|
# ReHiFace-S π€π€π€ |
|
|
|
## π Introduction |
|
|
|
ReHiFace-S, short for βReal Time High-Fidelity Faceswapβ, is a real-time high-fidelity faceswap algorithm created by Silicon-based Intelligence. By open-sourcing the capabilities of digital human generation, developers can easily generate large-scale digital humans who they want, enabling real-time faceswap capability. |
|
|
|
## πͺ Project features |
|
|
|
- Real-time on NVIDIA GTX 1080Ti |
|
- Zero-shot inference |
|
- High Fidelity faceswap |
|
- Support ONNX and live camera mode |
|
- Support super resulution and color transfer |
|
- Better Xseg model for face segment |
|
|
|
## π₯ **Examples** |
|
|
|
We show some faceswap examples. </br> |
|
<p align="center"> |
|
<img src="./assets/demo20.gif" alt="showcase"> |
|
<br> |
|
</p> |
|
<p align="center"> |
|
<img src="./assets/demo10.gif" alt="showcase"> |
|
<br> |
|
</p> |
|
|
|
## π§ Getting Started |
|
|
|
### Clone the code and prepare the environment |
|
|
|
- Python >= 3.9 (Recommend to use [Anaconda](https://www.anaconda.com/download/#linux) or [Miniconda](https://docs.conda.io/en/latest/miniconda.html)) |
|
- [PyTorch >= 1.13](https://pytorch.org/) |
|
- CUDA 11.7 |
|
- Linux Ubuntu20.04 </br> |
|
|
|
```bash |
|
conda create --name faceswap python=3.9 |
|
conda activate faceswap |
|
pip install -r requirements.txt |
|
``` |
|
|
|
## π Pretrained models |
|
|
|
Download all pretrained weights from [Google Drive](https://drive.google.com/drive/folders/1hVWFXPIDwACqoKKtgXAJubYC_H4k5njc?usp=drive_link) or [Baidu Yun](https://pan.baidu.com/s/1Bn47xOjZg-oU7_WyAHu3EQ?pwd=9bjo). We have packed all weights in one directory π. Download and place them in `./pretrain_models` folder ensuring the directory structure is as follows:</br> |
|
```python |
|
pretrain_models |
|
βββ 9O_865k.onnx |
|
βββ CurricularFace.tjm |
|
βββ gfpganv14_fp32_bs1_scale.onnx |
|
βββ pfpld_robust_sim_bs1_8003.onnx |
|
βββ scrfd_500m_bnkps_shape640x640.onnx |
|
βββ xseg_230611_16_17.onnx |
|
``` |
|
|
|
## π» How to Test |
|
|
|
```python |
|
CUDA_VISIBLE_DEICES='0' python inference.py |
|
``` |
|
Or, you can change the input by specifying the `--src_img_path` and `--video_path` arguments: |
|
```python |
|
CUDA_VISIBLE_DEICES='0' python inference.py --src_img_path --video_path |
|
``` |
|
|
|
### Live Cam faceswap |
|
|
|
You should at least run by NVIDIA GTX 1080Ti. </br> |
|
|
|
***Notice: The time taken to render to a video and warm up the models are not included.*** </br> |
|
|
|
Not support Super Resolution. |
|
```python |
|
CUDA_VISIBLE_DEICES='0' python inference_cam.py |
|
``` |
|
***Notice: Support change source face during live with 'data/image_feature_dict.pkl' !*** |
|
<p align="center"> |
|
<img src="./assets/cam_demo1.gif" alt="showcase"> |
|
<br> |
|
</p> |
|
<p align="center"> |
|
<img src="./assets/cam_demo2.gif" alt="showcase"> |
|
<br> |
|
</p> |
|
|
|
## π€ Gradio interface |
|
|
|
We also provide a Gradio interface for a better experience, just run by: |
|
|
|
```bash |
|
python app.py |
|
``` |
|
|
|
## β¨ Acknowledgments |
|
|
|
- Thanks to [Hififace](https://github.com/johannwyh/HifiFace) for base faceswap framework.<br> |
|
- Thanks to [CurricularFace](https://github.com/HuangYG123/CurricularFace) for pretrained face feature model.<br> |
|
- Thanks to [Xseg](https://github.com/iperov/DeepFaceLab/tree/master) for base face segment framework. |
|
- Thanks to [GFPGAN](https://github.com/TencentARC/GFPGAN) for face super resolution. |
|
- Thanks to [LivePortrait](https://github.com/KwaiVGI/LivePortrait) and [duix.ai](https://github.com/GuijiAI/duix.ai) for README template. |
|
|
|
|
|
## π Citation |
|
|
|
If you find ReHiFace-S useful for your research, welcome to π this repo. |
|
|