Unfolding 3DGS arxiv (opens in a new tab)
IEEE Transactions on Visualization and Computer Graphics

Unfolding 3D Gaussian Splatting via Iterative Gaussian Synopsis

Yuqin Lu1,2*Yihua Dai1,2*Yang Zhou1,2Guiqing Li1Shengfeng He2
1 South China University of Technology2 Singapore Management University

* Equal contribution

arxiv (opens in a new tab) Code Coming soon

Progressive level-of-detail representation

Truck · Tanks & Temples
3D Gaussian structureFull-detail 3D Gaussian structure of a blue truckTHE REPRESENTATION
Rendered viewFull-detail rendered view of the Truck sceneL3 / Full detail

Showing Truck · L3 · 50.7 MB

LEVEL OF DETAILCoarsest Finest

LOD visualization. Gaussian structure and rendered views at four LODs. The model is condensed from L3 to L0 during training; progressive rendering proceeds from L0 to L3.

Pre-rendered results at each level’s target resolution.

Abstract

3D Gaussian Splatting (3DGS) has become a state-of-the-art framework for real-time, high-fidelity novel view synthesis. However, its substantial storage requirements and inherently unstructured representation pose challenges for deployment in streaming and resource-constrained environments. Existing Level-of-Detail (LOD) strategies, particularly those based on bottom-up construction, often introduce redundancy or lead to fidelity degradation.

To overcome these limitations, we propose Iterative Gaussian Synopsis, a novel framework for compact and progressive rendering through a top-down “unfolding” scheme. Our approach begins with a full-resolution 3DGS model and iteratively derives coarser LODs using an adaptive, learnable mask-based pruning mechanism. This process constructs a multi-level hierarchy that preserves visual quality while improving efficiency.

We integrate hierarchical spatial grids, which capture the global scene structure, with a shared Anchor Codebook that models localized details. This combination produces a compact yet expressive feature representation, designed to minimize redundancy and support efficient, level-specific adaptation. The unfolding mechanism promotes inter-layer reusability and requires only minimal data overhead for progressive refinement.

Experiments show that our method maintains high rendering quality across all LODs while achieving substantial storage reduction. These results demonstrate the practicality and scalability of our approach for real-time 3DGS rendering in bandwidth- and memory-constrained scenarios.

Method

Iterative Gaussian Synopsis constructs a four-level hierarchy through top-down, mask-guided pruning of a full-fidelity anchor-based representation.

Iterative Gaussian Synopsis framework: hierarchical grids and an Anchor Codebook feed MLPs; learnable masks prune Gaussians to form coarser levels, which support progressive rendering
Framework overview. Mask-guided pruning derives coarser representations from a full-fidelity model; the resulting hierarchy supports progressive rendering.
01

Hierarchical representation

Multi-resolution hash grids encode global scene context, while a shared Anchor Codebook represents local features. Compact MLPs decode these features into Gaussian attributes.

02

Top-down Gaussian synopsis

Starting from the finest level, learnable masks and group sparsity regularization prune Gaussians and entire anchors. Coarser levels retain a subset of the hierarchical grids.

03

Level-specific adaptation

Coherent Basis Modulation adapts the shared codebook, and Level-aware Decoding fine-tunes the fusion and attribute-prediction MLPs for each LOD.

Hierarchical architecture
Hierarchical architecture combining multi-scale grid features with local features from a shared Anchor Codebook to predict Gaussian attributes
Hierarchical architecture. Multi-scale hierarchical grids and a shared Anchor Codebook encode global context and local features for Gaussian attribute prediction.

Quantitative Results

Comparison with 3DGS, Octree-GS, and LapisGS across four datasets and four levels of detail.

Quantitative comparison on Mip-NeRF360 at L3.
MethodPSNR ↑dBSSIM ↑LPIPS ↓Speed ↑FPSSize ↓MB
3DGS26.860.7930.226107795.26
Octree-GS27.100.7910.230114103.61
LapisGS26.190.7480.278701740.10
Ours Iterative Gaussian Synopsis27.170.7900.22411890.11

Model size ratio (3DGS / Ours): 8.8×

Mip-NeRF360 · L3 · 90.11 MB · 118 FPS

Evaluation uses held-out test views. All methods are evaluated at each level’s proportionally scaled target resolution; quality metrics should be compared within the same level. 3DGS and LapisGS sizes use their uncompressed explicit Gaussian representations. Baselines are reproduced under the paper’s evaluation settings; Octree-GS uses a fixed four-level configuration. See the paper for full protocols and incremental payloads.

Qualitative Results

Qualitative comparisons on indoor and large-scale scenes. Each row compares the same level of detail across methods, with close-up regions and per-scene quality and storage figures.

Qualitative comparisons of 3DGS, Octree-GS, LapisGS, and our method across L0 to L3, showing an indoor staircase and a city scene

Citation

If you use this work in your research, please cite the paper.

Download BibTeX
BIBTEX
@article{lu2026unfolding,
  title={Unfolding 3D Gaussian Splatting via Iterative Gaussian Synopsis},
  author={Lu, Yuqin and Dai, Yihua and Zhou, Yang and Li, Guiqing and He, Shengfeng},
  journal={arXiv preprint arXiv:2604.11685},
  year={2026}
}