An AI BIM model generator, read precisely, produces a concept-stage model in a BIM container: elements that know their class, openings that know which wall they cut, rooms with areas, quantities a schedule can read, and a spatial hierarchy, written as IFC4 that Revit and Archicad open as building elements rather than as one mesh. Nuit’s 3D mode does this; the published Mirror Lake House file has 1,079 elements, 5 IfcSpace rooms totalling 110.4 m² and 12 openings each related to its wall and its door or window. What no AI generates today is the information a BIM accumulates during development: families and types, build-ups, ratings, services, and more than one storey. The useful question is not “is it BIM” but “what in the file is real”, and this article answers that.
This is the semantics article: what “BIM from AI” means element by element. The mechanics of importing (Link versus Open, translator settings, what Revit does with a mesh wall) are in AI to Revit: An Early Concept Into BIM Without Redrawing, and the formats side is in Export an AI Architecture Model to IFC, DXF and GLB. Here we stay with the question in the title.
What “BIM” has to mean for the question to make sense
Building Information Modelling is not a file format and not a level of realism. It is geometry that carries information and relationships. Four properties make a model a BIM model rather than a 3D model:
- Classification. Each element has a class: wall, slab, door, window, roof, column, beam, space. A schedule is a query over classes; without them there is nothing to schedule.
- Relationships. An opening is cut from a specific wall; a door fills a specific opening; a room is bounded by specific walls; elements belong to a storey, which belongs to a building on a site. These are what make a model coordinated rather than coincident.
- Quantities and properties. Length, area, volume on each element; property sets that describe fire rating, thermal performance, and so on, as the project accumulates them.
- Editability by type. In a native BIM application, a wall has a type whose thickness and build-up can be changed once and applied everywhere.
An “AI BIM model generator” has to be tested against these four, not against how good the render looks. Most AI 3D tools fail the first property outright: a text-to-mesh generator produces a surface with no classes, and AI 3D Model Generator for Architecture explains why that is a different product. The rest of this article is about a tool that passes the first three and stops at the fourth.
How Nuit produces classified elements
The reason the IFC has classes is that the model is built with them. Nuit writes the building as a script: a parameter block and a build function that constructs the geometry element by element, tagging each with its IFC class as it is created. From the Mirror Lake House script:
aperture('Lake sliding doors', L, 0,
5.45, 2.44, 0.08, 'z', true)
tag(o, 'IfcSpace',
'Living room + open kitchen')An aperture is an opening in a wall and the door or window that fills it; the tag on a space names a room from which the floor area is computed. Because the classification happens at construction rather than being guessed afterwards from shapes, the exporter does not have to decide whether a box is a wall or a slab. It already knows. Text to 3D Building Model describes this generation in detail.
The consequence is an IFC4 file that a BIM application reads as a building. The published example, counted from the file at build time on the examples page:
| Class | Count | Reads in Revit / Archicad as |
|---|---|---|
| IfcWall | 12 | Walls |
| IfcSlab | 16 | Floors and roof decks |
| IfcRoof | 2 | Roofs |
| IfcColumn | 12 | Columns |
| IfcBeam | 205 | Structural framing |
| IfcMember, IfcPlate | 103, 8 | Framing, generic structural |
| IfcWindow, IfcDoor | 5, 7 | Windows, doors |
| IfcStair, IfcRailing | 6, 45 | Stairs, railings |
| IfcCovering | 566 | Cladding boards, finishes |
| IfcFooting | 9 | Foundations |
| IfcFurniture, IfcLightFixture | 12, 2 | Furniture, lighting |
| IfcGeographicElement | 57 | Site |
| IfcBuildingElementProxy | 7 | Generic models |
| IfcSpace | 5 | Rooms / spaces with area |
| Total | 1,079 |
Seven proxies out of 1,079 are elements the exporter could not classify within the schema; they keep their names and arrive as generic models. That is the correct behaviour for an exporter.
The relationships, which are the part people miss
A list of classified elements is a catalogue, not a model. The file becomes a model through the relationships it writes:
-
Openings and hosts: Each of the 12 openings in Mirror Lake House is an IfcOpeningElement with an IfcRelVoidsElement to the wall it cuts and an IfcRelFillsElement to the door or window that fills it. This is what lets Revit understand a window as in a wall. A door that is placed next to a wall rather than hosted in it would show up here as a door without a fill relationship; the completion check in the generator catches most of those, and the file shows the rest.
-
Spaces and boundaries: Each of the 5 IfcSpace elements has a name and a NetFloorArea quantity and is aggregated to the storey. The partitions that bound them are IfcWall elements built at generation; the doors between rooms are hosted in those partitions. This was made mandatory after early runs produced named rooms in one open volume. A model that says “five rooms” and has one room is not a BIM model, and now it cannot finish.
-
Spatial hierarchy: IfcProject, IfcSite, IfcBuilding, one IfcBuildingStorey, with elements contained in the storey through IfcRelContainedInSpatialStructure. This is the tree Revit’s project browser and Archicad’s navigator build from.
-
Quantities: IfcElementQuantity on every element with length, area and volume as appropriate. A Revit schedule of walls or rooms populates without any modelling. Standard property sets (Pset_WallCommon and the like) exist on the elements where the model had something to write: 15 sets in the example file. Most elements carry quantities but no Pset, because the concept has not decided the things a Pset records.
