Drag3D: DragGAN meets GET3D

This project extends the idea of DragGAN into the GET3D to enable interactive generation and drag editing of textured meshes.

A GUI is implemented for demonstration:

drag3d_car.mp4

drag3d_chair.mp4

drag3d_animals.mp4

Thanks @SteveJunGao for helping to create the demo video for animals!

Install

# download
git clone https://github.com/ashawkey/Drag3D.git
cd Drag3D

# dependency
pip install -r requirements.txt

# (optional) get a better font to display
wget https://github.com/lxgw/LxgwWenKai/releases/download/v1.300/LXGWWenKai-Regular.ttf

Download pretrained GET3D checkpoints from here and put them under ./pretrained_model.

Tested Environment

• Ubuntu 20 + V100 + CUDA 11.6 + torch 1.12.0
• Windows 10 + 3070 + CUDA 12.1 + torch 2.1.0

You need to have an OpenGL direct rendering enabled display to use the GUI.

The required GPU memory is about 4 GB.

NOTE: For Unix-based OS, we could only use the CUDA context of nvdiffrast in GUI, which seems to show some rendering artifacts when the mesh is too far from camera (triangles too small). Windows is recommended.

Usage

basics

# run gui
python gui.py --outdir trial_car --resume_pretrain pretrained_model/shapenet_car.pt

You need to click get to generate a 3D model first, and use geo or tex to resample geometry or texture.

Then, operate the GUI by:

• Left drag: rotate camera.
• Middle drag: pan camera.
• Scroll: scale camera.
• Right click: add / select source point.
• Right drag: drag target point.

After adding at least one point pair, click train to start optimization.

You can repeat these steps until getting satisfying shapes.

Finally, click save to export the current textured mesh.

bounding box mask loss (experimental)

In GUI, you could check bbox loss and adjust the bounding box to constrain the optimization, like the 2D mask loss.

The part outside the bounding box will be encouraged to remain unchanged.

The learning rate and loss weight can be adjusted in GUI to achieve a balance. A large lr (0.01) or smaller loss weight (1.0) maybe helpful if the point won't move after applying bbox loss.

Acknowledgement

• DragGAN:

  @inproceedings{pan2023draggan,
    title={Drag Your GAN: Interactive Point-based Manipulation on the Generative Image Manifold}, 
    author={Pan, Xingang and Tewari, Ayush, and Leimk{\"u}hler, Thomas and Liu, Lingjie and Meka, Abhimitra and Theobalt, Christian},
    booktitle = {ACM SIGGRAPH 2023 Conference Proceedings},
    year={2023}
  }

• GET3D:

  @inproceedings{gao2022get3d,
    title={GET3D: A Generative Model of High Quality 3D Textured Shapes Learned from Images},
    author={Jun Gao and Tianchang Shen and Zian Wang and Wenzheng Chen and Kangxue Yin
    and Daiqing Li and Or Litany and Zan Gojcic and Sanja Fidler},
    booktitle={Advances In Neural Information Processing Systems},
    year={2022}
  }