Learned seed-consistent base generator to replace the procedural prior at the hierarchy’s top level

Determine whether a seed-consistent learned generator (for example, InfinityGAN or another padding-free GAN) can effectively replace Perlin noise as the top-level prior in the Terrain Diffusion hierarchical pipeline for domains in which a procedural prior is not available, while preserving global coherence and user controllability of the generated terrain.

Background

Terrain Diffusion currently relies on a hierarchical stack where the top level provides coarse global structure that conditions lower-resolution and high-resolution stages. In the presented system, simple Perlin noise serves as this top-level prior because continental-scale structure is coarse and controllable, and it preserves seed consistency.

The authors note that in domains lacking such a procedural prior, it is unclear what learned alternative should provide the top-level conditioning. They suggest that seed-consistent padding-free GANs, such as InfinityGAN, could be candidates, but their own experiment with a small padding-free GAN yielded results comparable to Perlin noise with reduced controllability, leaving the choice of an effective learned replacement unresolved.