Abstract:
Animation Industry is facing a huge surge in demand because of its rapid growth over the years. This has been a huge problem in the creative process where animators of anime industry in particular, are finding it difficult to create new characters in a short amount of time, specially with female characters because their faces show more emotions than their male counterparts. The problem is that Anime is almost entirely drawn by hand and one scene can take 5-15 days to make, with the huge demand surge, this has been a growing problem for the animators in anime industry. The problem domain appears well-suited to Generative Adversarial Networks. GANs are deep learning algorithms that utilises Artificial Neural Networks for image generation. Recent advancements in Generative Adversarial Networks have allowed to generate state of the art fake synthesis human faces.
However, Generative Adversarial Networks have only been explored in animation domain. Few attempts have been made with older Generative Adversarial Networks algorithms to generate anime faces but those attempts have been futile because of reasons such as limited dataset, lazy training methods, and lazy implementations etc. that re- vealed huge gap in the related literature. Therefore, most improved and stable Generative Adversarial Networks algorithm was implemented, A Style-Based Generator Architecture for Generative Adversarial Networks (StyleGAN) for anime faces generation.
Extensive work was done to create a unique custom dataset of 217,800 anime face images from Danbooru2019 dataset, using two open source licensed applications. lbpcascadeanime face is used to crop anime faces from the dataset and waifu2x to upscale all cropped faces to 512 resolution images.
StyleGAN was trained using the unique dataset for 47 days, experimenting though a series of hyper-parameters. The results from final trained model were indistinguishable fake synthesis anime faces. A hypothesis was created to evaluate the images generated. Rating and Preference Judgement evaluation method was found through literature and mimicked through an online questionnaire quiz to test the hypothesis and evaluate the trained model. The results of the questionnaire was very surprising where the average score was 49%, indicating failure to distinguish between real anime faces drawn by animators and anime faces generated by the trained model with StyleGAN. The results answered the research question and validated the trained model and our research.
Final Grade: 70%
Updated dataset and other files that are mentioned on the appendix section:
All files (working code, results): https://drive.google.com/file/d/1NSZEZ3cfj-lsGxYxFNF4LsEbqRpSq5Ui/view?usp=sharing
Dataset: https://drive.google.com/file/d/1aXyMkCcgIOdcRLyDfmvJO6bCYbTEd_d8/view?usp=sharing
Snapshots (pictures used to make training montages): https://drive.google.com/file/d/16M-NyPOn43qHV9dNaUK4tvw8t2FC7l2Y/view?usp=sharing
Truncated pictures (randomly generated samples): https://drive.google.com/file/d/1yLM09cXQuomeqxy3wdMYLBduCNZVe0ZI/view?usp=sharing
Training montage videos: 10k iterations: https://drive.google.com/file/d/1UsyHGj5iIYn9yE4zRvqsIzQPlYipEbxr/view?usp=sharing
14k iterations: https://drive.google.com/file/d/1kJskVQkPyZdQZahbstgNwDpCaPKfbS-2/view?usp=sharing
49k iterations: https://drive.google.com/file/d/1G5ydRk6ypSlzVKNvPZhB9DuFLTmG4dl3/view?usp=sharing
61k iterations: https://drive.google.com/file/d/1ihdtB6Z8_d2eJYUho_FRNMJX2EaDCZaR/view?usp=sharing