BeautyGAN: Instance-level Facial Makeup Transfer with Deep Generative Adversarial Network

This image is from BeautyGAN

This is a modification of Offical Pytorch code for BeautyGAN. The main differences are:

• Dataloaders
• Identities Loss (VGG)
• Add Tensorboard

🌿 If you have trouble running this code, feel free to send me an issue. I'll try my best to help you 👍.

NOTE:: I got serveral checkpoint in Google Drive. The results are not good as stated in the main paper, but check it if you like.

For the best results, I recommend original tensorflow pre-trained model

First, clone this repo:

git clone https://github.com/thaoshibe/BeautyGAN-pytorch-reimplementation.git

&

cd BeautyGAN-pytorch-reimplementation

Install all required python packages:

• Via pip: pip install -r requirements.txt
• Via conda: conda env create -f environment.yml (then conda activate beautygan)

Please download the Makeup Transfer Dataset.

python train.py --data_path /path/to/dataset

For example: I downloaded Makeup Transfer Dataset to /home/ubuntu/makeup_dataset/. So the command will be python train.py --data_path /home/ubuntu/makeup_dataset/

• For Tensorboard: tensorboard --logdir runs, then open http://localhost:6006/

Results at epoch 80. From left to right: Original Image, Makeup Reference, Result

Tensorboard captured at epoch 0. The results are unpleasant at this time. Don't worry 🔥. To get the final model, I trained the model until epoch 200 (about ONE WEEEK in RTX 1080Ti)

Tensorboard Loss
Tensorboard Image. First rows are fake images (results). Second are original images (inputs). Third are recycle images (reconstructed images)

This code is heavily based on BeautyGAN - Offical Pytorch Implementation. Thanks wtjiang98 so much to make his work available 🙏🙏🙏