SIGGRAPH Asia 2026 Conference Proceedings
A practical application of CADKnitter: generating a complementary part that fits an existing CAD design.
1 — Create the environment (CUDA 12.1 PyTorch and the rest; use mamba if conda hangs on "Solving environment"):
conda env create -f environment.yml
conda activate cadknitter_env
2 — Install the vendored packages in third_party/. They compile against the installed PyTorch, so build isolation must be off:
pip install -e ./third_party/occwl-main
pip install --no-build-isolation -e ./third_party/pytorch3d-main
pip install --no-build-isolation -e ./third_party/chamferdist-master
If chamferdist fails to build, set CUDA_HOME to a CUDA 12.1 installation, or build it from source.
The full KnitCAD dataset (~115 GB) is hosted on Hugging Face at
Fsoft-AIC/knitcad. It is uploaded as a
single tar.gz archive split into four parts (knitcad.tar.gz.aa … knitcad.tar.gz.ad),
so the parts must be concatenated before extraction. The data split files are hosted there as well.
The repo ships with only one sample joint set in data/pkl/ for a quick smoke test.
Download the full dataset to train or evaluate.
1 — Download the archive parts (needs ~115 GB free, plus room for the extracted data):
huggingface-cli download Fsoft-AIC/knitcad --repo-type dataset \
--include "knitcad.tar.gz.*" "knitcad_data_split*.json" --local-dir ./knitcad_download
2 — Merge the parts and extract:
cat ./knitcad_download/knitcad.tar.gz.a? > ./knitcad_download/knitcad.tar.gz
tar -xzf ./knitcad_download/knitcad.tar.gz -C ./knitcad_download
3 — Replace data/pkl with the extracted folder and move the split files into data/:
rm -rf data/pkl
mv ./knitcad_download/knitcad data/pkl
mv ./knitcad_download/knitcad_data_split*.json data/
Once the parts are merged and extracted successfully, the archive parts and the merged
knitcad.tar.gz can be deleted to reclaim disk space.
1 — BrepGen checkpoints. Download the ABC pretrained weights from the original repo,
BrepGen — Pretrained Checkpoint,
and place these files in ./weights/.
2 — CADKnitter checkpoint. The fine-tuned surface-position network is on Hugging Face at
Fsoft-AIC/cadknitter:
huggingface-cli download Fsoft-AIC/cadknitter knitcad_ldm_surfpos.pt --local-dir ./weights
The resulting layout must match the paths in eval_config.yaml:
weights/
├── knitcad_ldm_surfpos.pt # CADKnitter (surface position)
├── abc_ldm_surfpos.pt # only needed to fine-tune (runs/train_ldm.sh)
├── abc_ldm_surfz.pt
├── abc_ldm_edgepos.pt
├── abc_ldm_edgez.pt
├── abc_vae_surf.pt
└── abc_vae_edge.pt
Train the latent diffusion model (change path to previously trained VAEs):
sh runs/train_ldm.sh
sh runs/sample.sh
This will load the settings in eval_config.yaml.
Create the ground-truth CAD pairs, then compute the metrics on the generated samples (results are saved to eval/eval_results/):
sh runs/create_eval_gt.sh
sh runs/eval.sh
To render turntable videos of CAD pairs, see README_vis.md.
Our codebase builds on BrepGen — many thanks to the authors for sharing their code, instructions, and pretrained weights.
If you use this repository, please cite CADKnitter and BrepGen, on which it builds:
@inproceedings{le2026cadknitter,
author = {Le, Tri and Nguyen, Khang and Huang, Baoru and Ta, Tung D and Nguyen, Anh},
title = {CADKnitter: Compositional CAD Generation from Text and Geometry Guidance},
booktitle = {SIGGRAPH Asia 2026 Conference Papers},
year = {2026},
}
@article{xu2024brepgen,
title={BrepGen: A B-rep Generative Diffusion Model with Structured Latent Geometry},
author={Xu, Xiang and Lambourne, Joseph G and Jayaraman, Pradeep Kumar and Wang, Zhengqing and Willis, Karl DD and Furukawa, Yasutaka},
journal={ACM Transactions on Graphics},
volume={43},
number={4},
pages={1--14},
year={2024},
doi={10.1145/3658129}
}