CLoD-GS: Continuous Level-of-Detail via 3D Gaussian Splatting

TL;DR

This work tackles the problem of achieving scalable, artifact-free level-of-detail in real-time rendering with 3D Gaussian Splatting. It introduces CLoD-GS, a single unified model that equips each Gaussian with a learnable distance-decay parameter $\sigma_{d,i}$ and uses a virtual distance scale $s_v$ alongside a coarse-to-fine training regime with a rendered-primitives regularization to learn a continuous spectrum of detail. By attenuating per-Gaussian opacity instead of removing primitives, the method generates smooth, pop-free transitions and reduces the final primitive count and memory footprint while delivering state-of-the-art fidelity across diverse real-world datasets. The approach is extensible to pre-trained compressed models (e.g., MaskGaussian) and carries practical implications for scalable, high-quality neural rendering in interactive applications.

Abstract

Level of Detail (LoD) is a fundamental technique in real-time computer graphics for managing the rendering costs of complex scenes while preserving visual fidelity. Traditionally, LoD is implemented using discrete levels (DLoD), where multiple, distinct versions of a model are swapped out at different distances. This long-standing paradigm, however, suffers from two major drawbacks: it requires significant storage for multiple model copies and causes jarring visual ``popping" artifacts during transitions, degrading the user experience. We argue that the explicit, primitive-based nature of the emerging 3D Gaussian Splatting (3DGS) technique enables a more ideal paradigm: Continuous LoD (CLoD). A CLoD approach facilitates smooth, seamless quality scaling within a single, unified model, thereby circumventing the core problems of DLOD. To this end, we introduce CLoD-GS, a framework that integrates a continuous LoD mechanism directly into a 3DGS representation. Our method introduces a learnable, distance-dependent decay parameter for each Gaussian primitive, which dynamically adjusts its opacity based on viewpoint proximity. This allows for the progressive and smooth filtering of less significant primitives, effectively creating a continuous spectrum of detail within one model. To train this model to be robust across all distances, we introduce a virtual distance scaling mechanism and a novel coarse-to-fine training strategy with rendered point count regularization. Our approach not only eliminates the storage overhead and visual artifacts of discrete methods but also reduces the primitive count and memory footprint of the final model. Extensive experiments demonstrate that CLoD-GS achieves smooth, quality-scalable rendering from a single model, delivering high-fidelity results across a wide range of performance targets.

Figures (23)

• Figure 1: Framework of the proposed methodology.
• Figure 2: Visual comparison at similar primitive counts. The number of Gaussians used and the corresponding PSNR are annotated in the bottom-right corner of each image. Our method preserves details and avoids the artifacts present in other methods, especially under complex lighting and texture conditions. Note that "scale = 1" and "scale = 3" denote a maximum virtual size $s_v$ of 1 and 3, respectively, during training; the same notation is used consistently throughout the remaining figures and captions.
• Figure 3: Quality vs. primitive count on the BungeeNeRF and Deep Blending datasets. The dashed lines indicate the maximum virtual scale used during training for each model, with the minimum scale always starting at 1. Our method (CLoD-GS) with varying virtual distance scale ranges ($s_v$) shows a more graceful quality degradation compared to baselines. Increasing the scale range for $s_v$ improves performance at lower detail levels without compromising maximum quality.
• Figure 4: Visual comparison of DLoD vs. CLoD strategies. The DLoD approach (the second column) uses two separate models, causing a visible quality jump at the boundary (red dashed line). Our CLoD approach (the left three columns) uses a single model with varying scale factors, resulting in a smooth, artifact-free transition.
• Figure 5: Metric curves for the DLoD vs. CLoD comparison. The plot shows the quality metrics across the three transition regions of the image. The DLoD strategy exhibits a sharp, discontinuous jump in quality at the model boundary, whereas our CLoD strategy shows a smooth and continuous progression.