TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 17—19 · POSITIONAL · HELD
Dreaming engines
Dreaming is research in week one, and it stands in for design in week eight.
Our position
Some uses of a generative model are meant to invent freely, with nothing held to a real input: mood, atmosphere, what-ifs, images you could not have drawn. The Lab calls these dreaming engines — our name for one end of a continuum, not a kind of machine. The continuum runs: text only → a reference image → image-to-image at rising strength → an edge or depth map → generation conditioned on real geometry. This card is the free end. Its strength: it takes you somewhere you would not have gone. Its danger: the new place can look finished when nothing was designed.
Why we hold it
Because students need a word for the moment, and the field's words — "unconditioned", "text-to-image" — name the setting, not the moment. Early in a project, with a question in hand, dreaming is research. Late, with a deadline, it stands in for design. Naming the two ends has, in our teaching, made students say which end they were at before they showed a picture.
The strongest objection
"Dreaming engines" and "grounded generators" are not accepted technical categories, and naming them as a pair can hide what is actually true: the same model, with the same settings, sits at every point of one continuum, and the difference between the two ends is a conditioning input and its strength. A student who learns two species may look for two tools. The honest teaching is the continuum itself, and this card's name is a convenience laid over it.
What would make us revise it
If students taught the pair turn out to reason less well about conditioning strength than students taught the continuum alone, the names go and the continuum stays. If the field settles on its own word for the free end, we adopt it. Reviewed every edition.
See it in the studio
Week one of a library brief: "what would a reading room feel like under a deep stone roof in the monsoon?" — twenty images, no plan, and a conversation about light. Week eight, the same prompt, dropped onto the sheet as "the concept". Same engine. Only the first was design.
Watch for this
Presenting a dream as a scheme. If you cannot say which question the image answered, it was not a study.
Try it
Make a study with a dreaming engine. Before you generate, write the question you are asking in one line. After, write one line on whether the images answered it. Keep both lines with the images.
Prove it
Name the moments in a project where free invention is the right tool, and show one study where you can say what question you asked it.
How it works
Technically, a dreaming engine is any generator run without conditioning beyond text: pure text-to-image, high denoising strength, no control image. The same model becomes a GROUNDED generator the moment you hold it to a plan or a photograph (TWO STARTS and CONTROL INPUT are the mechanisms). The pair names two ends of one continuum of conditioning, not two kinds of machine. Sahil's essay on the two families makes the case; this card states it as the Lab's position.
What this idea builds on
What this idea opens up
Sources
- generative-dichotomy
- ch03
- Dreaming Engines and Grounded Generators
- TU Delft, AI in Architectural Design
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.