Procedural generation of virtual pavilions via a deep convolutional generative adversarial network
Ziwei Chen, Desheng Lyu · Computer Animation and Virtual Worlds · 2022
Abstract Virtual pavilions can help spread culture and bring fun. A virtual pavilion needs a designed map, and terrain editors then manually layout each part of it. Procedural content generation via machine learning can quickly generate virtual pavilion maps to assist in virtual pavilion design. This article proposes adding a self‐attention module to some commonly used deep convolutional generative adversarial networks to generate virtual pavilion maps. A three‐dimensional(3D) virtual pavilion is built based on these maps, and interactive features are added to make it more experiential. Then the improved and original networks are mainly evaluated in generating maps that are solvable and similar to the training data for finding the best generator. The evaluation results show that our improved methods always perform better on each metric, and the WGAN with a self‐attention module is what we need.