Authoring multi-style terrain with global-to-local control

The appearance styles of natural terrains vary significantly from region to region in real world, and there is a strong need to effectively produce realistic terrain with certain style in computer graphics. In this paper, we advocate a novel neural network approach to the rapid synthesis of multi-style terrains that could directly learn and infer from real terrain data. The key idea is to explicitly devise a conditional generative adversarial network (GAN) which encourages and favors the maximum-distance embedding of acquired styles in the latent space. Towards this functionality, we first collect a dataset that exhibits apparent terrain style diversity in their style attributes. Second, we design multiple discriminators that can distinguish different terrain styles. Third, we employ discriminators to extract terrain features in different spatial scales, so that the developed generator can produce new terrains by fusing the finer-scale and coarser-scale styles. In our experiments, we collect 10 typical terrain datasets from real terrain data that cover a wide range of regions. Our approach successfully generates realistic terrains with global-to-local style control. The experimental results have confirmed our neural network can produce natural terrains with high fidelity, which are user-friendly to style interpolation and style mixing for the terrain authoring task.