How the IACG was made?

Example of a character generated by the IACG

How the IACG was made?

What is a GAN?

In June 2014, Ian J. Goodfellow together with a group of colleagues from the University of Montreal published a paper called "Generative Adversarial Nets", in which they propose a framework for generative models through an antagonistic process. The proposal is to confront two neural networks: one would be generative and the other discriminative. These two networks are antagonistic since they constantly compete in a zero-sum game, i.e., the gain or loss of one is compensated by the gain or loss of the other.

For further reading we recommend the following articles:

A Gentle Introduction to Generative Adversarial Networks (GANs) (Brownlee, 2019)
Generative Adversarial Nets (Goodfellow, Pouget-Abadie,Mirza, Xu, Warde-Farley, Ozair,... & Bengio, 2014).

What's in this website?

Here you will find 300+ images of characters generated by the IACG

What's the IACG?

The IACG is a GAN (Generative Adversarial Network) trained to generate anime-style characters. These characters can be used for different projects, such as video games, visual novels, RPGs, webcomics, or also as avatars in social media. They can be taken as inspiration or used freely.

How it was trained?

The StyleGAN2 ADA architecture proposed by Nvidia was used to train the network. With this architecture it was possible to train it with a limited number of images, this is usually a problem when training GANs, but thanks to the work done by Nvidia it is possible to achieve excellent results with only a few references.

If you want to know more about the work done by them I recommend you to visit their repository and read the paper.

NVIDIA Research Projects

StyleGAN2-ADA — Official PyTorch implementation

You didn't find the image you wanted?

Here you can find a 1200+ images of characters generated by the IACG

Click here for female characters

Click here for male characters

Page updated

Google Sites

Report abuse