-
Geometry representation: Where the model knows an element’s profile and direction (walls, slabs, columns, most beams), the exporter writes IfcExtrudedAreaSolid: 216 of them in the example. These are bodies a BIM application can treat as editable. Where it does not, it writes IfcTriangulatedFaceSet (875, mostly cladding boards, site and library furniture) with the correct class. Both are legitimate IFC; only the first is convenient to develop.
What the file does not contain, and will not
This is the section a BIM manager should read before anyone says “the AI did the model”.
-
Families and types: IFC has no notion of a Revit family or an Archicad favourite. Every element is an instance: a wall with a thickness, a door with an opening size. Your office’s wall types are not in the file and cannot be. Expect to swap.
-
Build-ups and materials: The published files carry no material assignments. Cladding is exported as separate IfcCovering elements, which is why that class is the largest in the table. The materials you see in the studio viewport, from a catalogue of 1,775 CC0 PBR scans plus a curated set, are a presentation choice in the tool, not an IfcMaterialLayerSet in the file.
-
Type objects and shared parameters: No IfcWallType, IfcDoorType or project parameters.
-
Ratings, performance, services: Nothing the concept did not decide. Fire ratings, U-values, structural sizing, ducts and pipes are design development.
-
Multiple storeys: One IfcBuildingStorey per export. A two-storey concept arrives on one level. This is the most significant current limit and it is stated on the product page rather than hidden.
-
Level of development: LOD 200 geometry and classification, LOD 100 information. A concept model in a BIM container, priced and presented as a concept.
If your definition of BIM requires the fourth property, editability by type within your office standards, then no AI generates BIM today, and the accurate phrasing is “an AI-generated concept model that imports as classified elements”. That phrasing is still a large saving over remodelling a render.
What this replaces
The days in which someone in the office rebuilds a chosen concept from images. Room areas, wall heights, opening positions, roof geometry and structural rhythm arrive as elements with the right classes and exact dimensions, because they are parameters; schedules populate and sections cut on the day of import. AI + Revit and Archicad covers where concept images sit in the same pipeline; the model is the step after. Design development, from the type library onward, stays your work.
How to read an AI-generated IFC before trusting it
Any IFC file, from any tool, can be read in a free viewer or in the BIM application itself. Five checks, in the order that catches the most:
- Count the spaces against the brief. Five asked for, five IfcSpace in the file. Four means the model merged two rooms; go back and instruct.
- Read the areas. NetFloorArea on each space. Mirror Lake House: 56.8 + 14.1 + 13.9 + 13.9 + 11.8 = 110.4 m² against a brief of “about 120”. The count is right and the total is 8% under, which is worth knowing before an area schedule reaches a client.
- Count the doors and check they are hosted. Seven IfcDoor, each with an IfcRelFillsElement into an opening. A door without one was drawn beside a wall.
- Check the storeys. One IfcBuildingStorey. If the project has two floors, the elements are on one level, and multi-storey work should be modelled and checked per floor for now.
- Look at the geometry types. Extrusions can be developed; meshes should be replaced. Structural and enclosure elements in the example are mostly extrusions; cladding and site are meshes.
The examples page runs the first four of these on the published files at build time, so the numbers on the page and in the download cannot disagree. Do the same on your own export before it goes to anyone.
Cost, access and where it fits
Generating a model from a brief costs roughly 60 to 150 credits with an estimate before the run; a prompt edit 5 to 25; sliders, direct manipulation, plans, sections and every export are free and unlimited. What a credit buys is on the pricing page. The 3D mode is open to every account from the Images | 3D switch in the studio, same as the images side, with 300 free credits.
Settle the concept in the studio, where an edit is cheap and every version is a branch; export the approved one; link it into Revit or Archicad; develop from there. Importing five versions of a concept into the BIM to compare them costs the days the tool was meant to save.
Download the Mirror Lake House IFC4, open it, and count.
Frequently Asked Questions
Can AI generate a BIM model?
A concept-stage model in a BIM container: classified elements with quantities, hosted openings, spaces with areas and a spatial hierarchy, exported as IFC4. Not families, types, build-ups, ratings, services or multiple storeys.
Can AI create an IFC file I can open in Revit?
Yes. Nuit’s IFC4 opens in Revit (Link or Open IFC) and Archicad. Mirror Lake House: 1,079 elements, 18 product classes, 5 rooms with areas, 12 hosted openings, one storey. Downloadable from the examples page.
What makes an AI-generated model a BIM model rather than a 3D model?
Classification and relationships: walls that know they are walls, openings related to the walls they cut, doors that fill openings, rooms bounded by walls with areas. Nuit tags elements as they are built and writes the relationships into the file.
Does an AI BIM model include rooms and areas?
Yes. IfcSpace with name and NetFloorArea. Mirror Lake: 56.8, 14.1, 13.9, 13.9, 11.8 m², total 110.4. Partitions and doors are built at generation and a completion check enforces enclosure.
What does not transfer into Revit or Archicad?
Families and types, build-ups and materials, project parameters, ratings and services, multiple storeys. Mesh elements arrive classified but not type-editable; extrusions can be developed.
Is an AI-generated BIM model accurate enough to use?
As a concept: dimensions are parameters and areas are computed. LOD 200 geometry, LOD 100 information. A starting point for schematic design, not a coordinated model.
How do I check an AI BIM model before trusting it?
Read the file: count spaces against the brief, read areas, confirm doors are hosted, check storeys, look at extrusions versus meshes. The examples page prints these from the file at build time.
Try Nuit free — 300 credits (12 images), no card required. Generate the concept, and when 3D access is enabled, export the IFC4 and open it in your own BIM before anyone calls it a model. Start your project →