# AI 3D House Design: Generate a Model of a House and Edit It by Words **AI 3D house design, done properly, produces a house you can measure, walk into room by room, change by words or sliders, and open in Revit, AutoCAD or SketchUp, rather than a picture of a house. Nuit's 3D mode generated Mirror Lake House from a two-sentence brief and four saved concept images: a 20 × 6 m single-storey house in charred larch, its living end cantilevered over a lake on steel columns, with five rooms totalling 110.4 m² built as real spaces with partitions and doors. One edit in words was free because it hit an existing parameter; another cost 11 credits because it added a terrace. This article walks through that house, from the brief to the exported files, as the worked example for anyone searching for how AI 3D house design actually works.** It is the residential companion to [AI House Design in 2026](/blog/ai-house-design-2026/), which covers concept images for houses. That article ends where this one starts: with a direction chosen and the question of what the house is as a building. For the mechanism of generation see [Text to 3D Building Model](/blog/text-to-3d-building-model/), and for how a parametric model differs from a rendered or meshed house, [AI 3D Model Generator for Architecture](/blog/ai-3d-model-generator-architecture/); here we stay with one house. --- ## Step 1: the brief, in numbers Mirror Lake House started, as most projects in the studio do, on the images side: exterior concepts generated from a brief, branched, a direction chosen, four images saved. Those images carried the form and the material: a long low volume, charred timber, a shallow gable, the living end held out over the water. The brief for the model, verbatim from the [examples page](/3d/examples/): > An elongated single-storey house on a rocky lake shore, as in the reference images: a long narrow volume with a low gable roof, its lake end cantilevered over the water on slender steel columns. About 120 m²: an open living room with kitchen at the lake end, a main bedroom, a second bedroom, a bathroom and an entrance hall at the land end. Plot 40×40 m with the lake, rocks and pines. Count what it specifies: one storey; five rooms, named, with a rough arrangement (living at the lake end, entrance at the land end); about 120 m²; a 40 × 40 m plot with water, rocks and trees. Everything about how the house looks is delegated to the four images with "as in the reference images". Numbers and program in words; form and materials in pictures. Describing the charred larch in the brief as well as showing it in the images gives the model two masters, and it compromises between them. [How to Write an Architectural Brief](/blog/how-to-write-architectural-brief/) covers the brief for images; for the model, shorten it. --- ## Step 2: generate In the 3D view, the saved images of the phase sit on the left; each has "Use in 3D chat", or drag it onto the composer. Up to six attach, resized to 1,600 px on the long side. The empty state shows the estimate (typically 60 to 150 credits; what a credit buys is on the [pricing page](/pricing/)) and the balance. Generate. What the model wrote for Mirror Lake House, reading its own parameters: - House 20 × 6 m; floor 1.4 m above the water; wall height 2.7 m; charred-timber cladding on a 150 mm module. - Gable roof at 14°, standing seam at 460 mm centres, 0.5 m eaves. - Steel columns 120 mm on concrete pads, lake support line at 8.5 m, diagonal braces to the girders. - Five rooms with partitions and doors between them; furniture from the asset library; 24 pines, rocks and the lake surface on the site. Each of those numbers is a parameter in the Parameters panel, grouped by IFC category, and each is a slider. --- ## Step 3: read the rooms before the picture The viewport is the least reliable place to check a house. Switch to the Interiors view, which lists the rooms from the IfcSpace elements with their computed areas: | 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 | | **Total** | **110.4 m²** | Five asked for, five built. The brief said about 120 m²; the model came out 8% under. That is the kind of number to know before a client sees an area schedule, and it is a slider away from 120 if the total matters: widen the house from 6 m, or lengthen it from 20. Each room is a real volume with partitions around it and a door into it, built at generation. Clicking a room in the Interiors panel places the camera inside, entering through the door. This was made mandatory after early runs produced five named rooms in one open box; a completion check now stops a run that declares rooms without enclosing them. [From an AI Floor Plan to a 3D Model](/blog/ai-floor-plan-to-3d-model/) has more on rooms and partitions when you start from a plan. Then the Plan view: the model cut at 1.2 m, with overall and wall-axis dimension chains and room labels, drawn by the same function the DXF export uses. Then, if you like, the sections. Only then the picture. --- ## Step 4: edit by words, and watch what each edit costs Two edits were made to Mirror Lake House, and they illustrate the two prices. "Extend the cantilever over the water by 2 m." The lake balcony depth was already a parameter, at 1.8 m. The instruction was routed to a slider move (1.8 → 3.8 m), the model was not called, and the edit was free. This is the "Before" branch in the viewer on the [3D page](/3d/). "Add a small terrace with a railing at the lake end, under the roof." No parameter can produce a terrace that does not exist. The model rewrote the build function: the roof now extends 1.8 m out over the water (PARAMS.roof.lakeOverhang in the final script), the terrace sits 1.8 m deep under it with a 1.05 m glass guard, and both are parameters. Two steps, 11 credits, estimate shown first. This is the "Lake terrace" branch, the final model. Both versions are kept. Every edit, by words, by slider or by dragging a wall in the viewport, is a checkpoint with its parameter delta, script hash and preview; branch from any of them. [AI Parametric Design Tool](/blog/ai-parametric-design-tool/) is about this