Deterministic procedural generation of mesh detail through gradient tiling

Modern virtual reality systems support scenery of both vast scale and intricate detail at levels beyond that which can be manually created by teams of human content creators. Scene geometry must instead be created through procedural geometry generation. Deterministic tiling strategies that synthesize content over unbounded regions are extended in this paper to resolve issues with structure at tile boundaries, and also achieve required level of detail. The process is demonstrated through its application to the real-time terrain synthesis domain. The synthesis process described retains properties of the content domain by directly placing tiles sampled from exemplar data. Inconsistencies at tile boundaries are resolved through a novel refinement involving the use of a gradient field representation for the tiles. The synthesis algorithm lays down multiple layers of tiles containing gradient components and the geometric structure is reconstructed from this gradient field. The reconstruction uses an effcient formulation of a run-time Poisson solver well suited to parallelization on the GPU. This approach reduces seams, recombines exemplar features to increase variety, and supports deterministic generation of localized regions on demand. The process is well suited to progressive refinement schemes with large tiles providing rapid detail amplification. This is validated in the case of terrain data by showing that this process is more effective at preserving drainage properties than other terrain synthesis approaches. Performance exceeds typical requirements, with the GPU based prototype capable of completely regenerating the entire visible region at interactive frame rates.