# AI Interior 3D Model: Furnishing Rooms with a Material and Asset Library **An AI interior 3D model, in Nuit's 3D mode, is a set of real rooms inside a building model: each room an IfcSpace with a name and a floor area, enclosed by partitions with a door, furnished from a library of CC0 models and finished from a catalogue of 1,775 PBR scans through a picker. Furniture is placed, not generated, which is why a chair is 0.8 m high and not 1.3. The model exports with the furniture as IfcFurniture in IFC4, as silhouettes on layer I-FURN in DXF, and as named instances in GLB. For a mood image of a room, an image tool is still faster; for whether the sofa fits between the door and the window, this is the tool.** This article covers the interior side of the model: rooms, furniture, materials and what leaves in the export. [AI Interior Design Tools for the Concept Stage](/blog/ai-interior-design-tools-concept-stage/) covers interior images; [AI 3D House Design](/blog/ai-3d-house-design/) covers the house the rooms sit in. --- ## Rooms first The model does not furnish a box. It builds rooms, and the furniture goes into them. At generation the model writes each room as an IfcSpace with a name, builds the partitions around it as IfcWall elements, cuts an opening in a partition and hosts a door in it. A completion check runs before the model is allowed to finish: a run that has named five rooms and built one open volume is sent back to build the partitions. This check exists because early runs did exactly that and looked finished. In the studio the Interiors phase in 3D reads the rooms from the IfcSpace elements. Mirror Lake House, the published example on the [examples page](/3d/examples/), lists five: | Room | Area | |---|---| | Living room + open kitchen | 56.8 m² | | Entrance hall | 14.1 m² | | Main bedroom | 13.9 m² | | Second bedroom | 13.9 m² | | Bathroom | 11.8 m² | Click a room and the camera goes inside, entering through the door, with a gap from the furniture so you are not standing in a wardrobe. Because the areas are computed from the geometry, a 13.9 m² bedroom is 13.9 m², and widening it by a slider changes the number and the partition at once. [From an AI Floor Plan to a 3D Model](/blog/ai-floor-plan-to-3d-model/) explains how a plan image sets the arrangement of these rooms. --- ## Furniture: a library, on purpose Text-to-mesh tools can generate a chair from a prompt, and the chair can look good. It is also, often, the wrong size, has a seat you cannot sit on, and weighs 200,000 triangles. For furnishing a room that has to be measured, a library is the better answer, and Nuit uses one. The catalogue is a few hundred models, all CC0 (no attribution required, commercial use and redistribution allowed): seating, tables, beds, shelves, lamps, potted plants, trees, ground cover, vases, wall decoration. Each is repacked to a compressed GLB with textures at 1K, placed on the floor at its origin and limited to 3 MB, with the author named in the manifest. BlenderKit was considered as a second source and dropped, because its terms forbid redistribution outside a project even for CC0 models. Two ways furniture gets into a room: At generation, the model reads the brief and references, queries the asset catalogue and places what the room needs: a sofa and a table in a living room, a bed and side tables in a bedroom. Mirror Lake House came back furnished; its IFC has 12 IfcFurniture and 2 IfcLightFixture elements. By hand, select an asset and replace it from the picker, which filters by furniture, planting, lighting, sanitary, decor, site and appliances, or place a new one. The picker has three tabs: the library, your own uploads, and upload. Your own GLB, up to 20 MB, comes in as a placeable asset; a short rights notice appears once before the first upload. Click-to-replace is free. What the library does not have, it does not have. A particular manufacturer's sofa is an upload, not a search. --- ## Materials: 1,775 scans and a picker Finishes come from a catalogue of PBR scans, and the picker is the part of the interior work you will touch most. The catalogue is 1,775 CC0 PBR material scans, the full set minus categories that are not architectural (snow, lava, food, decals, sci-fi), plus 23 curated entries that map the usual names (oak, plaster, brick) to specific scans. Each material carries albedo, roughness and normal maps at 1K, with ambient occlusion and metalness where the scan has them, and a real-world tile size read 's measured dimensions where they exist and from the category median where they do not. That last detail matters: a 300 mm tile reads as 300 mm on the wall, not as whatever the texture happened to be. The picker works like this: a search box over names and tags, a row of 39 category chips, a Recent strip with the project's last eight picks, and a grid of 96-pixel thumbnails paged 48 at a time so that opening it does not fire 1,800 image requests. The current material of the selected element is outlined. Click one and the viewport previews it on the target; confirm and the assignment is written into the model's script as an override on those elements. That makes it a checkpoint like any other edit: free, undoable, branchable. The model also picks from the same catalogue at generation, from the brief and references. Mirror Lake House arrived in charred timber because the images showed it. Override any element afterwards. Switch the viewport from Sketch to Materials to see the scans; Sketch is the default look and shows the geometry without them. --- ## Editing a room Everything in the room is either a parameter or an element in the script, and the [parametric editing model](/blog/ai-parametric-design-tool/) applies. Free: room dimensions where the model declared them, wall and opening sizes by slider or by dragging in the viewport, replacing an asset from the picker, applying a material, plans, sections, undo, branching, export. 