<p align="center"> <h1 align="center">GenCAD</h1> <h4 align="center">Image-conditioned Computer-Aided Design Generation with Transformer-based Contrastive Representation and Diffusion Priors</h4> </p> <p align="center"> <a href="https://openreview.net/pdf?id=e817c1wEZ6"> <img src="https://img.shields.io/badge/Paper-TMLR%202025-4b44ce.svg"> </a> <a href="https://arxiv.org/abs/2409.16294"> <img src="https://img.shields.io/badge/arXiv-2409.16294-b31b1b.svg"> </a> <a href="https://gencad.github.io/"> <img src="https://img.shields.io/badge/Project%20Page-Link-blue"> </a> </p>

📁 Dataset

Download from here and place it in the data/ directory.

📦 Pretrained Models

Download from here and place them in data/ckpt/.

🔧 Setup Options

First download the checkpoints and the dataset and put them in their respective directories.

Option 1: Docker (Recommended)

Clone the repo:

git clone https://github.com/ferdous-alam/GenCAD
cd GenCAD

Build the Docker image:
```
docker build -t gencad:latest .
```

Run a script, for example training CSR:

docker run -it gencad:latest conda run -n gencad_env python train_gencad.py csr -name test -gpu 0

For headless visualization (inference):

First, enter the container with GPU access and mount the appropriate folders:

docker run --gpus all \
  -v $(pwd)/data/images:/app/data/images \
  -v $(pwd)/assets:/app/assets \
  -v $(pwd)/results:/app/results \
  -it gencad:latest /bin/bash

Then inside the container, run:

xvfb-run --server-args="-screen 0 2048x2048x24" python inference_gencad.py -image_path data/images -export_img

Option 2: Manual (conda + pip)

Create and activate a virtual environment with GPU support:

conda create -n gencad_env python=3.10 -y
conda activate gencad_env

Install pythonocc-core using conda:

conda install -c conda-forge pythonocc-core=7.9.0

Install the rest via pip:
```
pip install -r requirements.txt
```

Now run training or inference:

python train_gencad.py csr -name test -gpu 0

🚀 Training

CSR Model

python train_gencad.py csr -name test -gpu 0

Optional checkpoint:

python train_gencad.py csr -name test -gpu 0 -ckpt "model/ckpt/ae_ckpt_epoch1000.pth"

CCIP Model

python train_gencad.py ccip -name test -gpu 0 -cad_ckpt "model/ckpt/ae_ckpt_epoch1000.pth"

Diffusion Prior

python train_gencad.py dp -name test -gpu 0 -cad_emb 'data/embeddings/cad_embeddings.h5' -img_emb 'data/embeddings/sketch_embeddings.h5'

🧪 Inference

For headless systems (e.g. servers):

xvfb-run python inference_gencad.py

🖼 STL Visualization

Convert STL to PNG:

python stl2img.py -src path/to/stl/files -dst path/to/save/images

📊 Evaluation

Coming soon.

GenCAD

Install / Use

README