[SIGGRAPH 2025] Drag-Your-Gaussian

Official implementation of the paper:
“Drag Your Gaussian: Effective Drag-Based Editing with Score Distillation for 3D Gaussian Splatting.”

😊 TL;DR

DYG allows intuitive and flexible 3D scene editing by enabling users to drag 3D Gaussians while preserving fidelity and structure.

🎥 Introduction Video

https://github.com/user-attachments/assets/1e484ff9-f44c-4995-a99d-453cf0f11f95

Visit our Project Page for more examples and visualizations.

🔧 Installation

Clone the repository:

git clone https://github.com/Quyans/Drag-Your-Gaussian.git
cd Drag-Your-Gaussian
git submodule update --init --recursive 

Create a new conda environment:

conda env create --file environment.yaml
conda activate DYG

📚 Data Preparation

Follow 3DGS for reconstruction.
We recommend setting the spherical harmonic degree to 0.

Alternatively, you can use our prepared example data.
Example structure (e.g., face scene):

└── data
    └── face
        ├── export_1
        │   ├── drag_points.json
        │   └── gaussian_mask.pt
        ├── image
        ├── sparse
        └── point_cloud.ply

🔄 Diffusion Prior

We use LightningDrag as the diffusion prior. Follow LightningDrag Installation Guide to download required models.
Organize them as follows:

└── checkpoints
    ├── dreamshaper-8-inpainting
    ├── lcm-lora-sdv1-5/
    │   └── pytorch_lora_weights.safetensors
    ├── sd-vae-ft-ema/
    │   ├── config.json
    │   ├── diffusion_pytorch_model.bin
    │   └── diffusion_pytorch_model.safetensors
    ├── IP-Adapter/models/
    │   ├── image_encoder
    │   └── ip-adapter_sd15.bin
    └── lightning-drag-sd15/
        ├── appearance_encoder/
        │   ├── config.json
        │   └── diffusion_pytorch_model.safetensors
        ├── point_embedding/
        │   └── point_embedding.pt
        └── lightning-drag-sd15-attn.bin

🚋 Training

🖥️ WebUI

Launch the WebUI:

python webui.py --colmap_dir <path_to_colmap> --gs_source <path_to_pointcloud.ply> --output_dir <save_path>

Example:

python webui.py --colmap_dir ./data/face/ --gs_source ./data/face/point_cloud.ply --output_dir result

You can train directly in the WebUI. Alternatively, after selecting drag points and masks, export the files and run:

python drag_3d.py --config configs/main.yaml                   --colmap_dir ./data/face/                   --gs_source ./data/face/point_cloud.ply                   --point_dir ./data/face/export_1/drag_points.json                   --mask_dir ./data/face/export_1/gaussian_mask.pt                   --output_dir result

📖 Citation

If you find our work useful, please cite:

@article{qu2025drag,
  title={Drag Your Gaussian: Effective Drag-Based Editing with Score Distillation for 3D Gaussian Splatting},
  author={Qu, Yansong and Chen, Dian and Li, Xinyang and Li, Xiaofan and Zhang, Shengchuan and Cao, Liujuan and Ji, Rongrong},
  journal={arXiv preprint arXiv:2501.18672},
  year={2025}
}

📄 License

This project is licensed under the CC BY-NC-SA 4.0.
The code is intended for academic research purposes only.

📬 Contact

For any questions or collaborations, feel free to contact:
📧 [email protected]