Step‑by‑Step Guide to Facial Expression Transfer with PaddleGAN

In this tutorial we demonstrate how to use Baidu's open‑source PaddleGAN project to perform facial expression transfer and generate a video.

First, clone the PaddleGAN source code from Gitee (or obtain it via the WeChat public account by replying “蚂蚁呀嘿”).

# From Gitee clone PaddleGAN code
git clone https://gitee.com/paddlepaddle/

Place the cloned PaddleGAN folder into a PyCharm project and open a terminal in that directory.

# Open the folder
cd PaddleGAN

# Install required Python packages
pip install -r requirements.txt -i https://mirror.baidu.com/pypi/simple
pip install imageio-ffmpeg -i https://mirror.baidu.com/pypi/simple

Create an output directory where the generated video will be saved.

# Create output folder
cd applications/
mkdir output

Install Baidu's PaddlePaddle and PaddleHub, which provide the deep‑learning framework and the face‑detection model.

pip install paddlepaddle -i https://mirror.baidu.com/pypi/simple
pip install paddlehub -i https://mirror.baidu.com/pypi/simple

Run the following command (adjust the paths to your own project) to transfer the facial motions from a driving video to a source image:

export PYTHONPATH=$PYTHONPATH:../PycharmProjects/mayiyahei/PaddleGAN && \
python -u tools/first-order-demo.py \
  --driving_video ../PycharmProjects/mayiyahei/MaYiYaHei.MP4 \
  --source_image ../PycharmProjects/mayiyahei/input.jpeg \
  --relative --adapt_scale

The command‑line arguments are:

driving_video : video whose facial expressions will be transferred.

source_image : the image that receives the transferred expressions.

relative : use relative key‑point coordinates (recommended).

adapt_scale : adapt motion scale based on the convex hull of key points.

After the video is generated (without audio), merge it with an audio track using ffmpeg:

ffmpeg -i MaYiYaHei.mp4 -i MYYH.mp3 -vcodec copy -acodec copy result.mp4

To create a GIF, run:

ffmpeg -ss 0 -t 8 -i result.mp4 -s 600x400 -r 15 result.gif

Note that the facial‑expression transfer can be time‑consuming, especially on CPU‑only machines.

Feel free to experiment with different source images and driving videos to produce personalized videos or GIFs.