Constraint-Aware Terrain Generation

In a mixed authoring workflow, a designer fixes the parts of a game level that matter, a road, a landmark, a race-track corridor, and the rest of the terrain has to be filled in automatically. The hard part is doing it fast, keeping the fixed points exact, and following the designer's slope preferences even where they conflict.

This started as my undergraduate thesis at Charles University, where I built the terrain tool in Unity and proposed the constraint-propagation idea at its core. My supervisor, Petr Šimůnek, then extended the method into a full paper, ran the experiments, and led the writing, so the author order reflects that he did the larger share.

The method propagates upper and lower height envelopes from the fixed points using slope-weighted path costs, then fills the gaps with a smooth base layer and bounded detail, clamping to the envelope so anchors stay exact. Every method compared keeps the anchors exactly, so the question is what happens between them.

On the benchmark it reaches the lowest absolute normalised mean-height bias of the methods evaluated: 0.01346 against 0.07112 for the blurred production base, a reduction of about 81%, and it runs between 23.9 and 2.1 times faster than the harmonic baseline depending on resolution.

Three things that number does not say. Inverse distance weighting scores 0.01351 on the same metric, so the margin over the simplest sensible alternative is about 0.4%. Total variation is meaningfully worse, 933.5 against 480.9 for the harmonic baseline, meaning the surface is rougher. And the stochastic "bounded detail" component measured a diversity of exactly zero under the parameterisation used, so it contributed nothing. A reviewer raised this and it is fair. It is one authored benchmark scenario with no ablation study.

It was accepted as a full paper for oral presentation at IEEE Conference on Games 2026 in Madrid, where 89 of 239 submissions were accepted.