HLoD: Screen‐Space Hierarchical Level‐of‐Detail Traversal for 3D Gaussian Splatting
Abstract While 3D Gaussian Splatting (3DGS) provides high‐quality novel‐view synthesis, its lack of an inherent hierarchy forces rasterization effort to scale with fine scene details rather than perceptually visible features. Existing Level of Detail (LoD) approaches typically address this by training discrete levels as independent models, relying on fixed selection thresholds that do not ensure smooth perceptual transitions. We introduce Hierarchical Level of Detail (HLoD) for 3DGS, a method that jointly optimizes Gaussian levels into a single, explicit parent‐child hierarchy. To maintain a valid tree topology during adaptive density control, HLoD performs hierarchy‐aware pruning and parent‐child regularization, combined with scale safeguards to ensure spatial consistency across frequency bands. Unlike prior methods, HLoD avoids fixed size thresholds per level, preserving the ability to assign point density based on local geometric detail. At runtime, an efficient GPU traversal dynamically collapses fine sibling groups into their trained coarse parents based on local pixel‐projection thresholds. Crucially, by integrating this traversal directly into the optimization process alongside strategies to balance photometric supervision between fine details and coarse nodes, HLoD maintains perceptually gradual LoD transitions that remain stable under inhomogeneous point distributions. Our evaluation demonstrates that HLoD achieves competitive reconstruction quality while enabling continuous control over the performance‐quality trade‐off along two axes: dynamically adjusting pixel‐projection thresholds for spatially local LoD, and selectively rendering hierarchical subsets for adaptive memory scaling. Achieving these seamless transitions requires minimal traversal overhead while allowing overall performance to scale dynamically with scene complexity. Finally, we establish that joint optimization is structurally essential, as post‐hoc hierarchical assignments fundamentally fail to preserve visual fidelity.
Authors
- Jan U. Müller (ORCID: https://orcid.org/0000-0002-9101-1033)
- Reinhard Klein (ORCID: https://orcid.org/0000-0002-5505-9347)
- Jakob Sommer (ORCID: https://orcid.org/0009-0008-4958-9107)
Institutions
- University of Bonn (DE)
- Lamarr Institute for Machine Learning and Artificial Intelligence (DE)
Publication Details
- Journal
- Computer Graphics Forum
- Published
- 2026-11-13
- DOI
- https://doi.org/10.1111/cgf.70636
- Primary Topic
- 3D Shape Modeling and Analysis
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- Land Nordrhein-Westfalen
- Bundesministerium für Bildung und Forschung