SpatialGS: Relative Spatial Regularization in Gaussian Splatting for Layout-guided text-to-3D generation

Abstract

Layout-guided text-to-3D generation has significantly advanced the controllable synthesis of multi-object scenes. However, existing methods still struggle to maintain consistent and precise spatial relationships, frequently exhibiting inaccurate relative positions and ambiguous object orientations. To bridge this gap, we introduce SpatialGS, a novel, plug-and-play framework that seamlessly incorporates relative spatial regularization into Gaussian Splatting. Specifically, we propose two core modules: Relative Position Regularization (RPR) and Relative Orientation Regularization (ROR), which explicitly enforce accurate inter-object spatial relationships during optimization. To further stabilize this generative process, we design an Alternating Global-to-Local Sampling Strategy (AGS) that harmonizes global scene context with local object details. Extensive experiments demonstrate that SpatialGS can be easily integrated into existing Score Distillation Sampling (SDS)-based pipelines, significantly outperforming state-of-the-art baselines in both spatial consistency and overall compositional quality to yield physically plausible 3D scenes.