Graphisoft · Archicad 27 · 2023
AI Visualizer
Generative AI for architects, back when it still needed a terminal.
What
Turns a simple mass model in Archicad into images of a building, for early visuals and ideas without detailed modeling.
My part
One of two designers on the interface. I did the UI ideation and prototyping, and we tested it together with 4 architects.
Result
Won the GRID hackathon (Business). Shipped 16 Nov 2023 as a free add-on for Archicad 27. Now a product line, up to 3.0 in Archicad 30.

01 — background
Stable Diffusion was usable, but only for enthusiasts
Early in 2023, generative AI was suddenly everywhere, and we wanted Archicad to be part of it from the start. Getting there wasn't simple. I set up Stable Diffusion on my own machine through the Terminal and AUTOMATIC1111, the go-to interface back then: a single page packed with technical parameters. Each image took several minutes, and you never knew what you'd get.
We wanted architects to skip all of that. The AI would live inside Archicad: block out a few volumes in the model, and get back a picture of a building. No extra app, no command line.

What Stable Diffusion looked like in 2023, and what we kept out of Archicad.
02 — from many to one
From a Creative Day to a shipped add-on in six months
Creative Day is a voluntary day organised by our design team. Four of us used it to see how far AI rendering could go in Archicad: images from sketches, from site photos and from mass models, facade variations, interiors. For the GRID hackathon we kept just one direction: turning a mass model into inspiration. It was the easiest to build and the one people would get the most out of. Reworking an existing render made the AI invent details that weren't there, while a handful of simple boxes could turn into something rich: realistic to build, and worth having.
Creative Day
8 AI ideas tested in a day.
Hackathon
4 days. Figma prototype + working demo. Won Business.
Design + testing
Narrowed the scope, tested with 4 architects.
Launch
Free experiment for Archicad 27.
↑ shipped!
Product line
Moved to the cloud, in paid plans.
03 — key decisions
Two or three clicks to an idea
The first concept wanted to cover everything: interiors, styles, weather, seasons, time of day. For version one we cut that back on purpose. The tool was meant to spark ideas, not to produce client-ready images, art styles or finely tuned results.
Cut it down to three controls
Prompt, shape fidelity and image size. The extras stayed on the list for later.
Name the control after the architect's question
“Creativity” describes the AI. Shape fidelity describes what the architect actually wants to know: how closely will the result follow my model?
Make an empty prompt still work
Every input is wrapped in a hidden prompt that already asks for a photo-quality building. An empty field still returns a building, and off-topic text can't pull the result off course.
Speak the architect's language
Model terms became plain words or disappeared behind good defaults. Developers got the exact model value behind every control.
Errors that say what to do next
The engine ran locally, could fail to start, and was sometimes shared. The dialog explains what went wrong and warns before anyone stops a shared engine.


GRID concept, May 2023 → shipped dialog, November 2023.
04 — testing
The help was there. Almost nobody saw it.
To get good results, the model needed things architects rarely add, like ground under the building and a tree for scale. The prototype explained this three ways: an illustration, an example prompt and hover tips. We tested it with four architects. The results landed; the help mostly didn't.
3 / 4
missed the hover tips completely.
1 / 4
followed the illustration.
3 / 4
misread shape fidelity. One dragged it to 0%, expecting a more building-like image.
They fixed the model, not the prompt
When a result missed, they edited the geometry instead of the text. Shape fidelity turned out to be the control that mattered most.
“This is cooler than the last one!”
A tester, after adding ground in her third try
We didn't fix it with better onboarding. The developers changed how the model view reaches the AI, so ground and trees were no longer needed. The tutorial never shipped, and the step people kept missing was simply gone.
05 — the reception
Out in the world, for free
On 16 November 2023, AI Visualizer went out as a free experiment. Graphisoft's announcement put it simply: “you don't need to learn prompt engineering.” AEC Magazine, Architosh and Engineering.com all wrote about it. Feedback from the community was mixed, and refreshingly honest.
Installing it was the hardest part
The engine took 16–20 GB, system drives filled up, and Mac installs failed. Plenty of people never saw a single image. That was the price of running AI locally in 2023.
Not everyone saw the value
Some dismissed it as a “3D mood board” and felt the images strayed too far from their model.
Put to work by professionals
One firm ran it on a live project, comparing the convenience of staying in Archicad with more powerful standalone tools.
Used exactly as intended
A teacher watched weak student models become “excellent results”. One user even suggested calling it “AI Inspiration”, which was the whole point.
06 — where it went
From experiment to product line
In Archicad 28 it moved to the cloud: nothing to install, included in paid plans. Our local version was retired at the end of 2024, but parts of our dialog are still there: Lock result, Iterations (default still 35), Prompt strength and the 800×600 default size. A new version has shipped with every release since, up to 3.0 in Archicad 30.
Its headline feature, Image from composition, is still our original idea: the model sets the brief, the prompt sets the style.
note:
I worked on the experimental first version, as an optional project next to my main work. Once it became a product, the team responsible for that area took it further.
07 — what I learned
What I took from it
Users steer with what they know
We designed around the prompt. Architects steered by modeling.
Visible isn't the same as seen
The advice was correct, and almost nobody read it. Removing the need for it worked better than explaining it.
Designing for AI, not with it
In 2023, AI was the material we shaped into a product, not a tool that helped us design.
08 — beyond the product
Taking it off the screen
Later we took the same technology off the screen and made it interactive: instead of typing a prompt, visitors arranged blocks and objects in front of a camera, and the AI turned them into architecture live.







