AI for Architecture: What Actually Works

Until 2023, the first question anyone asked a visualization studio was “how much does a render cost”. Today another question comes first: why would I hire a studio if AI already generates architectural images in seconds?

It is a fair question and it deserves an honest answer rather than a defensive one. Part of it is a genuine threat — for anyone selling a pretty picture as the final product, the bar has moved. The other part hides a misunderstanding about what a launch image has to be. And that is where the whole discussion is settled: between looking plausible and matching the construction documents.

In architecture, “AI” usually means two very different things, and conflating them causes most of the confusion:

Generative models, which produce a new image from a text description or from another image;
Models applied inside the technical pipeline, which speed up steps of a process whose result is still determined by a 3D model.

The first category is the one in the headlines. The second is the one that already changed studio work — quietly, and longer ago than most people assume.

What generative AI already does well

Worth starting with what works, because it is a lot:

Exploring massing and mood during concept. Dozens of facade, palette or atmosphere variations in minutes, so you can discard directions before modelling anything. The cost of being wrong drops to nearly zero, and this is exactly the phase where being wrong is cheap and useful.
Moodboards and expectation alignment. Instead of hunting for a reference that comes close to what the client has in mind, you generate the reference. The “what mood are you after” conversation gets specific in one meeting instead of three.
Denoising and upscaling inside the render itself. GPU render engines have shipped neural denoisers for several years — that is why a clean preview appears in seconds rather than minutes. That is already AI in the workflow; it simply is not marketed as such.
Post-production. Automatic selection and masking of sky, vegetation and people; removal of unwanted elements; frame extension. Retouching that used to be slow and manual is now routine.
Support work. Transcribing the briefing call, translating a materials schedule for an overseas client, drafting the first version of descriptive copy. Nothing glamorous — and real hours saved every week.

The common thread across all of these: AI operates on material that already exists, or on decisions that have not been locked yet. In none of them is it the source of geometric truth.

Where it still does not replace 3D rendering

Much of what is marketed today as AI for architecture is, in practice, an image generator with an architectural vocabulary. That is not an insult — it is a description of what the technology does. And what it does runs into five limits that matter precisely when a project becomes a sales asset.

1. Fidelity to the construction documents. A generative model is optimised to produce images that are plausible within the distribution it was trained on, not to reproduce one specific geometry. Even with depth-map or edge conditioning — the techniques that tie generation to a base drawing — what you get is silhouette similarity, not dimensional correspondence. Ceiling height, window width, sill height, number of treads: all of it emerges from what looks right, and what looks right has a wide margin.

2. Consistency across images. A render comes out of a three-dimensional scene: move the camera and it is still the same room, with the same cabinet in the same place. Two independent generations from the same prompt share no scene at all — the cabinet migrates, the floor tile layout changes, the light fixture becomes a different model. For a launch with a dozen images, floor plans and a tour, coherence between the pieces is not fussiness: it is what makes the set read as one property rather than an album of similar illustrations.

3. Specified materials. A materials schedule names products: that porcelain tile, that window line, that paint colour. A generative model delivers something like it. The gap between “like it” and “specified” is the gap between an image and a sales document.

4. Targeted revision. The client asks to raise the counter by five centimetres and change the sink. In a 3D scene that is an edit; in a generated image it is a new generation — and the new generation changes other things along with it, unannounced. An approval cycle runs on isolated changes, and isolated change is exactly what free-form generation does not offer.

5. Commercial consequence, not just aesthetics. Off-plan sales material is not free creative work. In Brazil, sufficiently precise advertising information forms part of the offer and binds whoever published it — consumer protection law treats the asset as a promise, not an illustration. Many jurisdictions take a comparable view. An image whose origin cannot be traced back to the approved project is a contractual risk before it is a matter of taste.

These limits become even clearer in deliverables that are not “an image”. A humanized floor plan needs correct areas and furniture at scale to do its job, which is to let someone judge whether the sofa fits. A 360 virtual tour requires a navigable scene that stays coherent in every direction at once. Neither is an image: both are projections of a metric model — and it is the model, not the image, that carries the information.

How to evaluate a tool before adopting it

The useful question is not “is this AI any good?” but “good for which step?”. Seven objective criteria, answerable from the documentation plus one test:

Control over the input. Does it accept your geometry — model, clay render, depth map — as a condition, or only text? A text-only tool is an ideation instrument, not a production one.
Reproducibility. Can you regenerate the same image changing one isolated parameter and holding the rest? Without that there is no approval cycle, only restarts.
Editability of the output. Does it deliver layers, masks and print resolution? Or a flattened file that only accepts retouching on top?
Fit with the workflow. Does it read the formats already moving between designer and studio, or does it require remodelling what is already done?
Rights and provenance. What is the commercial licence on the output? Does the tool state what it was trained on? For advertising that will run in paid media, the answer matters.
Confidentiality. Where does the project go when you press the button? An unannounced launch uploaded to a third-party service is a contractual problem before it is a technical one.
Honest total cost. Subscription plus learning time plus rework when the output is wrong. The third item usually decides it, and it is the one nobody publishes.

One test separates promise from delivery and costs nothing: run the tool against a completed project whose correct result you know by heart. If what comes back does not match what was built, you have just measured the error with the right ruler. A tool that scores well on criteria 1, 2 and 6 can enter production; one that fails them may still be excellent — for concept work, early on, where none of that is required.

What this changes for clients

If you commission visualization, four practical consequences:

Ask where AI enters the process and where it does not. A good answer is specific: which steps, with what controls, verified how. A vague answer in either direction — “we use none of it” or “we use AI for everything” — says more about the process than about the technology.
Require traceability. Every image going into sales material must be traceable back to the approved model. That is what holds the asset up if anyone questions what was promised.
Do not buy an image; buy correspondence. The 3D rendering that carries a launch is not the prettiest one: it is the one where the delivered apartment and the published image are recognisably the same thing.
A cheap tool does not make the process cheap. The cost of visualization was never in clicking render. It is in reading the construction documents, resolving conflicts between disciplines, running the revision rounds, and standing behind what the image claims.

AI has genuinely shortened the distance between an idea and a convincing image. It has not shortened the distance between a convincing image and a project that exists, was approved, and will be built — and that second distance is what a visualization studio is hired to cross.

GRON 3D has worked in architectural visualization since 2016, from Uberlândia, Brazil. If you want to discuss where automation helps your project and where it gets in the way, talk to our team.