Create a video warp sequence of human faces. Can be used, for example, to create a time-lapse video showing someone's face change over time. See demos here and here.

Supported on Python 3 and OpenCV 3+. Tested on macOS High Sierra and Mojave.

• OpenCV
  • For conda users, run conda install -c conda-forge opencv.
• Dlib
  • For conda users, run conda install -c conda-forge dlib.
• ffmpeg
  • For conda users, run conda install -c conda-forge ffmpeg.
• scipy
• numpy
• matplotlib
• pillow

For brew users:

brew install opencv ffmpeg

1. Clone the repo.

git clone https://github.com/andrewdcampbell/face-movie
cd face-movie

2. Download the trained face detector model from here. Unzip it and place it in the root directory of the repo.

curl http://dlib.net/files/shape_predictor_68_face_landmarks.dat.bz2 > shape_predictor_68_face_landmarks.dat.bz2
bunzip2 shape_predictor_68_face_landmarks.dat.bz2

3. Create a virtualenv.

python3 -m venv venv
source venv/bin/activate
python3 -m pip install -r requirements.txt

1. Make a directory <FACE_MOVIE_DIR> in the root directory of the repo with the desired face images. The images must feature a clear frontal view of the desired face (other faces can be present too). The image filenames must be in lexicographic order of the order in which they are to appear in the video.

2. Create a directory <ALIGN_OUTPUT>. Then align the faces in the images with

python face-movie/align.py -images <FACE_MOVIE_DIR> -target <BASE_IMAGE> 
                           [-overlay] [-border <BORDER>] -outdir <ALIGN_OUTPUT>

The output will be saved to the provided <ALIGN_OUTPUT> directory. BASE_IMAGE is the image to which all other images will be aligned to. It should represent the "typical" image of all your images - it will determine the output dimensions and facial position.

The optional -overlay flag places subsequent images on top of each other (recommended). The optional -border <BORDER> argument adds a white border <BORDER> pixels across to all the images for aesthetics. I think around 5 pixels looks good.

If your images contain multiple faces, a window will appear with the faces annotated and you will be prompted to enter the index of the correct face on the command line.

At this point you should inspect the output images and re-run the alignment with new parameters until you're satisfied with the result.

3. Morph the sequence with

python face-movie/main.py -morph -images <ALIGN_OUTPUT> -td <TRANSITION_DUR> 
                          -pd <PAUSE_DUR> -fps <FPS> -out <OUTPUT_NAME>.mp4

This will create a video OUTPUT_NAME.mp4 in the root directory with the desired parameters. Note that TRANSITION_DUR and PAUSE_DUR are floating point values while FPS is an integer.

You may again be prompted to choose the correct face.

4. (Optional) Add music with

ffmpeg -i <OUTPUT_NAME>.mp4 -i <AUDIO_TRACK> -map 0:v -map 1:a -c copy 
       -shortest -pix_fmt yuv420p <NEW_OUTPUT_NAME>.mov

The images used to produce the first demo video are included in demos/daniel. Step 2 should produce something like the aligned images shown in demos/daniel_aligned. Step 3 should produce something like the video daniel.mp4.

• If you have photos that are already the same size and all of the faces are more or less aligned already (e.g. roster photos or yearbook photos), you can skip steps 1 and 2 and just use the directory of photos as your aligned images.

• If you want to add some photos to the end of an existing morph sequence (and avoid rendering the entire video again), you can make a temporary directory containing the last aligned image in the movie and aligned images you want to append. Make a morph sequence out of just those images, making sure to use the same frame rate. Now concatenate them with

ffmpeg -safe 0 -f concat -i list.txt -c copy merged.mp4

where list.txt is of the form

file 'path/to/video1.mp4'
file 'path/to/video2.mp4'
...

• For adding music longer than the video, you can fade the audio out with

VIDEO="<PATH_TO_VIDEO>"
AUDIO="<PATH_TO_AUDIO>" # can also drag and drop audio track to terminal (don't use quotes)
FADE="<FADE_DURATION_IN_SECONDS>"

VIDEO_LENGTH="$(ffprobe -v error -select_streams v:0 -show_entries stream=duration -of csv=p=0 $VIDEO)"
ffmpeg -i $VIDEO -i "$AUDIO" -filter_complex "[1:a]afade=t=out:st=$(bc <<< "$VIDEO_LENGTH-$FADE"):d=$FADE[a]" -map 0:v:0 -map "[a]" -c:v copy -c:a aac -shortest with_audio.mp4

You can also use the code to create a face average. Follow the same steps 1) - 2) as above. You probably don't want to overlay images or use a border, however. Then run

python face-movie/main.py -average -images <ALIGN_OUTPUT> -out <OUTPUT_NAME>.jpg

A small face dataset is included in the demos directory.

The computed average male and female face are shown below.

• Facial landmark and image alignment code adapted from https://matthewearl.github.io/2015/07/28/switching-eds-with-python/.
• ffmpeg command adapted from https://github.com/kubricio/face-morphing.
• Affine transform code adapted from https://www.learnopencv.com/face-morph-using-opencv-cpp-python/.