In a converted warehouse studio in SoMa, Autodesk Research released a project called Project Dreamcatcher — a generative design system that had been quietly tested on aerospace components but was now aimed, nervously, at architecture. The team's lead researcher, Mark Davis, remembered the first time a client saw an AI generate a residential floor plan in real time: the client went silent for nearly a minute, then asked, "Did it know the sun moves?"
The answer was: not exactly. But it was learning. Dreamcatcher consumed thousands of floor plan images, extracted spatial relationships, and produced layout candidates ranked by stated objectives — natural light exposure, circulation efficiency, structural bay regularity. It was the first commercially demonstrated system to treat the floor plan not as a drawing but as a configuration problem with computable constraints.
Before generative tools, architects produced floor plans through a well-understood iterative sequence: bubble diagrams to establish adjacencies, blocked massing to test proportions, then detailed drafting in AutoCAD or Revit. The intelligence was entirely human; the software was a precision pencil. A competent drafter could produce a buildable schematic plan in one to two days. A principal architect would review, redline, and return it for revision — sometimes five or six rounds before a client saw anything.
The bottleneck was not creativity but enumeration. An architect exploring a 3-bedroom, 2-bath residential layout has hundreds of topologically distinct configurations when you account for room adjacency, corridor placement, service core location, and fenestration. No human team explores more than a fraction of them. This is precisely the gap that machine generation was designed to close.
Early computational layout experiments in the 1970s — notably William Mitchell's shape grammar work at MIT — showed that a small set of formal rules could generate thousands of syntactically valid architectural plans. Mitchell's 1977 book Computer-Aided Architectural Design laid the theoretical groundwork, but the hardware wasn't there. The idea sat dormant for almost forty years.
The true breakthrough came when researchers stopped trying to encode architectural rules and started training neural networks on observed outcomes. In 2019, a team at Tsinghua University published HouseGAN — a graph-conditioned generative adversarial network that accepted a room-adjacency graph as input and produced photorealistic floor plan images as output. The model was trained on 145,000 floor plans scraped from real estate listings across China and the United States.
HouseGAN's architecture was novel in a specific way: the graph-conditioned discriminator forced the generator to respect topological constraints — bedroom must be adjacent to bathroom, kitchen must connect to dining — while still allowing geometric freedom. The result was plans that looked like they'd been drawn by a human who had studied thousands of apartments. They were imperfect but recognizable.
Simultaneously, Stanislas Chaillou at Harvard GSD was running his own experiments using a U-Net architecture trained on Parisian Haussmann apartment plans. His 2019 thesis, published as ArchiGAN, demonstrated that a GAN could not only generate plans but interpolate between architectural styles — producing a floor plan that was "70% Haussmann, 30% modernist" by navigating latent space. Chaillou showed the results at the Fondation d'entreprise Hermès and the response, he recalled, was equal parts fascination and mild horror.
The fundamental shift in AI floor plan generation was moving from rule-based enumeration (encode what architects know) to distribution learning (learn what architects have done). The first approach requires expert knowledge encoding; the second requires data. By 2018, architectural data — scraped from real estate listings, BIM repositories, and open municipal building permit archives — had become abundant enough to make the second approach viable.
The word "generate" conceals substantial technical complexity. In the context of AI floor plan tools, generation means one of three distinct processes, and confusing them leads to real mistakes in practice.
Each method has different failure modes. Parametric generation can get stuck in local optima and produce plans that satisfy all stated constraints but feel spatially awkward — what architects call "algorithm breath." GAN-based methods can produce topologically invalid plans with rooms that don't connect or corridors that dead-end. Diffusion models produce visually compelling images that may have no buildable structural logic at all.
Understanding which generation method underlies a given tool is not academic: it determines what you can trust, what you must verify, and how you integrate the output into a real design workflow.
