Duplex-GS: Proxy-Guided Weighted Blending for Real-Time Order-Independent Gaussian Splatting

TL;DR

D Duplex-GS is proposed, a dual-hierarchy framework that integrates proxy Gaussian representations with order-independent rendering techniques to achieve photorealistic results while sustaining real-time performance and develops a physically inspired weighted sum rendering technique that simultaneously eliminates "popping" and "transparency"artifacts, yielding substantial improvements in both accuracy and efficiency.

Abstract

Recent advances in 3D Gaussian Splatting (3DGS) have demonstrated remarkable rendering fidelity and efficiency. However, these methods still rely on computationally expensive sequential alpha-blending operations, resulting in significant overhead, particularly on resource-constrained platforms. In this paper, we propose Duplex-GS, a dual-hierarchy framework that integrates proxy Gaussian representations with order-independent rendering techniques to achieve photorealistic results while sustaining real-time performance. To mitigate the overhead caused by view-adaptive radix sort, we introduce cell proxies for local Gaussians management and propose cell search rasterization for further acceleration. By seamlessly combining our framework with Order-Independent Transparency (OIT), we develop a physically inspired weighted sum rendering technique that simultaneously eliminates "popping" and "transparency" artifacts, yielding substantial improvements in both accuracy and efficiency. Extensive experiments on a variety of real-world datasets demonstrate the robustness of our method across diverse scenarios, including multi-scale training views and large-scale environments. Our results validate the advantages of the OIT rendering paradigm in Gaussian Splatting, achieving high-quality rendering with an impressive 1.5 to 4 speedup over existing OIT based Gaussian Splatting approaches and 52.2% to 86.9% reduction of the radix sort overhead without quality degradation.

Figures (13)

• Figure 1: Motivation of Duplex-GS. (Top) 3DGSs are rendered via $\alpha$-blending for a long tile-Gaussian sequence. However, sudden changes in viewing angle can cause abrupt changes in the blending order and the corresponding weights, resulting in noticeable popping artifacts. (Bottom) The sort-free LC-WSR method eliminates the sorting stage through WSR paradigm. However, the absence of physical constraints leads to transparency artifacts and efficiency degradation, as blending continues unnecessarily for opaque surfaces.
• Figure 2: Overview of the proposed Duplex-GS. Our approach is built on a dual-hierarchy structure that introduces ellipsoidal cells as proxies for local Gaussians. (a) Explicit Cell Proxies. Unlike abstract anchor points, our geometrically defined proxy cells are directly utilized for rasterization, guiding the subsequent Gaussian blending process. (b) Hybrid Rendering Paradigm. By synergistically combining $\alpha$-blending with physically grounded WSR, our method simultaneously eliminates both popping and transparency artifacts. (c) The integration of cell rasterization (Sec. \ref{['sec:layerI']}) and physically grounded WSR (Sec. \ref{['sec:layerII']}) achieves significant improvements over prior OIT-based Gaussian splatting methods in both rendering accuracy and computational efficiency.
• Figure 3: Illustration of the proposed Duplex-GS pipeline. (Layer I) Ellipsoidal cells act as proxies to spatially organize local Gaussians. This structure enables efficient cell-level rasterization, significantly reducing the overhead of view-adaptive global sorting. (Layer II) Intersected cells are decoded into Gaussians and blended with the proposed physically grounded WSR. The final color is computed by first determining cell-level weights via $\alpha$-blending, then aggregating the Gaussians inside each cell in an order-independent WSR manner.
• Figure 4: Comparison between rasterization of GS and cell. Unlike the original $\alpha$-blending which processes each Gaussian individually, our method performs rasterization and early termination at the coarser cell level, where each cell encapsulates multiple Gaussians. The order confusion caused by overlap among cells is discussed in Sec. \ref{['sec:V-A']}.
• Figure 5: Comparison of three different transmittance: LC-WSR (Eq. \ref{['eq:LC-WSR']}), vanilla $\alpha$-blending (Eq. \ref{['eq:transmittance']}) and ours (Eq. \ref{['eq:dulpex_T']}). By reintroducing physical constraints, our model produces a transmittance curve similar to that of $\alpha$-blending, establishing it as a reliable criterion for early termination. Furthermore, the proposed transmittance is not strictly monotonic with depth, confirming the local order-independent nature of Gaussian blending in our framework.