Multilayer Map Generation Using Attribute Loss Functions
Runze Tang, Penny Sweetser · 2023
Procedural Content Generation via Machine Learning (PCGML) has been studied to generate terrain maps, but many studies focus on height maps and lack human control. We propose a method based on Generative Adversarial Networks (GANs) to generate multilayer maps of terrain with statistical attributes as inputs to introduce more human control. Since the discriminators used in GANs are difficult to evaluate and lack transparency, we propose attribute loss functions, which work as a supervised approach to evaluate the statistical attributes of generated maps directly using differentiable functions for backpropagation. We tested combinations of two model architectures and different conditional normalisation methods and analysed their characteristics. We found that CGAN architecture with batch normalisation worked well in general, while SPADE block introduced more fragments, and channel-wise normalisation satisfied input conditions better but lost distribution diversity and inter-layer relationships.