Design2GarmentCode: Turning Design Concepts to Tangible Garments Through Program Synthesis

arXiv | Project Page

Feng Zhou, Ruiyang Liu, Chen Liu, Gaofeng He, Yong‑Lu Li, Xiaogang Jin, Huamin Wang. CVPR 2025 .

we propose a novel sewing pattern generation approach Design2GarmentCode based on Large Multimodal Models (LMMs), to generate parametric pattern-making programs from multi-modal design concepts

Installation

1. Clone the repository

git clone https://github.com/your-org/design2garmentcode.git   # ← replace with the real URL
cd design2garmentcode

2. Create the Conda environment

An environment.yml file is provided in the project root with all required Conda and PyPI dependencies (Python 3.9.19, Torch 2.4.0 + CUDA 12.1, etc.).

conda env create -f environment.yml
conda activate d2g
python -m pip install --upgrade pip          # optional: upgrade pip

3. (Optional) Enable 3‑D simulation

If you need local cloth simulation and 3‑D visualization, follow the installation instructions for GarmentCode Warp Simulator:
https://github.com/maria-korosteleva/NvidiaWarp-GarmentCode

---

4. Language‑Model API

Design2GarmentCode communicates with large multimodal models.
Follow the steps in the given order:

1. Provide API credentials
   • Environment variable (recommended) – defaults to ChatGPT‑4o

     export OPENAI_API_KEY="sk‑..."
     

   • Edit system.json (project root) – manually specify api_key, base_url, and model if you prefer a file‑based approach.
2. Download the required models:

   • First, download the base model Qwen2-VL-2B-Instruct.
     Place the entire folder at:
     lmm_utils/Qwen/Qwen2-VL-2B-Instruct/

   • Next, download the fine-tuned weights file model.pth,
     and place it in:
     lmm_utils/Qwen/qwen2vl_lora_mlp/

---

Quick GUI Demo

python gui.py         

• Input: Free‑form prompt or an image/sketch
• Output: GarmentCode JSON, preview image, and (optionally) physics simulation

---

1. Text-Guided Pattern Generation

• Go to the PARSE DESIGN tab.
• In the input box at the bottom ("Describe your design..."), type a natural language description of the garment.
  e.g., a T-shirt
• Click SEND to generate patterns based on your description.

---

2. Image-Guided Pattern Generation

• Click the upload icon inside the input box to upload a reference image or sketch.
• Once the image is uploaded, click SEND to parse the design and generate corresponding patterns.

---

3. Modify Patterns in the GUI

Once a pattern is generated, you can refine it directly inside the GUI:

1. Focus the input box at the bottom.
2. Type modify: <your-instruction>
   • e.g., modify: make sleeves shorter
3. Press Enter – the system will regenerate the pattern accordingly.

---

Batch Inference

1. Text Guided Generation

Use test_text_batch.py to process a list of text descriptions from a JSON file.

python lmm_utils/test_text_batch.py \
  --input assets/test_text/examples.json \
  --output assets/test_text_result \
  --sim 

• --input: Path to your input JSON file containing multiple garment description texts.
• --output: Directory where the output .json files will be saved.
• --sim: Enable or disable physical simulation output. Supports physical simulation (enabled by default in script).

---

2. Image Guided Generation

Use test_picture_batch.py to process all image files in a directory.

python lmm_utils/test_picture_batch.py \
  --input assets/test_img/examples \
  --output assets/test_image_result/examples \
  --sim 

• --input: Folder containing multiple image files.
• --output: Output folder where results will be saved.
• --sim: Enable or disable physical simulation output.

---

Get 3D Garment Patterns

1. Generate from a pattern.json

After generating the pattern data, you can simulate the corresponding 3D output directly from the pattern's JSON file.

python test_garment_sim.py --pattern_spec $INPUT_JSON 

2. Generate from gui

You can also run the simulation directly on the GUI to obtain 3D data.

python gui.py 

Citation

If you find this work useful, please cite:

```bibtex
@article{zhou2024design2garmentcode,
  title={Design2GarmentCode: Turning Design Concepts to Tangible Garments Through Program Synthesis},
  author={Zhou, Feng and Liu, Ruiyang and Liu, Chen and He, Gaofeng and Li, Yong-Lu and Jin, Xiaogang and Wang, Huamin},
  journal={Computer Vision and Pattern Recognition (CVPR)},
  year={2025}
}