Bridging geometry-coherent text-to-3D generation with multiview diffusion priors and Gaussian Splatting

Score distillation sampling leverages pretrained two-dimensional diffusion models to advance text-to-three-dimensional (3D) generation; however, it neglects multiview correlations, which leads to geometric inconsistencies and multiface artifacts. Herein, we propose coupled score distillation (CSD), a framework that couples multiview joint distribution priors to ensure geometrically consistent 3D generation while enabling the stable and direct optimization of 3D Gaussian Splatting (3D-GS). Specifically, we reformulate optimization as a multiview joint optimization problem and derive a gradient-based update rule that effectively couples multiview priors to guide optimization across different viewpoints while preserving the diversity. We further propose a pipeline that directly optimizes 3D-GS from random initialization and refines a deformable tetrahedral grid initialized from 3D-GS to generate consistent and high-quality 3D content. Extensive quantitative and qualitative experiments demonstrate that CSD achieves superior geometric and semantic consistency, stronger optimization robustness, and better diversity. Code is available at https://github.com/FengY3337/Coupled-Score-Distillation.