Scan a building and you get a point cloud. Convert that point cloud and you usually get a mesh. A mesh is a shape. It is not a wall, it is not load bearing, it is not made of anything, and you cannot open it in Revit and change its height.
That gap is why scan-to-BIM is still largely a manual modelling job with a very expensive photograph sitting next to it. This paper attacks the gap by changing what the model outputs.
Output a program, not a shape
BIMScript treats reconstruction as writing a parametric command program. Rather than predicting geometry directly, it predicts the sequence of instructions that would create that geometry in an authoring tool. The output is inherently editable because it is a set of commands, not a surface.
The commercially important detail is one-to-one command-to-Revit-object mapping, validated through add-ins with IFC4 export. That is the difference between a research artefact and something a coordinator could actually receive.
The output is not a picture of a wall. It is the instruction that makes a wall.
It also labels what things are made of
The second contribution matters more than it sounds. The layout language is extended with per-element material and condition attributes, supervised by vision-language-model material passports across 100,000 synthetic scenes and 1.9 million pseudo-labelled elements.
Material is the property that unlocks everything downstream. Without it you cannot price the element, cannot calculate its embodied carbon, cannot assess its thermal behaviour and cannot plan its replacement. A geometrically perfect model with no materials answers almost no useful question.
The speed work is unglamorous and worth noting anyway. Kernel-launch overhead was eliminated with CUDA-graph decoders, taking a decode step from 6.4 to 1.9 milliseconds, and a grammar-parallel draft-and-verify scheme exploits the fact that the entity schema is deterministic. When you know the grammar of the output, you can guess ahead and check.
The 5cm problem
Token grids quantise. The paper reports a 5cm granularity limit, which is fine for a demo and useless for a model anyone will build from. Their fix is a hybrid discrete-continuous decoder head with training-free geometric snapping.
Read that as an admission and a solution in the same sentence. Discrete tokens gave them tractability and cost them precision, so they bolted a continuous head on to get the precision back. Anyone building in this space will hit the same wall.
What this means for a practice
- Ask what the deliverable is. Scan-to-BIM vendors sell a service. The question that separates them is whether you receive editable parametric objects or a mesh you will remodel by hand.
- Materials are the value, not the geometry. If the output has no material attribution, you have bought a shape and will pay someone to add the information that mattered.
- Synthetic training data is now normal here. 100,000 synthetic scenes is cheaper than any equivalent real capture programme, and it is why this area is moving quickly.
- Precision claims deserve a follow-up question. Ask what the underlying grid resolution is before anyone quotes you a tolerance.