Zhixiang Wang1,2, Yu-Lun Liu3, Jia-Bin Huang4, Shin'ichi Satoh2,1, Sizhuo Ma5,
Guru Krishnan5, Jian Wang5
1University of Tokyo 2NII 3NYCU 4University of Maryland, College Park 5Snap Inc.
- 2025.02.04: ✅ Inversion code release.
- 2024.01.03: 🚀 Paper is accepted to IJCV.
- Full system release
- Killer video release
- Hugging Face demo release
- Clone EG3D and ensure
Deep3DFaceRecon_pytorchproperly initialized
git clone https://github.com/NVlabs/eg3d.git
cd eg3d
git submodule update --init --recursive- Download the checkpoint on FFHQ from NGC Catalog
wget --content-disposition 'https://api.ngc.nvidia.com/v2/models/org/nvidia/team/research/eg3d/1/files?redirect=true&path=ffhq512-128.pkl' -O ffhq512-128.pkl
cd ..- Clone this repo and install environment
git clone git@github.com:lightChaserX/DisCO.git
cd DisCO
conda env create -f environment.yml- Modify
example_configs/config.pyaccordingly, including the path to input images, pre-trained model weight, etc - Process your data
python process_in_one_time.py example_configs/config.py
- Run the following command
python run_in_one_time.py
If you find our work useful, please kindly cite as:
@article{wang2024disco,
title={DisCO: Portrait Distortion Correction with Perspective-Aware 3D GANs},
author={Wang, Zhixiang and Liu, Yu-Lun and Huang, Jia-Bin and Satoh, Shin'ichi and Ma, Sizhuo and Krishnan, Guru and Wang, Jian},
journal={International Journal of Computer Vision},
year={2024}
}
This code is built on the following code base EG3D, PTI, STIT, and 3d-photo-inpainting. Our functionality also depends on MiDaS for depth estimation, MODNet for image matting, SD1.5 or DALLE2 for background inpainting.