TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · The Brief · AGE 18—20 · POSITIONAL · METHOD

Brief spatially, verify spatially

A spatial clause in the brief asks, and the check on the output enforces it.

When to use

Any brief with a spatial clause: a count ("three bays"), a relation ("the kitchen opens onto the court"), an adjacency, a left/right/above/below. That is nearly every architectural brief.

The method

Before you run the brief, write one check for each spatial clause in it. "Three bays": count the bays. "Kitchen opens onto the court": find the kitchen, find the court, find the opening. Run the brief. Run the checks one by one on the output itself, not on your impression of it. Record which clauses held and which the model dropped. Put the dropped ones back into the brief, stated again — or move them into a grounded step that can hold them. A spatial clause in the brief asks; the check you run on the output is what enforces it.

Watch for this

Checking the image for beauty and forgetting to check it for truth. An image can be compelling and have four bays where you asked for three. The eye forgives counts. A client's surveyor does not.

The Lab's note

The weakness is measured; the loop is ours. Benchmarks establish that current image models drop spatial clauses. Writing the check before you generate, and running it on the output rather than on your impression of it, is the Lab's method for living with that — not the only one. Others check afterwards, or move every spatial clause into a conditioned step from the start. Ours is the one that leaves a record of which clause failed.

Try it

Write a brief with three spatial clauses and, beneath it, the three checks. Run and check. Most first attempts find at least one dropped clause. Re-anchor it and run again.

Prove it

Catch one spatial instruction the model dropped, using a check you wrote before you saw the output.

How it works

Benchmarks built to test spatial language in image generation show current models failing often on exactly these clauses: relative position, counting, several objects placed together. Part of the cause is the training data. Captions rarely say where things sit in relation to each other, so the models learned how things look better than how they are arranged. Re-captioning images with clear spatial language improves the results. The numbers will move: the weakness is measured today and is being worked on. The loop — brief spatially, verify spatially — holds whatever the numbers do.

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.

Age grows from 11 at the centre to 22 at the edge, and six sectors show the learning strands. Tab into the map and the arrow keys step from idea to idea, following the links where there is one. Enter opens the idea under the cursor, and E reads out its links and the reason recorded on each. Press slash for Search, question mark for the full key list, and Escape to leave. Open Ideas for the complete readable list, including what each idea builds on and what it opens up.

LOGIKA · RBDS AI LAB INDIA
ON-RAMP · AGE 11 · FIRST ENCOUNTERS, NOT GATES — IDEAS · — DEPENDENCIES
DONE
OPENS NEXT
SOLID — STANDS ON