William Mitchell's 1977 shape grammar work at MIT established the theoretical basis for computational floor plan generation. His core claim — that architectural plans follow a generative grammar analogous to linguistic syntax — is the intellectual ancestor of every GAN, diffusion model, and large language model that generates spatial layouts today. The grammar is now learned rather than encoded, but the underlying claim has proven correct.
You've learned that "generating" a floor plan means very different things depending on whether a parametric, GAN-based, or diffusion-based system is involved. In this lab, explore the practical implications with your AI tutor. Ask about failure modes, real tools, or how you'd choose between approaches on a real project.
At ETH Zürich's Chair of Architecture and Digital Fabrication, doctoral researcher Liane Müller was debugging a floor plan generator that kept producing kitchens adjacent to master bedrooms with no intervening corridor. The model had been trained on 80,000 American apartment plans, and in a statistically significant portion of them, open-plan layouts placed kitchen and sleeping zones in close proximity. The model was not wrong by its training data. But it was wrong by every client who valued sonic privacy.
Müller's solution was not to retrain the model. It was to enforce a hard adjacency constraint graph that sat between the generator and the output: a set of room-pair rules that could never be violated regardless of what the generator produced. Kitchen and master bedroom could not share a wall. Bathroom could not open directly into a dining room. The generator still controlled geometry; the graph controlled topology. The fix took a weekend. It also revealed the deeper principle: in AI floor plan generation, topology and geometry must be governed separately.
A room adjacency graph is the mathematical backbone of most serious AI floor plan systems. It is a graph G = (V, E) where vertices V represent rooms or spatial zones and edges E represent adjacency relationships — "these two spaces share a wall or a direct door connection." The graph encodes topology, not geometry: it says what connects to what, not how big or where.
This distinction matters enormously. Two floor plans can have identical adjacency graphs — bedroom connects to bathroom connects to corridor connects to living room — and yet look completely different in drawn form because the rooms are different sizes, oriented differently, or organized around different circulation spines. The graph is the structural DNA; the geometric drawing is its expression.
In the HouseGAN system, the input is exactly this: a room adjacency graph, specified by the user or generated procedurally, which the model uses to condition its output. The model has learned, from 145,000 training examples, what typical geometries correspond to each adjacency pattern. It samples from that learned correspondence to produce drawn plans.
Maket.ai, a San Francisco–based startup founded in 2022, built its core product around a constraint-first interface for residential floor plan generation. Rather than asking users to describe spaces in natural language or sketch bubbles, Maket presented a structured input form: room types, quantity, minimum area per room, adjacency requirements, and site boundary. Behind the form was a diffusion-based generator conditioned on the structured input.
In beta testing with 200 residential architects in 2023, Maket found that architects spent 40% more time on constraint specification in their system than on equivalent tasks in Revit — but produced three times as many initial layout alternatives for client review. The constraint overhead was front-loaded; the downstream variation was the payoff.
The critical finding was about failure modes. When architects specified incomplete constraint graphs — omitting adjacency requirements for utility spaces like laundry or mechanical — the generator placed these rooms in geometrically convenient but operationally problematic locations. Laundry rooms ended up adjacent to primary bedrooms; mechanical shafts ran through structural bays. The generator was not "wrong" in any absolute sense; it simply filled underspecified space with plausible-but-inconvenient guesses.
This is the fundamental discipline that AI floor plan generation demands: the quality of the constraint graph determines the quality of the output. Garbage in, elegant but dysfunctional plans out.
Before using any AI floor plan generator, produce a complete room adjacency graph for your program — even a hand-drawn bubble diagram. Identify which adjacencies are hard constraints (non-negotiable) and which are soft (preferred). Underspecified adjacency relationships will be filled by the model's training distribution, which may not match your project's functional requirements.
The most persistent failure mode in AI-generated floor plans — documented across HouseGAN, ArchiGAN, and commercial tools through 2024 — is circulation logic. Models trained on existing plans absorb the room configurations of those plans but frequently fail to reproduce the underlying circulation hierarchy: primary corridor → secondary corridor → room entry sequence that makes buildings navigable and fire-code compliant.
