TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Judgment · AGE 18—21 · FACTUAL · EVOLVING
Styles known unevenly
The most likely image of a rare style is a blend of nearby familiar ones.
The dilemma
You ask for "a Gothic cathedral interior" and the result may pass a historian's glance. You ask for "a Hoysala temple mandapa" and you may get the right mood with the details wrong: the pillar count, the plinth, the way the ceiling is built. Same tool, same confidence. How much of its history do you trust? The only way to know is to test it.
The choices
Trust the model's style knowledge evenly, because it sounds even. Or test it. Prompt a heavily photographed style and a lightly photographed one, compare each against a real building, and map where the knowledge is solid and where it is bluffing.
The consequence
Trust it evenly and the under-photographed traditions, which include many of India's regional ones, get reinvented by the machine and passed off as reference. Students "learn" a regional style from a likeness of it. Test it and you know where the tool is a quick way into a tradition and where it is a confident stranger. Both are useful, as long as you know which one you are talking to.
The case
A history assignment: draw a comparative plate of a Chalukyan temple and a Palladian villa. A student uses an image model for both as a starting reference. The Palladian plate is close enough to correct the proportions against. The Chalukyan one has a stepped shikhara from the wrong region and a pillared hall with the wrong bay rhythm. Correcting it took the student longer than drawing from the survey book would have. The tutor's question was why the two results differed. The answer is the number of photographs.
Try it
Prompt a canonical style and a lesser-known regional one, and compare each result against a measured drawing or a good photograph. Then prompt the regional one three ways — in English ("Hoysala temple mandapa"), in Romanised Kannada ("Hoysala devalayada mantapa") and in Kannada script ("ಹೊಯ್ಸಳ ದೇವಾಲಯದ ಮಂಟಪ") — and compare the three. A name the model knows in one script and not another tells you which captions it learnt from. Then find one case where the model got the look of a style and missed the logic behind it: the structural reason, the climatic reason, the ritual reason.
Take it to crit
Can the student tell where the model's style knowledge is solid and where it is bluffing, and prove it with a real building beside the output? Ask them which Indian traditions they would never use it for, and why.
How it works
How well a model knows a style depends on how many labelled images of it the model saw. Canonical European styles are photographed, captioned and written about in vast quantity. Many Indian regional traditions are not, or are captioned vaguely. Where the examples are thin the model does what it always does, which is produce the most likely image, and the most likely image of a rare style is a blend of nearby familiar ones. A 2023 study tested text and image models on architectural styles and asked where they distinguish correctly and where they fabricate. It also looked at how practitioners actually prompt these tools. What comes back is resemblance, not reasoning about why a style is the way it is. The Lab has not yet published a controlled run of this comparison on Indian traditions; the TRY IT below is that test, at your own desk. EVOLVING: better-captioned datasets and regional fine-tuning could change the map, and the card is reviewed as they appear.
What this idea builds on
What this idea opens up
- Nothing yet names this as a foundation.
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.