Trained a GAN to generate images of realistic terrain and associated height maps for rendering in 3D based on a user created segmentation map. Details, code, dataset on my git:

https://github.com/tpapp157/SPADE-Terrain-GAN

The GAN itself uses the NVidia SPADE network as a start point and makes quite a few modifications to better suit the application. Next steps, I’d like to explore incorporating some of the features of the recent Style-GAN 2. Open to any thoughts or suggestions.