Multilayer Map Generation Using Attribute Loss Functions

This research introduces a novel method for procedural terrain generation by utilizing Generative Adversarial Networks (GANs) combined with custom attribute loss functions. The primary objective is to provide game developers with greater human control over the generation of multilayer maps, specifically height and segmentation layers, by using statistical attributes as input conditions. By implementing differentiable attribute loss functions, the model can evaluate generated maps against specific statistical targets—such as terrain category proportions or altitude distributions—directly during the training process, functioning similarly to supervised learning.

The study evaluates various model architectures, including Conditional GANs (CGAN) and GauGAN, alongside different normalization methods such as batch normalization, channel-wise normalization, and modified SPADE blocks. The dataset consists of 5,000 samples of 128x128 pixel height and segmentation maps. Quantitative analysis was performed using the Normalized Relative Discriminative Score (NRDS), while qualitative assessments categorized the outputs into three distinct types based on visual fidelity and inter-layer relationships.

Key findings indicate that the CGAN architecture with batch normalization generally produces the most coherent maps. While channel-wise normalization effectively satisfies input conditions, it often sacrifices the diversity of ground categories. Conversely, element-wise normalization via SPADE blocks tends to introduce excessive fragmentation, which may be undesirable for standard terrain but potentially useful for specific environmental features like marshes. The research concludes that while NRDS is a useful metric, it primarily detects abnormal patterns rather than high-level structural relationships. The proposed attribute loss functions offer a robust, transparent mechanism for guiding GAN training, providing a scalable approach for developers to exert precise control over procedurally generated game environments.