Generating part-aware 3D assets from a single image has become increasingly capable, but visual completeness does not guarantee that the resulting parts can actually be assembled. Existing methods often produce neighboring parts that interpenetrate, lack valid connections, or collapse under gravity, limiting their usefulness for editing, articulation, simulation, and fabrication.
To address this, researchers propose SNAP3D, a physics-guided framework that improves single-image part-aware 3D generation by enforcing physically compatible geometry and stable connections. The method resolves inter-part penetration, recovers a contact graph between neighboring parts, and introduces parameterized connectors at their contact surfaces.
Using feedback from physical simulation, SNAP3D refines connector placement, orientation, and dimensions to improve assembly stability while preserving the generated geometry. The authors also introduce a physics-based evaluation protocol that complements conventional geometric metrics by directly testing assembly validity and stability under gravity.
Experiments comparing SNAP3D against multiple part-aware 3D generators show substantial improvements in physical realizability and stability while maintaining geometric quality. The resulting parts were additionally validated through 3D printing and real-world assembly, demonstrating practical viability.