GANime
Projects | | Links: GitHub Repository
Convolutional Generative Adversarial Network which synthesizes fake anime characters.
During my internship at UC Berkeley, I wanted to better acquiant myself with the models I was working with. At the time, I was learning a lot fast and had ideas I wanted to try out, but never had the chance too. I started this project to take existing more modern (mostly post 2000) and train a model to produce new ones. After creating my algorithm to scrape the website My Anime List for images, the project was unfortunately benched.
In the Fall semester of my senior year at OSU, I was in the process of taking an ML course and needed to pick a topic for a final project. With the data already downloaded (although not cleaned), I thought it was a great opportunity to revisit the project. I cleaned the data by using a previously trained facial detection model to confirm whether an image had a detectable face. This was not a perfect method, but it cut out a lot of grunt work on my end and cleaned the data enough. Armed with much more knowledge about deep networks than I had when I first concieved the project, I began constructing the model.
The weapon of choice for the task was originally going to be a Diffusion-GAN, introduced that summer in a paper by (XXX). Instead I opted for a more “stock” deep convolutional generative adversarial network (DCGAN). Some of the layers within this standard model were adjusted though. I wanted to use many of the new tools that had been being published, such as spectral normalization.
Within a month, I had a working model that produced decent results. Nothing spectacular, but for the complexity and quality of data (which wasn’t great) I used it wound up well! The code for the model is available on GitHub to learn from. Here is a gif showing the training of a set of images.
