HairstyleGAN

HairstyleGAN is Graduation Assignment in FPT University of Ta Dang Khoa and Pham Cao Bang

Note

I have updated from using stylegan to using stylegan2-ada-pytorch. But I have not updated the stylegan-encoder yet. You can use older commit for better result.

Directory layout

.
├── configs         # store all configs
├── core            # model, loss, optimizer, metrics, dataloader, ...
├── data
├── docs
├── exp             # weights, processed data features, ...
│   ├── features
│   └── weights
├── modules         # pre-process, post-process, ...
├── service         # REST API app
├── tools           # scripts for train, test, ...
└── utils

Generate data

python tools/generate.py --outdir=data/raw_images/ --trunc=1 --seeds=1-100 --network=https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/ffhq.pkl

Align images

mkdir exp/features/aligned_images
mkdir exp/features/landmarks
mkdir exp/features/aligned_landmarks
python tools/align_images.py data/raw_images/ exp/features/aligned_images/ --landmarks_path=exp/features/landmarks/ --aligned_landmarks_path=exp/features/aligned_landmarks/

Merge hair

mkdir exp/features/merge_hair_images
python tools/merge_hair.py --face_dir exp/features/aligned_images/ --hair_dir data/hairs/ --merge_hair_dir exp/features/merge_hair_images/ --face_landmark_dir exp/features/aligned_landmarks/ --hair_landmark_dir data/hair_landmarks/

Generate merge hair-face images

python tools/generate_merge_hair_images.py --input_dir exp/features/merge_hair_images/ --output_dir exp/results/

Result

License

MIT