TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Judgment · AGE 18—21 · POSITIONAL · HELD
Variation is not difference
Re-rolling gives you variation, and changing the brief gives you difference.
The dilemma
You need alternatives for a jury: real ones, schemes that would take the project somewhere else. The tool gives you eight images that all look different: warmer, cooler, a taller roof, a wider court. Are these alternatives, or one idea dressed eight ways?
The choices
Count surface differences and call them options. Or test each pair with one question: if I chose this one instead of that one, would the project go somewhere else? A different section, a different structure, a different way of living in it? If not, it is variation. If yes, it is difference.
The consequence
Present variation as options and a sharp juror sees one scheme with a colour study attached. Worse, you believe you explored when you only went round in a circle. Find the real differences and you may have only two. But two genuine forks are worth more than eight versions of one idea, and now you know the next difference has to come from the brief, not from the re-roll button.
The case
A student brings five "options" for a hostel block in Belagavi. All five have the same double-loaded corridor. Only the skin changes. The tutor draws the section of each on the board: five identical sections. The student had been re-rolling. The sixth option, made after the brief was changed to "single-loaded, open to the west wind," was the first real alternative, and it took one cycle.
Our position
Re-rolling gives you variation: the model's own range around one reading of your brief. Changing the brief is what gives you difference. A set of variations is not a set of alternatives, and a jury should never be shown one as the other.
Why we hold it
A generated image is a sample from the model's reading of your words. Sampling again gives you another point in the same neighbourhood. To leave the neighbourhood you have to change the words. That is how these systems work, and it is why the Lab teaches brief revision rather than seed-scrolling.
The strongest objection
Variation is sometimes exactly what a project needs. Late in a scheme, when the parti is fixed and the question is material, proportion or light, the "same idea in new clothes" is the right output. Calling it lesser is a bias towards early-stage thinking. Difference is not always the more valuable thing.
What would make us revise it
This card already accepts the late-stage case. We would rewrite it if studio evidence showed students reliably telling the two apart without being taught, which would mean the distinction is obvious in practice and the card is solving a problem nobody has. So far the logs say otherwise.
Try it
Sort a set into "only varies" and "truly differs," and defend where you drew the line. Then find two outputs that look different but carry the same design decision, and two that look alike but differ in intent. The second pair is harder to find, and more useful.
Take it to crit
Ask the student which outputs are real alternatives and which are the same image in new clothes. Then ask them to draw the section of two "different" ones. The two sections will settle it.
How it works
Each generation is a draw from a distribution shaped by your prompt and the model's training. Draws cluster. Eight draws from one prompt give you eight points in one cluster: different in detail, alike in what they assume. An experimental study of text and image models on design tasks found the same pattern: repeated sampling tends to narrow rather than widen. How far it narrows depends on the prompt, the settings and the model, so read it as a tendency to test, not a law. Changing the prompt moves the cluster. Changing the model moves it in a different way again.
What this idea builds on
What this idea opens up
- Nothing yet names this as a foundation.
Sources
- disciplined-iteration-log
- dreaming-engines-and-grounded-generators
- essential-ai-concepts
- Understanding Design Fixation in Generative AI
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.