In the past we played a lot of pen and paper with our friends. Some of them could draw very well and designed character avatars for all of us. When playing, our game master always had some pictures of characters at hand showing non-player characters (NPCs) for better imagination. Additionally, we saw some RPG streams on YouTube where the players have really cool avatars of their characters as their profile pics while playing and important NPCs were always represented by some drawn character.
To give all players without art skills access to thousands of free to use fantasy avatars we decided to create our own StyleGAN2 model. We hope that the community will use our avatars and have a lot of fun playing!
The images used for training are under licenses, that enable free use even for commercial purposes with no attribution required. Additionally, we asked a lot of artists if I can use their images for the creation of this model. When training, the AI extracts only information out of the pictures and learns how to draw them. The images are not getting modified itself. It is a new creation with the necessary individuality. In addition the original images aren´t published in any form. But we want to honor and respect the artists who have spent years of work to practise to become experts. So, we decided to ask for permission and received positive feedback by some great professional artists. To further improve the model and add races like orcs, trolls, gnomes, etc., we will need tons of images. We continue to ask a lot of different artists for permission to add their images to our training set. Our training set will never be published, as keeping their images private understandably is a key requirement for the artists. This process is quite time consuming, and we hope you understand that we are limited to mostly human characters for now. Additionally, the model still is a little "colour-blind" which means most of the images look quite caucasian. Reason for this is also the missing variety in training data. We hope we can solve this with more training data as well. To generate these images NVIDIA Stylegan2 and the anime faces model of gwern.net was used. The model itself is copyrighted to the creators of the model. Fortunately, gwern distributed the model with a CC-0 license, so we can use it legally for transfer learning. All resulting images are not created by a human and in a highly transformative way based on a dataset with thousands of images. Therefore, we cannot have any copyright on the images by ourselves.
For a quick start you need a google account, some space in your google drive and this colab notebook. To open the colab notebook. Just click on "open in colab".
Before starting you have to make sure that a GPU runtime environment is selected.
Now Download the model file and place it in your google drive in the folder Networks/fantasy.pkl. (of course it is possible to use another naming or location, but you will need to adjust the settings when writing images in the colab notebook to find your model) Now follow the instructions in the colab notebook and press the play button one block after another.
We are always happy if you send me some feedback!
A huge thanks you to www.gwern.net for the great work and tutorials on the anime dataset. Another huge thanks to Derrik Schultz and the team from "Artificial Images". The videos helped us a lot to develop the models. Check out their youtube channel for awesome videos.