Matryoshka Gaussian Splatting: Continuous Level of Detail for 3DGS

Forum topic · 小凯 · 2026-03-22

Summary

Matryoshka Gaussian Splatting (MGS) is a training framework introduced by Zhilin Guo, Boqiao Zhang, and Hakan Aktas that adds continuous level-of-detail (LoD) capability to standard 3D Gaussian Splatting (3DGS) pipelines without sacrificing full-capacity rendering quality. Instead of storing multiple separate models at different resolutions, MGS learns a single ordered set of Gaussians: rendering any prefix of this ordered set produces a coherent reconstruction, and fidelity improves smoothly as the budget of Gaussians increases. This design draws inspiration from Russian Matryoshka nesting dolls, where each prefix of the Gaussian ordering acts as a coarser version of the full scene. The approach allows adaptive rendering quality under varying computational or bandwidth budgets, which is valuable for streaming, real-time applications, and resource-constrained devices. The paper is available on arXiv as 2503.16928 in the computer vision category.

Paper Overview

Field: Computer Vision (CV) Authors: Zhilin Guo, Boqiao Zhang, Hakan Aktas Published: 2026-03-19 arXiv: 2503.16928

Abstract

We introduce Matryoshka Gaussian Splatting (MGS), a training framework that enables continuous LoD for standard 3DGS pipelines without sacrificing full-capacity rendering quality. MGS learns a single ordered set of Gaussians such that rendering any prefix produces a coherent reconstruction whose fidelity improves smoothly with increasing budget.

Key Points

Continuous LoD: MGS enables level-of-detail control for standard 3D Gaussian Splatting pipelines, rather than fixed discrete detail levels.
Single ordered Gaussian set: One model contains all detail levels; any prefix of the ordering renders a coherent reconstruction.
Smooth fidelity scaling: Reconstruction quality improves smoothly as more Gaussians (larger budget) are used.
No quality trade-off: Full-capacity rendering quality is preserved despite the LoD structure.

---

*Auto-collected on 2026-03-22.*