Skip to content

Latest commit

 

History

History
14 lines (11 loc) · 1.46 KB

README.md

File metadata and controls

14 lines (11 loc) · 1.46 KB

Anime GAN

By: Shockblack/Aiden Zelakiewicz logo

This repository houses the model architecture I used to create my Generative Adversarial Network to synthesize anime faces! I have seen other models do something similar, but they use only characters from the 90s, so I wanted to spice it up with newer images. The characters were all scraped from MAL and detected for facial features using a Faster-RCNN from user @qhgz2013 on GitHub. The images were downloaded in their original size from MAL, but cropped to 64x64 to make it easier for training, like so.

drawing

Images used in training

The model uses a convolutional framework similar to the DCGAN framework, but utilizes spectral normalization on the convolution layers. This was originally was going to use the LSGAN loss function, but I broke the model at some point so just fell back to BCE because I couldn't figure out what happened. The model can be found in model.py with their components in blocks.py. Train.py houses the main training loop for the whole model. Here is a silly goofy gif of the training of 300 epochs. The images become cursed after ~100 epochs, but I think it's still interesting to include. training