Test‑time geometry constraints tighten vision model predictions

Vision foundation models can predict depth, pose, and point clouds in a single forward pass, yet they...