Creating Anime Characters with a DCGAN

Description: In this project, I train a DCGAN model to generate anime character faces such as those from Danbooru. In general, I scraped anime images from Danbooru, performed face detection and cropped the images of the faces to 64x64x3 images, and trained a Deep Convolutional Generative Adversarial Network to generate new images similar to the input facess. In training the model, I suffered from mode collapse, which is the case when the generator’s output lacks diversity given any random noise input. To combat this, I used label smoothing, label flipping, and reduced the complexity of the discriminator network. The model was trained in Google Cloud using a single NVIDIA Tesla P4 GPU and 16 cores with 64GB RAM.

The Github repository for this project can be found here.

Avatar
Ancil Crayton
Senior Research Scientist

My research interests lie at the intersection of machine learning, economic analysis, and public policy.