In 2022, researchers at Georgia Tech's School of Architecture published an analysis of 500 AI-generated residential plans from four different systems. Their finding: 68% of generated plans had at least one room accessible only through another habitable room — what building codes call a "non-code-compliant room access" violation. The models had learned that bedrooms are usually near bathrooms, but had not internalized the principle that every habitable room must have independent corridor access.
The fix, implemented in later systems like LayoutGPT (2023, Stanford/Microsoft Research), was to encode circulation as an explicit node type in the adjacency graph — treating corridors and thresholds as first-class spatial elements rather than inferred gaps between rooms. When the model was forced to plan circulation explicitly, code compliance violations dropped to 12%.
Circulation is not the residual space between rooms. It is the primary organizing system of a floor plan. Any AI floor plan tool that does not treat corridors, thresholds, and stair landings as explicit nodes in its constraint graph will systematically underperform on navigability and code compliance. Always audit AI-generated plans for circulation logic before presenting to clients.
In this lab, you'll work with the AI tutor to construct a room adjacency graph for a real residential program. Practice distinguishing hard from soft constraints and identify where circulation nodes must be made explicit.
When Midjourney v5 was released in March 2023, architectural firms across Manhattan began testing it for floor plan generation within days. SHoP Architects shared early experiments internally: given the prompt "luxury residential tower floor plate, corner unit, two bedrooms, Hudson River views," the model produced images of extraordinary visual sophistication. The plans looked, at a glance, like something a Pritzker Prize winner might have sketched on a napkin.
A junior designer at SHoP, Priya Nair, was tasked with actually measuring one of them. She overlaid a scale grid. The master bedroom was 9 feet wide — barely enough for a double bed and circulation. The bathroom had no plumbing wall. The "Hudson River" orientation had the living room facing northeast. The plan was spatially hallucinated: visually fluent, programmatically incoherent. SHoP's design technology lead reportedly wrote a memo that week: "Use Midjourney for inspiration. Do not use it for anything a contractor will see."
Diffusion models like Stable Diffusion, Midjourney, and DALL-E 3 generate images by iteratively denoising a random signal conditioned on a prompt. They do not reason about space; they reason about the visual statistics of images described by text. This is a fundamental distinction with serious architectural implications.
A diffusion model trained on architectural imagery has learned that "floor plan" images tend to contain thin black lines on white backgrounds, labeled rectangles, and dotted threshold indicators. It has learned the visual grammar of plan drawings. What it has not learned is what those drawings mean — which walls are load-bearing, which rooms require minimum clear dimensions, which corridors must meet egress width requirements.
The result is what researchers at MIT's Computational Design Lab called "architectural hallucination" in a 2023 paper: generated plans that are visually indistinguishable from drawn plans but contain spatial impossibilities that would be immediately apparent to any licensed architect reviewing for code compliance. In a study of 200 diffusion-generated residential plans, MIT found that 91% contained at least one room with sub-minimum code dimensions — rooms smaller than 70 square feet, corridors narrower than 36 inches, or bathrooms with zero plumbing clearance.
Diffusion models perform visual pattern completion, not spatial reasoning. They produce images that match the statistical appearance of floor plans. They do not verify dimensions, structural logic, code compliance, or buildability. Every diffusion-generated floor plan must be treated as a visual sketch requiring complete architectural re-examination, not as a schematic design document.
The challenge of using diffusion models for floor plan work is not the model's capability limit — it's the mismatch between natural language and spatial specification. Natural language is excellent at conveying aesthetic intent ("warm, Japandi-influenced," "open and airy") but extremely poor at conveying dimensional and topological precision ("master bedroom minimum 180 sq ft with 9-foot ceiling and direct bathroom access without crossing a shared corridor").
In 2023, researchers at ETH Zürich published a systematic study of text-to-floor-plan prompts across GPT-4V and Stable Diffusion XL. They found that prompts including explicit room counts and adjacency specifications improved topological validity by 34%, while prompts including dimensional targets improved dimensional accuracy by only 8% — the models had largely failed to internalize metric space from their training data.
