PyTorch implementation of PSGAN: Pose and Expression Robust Spatial-Aware GAN for Customizable Makeup Transfer, still in construction...
A makeup-transfer App MagicMirror is developed by Fengwei Zhang.
Here are some exemplar results.
When source image and target image are both from the makeup dataset.
When the source image is from the makeup dataset and the target image is from weibo.
When we try to transfer the makeup style from the makeup dataset to an makeup image from weibo.
-
Prerequisites
- PyTorch
- Python 3.x with matplotlib, numpy and scipy
-
Dataset Dataset can be found in project page: http://colalab.org/projects/BeautyGAN
Before training, you should generate train/test split labels using data_preparation/generate_labels.py. What you should do is just modify the data path in generate_labels.py.
The training setting is the same as BeautyGAN. The implementation of PSGAN is still incomplete. I still have some problems in implementing the AMM with 68 landmarks detector.
However, the incomplete results is satisfying.
The training example of 50th epoch is as below:
The training example of BeautyGAN in 200th epoch is as below:
Though the implementation of PSGAN is still incomplete, it's obvious that PSGAN is pose and expression robust for makeup transfer.
The code is built upon BeautyGAN, thanks for their excellent work!