Given only 2D inputs, our method reconstructs coherent 3D with consistent global structure across diverse shapes. The results demonstrate strong shape completeness and smooth surfaces, even under limited input cues.

We evaluate on 3D and 2D inputs, measuring both whole-asset and part-level performance.

• 3D input: Competitive on whole assets, but weaker on part-level metrics compared to clustering-based methods.
• 2D input: Best performance on whole assets, but noticeable drop on part-level quality.

Takeaway: Our method excels at global shape reconstruction, while local part consistency remains the main limitation.