giaanthunder/facenix

Facenix project for face image modification

Facenix project

Facenix is a face attribute manipulating application. This app is deployed by using Django web framework with STGAN and StyleGAN models as backend.

Video demonstration: Youtube

Usage

• Environment

  • Python 3.6

  • TensorFlow 2.1.0

  • OpenCV, Django, scikit learn,...

  • We recommend Anaconda or Miniconda, then you can create the environment with commands below

    conda create -n tf2 python=3.6
    conda activate tf2
    pip install --upgrade pip
    pip install tensorflow==2.1.0
    pip install tensorflow_addons
    pip install django
    pip install dlib
    pip install opencv-python
    pip install sklearn
    pip install pillow
    pip install requests matplotlib
    

• Run web application All commands are run from facenix directory after "git clone"

  • Download pre-trained models
    • weights.zip Google Drive
  • Unzip and copy all folders into facenix/

    unzip weights.zip

  • Run web application

    cd web_app
    ./runserver.sh

  • Open web browser. Go to http://127.0.0.1:8000/. After the web page is loaded, you can try your own sample. If you would like to run web server on specific IP and port, you can pass it as arguments.

    e.g.
    ./runserver.sh 192.168.1.254:8888

• Re-training models

  • Make sure to clean all previous training

    ./clean_all.sh

  • STGAN dataset preparation: CelebA aligned

    • download the dataset

      • img_align_celeba_crop_128.zip (move to facenix/data/img_align_celeba.zip): Google Drive
      • list_attr_celeba.txt (move to facenix/data/list_attr_celeba.txt): Google Drive

    • unzip the dataset

      cd data
      unzip ./img_align_celeba.zip

  • StyleGAN dataset preparation: CelebA-HQ

    • download the dataset

      • CelebAMask-HQ.zip (move to facenix/data/CelebAMask-HQ.zip): Google Drive

    • unzip the dataset

      cd data
      unzip ./CelebAMask-HQ.zip

  • Train STGAN

    cd stgan
    python train.py --experiment_name origin

  • Train StyleGAN

    • Train generator

      cd stylegan
      python train.py --experiment_name origin

    • Train classifier

      cd classifier
      python train.py

    • Train attribute vectors

      cd att_vector_finder
      python sample_maker.py
      python vector_finder.py

• Samples STGAN

• Samples StyleGAN

Known issues

• Attribute editing is not always successful.
• When editing with StyleGAN, the same picture could be led to different results each time the image is uploaded.
• The web server can serve 1 user each time due to high computing cost.

Reference

• STGAN
• StyleGAN