Anime Face Customization

Generation of Anime faces through the latent space of a Variational Autoencoder (VAE).

Built with Pytorch, Using Danbooru2018 Dataset.

Example

Requirements:

• Pytorch
• Visdom
• lbpcascade_animeface
• Collection of anime-style images (Danbooru2018)

Insights:

• When using plain VAE, balancing the KL Divergence and the MSE Loss is hard. Images tend to be blurry.
• With RMSprop and fine tuning convergence tends to be really slow at the end

References:

• https://www.gwern.net/Danbooru2018
• https://github.com/ecrows/danbooru-faces
• https://github.com/nagadomi/lbpcascade_animeface
• https://github.com/wuga214/IMPLEMENTATION_Variational-Auto-Encoder