Diversity Scale: An Evaluation Algorithm for Terrain Generators

Key insights

7 takeaways

01
The Diversity Scale is an objective algorithm that returns a value between 0 and 1 measuring how diverse the feature combinations of a procedurally generated terrain are at a given area size, addressing a gap where prior approaches were designer-driven or evaluated only a single feature with genetic algorithms.
02
Terrain diversity is framed as critical for games built on discovery and near-infinite maps, where the sense of exploration depends on continually encountering terrain that feels new.
03
The method works by computing average values of predefined features (such as temperature, material composition, relief smoothness, vegetation, buildings and roads, and precipitation) across areas of a chosen size, applying Shannon entropy to their distributions, and aggregating per-feature results with weights reflecting each feature's importance to the game.
04
When comparing multiple terrains, the algorithm only measures features present on all of them, a constraint that shapes how cross-terrain comparisons can be made.
05
The reported evaluation used four synthetic terrains of 2000x2000 height points and average height as the sole feature, so it demonstrates the metric's behavior rather than validating it across diverse generators.
06
Diversity scores show small discontinuities where the algorithm switches between grid sizes, and accuracy is reported to improve for larger terrains.
07
The authors propose using the scale to guide selection among terrain generators and to determine a minimum terrain size that still yields acceptable diversity.

Summary

This paper introduces the Diversity Scale, an objective evaluation algorithm for comparing the feature diversity of procedurally generated terrains at varying spatial scales. The authors motivate the work by arguing that terrain diversity matters for games built on discovery and near-infinite maps, where the sense of exploration depends on continually encountering terrain that feels new. They position the method against existing approaches, which are either designer-driven or rely on genetic algorithms that evaluate only a single terrain feature.

The algorithm takes a predefined set of quantifiable terrain features, such as temperature, material composition, relief smoothness, vegetation, buildings and roads, and precipitation, and returns a value between 0 and 1 representing the diversity of feature combinations at a given area size. For each feature, it computes average values across areas of a chosen size, applies Shannon entropy to the distribution of those averages to capture locality-independent diversity, and aggregates the per-feature results using weights reflecting each feature's importance to the game. When multiple terrains are compared, only features present on all of them are measured.

The reported results come from four generated terrains of 2000x2000 height points, evaluated using average height as the sole feature. Diversity scores were plotted across area sizes, with small discontinuities appearing where the algorithm shifts between grid sizes, and the authors note that accuracy improves for larger terrains. The evaluation is limited to a single feature and four synthetic maps, so the paper demonstrates the metric's behavior more than its validation across diverse generators. The authors suggest the scale can guide terrain generator selection and help determine a minimum terrain size that still yields acceptable diversity.