GSWT

GSWT: Gaussian Splatting Wang Tiles

SIGGRAPH Asia 2025, Hong Kong

Yunfan Zeng1,

Li Ma2,

Pedro V. Sander1

1The Hong Kong University of Science and Technology, Hong Kong SAR, China
2Eyeline Labs, USA

Paper (TBD)

Supplementary (TBD)

Code

Dataset

Demo

Abstract

3D Gaussian Splatting (3DGS) has shown strong capability in reconstructing and rendering photorealistic
3D scenes with high efficiency.
However, extending 3DGS to synthesize large-scale or infinite terrains from a single captured
exemplar—remains an open challenge.
In this paper, we propose a tile-based framework that addresses this problem. Our method builds on Wang
Tiles, where each tile encodes a local field of Gaussians with boundary constraints to ensure seamless
transitions. This enables stochastic yet continuous tiling of Gaussian fields over arbitrary surfaces,
allowing for procedural generation of expansive terrains with high spatial diversity. Furthermore, we
introduce several rendering optimizations tailored to the unique characteristics of 3DGS Wang tiles,
achieving real-time rendering of large-scale 3DGS terrains.

Pipeline

Given multi-view images of an exemplar scene, our goal is to construct Gaussian Splatting Wang Tiles
(GSWT) that can be tiled on arbitrary surfaces and rendered in real time with our novel GSWT
renderer.
An overview of the entire pipeline is illustrated below. We begin by
reconstructing the 3DGS exemplar at multiple LODs. For each level, we generate a set of Wang Tiles by
sampling the edge and center patches and applying a semantic-aware graph cut algorithm. Prior to
rendering, we pre-sort each tile for efficient sort-free splatting, and during runtime, we perform
tiling on the fly, allowing efficient GSWT-based terrain synthesis and rendering.

(a) Given the input images, we construct the exemplar multiple times with different Level of Detail
(LOD).
(b) We construct the tile set and preprocess it before rendering.
(c) The surface is tiled at run-time on the worker thread, while the main thread renders each frame.

Full Demo
TBD

BibTeX

Copy

@inproceedings{Zeng:2025:gswt,
author = {Zeng, Yunfan and Ma, Li and Sander, Pedro V.},
title = {GSWT: Gaussian Splatting Wang Tiles},
year = {2025},
publisher = {Association for Computing Machinery},
booktitle = {SIGGRAPH Asia 2025 Conference Papers},
location = {Hong Kong, China},
series = {SA '25}
}

This page was built using the Academic Project Page Template which was adopted from the Nerfies project page.
This website is licensed under a Creative
Commons Attribution-ShareAlike 4.0 International License.

The presented work, authored by Yunfan Zeng, Li Ma, and Pedro V. Sander, details a novel approach to procedural terrain generation utilizing Gaussian Splatting (3DGS) and a Wang Tile framework, termed GSWT (Gaussian Splatting Wang Tiles). The core challenge addressed is the efficient synthesis of large-scale, potentially infinite terrains from a single exemplar scene, a limitation of standard 3DGS implementations. The research proposes a tile-based system designed to overcome this constraint, enabling dynamic and spatially diverse terrain creation.

The methodological foundation of GSWT rests on the concept of Wang Tiles – discrete, geometrically defined units that can be arranged to tile a surface without gaps or overlaps. Each Wang Tile, within the GSWT framework, encodes a localized Gaussian field, incorporating boundary constraints to maintain seamless transitions between adjacent tiles. This stochastic yet continuous tiling approach facilitates the procedural generation of expansive terrains, characterized by inherent spatial variety.

The proposed pipeline begins with the reconstruction of the exemplar scene at multiple levels of detail (LODs). This multi-LOD approach is crucial for optimizing both the initial reconstruction and subsequent rendering processes, allowing for a trade-off between accuracy and computational cost. The system then generates a tile set, a collection of Wang Tiles, through a process that integrates a semantic-aware graph cut algorithm. This algorithm is triggered by sampling both the edge and center patches of the exemplar, suggesting an intelligent method for defining the boundaries and shapes of the tiles. Following this initial generation, a preprocessing step is undertaken, specifically sorting the tiles prior to runtime. This sort-free splatting technique, a key optimization, is implemented to dramatically improve rendering efficiency. At runtime, the tiles are applied to the target surface, achieved through parallel execution on a worker thread, while the main thread performs the actual frame rendering. This parallel architecture is central to the real-time performance of the system.

The system's design directly leverages the strengths of 3DGS while mitigating its limitations concerning large-scale terrain synthesis. The Wang Tile approach ensures continuity and seamless transitions, which are often problematic when applying standard 3DGS directly to vast areas. The inclusion of semantic-aware graph cuts indicates a deliberate attempt to imbue the generated terrain with meaningful structural characteristics, rather than purely random variations. Furthermore, the multi-LOD reconstruction and sort-free rendering contribute to the system's capacity for real-time performance. The integration of parallel threads—worker threads for tiling and the main thread for rendering—represents a key architectural element for optimizing execution speed, crucial for the real-time synthetic terrain generation intended by the research.

The referenced paper (Zeng et al., 2025) builds upon the established foundation of 3DGS, repackaging it within a fundamentally different structural approach. The use of a formal citation (Zeng et al., 2025) highlights the research as a contribution within the broader field and provides a readily available reference for further study. The detailed information provided regarding the system design, including the multi-LOD reconstruction, the sorting of tiles and the utilization of parallel threads, suggests a sophisticated and rigorously engineered solution to the problem of large-scale procedural terrain generation. The system’s overall architecture is specifically tailored to exploit the performance characteristics of 3DGS, aiming to achieve both visual fidelity and real-time rendering capabilities.