/cartoonGAN_tensorflow2

Training Generative Adversarial Network over own-made generated cartoon dataset

Primary LanguagePython

cartoonGAN with tf2

Sapienza, MARR, Elective in Artificial Intelligence project: training cartoonGAN with an own implemented dataset

original paper http://openaccess.thecvf.com/content_cvpr_2018/papers/Chen_CartoonGAN_Generative_Adversarial_CVPR_2018_paper.pdf

related blog https://leemeng.tw/generate-anime-using-cartoongan-and-tensorflow2-en.html

Approach

The work consisted of a process of re-elaboration and re-implementation of the gan presented by https://github.com/mnicnc404/CartoonGan-tensorflow. Finally she was trained with colab on a new set of images to then observe the results.

NB: Our method takes unpaired photos and cartoon images for training: we captured key frames from the Cartoon film ”Spirited Away” of Hayao Miyazaki.

Documentation

You can read our final technical documentation about this project here DOCUMENTATION

Results

After some problems encountered (such as enormous training time, OOM exhausted resource, network instability ...), we found good results thanks to tensorboard, shown here RESULTS