Jingwei Xu^1* · Chenyu Wang^2* · Zibo Zhao¹ · Wen Liu³ · Yi Ma⁴ · Shenghua Gao^4†


ShanghaiTech University | Transcengram | DeepSeek AI | University of Hong Kong
(* denotes equal contribution, † denotes the corresponding author)

• [2025.05] 🔥🔥🔥 Release the Evaluation Code on the proposed new metrics
• [2025.05] 🔥🔥🔥 Release the Omni-CAD dataset
• [2024.11] 📢📢📢 Release arXiv paper and init project

• Release the Inference Code
• Release the Training Code
• Release the Evaluation Code on the proposed new metrics
• Release the Omni-CAD dataset
• Release Project Page & Technical Report

Please view our project page for more information.

Our preprocessing code is built upon DeepCAD.

git submodule update --init --recursive
conda create -n DeepCAD python=3.8
cd ./3rd_party/DeepCAD
pip install -r requirements.txt
conda install -c conda-forge pythonocc-core=7.8.1

(model json description and the text caption)

mkdir data
cd ./data
wget https://huggingface.co/datasets/jingwei-xu-00/Omni-CAD/resolve/main/Omni-CAD.zip
unzip Omni-CAD.zip
cd ..

The directory structure should look like this:

└── data
    └── Omni-CAD
        └── json
            ├── 0000
            ├── 0001
            └── ...
        └── txt
            ├── 0000.json
            ├── 0001.json
            └── ...

cd ./3rd_party/DeepCAD/utils 
python3 ./export2step.py --src ../../../data/Omni-CAD/json --form json --num -1 [--output_form ply]
cd ../../..

After the .step file export is complete, run:

cd ./data/Omni-CAD
mv json_step step
# [mv json_ply ply]
cd ../..

The directory should look like:

└── data
    └── Omni-CAD
        └── json
            ├── 0000
            ├── 0001
            └── ...
        └── step
            ├── 0000
            ├── 0001
            └── ...
        └── txt
            ├── 0000.json
            ├── 0001.json
            └── ...

cd ./3rd_party/DeepCAD/utils
python3 export2pc.py --data_root ../../../data/Omni-CAD --normalize --output_format npy 
cd ../../..

The directory structure should look like this:

└── data
    └── Omni-CAD
        ├── json
        ├── step / (ply)
        ├── txt
        └── pcd

There are multiple options for the image rendering with the models.

1. Blender
2. Mitsuba3
3. Open3D
4. PythonOCC
5. ...

Here we provide a script to utilize the PythonOCC to render the .step file.

cd ./scripts
python3 occ_render.py --src ../data/Omni-CAD/step --num -1
cd ..

You may use .ply file format to render the image with other renderers for simplicity.

cd ./data/Omni-CAD
mv step_img img
cd ../..

Finally, the directory structure should look like this:

└── data
    └── Omni-CAD
        ├── json
        ├── step / (ply)
        ├── txt
        ├── pcd
        └── img

We state our plan at #1.

In this paper, we proposed new metrics Segment Error (SegE), Dangling Edge Length (DangEL), Self Intersection Ratio (SIR), and Flux Enclosure Error (FluxEE).

Example: blue lines denote dangling edges bounded by only one face. Our metric DangEL counts the sum of their lengths.

If you want to evaluate them in your own project, please visit CAD-MLLM-metrics for the code and guidance.

We acknowledge the support of the following projects to our data preprocessing: DeepCAD.

@misc{xu2024CADMLLM,
      title={CAD-MLLM: Unifying Multimodality-Conditioned CAD Generation With MLLM}, 
      author={Jingwei Xu and Chenyu Wang and Zibo Zhao and Wen Liu and Yi Ma and Shenghua Gao},
      year={2024},
      eprint={2411.04954},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}