BlendFace (ICCV2023)

Overview The official PyTorch implementation for the following paper:

BlendFace: Re-designing Identity Encoders for Face-Swapping,
Kaede Shiohara, Xingchao Yang, Takafumi Taketomi,
ICCV 2023

Attention

This project is only for research purpose. Please do not apply it to illegal and unethical scenarios.
The code is distributed under the CC BY-NC-SA 4.0 license.

Changelog

2023/09/22: Released training code for BlendFace.
2023/09/09: Released demo code and model for face-swapping.
2023/07/21: Released demo code and pretrained models.

Recomended Development Environment

GPU: NVIDIA A100
CUDA: 11.4

Installation

Docker

(1) Pull a docker image from docker hub:

docker pull pytorch/pytorch:1.13.1-cuda11.6-cudnn8-runtime

(2) Replace the absolute path to this repository in ./exec.sh.
(3) Execute the image:

bash exec.sh

Pretrained Models

We provide trained models for ArcFace and BlendFace and place them to checkpoints/.

Demo

We provide a demo code to compute identity similarities of ArcFace and BlendFace for an actual positive pair (images/anchor.png and images/positive.png), negative pair (images/anchor.png and images/negative.png), and pseudo-positive pair (images/anchor.png and images/swapped.png).

python3 demo.py

The result will be displayed as follows:

> ArcFace| Positive: 0.7967, Negative: 0.0316, Swapped: 0.6212
> BlendFace| Positive: 0.8186, Negative: -0.0497, Swapped: 0.8015

It can be seen that ArcFace underestimates the similarity for the pseudo-positive pair while BlendFace predicts properly it, which indicates BlendFace mitigates the biases while keeping the discriminativeness for the negative sample.

Face-Swapping

We also provide code for face-swapping (AEI-Net + BlendFace). The pretrained checkpoint is here.
Move to /workspace/swapping.

cd /workspace/swapping

Swap target face with source face:

python3 inference.py \
    -w checkpoints/blendswap.pth \ # path to checkpoint
    -s examples/source.png \ # path to source image
    -t examples/target.png \ # path to target image
    -o examples/output.png # path to output image

Note: source images should be aligned following InsightFace and target images should be aligned following FFHQ.

Training BlendFace

Please see training/README.md.

Acknowledgements

We borrow some code from InsightFace and FaceShifter (unofficial).

Citation

If you find our work useful for your research, please consider citing our paper:

@InProceedings{Shiohara_2023_ICCV,
    author    = {Shiohara, Kaede and Yang, Xingchao and Taketomi, Takafumi},
    title     = {BlendFace: Re-designing Identity Encoders for Face-Swapping},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {7634-7644}
}