The study's most actionable finding: structured prompts outperformed natural language prompts. A prompt formatted as a structured list — "Program: [bedroom: 2, bathroom: 2, kitchen: 1, living: 1]. Adjacencies: [bedroom1-bathroom1, kitchen-dining, living-entry]. Style: contemporary open plan." — produced topologically valid outputs 58% of the time versus 31% for equivalent natural language descriptions.
By late 2023 and into 2024, a new generation of tools began hybridizing diffusion models with constraint enforcement layers. Architect AI (formerly ARCHITECTai) coupled a Stable Diffusion XL backbone with a post-processing constraint validator that rejected outputs with rooms below code minimum dimensions and flagged circulation failures for human review. The validator ran in under 2 seconds and reduced architectural hallucination in accepted outputs to approximately 18%.
TestFit, a site planning and floor plate optimization tool used by major developers including Related Companies and Hines, took a different approach: it used parametric logic for structural and code-compliance validation and diffusion-based aesthetic refinement only at the presentation layer. The bones were parametric; the skin was diffusion-generated imagery. This architectural division of labor — parametric for function, diffusion for communication — is emerging as the dominant professional hybrid approach as of 2024.
The practical implication: no serious commercial tool in 2024 uses raw diffusion generation for buildable floor plans. The diffusion component always operates behind a validation or filtering layer. Architects using standalone image-generation tools without such layers are working outside the professional standard, not inside it.
The emerging professional norm is a hybrid workflow: structured constraint graphs drive parametric layout generation for functional and code-compliant spatial organization, while diffusion models produce presentation-quality imagery for client communication. These are two separate tasks requiring two separate tools. Treating diffusion output as a design document — rather than as a communication artifact — is the most common source of AI-generated error in contemporary architectural practice.
Diffusion model outputs depend heavily on prompt structure. In this lab, work with the AI tutor to transform natural language architectural briefs into structured spatial prompts, and learn to predict where diffusion-generated plans are most likely to hallucinate.
MVRDV, the Dutch practice known for densely programmed mixed-use buildings, ran an internal experiment in early 2023. Using a combination of Autodesk Forma for site-level massing and a custom HouseGAN derivative for unit layout generation, they produced 1,200 apartment unit floor plan variants for a 240-unit social housing project in Amsterdam Noord over a single weekend. The project architect, Winy Maas's team lead Fokke Moerel, later described the experience: "We had more options on Monday morning than we could look at in a month. The problem was no longer generation. It was evaluation."
The team developed what they called a "triage protocol" — a four-stage filter that reduced 1,200 variants to 18 finalist schemes in two days. Stage one: automated dimensional compliance check (area, ceiling height, window-to-floor ratio). Stage two: automated circulation audit (every room independently accessible). Stage three: human review for spatial quality — light, proportion, livability. Stage four: structural feasibility scan against the project's concrete frame grid. Eleven of the 18 finalists passed all four stages. Two became the basis of the submitted design.
MVRDV's triage approach crystallizes a principle that has emerged across multiple major firms experimenting with AI floor plan generation: the evaluation framework is more valuable than the generation tool. Generating 1,000 plans is easy; knowing which 10 are worth developing is the hard architectural judgment that AI has not yet automated.
A practical triage protocol for AI-generated floor plans operates across four dimensions:
1. Dimensional Compliance. Every generated plan should be run against project-specific and code-mandated dimensional minimums before human review. This is fully automatable using parametric scripts in Rhino/Grasshopper or Dynamo for Revit. The Georgia Tech study found this filter alone eliminates 60–70% of diffusion-generated outputs. For GAN-based outputs, it typically eliminates 20–30%.
2. Topological Validity. Verify that every habitable room has independent access from a corridor or common space, that no room requires traversal of another habitable room, and that required adjacencies from the brief are satisfied. This is also automatable but requires that the brief's adjacency graph has been formally specified.