5 to 25 credits, estimate first: instructions that need the model. "Add a walk-in wardrobe off the main bedroom" (new partition, new door, new space). "Move the bathroom door to the hall side." "Replace the dining table with a longer one for eight" is borderline: if the picker has the table, it is a free replacement; if the model has to reposition chairs around it, it is a small call. The most useful habit is to branch before a large change. Two furniture arrangements of the same living room, side by side with the same area, is a better client conversation than one arrangement and a memory of the other. --- ## What leaves in the export The interior travels differently in each format, and it is worth knowing before the first import. - IFC4: Rooms as IfcSpace with NetFloorArea; partitions as IfcWall; doors as IfcDoor hosted in openings; furniture as IfcFurniture and lights as IfcLightFixture, with mesh geometry (IfcTriangulatedFaceSet) and the correct class. Revit shows them as furniture and lighting families of generic type. No material assignments travel; the published files carry none, and finishes in the BIM sense are design development. [AI BIM Model Generator](/blog/ai-bim-model-generator/) lists what the file carries element by element. - DXF: The plan at 1.2 m with room outlines on A-AREA and labels with area on A-ANNO-TEXT; each furniture asset as one closed silhouette (a convex hull of the instance) on I-FURN, lights on E-LITE, sanitary on P-FIXT, plants on L-PLNT. Silhouettes, because drawing every asset edge produced a 232 MB file; with silhouettes the Mirror Lake DXF is 1.6 MB. [What's Inside an AI-Generated DXF](/blog/dxf-layers-ai-export-what-architects-get/) has the full layer table. - GLB / OBJ: Every asset as a named instance with its materials; scan materials arrive as colour maps at best. Hide the planting group in SketchUp while you work; the 24 pines on the Mirror Lake site are 27,000 triangles each. --- ## When to render instead An interior rendering from Nuit's Interiors image mode, or from any image tool, gives atmosphere, light and finish in seconds, with a level of visual richness the model's viewport does not attempt. For a mood, for a client who wants to feel the room, for a moodboard, render. [AI Prompts for Interior Design: Light and Texture](/blog/ai-prompts-interior-design-light-texture/) is about getting that right. The model answers a different question: does it fit. Room dimensions, door swing, the distance from sofa to screen, whether a 2.4 m table goes in a 3.6 m dining space with chairs pulled out. Those are numbers, and the model has them; a rendering does not. In practice the two are used together. Generate the interiors as images from the brief, choose a direction, attach it to the 3D chat as a reference, and the model furnishes the room in that direction with library assets and catalogue materials. The image sets the mood; the model checks the measurements. --- ## Limits Concept stage: no joinery, no ceiling design, no services, no lighting calculation. The library is a few hundred models; bespoke pieces are uploads. Materials are viewport finishes, not IfcMaterialLayerSet in the export. One storey per IFC and DXF export. No DWG, SKP or Revit family import, no plugin. The 3D mode is open to every account from the Images | 3D switch in the studio, same as interior images, with 300 free credits. Generating the model costs 60 to 150 credits, a model-assisted edit 5 to 25, and everything by hand is free; plans are on the [pricing page](/pricing/). Download the [Mirror Lake House GLB](/3d/examples/) and open it in Blender or SketchUp: the 12 pieces of furniture are there by name, in the rooms they belong to. --- ## Frequently Asked Questions ### Can AI generate a 3D interior model? Yes, as rooms inside a building model: IfcSpace with area and a door, furniture from a CC0 library, finishes from 1,775 PBR scans, exporting to IFC4, DXF and GLB with the furniture. ### Where does the furniture come from? A library of CC0 models, placed by the model at generation or picked by hand, plus your own GLB uploads up to 20 MB. Not generated, because library pieces are correctly scaled. ### How are materials applied? Through a picker over 1,775 material scans plus 23 curated entries: search, categories, recent picks. The assignment is an override in the script: free, undoable, branchable. The model also picks materials at generation. ### Does the interior model have real rooms with doors? Yes. Partitions and a hosted door per room are built at generation and enforced by a completion check. Mirror Lake House: five rooms, 110.4 m², seven doors. ### Can I export an AI interior model with its furniture? Yes. IfcFurniture and IfcLightFixture in IFC4 (12 and 2 in Mirror Lake House), silhouettes on I-FURN, E-LITE, P-FIXT and L-PLNT in DXF, named instances in GLB. ### Is an AI 3D interior better than an AI interior rendering? Different questions. Render for mood and finish; model for dimensions, fit and export. Use the image as a reference for the model. ### What are the limits? Concept stage only, library furniture, viewport-only materials, one storey per export, no CAD or BIM import, no plugin. --- **Try Nuit free — 300 credits (12 images), no card required.** Generate the interiors as images first; when 3D access is enabled, furnish the rooms from the library and check what fits. [Start your project →](https://nuit.archi)