TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 18—20 · FACTUAL · EVOLVING
The spatial instruction test
Generate ten seeds of a prompt with stated positions and score each relation and count.
When to use
Before you rely on any image model to place things where you say — and whenever a new version arrives and someone claims it "handles layout now".
The method
Write a prompt with three or four named elements in stated positions: "courtyard left of the entry, service block behind, water tank on the roof to the right". Generate ten images from ten different seeds. For each image, score each element on four things. Present? Relation satisfied — actually left of the entry, actually behind? Count right — one tank, not three? Viewpoint lets you judge the relation at all? If the viewpoint does not, mark that element unjudgeable, not wrong. Tally per element and per relation, as a rate over the ten seeds. Then swap the order you name things — put "service block" first — and run ten more. If the tally shifts towards whatever you named first, you have found first-object bias. Write the scores down with the model's name, version and the date.
Watch for this
Scoring presence when you meant to score relation. A model can put all four things in the wrong places and pass a presence tally. And scoring by "looks about right". Score each element, each relation, each image. A test without a tally is only an impression.
Try it
Run the twenty images today on whatever tool you use. Keep the sheet. Run the same twenty on the next version. Now you can say "improved" with a number instead of a feeling.
Prove it
Design a small test that proves where a model's spatial placement breaks, scored on relation and count rather than presence alone, and show the first-object bias by changing which element you name first.
How it works
Public benchmarks do exactly this at scale. VISOR scores objects and their stated relations. T2I-CompBench++ covers spatial and other kinds of composition. Both report that current models struggle with prompts that place several objects. Your ten-image version is the same logic at studio scale. EVOLVING: the scores move with every release. The method stays.
What this idea builds on
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.