Hacker News Reader

> This is a fun project I've been working on recently. The idea of the cards started when I saw how an AI could generate realistic portraits of people. I tried to apply similar techniques to generating fantasy characters, and got amazing results.

Can you share a bit more about the art generation? Are you using StyleGAN or some other architecture? Are you hand-selecting the art from among generated images? What training dataset did you use or create?

View on HN

Sure! I'm using a network very similar to StyleGAN. After generating a batch, I review the images and weed out maybe 10% of the really bad ones. I was actually surprised at how well it worked.

For training, I spent a lot of time scraping images. I couldn't find any good dataset for this purpose. Controlling for quality is pretty tough, and I didn't spend nearly as much time as I should. And it's evident in the result. Some of the generated arts have pretty low pixel quality if you look at the original image. Luckily, it looks ok on the cards.

Reply on news.ycombinator.com