PVSR: Single-Image 3D Tabletop Reconstruction with Physical Stability and Semantic Preservation

Reconstructing a 3D tabletop scene from a single image is challenging because errors in object assets and pose estimates can lead to misplaced support, floating objects, and collisions. Object-level spatial relations describe how objects are related, but do not identify the local geometric regions on complex objects that realize those relations, leaving pose refinement without explicit geometric targets. We present PVSR (Part-level Visual-grounded Spatial Relations), a framework for part-level relation grounding and solving in single-image tabletop reconstruction. PVSR first constructs an object-level SceneTree from the reference image and extracts addressable local geometric regions from the reconstructed assets. It then grounds each relation edge to the geometric regions on the child and parent objects that participate in the relation, and uses them to formulate relation-conditioned layout constraints for refining the initial poses. We further introduce a relation-aware physical optimization module that uses grounded part-level relations to guide collision removal and improve scene stability. We evaluate PVSR on the TabletopGen test set and a curated benchmark we construct for part-level spatial relation modeling in single-image tabletop reconstruction, covering both end-to-end reconstruction and controlled relation evaluation. Results show that PVSR produces more accurate relational layouts in structurally complex scenes and improves physical stability in simulation while maintaining visual consistency with the reference image.