TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Judgment · AGE 18—21 · POSITIONAL · HELD
Atmosphere is not information
The damage happens at the handover, when a mood image gets read as a drawing.
The dilemma
A render of a reading room: dust in the light, a worn timber floor, a deep window seat. Everybody loves it. Is it good? You cannot answer that until you ask an earlier question: what is this image for? To set a mood, or to tell someone what to build?
The choices
Judge the image on feel. Or sort it first (is this mostly atmosphere, mostly information, or one of the rare images that is both?) and then judge it against its own job. An atmosphere image is judged on whether the mood is right and specific. An information image is judged on whether a contractor could act on it.
The consequence
Mix the two up and mood gets mistaken for finished design. The board goes to the client with feeling standing in for facts, and the facts get invented later by whoever builds it. Keep them apart and each image is allowed to be what it is. The mood image sets the direction, the information image gets drawn, and nobody takes measurements off the wrong one.
The case
A student presents a "final" image of a café in Mangaluru. Asked what in it could be built, she starts listing: the window seat (but how deep?), the floor (timber, but which timber, laid on what?), the ceiling (there is light, but no visible source). In two minutes the image has been sorted correctly: nearly all atmosphere, almost no information. It is a good atmosphere image. It was presented as something else.
Our position
Atmosphere and information are two different jobs an image can do, and a designer must know which job each image on the wall is doing. The Lab's dreaming engines tend to produce the first. Grounded generators tend to serve the second. Neither is lesser. The error is confusing them.
Why we hold it
Mood is where design starts. Information is where it ends up. Both are needed, and the damage happens at the handover, when a mood image gets read as a drawing. The Lab has watched that handover fail in practice often enough to teach the sorting before anything else about images.
The strongest objection
The great architectural images carry both at once. A Binet photograph of a Zumthor interior is atmosphere and a lesson in construction in one frame. A strict sorting trains students to expect the split, and may lower their ambition for the image that refuses it. The rare "both" is the goal, not an exception to be filed away.
What would make us revise it
If generated images routinely began to carry reliable buildable information (dimensions that hold, junctions that resolve, light with a source), the sorting would become a check for the rare miss rather than the default reading. We test this every edition by measuring off the prettiest frame with a scale rule. So far the scale rule disagrees with the render.
Try it
Sort a set by whether each image carries mostly atmosphere or mostly information. Then take one atmospheric image and state exactly what buildable information it has, and what it does not. Be precise: "a window seat, depth unknown, sill height unknown, no structure shown."
Take it to crit
When the student presents a beautiful image, ask what in it could be built. Watch whether atmosphere was standing in for information, and whether the student knew which one they were showing.
How it works
Image models learn from photographs, and photographs carry mood well and construction poorly, because the plan is not in the picture. So an engine tuned for evocative output produces atmosphere easily and information only by accident. Grounded generators, the ones that work from your own drawing, model or reference, hold information better because the information came from you. Knowing which engine made the frame tells you which job it is probably doing.
What this idea builds on
What this idea opens up
Sources
- dreaming-engines-and-grounded-generators
- dreaming-vs-grounded
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.