GSWT: Gaussian Splatting Wang Tiles (yunfan.zone)

A team of researchers has introduced Gaussian Splatting Wang Tiles (GSWT), a groundbreaking method that enhances 3D Gaussian Splatting (3DGS) for efficiently generating vast, photorealistic terrains. This tile-based framework leverages Wang Tiles, which encode local Gaussian fields with boundary constraints, facilitating seamless transitions and allowing for the procedural generation of expansive terrains. The significance of this development lies in its ability to synthesize large-scale or infinite environments from just a single captured exemplar, addressing a persistent challenge in the AI/ML community related to terrain modeling. The proposed pipeline begins with reconstructing 3DGS exemplars at multiple levels of detail (LODs) from multi-view images, followed by the generation of Wang Tiles through a semantic-aware graph cut algorithm. Key optimizations enable real-time rendering and on-the-fly tiling of the terrain, dramatically enhancing efficiency. By implementing efficient sort-free splatting and managing rendering workflows between threads, this method opens new avenues for procedural generation and rendering in computer graphics, potentially impacting fields such as video game development and virtual reality.