/CrossModalAlign

Primary LanguagePython

Global Direction

  • Prepare

    1. Download npy/ffhq, styleGAN2 pretrained model, ArcFace pretrained model, Face segmentation pretrained model and test faces from https://drive.google.com/drive/folders/1LXGi5WF2uxRs0gRICDSkhhfYynTs2haL?usp=sharing

    2. Create Docker Contrainer

      cd global bash ./docker/docker.sh

    • StyleGAN2 & ArcFace : ./pretrained_models/ 아래에 위치함
    • test_faces.pt: Optimization & Global에서 사용 -> 파일 위치 올바르게 바꾸기

실험: StyleCLIP baseline

  
  cd global
  python global.py --method "Baseline" --num_test 10 --topk 50

실험: StyleCLIP Ours

  
  cd global 
  python global.py --method "Random" --num_test 100 --topk 50
  • num_test: Number of test faces to use
  • num_attempts: Number of iterations (check diversity)
  • topk: Number of channels to change

model.py : RandomInterpolation defined

Extracts core semantics, unwanted semantics from target and source positives from the source
Use probabilistic approach to sample and create updated final target embedding