GoDe: Gaussians on Demand for Progressive Level of Detail and Scalable Compression

TL;DR

This work tackles real-time novel view synthesis with 3D Gaussian Splatting by introducing GoDe, a hierarchical and scalable framework that enables progressive LoD and multi-rate compression without retraining. It builds a base Gaussian layer and enhancement layers through gradient-informed masking, paired with quantization-aware fine-tuning and per-level compression to navigate the rate-distortion curve. GoDe delivers smooth LoD transitions via opacity interpolation and supports adaptive rendering under varying budgets, while maintaining competitive RD performance across neural-based and explicit 3DGS baselines. Empirical results demonstrate substantial storage reductions and notable FPS gains, highlighting GoDe's practicality for bandwidth- and compute-constrained scenarios.

Abstract

3D Gaussian Splatting enhances real-time performance in novel view synthesis by representing scenes with mixtures of Gaussians and utilizing differentiable rasterization. However, it typically requires large storage capacity and high VRAM, demanding the design of effective pruning and compression techniques. Existing methods, while effective in some scenarios, struggle with scalability and fail to adapt models based on critical factors such as computing capabilities or bandwidth, requiring to re-train the model under different configurations. In this work, we propose a novel, model-agnostic technique that organizes Gaussians into several hierarchical layers, enabling progressive Level of Detail (LoD) strategy. This method, combined with recent approach of compression of 3DGS, allows a single model to instantly scale across several compression ratios, with minimal to none impact to quality compared to a single non-scalable model and without requiring re-training. We validate our approach on typical datasets and benchmarks, showcasing low distortion and substantial gains in terms of scalability and adaptability.

Figures (12)

• Figure 1: Our method builds a progressive Gaussian hierarchy, enabling adaptive LoD. Combined with compression, it supports multiple rates (eight in the figure), from high to low compression. The model can dynamically adjust based on criteria such as a target minimum FPS.
• Figure 2: Graphical overview of our method GoDe. 1) Starting from a pre-trained model, we build a progressive hierarchy of Gaussians through iterative gradient-informed masking. 2) A quantization-aware fine tuning stage enhances the model's quality and reduces its memory footprint. 3) The model is further compressed through entropy coding.
• Figure 3: Rate/distortion (above) and PSNR/FPS (below) curves on the three main datasets. Our approach achieves competitive results with both neural-based Gaussian models and explicit ones while maintaining significantly higher speed at comparable quality. It is also the only method that achieves scalability without requiring ad-hoc training.
• Figure 4: Visualization of LoDs. For visual purposes, we show the first three LoD levels and the last one. For each LoD we provide the average FPS, PSNR, primitives count W (in thousands), and size. Starting from top to bottom we have: kitchen, counter, drjohnson and truck scenes.
• Figure 5: Continuous transitions between two adjacent levels, with eight linear interpolating factors on Garden scene.