3. Structural Feasibility. Verify that the plan is compatible with the structural system under consideration — that walls align with bearing lines, that spans are within structural norms, that columns don't bisect rooms. This requires human judgment or structural analysis software. It cannot be delegated to the floor plan generator.
4. Spatial Quality. Human review for the qualitative properties that make space livable and architecturally worthy: proportional relationships, natural light distribution, visual connections, spatial hierarchy. This is, and will remain for the foreseeable future, an exclusively human judgment.
Treat AI floor plan generation as producing a large candidate set for human curation, not a small set of finished proposals. The architectural value is in the combination of AI's breadth (hundreds of options) and human judgment's depth (discerning quality within those options). Firms that use AI to generate one or two plans and present them directly to clients are using the technology incorrectly.
No AI-generated floor plan — from any system currently available — produces output that is directly usable as a construction document or permit drawing. The gap between a generated plan image and a buildable set of drawings involves multiple professional translation steps.
The most efficient reconstruction workflow documented in practice, as reported by firms including Gensler, HOK, and Perkins+Will in their 2023 AI workflow surveys, follows this sequence:
Step 1: Geometric Reconstruction. Trace the selected AI plan in CAD (AutoCAD or Revit) using the generated image as a reference underlay. Do not attempt to import AI-generated images directly as CAD geometry — the lines will not align to a structural grid or coordinate system.
Step 2: Grid Alignment. Align the traced plan to the project's structural grid. This almost always requires adjusting room dimensions from the AI output. The goal is preserving the topological relationships (adjacency structure) of the AI plan while correcting the geometry to structural reality.
Step 3: Dimensional Correction. Verify and correct all room dimensions against code minimums, client brief requirements, and furniture planning standards. This is where the spatial quality of the AI plan is either confirmed or substantially redesigned.
Step 4: Wall Type Assignment. Assign structural, partition, and wet-wall types to all walls in the reconstructed plan. Specify door and window locations with code-compliant dimensions. This is fully within standard architectural drafting practice and requires licensed judgment.
Step 5: Systems Coordination. Coordinate the reconstructed plan with MEP (mechanical, electrical, plumbing) requirements — locating wet walls for plumbing, ensuring HVAC duct space is accounted for, confirming electrical panel and service core locations.
The question architects most frequently ask about AI floor plan generation is: "How much time does it actually save?" The answer from firms with documented workflows is more nuanced than marketing materials suggest.
A 2024 survey of 85 architectural firms by the American Institute of Architects Technology Practice Group found the following measured time savings and losses from AI floor plan integration:
Time saved: Schematic alternative generation (reduced from 3–5 days to 4–6 hours for a typical residential program). Client presentation of options (more alternatives, faster). Early-stage brief exploration and client expectation-setting.
Time added: Constraint specification and adjacency graph preparation (new workflow step, 2–6 hours). Triage and evaluation of generated candidates (new workflow step, 4–12 hours depending on candidate volume). Geometric reconstruction from selected AI plans (roughly equivalent to conventional drafting from a bubble diagram).
The net finding: firms using AI floor plan generation saved an average of 1.2 days on schematic design phases compared to conventional workflows, with the caveat that this gain was only realized when a disciplined constraint-first workflow was followed. Firms that used AI without formal constraint specification and triage protocols reported longer schematic design phases due to the time spent evaluating and reconstructing unsuitable AI outputs.
AI floor plan generation saves time in schematic design only when practiced with discipline: specify constraints first, triage generated outputs systematically, and reconstruct selected plans into CAD with structural and dimensional correction. The net time gain is approximately 1.2 days per project at schematic phase — meaningful but not transformative. The transformative gain is in the breadth of spatial exploration that was simply impossible before AI. Architects can now test ideas that time constraints previously made unreachable.
You now understand how to generate AI floor plans and what to do with them. In this lab, work through applying the triage protocol to a realistic scenario — and plan a reconstruction workflow from AI candidate to permit-ready document.