TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Foundations · AGE 16—18 · FACTUAL · EVOLVING
The model has no ruler
The model has seen a 3-metre wall in photographs and has never measured one.
The idea
A generative image model has no units and no fixed scale. It has seen what "a 3-metre wall" looks like in photographs; it has never measured one. Pixels carry likeness, not dimensions, so nothing guarantees a relationship between a generated image and real measurements. Benchmarks that ask models to rebuild a flat's room layout from photographs find most of them at or below a random baseline and all of them well under people. The picture can be beautiful. The ruler is simply not in the machine.
Why it matters
The moment anyone — you, a client, a contractor — reads a generated image as measurable, fiction enters the drawing set. Knowing there is no ruler keeps generated imagery where it belongs: atmosphere, likeness, exploration.
See it in the studio
A client scales off a generated "plan" with a real scale rule and asks why the bedroom is 2.1 metres wide. Nobody drew that bedroom. The machine painted something that looks like a plan, and the scale rule measured paint.
Watch for this
Generated drawings entering any set that someone downstream might measure. The danger is not the image. It is the image crossing into the drawing set without being marked. A caption asks; the set's own controls enforce — a separate folder for generated imagery, a "not for construction" stamp on the sheet, and a check before anything is issued.
Try it
Generate a plan, print it, and measure five things with a scale rule. Write down the width of a door, a corridor, a stair. Judge whether that building could be walked through.
Prove it
Explain why a model fluent in building photographs still cannot hold dimensions, and what that permits and forbids in your workflow.
How it works
Part of the failure is the data: researchers who studied spatial consistency found that training captions rarely carry precise spatial relations, and re-captioning six million images improved it. Part is the representation: a raster image is not geometry, and the idea pixels are not geometry, beside this one, explains why. Dedicated training helps, and the leaderboard has moved since the first paper; the gap to people stands. EVOLVING: this card carries a review date for that reason.
What this idea builds on
What this idea opens up
Sources
- Petersson et al., 2025 — Blueprint-Bench (arXiv:2509.25229)
- TEC.U
- essential-ai-concepts
- SPRIGHT
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.