The creators of the innovative life simulator inZOI, who previously announced the implementation of generative neural networks, have presented a detailed breakdown of their AI training methodology in response to requests from the gaming community.

For the text-to-image conversion technology, a training dataset of 20 million visual materials distributed under the free Creative Commons license with rights for commercial use and modification was utilized. Additionally, the developers integrated the Common Canvas method from MOSAICML and created specialized filtering algorithms to optimize the source dataset.

The neural network responsible for 3D modeling was trained on a collection of 46 thousand professionally created three-dimensional objects. The database included both the studio's original developments and verified models with open licenses. This solution enables players to create technical sketches directly within the game environment.

The video-to-animation conversion system was prepared using 1.7 million motion samples. The training dataset combined the team's unique developments with verified open sources. The result is natural character animation encompassing all nuances—from everyday walking to expressive gestures and complex dance moves.

In conclusion, the developers particularly emphasized their commitment to open-source principles and ethical standards for AI application in the gaming industry.

Main image: wallpapercave.com

Comment on this