Virtualized 3D Gaussians (V3DG)

Flexible Cluster-based Level-of-Detail System for Real-Time Rendering of Composed Scenes

(SIGGRAPH 2025 Conference Paper)

Xijie Yang^1,2, Linning Xu³, Lihan Jiang^4,2, Dahua Lin^3,2, Bo Dai^5,6

¹ Zhejiang University, ² Shanghai Artificial Intelligence Laboratory, ³ The Chinese University of Hong Kong, ⁴ University of Science and Technology of China, ⁵ The University of Hong Kong, ⁶ Feeling AI

V3DG achieves real-time rendering of massive 3D Gaussians in large composed scenes through a novel LOD approach. Inspired by Nanite in Unreal Engine 5, V3DG processes detailed 3D assets into clusters at various granularities offline, and selectively renders 3D Gaussians at runtime —flexibly balancing rendering speed and visual fidelity based on user-defined tolerances.

Framework

Our virtualized LOD system features the offline build stage to prepare clusters at various levels of detail, and the online selection stage to select clusters at appropriate levels of detail given rendering conditions.

Local Splatting for Simplification

To preserve visual quality while reducing the number of 3D Gaussians by half locally in each cluster group, we develop the local splatting method, leveraging the differentiable feature of 3D Gaussians. Randomly generated pseudo-views are positioned to face the center of the cluster group, used to iteratively optimize the properties of the newly-created 3D Gaussians. This process distills the appearance of original 3D Gaussians layer by layer.

Results

The more the simplification iteration, the similar the renders with vanilla 3DGS.

Users can adjust footprint tolerances to balance visual quality and rendering speed.

Our LOD method achieves better quality than baselines.

Text Citation

Xijie Yang, Linning Xu, Lihan Jiang, Dahua Lin, and Bo Dai. 2025. Virtualized 3D Gaussians: Flexible Cluster-based Level-of-Detail System for Real-Time Rendering of Composed Scenes. In Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers (SIGGRAPH Conference Papers ’25), August 10–14, 2025, Vancouver, BC, Canada. ACM, New York, NY, USA, 11 pages. https://doi.org/10.1145/3721238.3730602

BibTeX Citation

@inproceedings{Yang2025V3DG,
    author = {Yang, Xijie and Xu, Linning and Jiang, Lihan and Lin, Dahua and Dai, Bo},
    title = {Virtualized 3D Gaussians: Flexible Cluster-based Level-of-Detail System for Real-Time Rendering of Composed Scenes},
    booktitle = {ACM SIGGRAPH 2025 Conference Papers},
    year = {2025},
    doi = {10.1145/3721238.3730602},
    url = {https://doi.org/10.1145/3721238.3730602},
    publisher = {Association for Computing Machinery},
    series = {SIGGRAPH '25}
}