CartoonGAN

Within this repo, I try to implement a GAN network which is primarily based on cartoon GAN [Chen et al., CVPR18].

I created a github page for detailed documentation, please see https://tobiassunderdiek.github.io/cartoon-gan/ for details.

Usage

Step 1: Generate datasets

All scripts to create the images are resumeable. It is possible to run make cartoons and make photos in parallel by calling them manually in separate terminals.

Cartoon images

• download all_data.csv from safebooru dataset [2]
• point to all_data.csv in PATH_TO_SAFEBOORU_ALL_DATA_CSV of cartoon_image_downloader.py
• run make install to install necessary libraries
• run make cartoons to download configurable amount of medium size images

Edge-smoothed version of cartoon images

• run make cartoons-smooth to create the images

Photos

• download and unzip coco annotations from [3]
• configure annotations dir location in PATH_TO_COCO_ANNOTATIONS_ROOT_FOLDER of photo_downloader.py
• run make photos to download configurable amount of photos of persons

Step 2: Train model

All the steps are described in a jupyter notebook on colab, please see here for details.

Step 3: Test

• run make install-transform
• download pre-trained weights, they are available for download as part of the release here..
• run make transform IMAGE=some_example_image_path

Additional information about how to load the pre-trained weights and transform images can be found in the project documentation here: https://tobiassunderdiek.github.io/cartoon-gan/ .

Credits

Thanks to the authors [Chen et al., CVPR18] of the paper for their great work.

References

[Chen et al., CVPR18] http://openaccess.thecvf.com/content_cvpr_2018/papers/Chen_CartoonGAN_Generative_Adversarial_CVPR_2018_paper.pdf

[2] https://www.kaggle.com/alamson/safebooru/download

[3] http://images.cocodataset.org/annotations/annotations_trainval2017.zip