GSCompleter: A Distillation-Free Plugin for Metric-Aware 3D Gaussian Splatting Completion in Seconds

arXiv cs.CV / 4/23/2026

📰 NewsTools & Practical UsageModels & Research

共有:

Key Points

The paper addresses a key limitation of 3D Gaussian Splatting (3DGS): performance drops sharply in sparse-view extrapolation, producing geometric holes and visual artifacts.
It proposes GSCompleter, a distillation-free plugin that replaces the unstable “Repair-then-Distill” workflow with a more stable “Generate-then-Register” pipeline.
GSCompleter first generates plausible metric-aware 2D reference images and lifts them into 3D using a robust Stereo-Anchor mechanism.
It then integrates the synthesized primitives into the global scene via a Ray-Constrained Registration strategy, enabling fast completion in seconds.
Experiments on three benchmarks show improved completion quality and efficiency, including new state-of-the-art results and benefits for multiple baselines.

Abstract

While 3D Gaussian Splatting (3DGS) has revolutionized real-time rendering, its performance degrades significantly under sparse-view extrapolation, manifesting as severe geometric voids and artifacts. Existing solutions primarily rely on an iterative "Repair-then-Distill" paradigm, which is inherently unstable and prone to overfitting. In this work, we propose GSCompleter, a distillation-free plugin that shifts scene completion to a stable "Generate-then-Register" workflow. Our approach first synthesizes plausible 2D reference images and explicitly lifts them into metric-scale 3D primitives via a robust Stereo-Anchor mechanism. These primitives are then seamlessly integrated into the global context through a novel Ray-Constrained Registration strategy. This shift to a rapid registration paradigm delivers superior 3DGS completion performance across three distinct benchmarks, enhancing the quality and efficiency of various baselines and achieving new SOTA results.