TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Judgment · AGE 18—21 · POSITIONAL · HELD

Visual forensics

Every generated image is a witness to be cross-examined, not a reference to be trusted.

The dilemma

An image is in front of you that you want to be true. A long span, a thin slab, a wall of glass meeting a stone roof with no visible frame. You can enjoy it, or you can cross-examine it. You can do both in the same minute. Only one of them is your job.

The choices

Ask "do we like this?" Or ask the Lab's forensic question, what would fail first if we tried to build this?, and go looking. Where is the load path? Where is the light coming from, and do the shadows agree? Is this material behaving the way that material really behaves? Does this junction resolve, or does the image go blurry exactly where the detail should be?

The consequence

Skip the forensics and the impossible enters the project as an expectation: the client's, the student's, sometimes the tutor's. It is discovered weeks later, at coordination or costing, by someone who never made the promise. Run the forensics and the image stays in its place: a useful witness, cross-examined, never allowed to sign anything.

The case

An image of a library in Mysuru: a stone vault, a roof-light running its full length, a reading balcony cantilevered from one side. The student goes looking. The vault has a roof-light where its compression ring should be. The balcony has no beam, no bracket, no thickening. Only the rendering is holding it up. The shadows say late afternoon. The roof-light says noon. Three findings in four minutes. The image is still beautiful. It is also now filed where it belongs.

Our position

Every generated image is a witness to be cross-examined, not a reference to be trusted. Visual forensics is the first stage of the Lab's working protocol, diagnosis before direction, and it exists to stop fantasy getting into the project.

Why we hold it

Images now reach the studio before the brief does. Clients arrive with renders forwarded on WhatsApp. Those images set expectations, and a studio that cannot take them apart on sight will spend weeks trying to build something that should never have entered the work. The habit is a professional defence, not an academic nicety.

The strongest objection

Running forensics on every image turns the studio into a checking department and trains students to look for failure before they look for possibility. The best early-stage images are supposed to be structurally impossible. That is what a dream is for. A forensic reflex applied too early can cost the project the idea it was about to have.

What would make us revise it

The Lab already separates dreaming from grounding, and forensics belongs at the boundary between them, not inside the dream. We would revise the card if we saw forensics-trained students producing timid early work (fewer risks, safer images) across a whole cohort. We watch the log for exactly that.

Try it

Mark up a render. Point to every part of the structure and lighting that could not be real, with a reason for each. Then take three good-looking images and rank them by how much of each one would survive an engineer's first look.

Take it to crit

Hand the student a beautiful image. Time how long it takes them to find the load path that is not there. Under a minute means they have the habit. Over five minutes means they were enjoying it.

How it works

Image models learn surfaces, not statics. They know what a cantilever looks like from a thousand photographs. They do not know why any of those cantilevers stood up. So the model produces the look of structure with no structure underneath, and it produces light by resemblance: a shadow where shadows usually go, not where this particular sun would put it. Those two facts predict the two commonest findings, missing load paths and light that contradicts itself. Junctions blur for a related reason. The photographs the model learnt from rarely show the detail, so the model goes vague where it has no pattern to draw from.

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