Graphisoft · AIA 2024 Washington DC · Media Architecture Awards Budapest
AI Design Live Experience
Build it with your hands, and watch AI turn it into architecture.
What
An interactive installation where visitors arrange blocks and everyday objects in front of a camera, and generative AI turns the composition into architectural imagery in real time.
My part
Concept, prototyping, visual direction and refinement: tuning transitions, visual effects and the predefined prompts in TouchDesigner.
Result
Shown at the AIA Conference on Architecture & Design 2024 in Washington, DC, and the 20th Media Architecture Awards in Budapest.

01 — context
Model-making, meet generative AI
Early-stage architectural design is tactile and experimental: blocks, boxes, quick models. We wanted to bring that spirit into generative AI. I worked on the concept with two other designers, including my mentor, and with technical teammates who built the pipeline: Stable Diffusion generates the images, TouchDesigner runs the workflow and the visual effects.
Visitors arrange objects, and the AI turns them into architecture live.
02 — the problem
Make AI feel like a material, not a prompt box
The experience had to work for two very different audiences at the same time.
Architects
Industry professionals at the AIA conference, who know exactly what good architectural imagery looks like.
The public
Visitors at the Media Architecture Awards with no AI or design background, who needed to get it in seconds.
Live
Every change on the table had to show up on screen quickly, while still looking like architecture.
03 — key decisions
Hands first, prompts in the background
Physical objects as the input
Instead of typing, visitors arrange wooden blocks, geometric shapes, boxes, cups or their own items in front of a live camera. The composition becomes the starting point for the image, so moving one object changes the result. It brings back the experimental feel of early model-making.
Short prompts, rich results
Visitors only had a short input, like "medieval architecture" or "futuristic cityscape". I helped refine the predefined prompts and the underlying architectural prompt behind them, so a few words were enough for rich, coherent architectural imagery.
Fluid, but still convincing
The images had to change smoothly as visitors moved objects, and still look good at every moment. In TouchDesigner, I worked on the balance between fluid transitions and visually compelling results, and on the mix of visual effects.
04 — process
Prototype, test, refine
Concept
Defined the experience with the design team: tactile input, AI output, open to anyone.
Prototype
First tested in ComfyUI with Lego models, then moved the workflow to TouchDesigner with the technical team for the live version.
Test
Internal tests, then tests with visitors.
Refine
Iterated on prompts, effects and transitions to make the results more engaging and the possibilities easier to explore.

An early prototype in ComfyUI: a Lego model in, an aerial view of a building out.
05 — the result
Two stages, one conversation
AIA 2024
AIA Conference on Architecture & Design, Washington, DC
Media Architecture Awards
The 20th edition, at Uránia in Budapest
Visitors experimented with different arrangements and asked how the system worked. The installation showed AI-assisted visualization not just as an image generator, but as an interactive medium for creative exploration, and it started conversations about the future of architectural design.

At AIA 2024 in Washington, DC: visitors built with blocks, and the screen turned them into architecture.
06 — what I learned
Designing for play
The input shapes the experience
Arranging objects instead of typing made the AI something people could play with. The interface was the table.
Prompts are design material
Much of the quality came from prompts visitors never saw. Writing and tuning them was design work, not setup.
Real time is a balance
Fluid and impressive pull against each other. Finding the balance took testing with visitors, not just tuning on our own screens.





