Not a medical device. These weights have no FDA / EU MDR / TFDA clearance and have not undergone clinical validation. They must not be used to plan, manufacture, or deliver orthodontic treatment for any patient.
Files
| File | Size | Replaces in TRELLIS.2 |
|---|---|---|
dental_ss_flow_64.pt | 4.9 GB | Stage 1 — sparse-structure flow denoiser @ 64³ |
dental_shape_flow_1024.pt | 4.9 GB | Stage 2 — shape SLat flow denoiser (loaded into the shape_slat_flow_model_512 slot, voxel resolution 1024) |
Each file is a plain PyTorch state_dict (≈1.29 B parameters) holding the EMA weights at
step 10 000. Everything else in the pipeline — the VAEs, the O-Voxel decoder, the texturing
head — is stock TRELLIS.2 and is pulled from microsoft/TRELLIS.2-4B at runtime.
Usage
git clone --recursive https://github.com/JeremyHo1123/DCTrellis2.git
cd DCTrellis2
# follow docs/INSTALLATION.md to build the environment first
hf auth login
hf download Jeremy19991123/DCTrellis2-dental --local-dir weights
mkdir -p results/dental_ss_flow_64/ckpts results/dental_shape_flow_1024/ckpts
cp weights/dental_ss_flow_64.pt results/dental_ss_flow_64/ckpts/denoiser_ema0.9999_step0010000.pt
cp weights/dental_shape_flow_1024.pt results/dental_shape_flow_1024/ckpts/denoiser_ema0.9999_step0010000.pt
python app_2.py # must be launched from the repo root
app_2.py resolves these as relative paths from the repo root; change the SS_CKPT /
SHAPE_CKPT constants near the bottom of the file if you want them elsewhere.
Training
| Base checkpoints | stock TRELLIS.2 Stage-1 and Stage-2 denoisers (MIT) |
| Data | ~1 049 single-arch dental STL casts from Teeth3DS+, rendered under domain randomisation |
| Optimiser | AdamW, lr = 1e-5, weight_decay = 0.01, betas = (0.9, 0.95) |
| Steps | 10 000, batch size 1 per GPU |
| EMA | 0.9999 (the released checkpoints are the EMA weights) |
| Gradient clipping | AdaptiveGradClipper, max_norm = 1.0, 95th percentile |
| Precision | bf16 |
| Hardware | 2 × NVIDIA RTX A6000 48 GB, CUDA 12.4 |
The exact configs live in configs/dental_finetune/ in the GitHub repo, along with the
commands needed to reproduce both stages.
Hardware for inference
Tested on an RTX A6000 (48 GB). Reconstruction takes 1–2 minutes per photo; the aligner generation stage is CPU-bound and takes several minutes. Roughly 50 GB of disk is needed in total (18 GB of stock TRELLIS.2 HF cache + 9.8 GB for these weights + the conda env).
Licensing
These weights are licensed CC BY-NC-ND 4.0, inherited from the Teeth3DS+ training data:
- BY — attribution required (see Citation).
- NC — non-commercial use only; no commercial product, service, or clinical offering.
- ND — do not distribute modified or further fine-tuned versions of these weights.
The DCTrellis2 source code is separately MIT-licensed. Full details in NOTICE.md.
Teeth3DS+ is not redistributed here in any form. To reproduce the fine-tune you must obtain it yourself from OSF and accept its licence directly.
Citation
[!IMPORTANT] Citation is a condition of use.
The paper describing this work is under preparation and the citation will be posted here as soon as it is available. Until then, please open an issue on the GitHub repo or start a discussion here, and the current reference will be sent to you.
You must also cite the Teeth3DS+ dataset and the TRELLIS.2 backbone:
@article{ben2022teeth3ds,
title = {Teeth3DS: a benchmark for teeth segmentation and labeling from intra-oral 3D scans},
author = {Ben-Hamadou, Achraf and Smaoui, Oussama and Chaabouni-Chouayakh, Houda and
Rekik, Ahmed and Pujades, Sergi and Boyer, Edmond and Strippoli, Julien and
Thollot, Aur{\'e}lien and Setbon, Hugo and Trosset, Cyril and others},
journal = {arXiv preprint arXiv:2210.06094},
year = {2022}
}
@article{ben20233dteethseg,
title = {3DTeethSeg'22: 3D Teeth Scan Segmentation and Labeling Challenge},
author = {Ben-Hamadou, Achraf and Smaoui, Oussama and Rekik, Ahmed and Pujades, Sergi and
Boyer, Edmond and others},
journal = {arXiv preprint arXiv:2305.18277},
year = {2023}
}
Disclaimer
Research code and research weights. Provided "as is", without warranty of any kind. The author accepts no liability arising from their use.