StructBridge: Structure-Grounded 3D Indoor Object Generation via 3D Latent Diffusion Bridge and Normal Refinement

Generating 3D indoor objects under fine-grained structural control is crucial and has broad applications in robotics, gaming, and simulation. Such fine-grained structural control encompasses an object’s spatial layout, geometric icons, and relative size, which together constitute the object’s core structural properties. However, existing controllable 3D generation methods do not take structural control as a primary objective, and they commonly rely on suboptimal conditions in representing structure or inadequate algorithms for maintaining structural consistency. Therefore, they fail to achieve fine-grained structural control during generation. To address this, we present structure-grounded 3D object generation, a novel perspective on 3D object generation, which decomposes 3D shapes into structural configurations and geometric details, directly conditions on the object’s structural configurations and synthesizes realistic geometric details to produce high-quality 3D objects. Building on this perspective, we introduce StructBridge, which first employs a diffusion bridge in the 3D latent space to transform comprehensive structural conditions into structurally accurate coarse meshes, followed by a structure-preserving normal refiner that enriches geometric details while preserving structure. We conduct extensive experiments and comprehensive comparisons with various conditional generation methods, demonstrating that StructBridge achieves state-of-the-art performance in both generation quality and structural control capabilities.