A step toward procedural terrain generation with GANs

Here is a side project I did lately which has been accepted as a poster in the VGML (Video Games and Machine Learning) workshop at ICML. We train two generative adversarial networks (GANs) to synthesise heightmaps and their corresponding satellite images from scratch.

Arxiv: https://arxiv.org/abs/1707.03383 (to be updated with better figures)

Code: https://github.com/christopher-beckham/gan-heightmaps

Interpolation between 100 randomly generated heightmaps (and their translations to textures)
The same thing but in 3D (sans texture, we just use a default colourmap)

Reblog: Sam Snider-Held also did procedural heightmap generation using style transfer techniques, and you can see his work here:


I wholeheartedly agree that generative modelling and style transfer techniques can do a lot of good in the realm of video games (and movies).