[NeurIPS 2025] GeoCAD: Local Geometry-Controllable CAD Generation with Large Language Models
Zhanwei Zhang, Kaiyuan Liu, Junjie Liu, Wenxiao Wang, Binbin Lin, Liang Xie, Chen Shen, Deng Cai.

Local geometry-controllable computer-aided design (CAD) generation aims to modify local parts of CAD models automatically, enhancing design efficiency. It also ensures that the shapes of newly generated local parts follow user-specific geometric instructions (e.g., an isosceles right triangle or a rectangle with one corner cut off).

Installation

• Create a conda environment and install all the dependencies

conda env create -f environments.yaml

• After installation, activate the environment with

conda activate <env>

Data preparation

• Download the processed data by SkexGen: Google Drive link.

gdown --id 1so_CCGLIhqGEDQxMoiR--A4CQk4MjuOp

• Convert the dicts to the sequences. Note train.pkl, val.pkl and test.pkl should be converted separately.
• circle_type: [ udlr, ldru, diam, or ].

python3 utils/convert.py --in_path <in_path_name> --out_path <out_path_name> --circle_type <circle_type_name>

For example:

python3 utils/convert.py --in_path ./cad_data/train_deduplicate_s.pkl --out_path ./cad_data/processed_data/train_ds.pkl --circle_type ldru

Please see the loop captions in the ./cad_data/processed_data directory.

Training

Since the pre-training phase is optional and straightforward to implement, we only release the training code in this work.

• Run training with multiple GPUs. Change num_processes in ds_config.yaml to specify how many GPUs will be used.

CUDA_VISIBLE_DEVICES=<gpu_ids> accelerate launch --config_file ds_config.yaml finetune.py --run-name <run_name> --data-path <data_path> --eval-freq 200000 --save-freq 50000 --model-name <model_name>

• Run training with single GPU.

CUDA_VISIBLE_DEVICES=<gpu_id> python3 finetune.py --run-name <run_name> --data-path <data_path> --eval-freq 100000 --save-freq 25000 --model_name <model_name>

Inference

Notably, as the original test set contains substantial noise, it is essential to filter out invalid CAD sequences from the test set before inference.

CUDA_VISIBLE_DEVICES=<gpu_id> python3 sample.py --model_path <model_checkpoint_path> --num_samples <num_samples> --model_name <model_name> --mask_type loop

The output should be a jsonl file, where each line is a string representing a CAD design.

Visualization

• Step 1: parse the generated string to CAD obj. The in_path should be set the same as the out_path in the inference.

python3 utils/parser.py --in_path <in_path> --out_path <out_path>

• Step 2: convert generated CAD obj to stl format. Use timeout command to prevent occ hanging. The data_folder should be set the same as the out_path in step 1.

timeout 180 python3 utils/visual_obj.py --data_folder <data_folder>

• Step 3: render and visualize to images. The input_dir should be set the same as the data_folder in step 2. Note that this step only succeeds on Windows now.

python3 utils/cad_img.py --input_dir <input_dir> --output_dir <output_dir>

Citation

If you find our work useful in your research, please cite our paper:

@article{zhang2025geocad,
  title={GeoCAD: Local Geometry-Controllable CAD Generation with Large Language Models},
  author={Zhang, Zhanwei and Liu, Kaiyuan and Liu, Junjie and Wang, Wenxiao and Lin, Binbin and Xie, Liang and Shen, Chen and Cai, Deng},
  journal={arXiv preprint arXiv:2506.10337},
  year={2025}
}

Acknowledgement

Our code is partially based on FlexCAD, Skexgen and Crystal-text-llm. We appreciate all the contributors for their awesome work.