An Adjectival Interface for Procedural Content Generation

 

This paper presents a novel interface for generating procedural models, textures, and other content, motivated by the need for interfaces that are simpler to understand and more rapidly utilize. Instead of directly manipulating procedural parameters, users specify adjectives that describe the content to be generated. By making use of a training corpus and semantic information from the WordNet database, our system is able to map from the set of all possible descriptions, adjective space, to the set of all combinations of procedural parameters, parameter space. This is achieved through a modification to radial basis function networks, and the application of particle swarm optimization to search for suitable solutions. By testing with three very different procedural generation systems, we demonstrate the wide applicability of this approach. Our results show that non-technical users not only prefer an adjectival interface to one offering direct control over the procedural parameters, but also produce content that more closely matches a given target.