Hybrid Gaussian Wang Tiles for Class-aware Authoring and Rendering

We present a framework for synthesizing large, richly detailed scenes from multiple Gaussian Splatting exemplars using a class-aware Wang tiling formulation, building on Gaussian Splatting Wang Tiles as the underlying representation. The method introduces a multi-class representation that supports both hard class regions and hybrid tiles that mix Gaussians from different classes at fine spatial scales, enabling smooth and organic transitions between classes. Hybrid tiles are generated at runtime by a priority-based selection scheme that selects and blends tile contents based on class importance and spatial context. To preserve real-time performance in large multi-class scenes, we introduce a CPU–GPU hybrid sorting and caching strategy that balances workload and improves frame stability. The system also includes a new interactive authoring tool for real-time editing of scene structure and class distributions. Finally, we demonstrate that the approach can be applied to arbitrary meshes with polycube mapping, turning the tiled Gaussian fields into high-frequency 3D textures with controllable, artist-directed variation, providing rich visuals and great flexibility.