From an input mesh used as conditioning, the model predicts a 3 vector displacement field to the recto surface of the winding in front of and behind the conditioning. Model was trained at 4.8um resolution, but performs ok at 2.4um and 9.6um (and likely resolutions in between)