editing model in full. What sliders do for free on a house like this: wall height, footprint, roof pitch, eaves, seam spacing, deck depth, rail height, opening widths and heights, room dimensions where the model declared them. What words do for 5 to 25 credits: add or remove a room, move the entrance, add a porch, change the roof form, add a carport. What is closer to a new model than an edit, and should be a branch: a second storey, a different plan type. --- ## Step 5: materials The house arrived in charred timber because the references showed it. The materials themselves come from a catalogue of 1,775 CC0 PBR material scans plus a small curated set: the model picks from it at generation, and you can override any element from a picker with search, categories and the project's recent choices. Pick a material, pick the elements it applies to, done; it is a free edit and a checkpoint. Switch the viewport from Sketch to Materials to see it. For a house, the useful comparison is two branches with the same geometry and different cladding, viewed from the same angle. [AI Interior 3D Model](/blog/ai-interior-3d-model-furniture/) covers materials and furniture in the rooms. --- ## Step 6: export, and open it in your own software Export is free and unlimited, from any checkpoint, in three formats. The Mirror Lake House files are published on the [examples page](/3d/examples/) and the counts below are read from them at build time. - IFC4 for Revit or Archicad: 1,079 elements: 12 walls, 16 slabs, 2 roof elements, 12 columns, 205 beams, 5 windows, 7 doors, 5 spaces with areas, plus railings, stairs, footings, cladding boards, furniture and site. Twelve openings each related to the wall they cut and the door or window that fills them. One storey. [AI to Revit](/blog/revit-ai-ifc-concept-to-bim/) walks through the import and what does not transfer (families, types, materials). - DXF for AutoCAD: One file, seven views in model space: plan cut at 1.2 m with dimensions and room labels, four elevations, two sections, on AIA-style layers mapped from each element's class (A-WALL, A-GLAZ, S-BEAM and so on; 24 declared, 19 used by this house). Trees and furniture as silhouettes. [What's Inside an AI-Generated DXF](/blog/dxf-layers-ai-export-what-architects-get/) lists every layer and every limit. - GLB or OBJ for SketchUp or Blender: The whole scene as named groups: house, deck, columns, roof, each room volume, lake, rocks, the 24 pines. [SketchUp + AI](/blog/sketchup-ai-alternatives-2026/) covers the import settings. Which format for what is the subject of [Export an AI Architecture Model to IFC, DXF and GLB](/blog/ai-architecture-model-export-ifc-dxf-glb/). --- ## What kind of house this works for It works well for single-family houses, villas, cabins, pavilions, ADUs and holiday houses: single-storey today, with a clear program and a site that can be described in a sentence. Anything where the client wants to see the house and the office wants to develop it without remodelling from images. [AI for ADU and Tiny House Design](/blog/ai-adu-tiny-house-design/) is the small end of this range. Two-storey houses need care. Plans in the studio are per floor, and the model can be generated per floor, but IFC and DXF export a single storey, so the floors arrive as separate files. Sloping sites: describe the levels in numbers. It does not fit a renovation of a surveyed house (start from the survey in your CAD tool; Nuit takes images as references, not DWG or SKP as geometry). Anything past schematic design. Structural or energy calculation: areas are computed, loads are not. The limits in one list: concept stage only; one storey per export; no plugin; no DWG or SKP import; no construction documentation; no calculations. --- ## What it cost One model run in the 60 to 150 credit range; two edits at 0 and 11 credits; every slider, material choice, plan, section and export at nothing. The files are published so the result can be checked rather than believed: download the [Mirror Lake House IFC4, DXF and GLB](/3d/examples/), open them in your own software and count the rooms. --- ## Frequently Asked Questions ### Can AI generate a 3D model of a house? Yes, and the useful kind is parametric. Nuit wrote Mirror Lake House from a two-sentence brief and four images: 20 × 6 m, five rooms totalling 110.4 m², walls, doors, roof and spaces as classified elements, exportable to IFC4, DXF, GLB and OBJ. ### How do I design a house in 3D with AI? Brief in numbers, concept images first, attach two or three to the 3D chat with the brief, generate, adjust by sliders and words, branch to compare, export the chosen one. ### Can I edit the AI house model after it is generated? Yes. On Mirror Lake House one edit was free (existing parameter) and one cost 11 credits (new terrace). Sliders, drag, materials, plans, sections and export are free. Both versions are branches. ### Does the AI house model have real rooms and doors? Yes. Five IfcSpace with areas (56.8, 14.1, 13.9, 13.9, 11.8 m²), partitions built at generation, seven hosted doors. The Interiors view enters each room through its door. ### Can the AI house design be opened in Revit, AutoCAD or SketchUp? Yes: IFC4 (1,079 elements) for Revit and Archicad, DXF with seven views on AIA-style layers for AutoCAD, GLB or OBJ with named groups for SketchUp and Blender. The files are published to test first. ### Is AI 3D house design accurate enough for a real project? For concept stage: dimensions are parameters, areas are computed. Approximate targets stay approximate. No details, services or documentation; one storey per export. ### How much does it cost to generate a 3D house model with AI? Roughly 60 to 150 credits per model with an estimate first; 5 to 25 per model-assisted edit; everything by hand and every export free. Free accounts start with 300 credits (12 images); 3D is open to every account. --- **Try Nuit free — 300 credits (12 images), no card required.** Generate the house as images first; when 3D access is enabled, attach them to the brief and get the model back with the rooms in it. [Start your project →](https://nuit.archi)