Skip to main content
Date
2021-07-14
World-GAN

Artificial intelligence can also be creative. So-called Generative Adversarial Networks (GAN) create music and photorealistic images. This creativity is also used in the video game sector: Researchers at L3S show that World-GAN is also suitable for generating game worlds in Minecraft.

So far, it has already been possible to develop 2D environments from AI-supported generators, for example for the game Super Mario Bros. World-GAN is now also capable of creating 3D worlds of any size. The AI system learns based on human-made or generated Minecraft structures and creates 3D worlds in a comparable style. This could create more varied and detailed worlds that catch up with human-created worlds in terms of quality.

This is made possible by "block2vec" representations. They make World-GAN independent of the number of different blocks, which can vary greatly in Minecraft, and enable the generation of larger levels. There is also a so-called discriminator that can detect insufficient forgeries and thus spur the generator on to better performance.

The entire paper is available here