Tools used while creating the project:
- Python 3.10
- Jupyter notebook (kaggle GPU)
You can find all about basic libraries and modules used in the REQUIREMENTS.txt file.
To play with network using GUI you can just run 'main.py' script file in the project directory using python (windows 10):
your\folder\with\main\file> python main.py
The whole work is based on "Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Network 2017" paper: (https://arxiv.org/pdf/1703.10593.pdf)
The implementation of the network started with the Monet2Photo task - to teach the network to translate photos into the style of Monet's paintings / make Monet's paintings photorealistic. Work process is described in the cyclegan_demo.ipynb notebook.
The main task of A_CGAN is to transform a real face into corresponding Japanese animation style face while preserving the original features. The basic training process and architecture of the network is described in the AniCycleGAN.ipynb file.
The network was originally trained on the Flickr Faces dataset (https://www.kaggle.com/xhlulu/flickrfaceshq-dataset-nvidia-resized-256px). However, due to the variety of people's gender/age, as well as differences in their scale, the network was NOT able to correctly carry the style.
After some researching, the selfie2anime dataset (https://www.kaggle.com/arnaud58/selfie2anime) was found to be perfect for the task. Training was done using the kaggle platform's dedicated Nvidia Tesla K80 graphics card. Full training time took around ~30 hours.
In the final version in 80% of cases the network successfully transfers the style, as well as in good lighting and photo rotation correctly detects the face and modifies it. However, sometimes you can find non-critical artifacts in generated faces - white (sometimes black) 'spots'. In Monet2Photo this issue practically disappeared by the end of network training.
With input image having small size faces, there is an usual style transfer, which can be used as well:
Ani_CycleGAN network can also be used as a SuperResolution transformer for faces, that are Already anime:
In conclusion, I would like to show you what all the work I have done turned out to be. A simple but quite functional GUI was created for processing and further saving an image. The application also has a brief instruction (the '?' button)
