GAD Editorial · Published · · English study · Chinese summary
An existing interior becomes easier to discuss when its walls and openings are represented as editable relationships. The difficult question is how much confidence to place in those relationships when they have been inferred from a capture. For an adaptive-reuse team, a missing opening or an incorrectly connected wall can matter more than a visually convincing room.
SceneScript, published as research in 2024, represents reconstructed scenes through a sequence of structured commands. Its vocabulary includes commands for walls, doors, windows and bounding boxes. This gives a reviewer a different kind of evidence from a rendered view: the proposed components and their parameters can be inspected. Research paper
The release should be read precisely. The public inference models accept Project Aria semi-dense point clouds, and the repository specifies non-commercial terms. The released implementation requires CUDA and reports Linux-only testing. Official repository
The paper describes five-centimetre geometric tokenization and acknowledges difficulty reaching millimetre precision. These are research resources, not evidence of a survey-grade, one-click BIM conversion service. Precision limitations
Decide what the reconstruction may answer
GAD proposes using a structured reconstruction as a room hypothesis. Before any appropriately licensed evaluation, define the question: can the capture help establish a provisional room arrangement, identify uncertain openings or organise a follow-up survey? Do not begin by treating every generated dimension as an existing-condition measurement.
Create a capture coverage plan tied to named rooms. Record inaccessible areas and surfaces hidden by furniture, linings or equipment. Keep those gaps attached to the interpretation rather than filling them with apparently complete geometry. An unobserved boundary should remain a question even when the output supplies a plausible wall.
Check the spatial relationships first
Review each room with the capture beside the structured output. Inspect whether walls meet where expected, openings belong to the correct wall and adjacent spaces connect consistently. A door-sized rectangle is not enough: the relationship between that opening, its host and the space beyond it must also make sense.
Prepare a disagreement list rather than immediately editing away all differences. Distinguish capture failure, an uncertain model interpretation and a condition that genuinely needs site investigation. This keeps a repairable software result from becoming an undocumented claim about the building.
Establish a separate dimensional record
Choose control dimensions according to the intended design use and have them checked through an appropriate survey process. Document their source, datum and tolerances. Compare critical clearances and opening locations directly; a room that looks plausible overall may still be unsuitable for setting out work.
If the team later rebuilds accepted components in its authoring model, record which properties come from verified measurements and which remain inferred. Check levels, coordinates and object relationships after transfer. No native Revit or IFC export capability is assumed here.
A useful handover contains the capture reference, the proposed room structure, checked control information and an unresolved-condition list. Keep the original hypothesis so later discoveries can be traced. The outcome is a better targeted verification exercise, with research output helping the team decide what to inspect rather than silently becoming the survey record.