Improving geometry and back-view reconstruction in SAM3D

Hello Team,

Great work on SAM3D — results look strong when the visible views are available.

I have a simple question:
When the back view of an object is not visible, geometry quality drops (flattening or incorrect shape).

Is it possible to:

Improve results by providing additional images (e.g., a back-view image) at inference time?

Or fine-tune SAM3D to better handle multi-view inputs for improved geometry and shape completion?