Question: How does Move differ from SceneConductor?

SceneConductor corrects poses by deterministic optimization without physics. In SceneRig, an agent chooses each correction, and every edit, whether a bounded Move or a direct code edit, is checked in simulation. Below, the agent switches to a code edit after physics rejects a Move.

Example: placing a laptop on its stand

Input photo
Input photo: the laptop rests on a black stand.
Before — laptop under the stand, too small
Before pose refinement: the laptop is under the stand and too small.
After — on the stand, settled in simulation
After the code edit: the laptop rests on the stand.
  1. Move (resize): rejected by physics. The enlarged laptop tips over in simulation (73° from upright), so the change is reverted.
  2. Code edit: the agent enlarges the laptop 1.28× and places it on the stand. The laptop is settled in simulation and shown to the agent, which keeps it: silhouette overlap with the photo rises from 0.54 to 0.64 IoU.
  3. Refinement: a Move on its position and a second code edit on its size, both settled in simulation, raise the overlap to 0.75 IoU. In the final whole-scene simulation the laptop moves only 0.2 mm.