StyleTerrain: A novel disentangled generative model for controllable high-quality procedural terrain generation

Abstract

Terrain is a vital element in construction of virtual scene in the digital era. Despite considerable progress has been made in Generative Adversarial Network (GAN) based terrain modeling methods, their quality and controllability still cannot meet up the requirements of many emerging industries. The present work proposes a novel disentangled generative model, named as StyleTerrain, for achieving controllable high-quality terrain generation. It introduces disentangled representation learning into GAN-based terrain modeling methods for the first time. The model has been evaluated quantitatively. The results show a significantly short perceptual path length and the effectiveness of the disentanglement mechanism in controllable terrain generation, indicating that latent space disentanglement is a promising future direction for achieving generation controllability in GAN-based terrain modeling methods.