Figure 1: Feed-forward 3D foundation models fail on transparent objects in two opposite directions.
(a) L1 depth error for a transparent object under zero-shot prediction, LoRA, and Ours. The zero-
shot prediction is dominated by overestimation associated with background leakage. LoRA reduces
the error, but underestimation around the concave interior remains, whereas our method corrects both
error directions. (b) The two error directions differ in reference availability. Background leakage
can be related to an observable background surface, whereas a filled-in concavity has no analogous
geometric reference for recovering the missing depth.