TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 17—19 · POSITIONAL · HELD
Grounded generators
Grounding adds no design thinking, so a weak drawing comes back well-lit.
Our position
A grounded generator is held to a real input: a plan, a site photograph, a massing model, a control image. This is the conditioned end of the continuum DREAMING opens — reference → image-to-image strength → edge or depth map → real geometry. Its strength is that it stays close to what exists. Conditioning raises adherence to your plan or photograph; it does not make the geometry invariant, so you check it. The catch: the input has to be worth staying close to. Grounding adds no design thinking. It keeps whatever the input carries and dresses it up. A weak drawing comes back well-lit.
Why we hold it
Grounding is the difference between a picture of a building and a picture of your building. It is also the easiest way to give a weak idea a finished face. Naming this end of the continuum makes the student ask, before the render, whether the drawing underneath has been judged.
The strongest objection
No control mechanism warrants "the walls stay where you put them". Conditioning increases adherence; it does not lock geometry, and adherence varies with the conditioning scale, the pre-processing, the model, the prompt and the sampler. A category called "grounded" invites a student to trust the geometry because the input was real, which is the one thing a generated image never earns. The same objection as DREAMING also applies: this is a name for one end of a continuum, not a kind of engine, and the honest teaching is the continuum.
What would make us revise it
A conditioning method that provably preserves the input geometry — a render path rather than a generator — would make "grounded" a literal description and this card a FACTUAL one. Evidence that the name makes students trust conditioned output more than they should would make us drop the name and keep only the continuum. Reviewed every edition.
See it in the studio
A plan with a 900 mm corridor and a bedroom that cannot take a bed, rendered through a control image into a warm, populated interior. The jury sees the warmth first. The plan is unchanged.
Watch for this
Grounding on a drawing you have not yet judged. The render will not judge it for you.
Try it
Take one idea. Run it once free (text only) and once grounded (the same prompt plus your drawing as control). Put them side by side and list what stayed the same. That list is what your drawing actually decided.
Prove it
Explain what a control input holds close and what it still leaves free, using your own two runs — and say why "holds close" is the honest phrase rather than "locks".
How it works
Grounding is implemented as conditioning. Image-to-image starts from your picture (TWO STARTS). Control methods feed edges, depth or a plan alongside the prompt at every step (CONTROL INPUT). A written instruction — "keep the walls where they are" — asks; the conditioning input and its strength are what enforce, and even they enforce adherence, not a lock. The same model dreams or grounds depending on what you hold it to. The pairing with DREAMING comes from Sahil's essay on the two families; 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
- Zhang et al., ControlNet
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.