TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 19—21 · FACTUAL · EVOLVING

Teaching the model your language

Training is a commitment, so use up the zero-training paths first.

When to use

When you have tried prompting, conditioning and reference images, the model still cannot produce your studio's visual language, and you hold enough rights-cleared, consistent work to teach it.

The method

Audit the archive first. Which images actually carry the signature — typology, light, composition — and which are noise? Caption each one carefully. The captions are a large part of what the model learns to associate with the images; the learning comes from the pairing of each example, its caption and the parts of the model being adapted. Start with the lightest method the tool offers before any full fine-tune. Today that is usually a LoRA: a small adapter trained beside a frozen model. Train, then test on prompts the model never saw in training. Judge on work near your own typology. The signature carries best where the target is close to the training set. Record what the trained model can now do and what it still cannot.

Watch for this

Training to fix what a reference image would have fixed. Most studios should use up the zero-training paths first: conditioning, references, retrieval. Training is a commitment. It needs rights, compute and upkeep, and the model ages.

Try it

Choose fifty images of your own work. Before training anything, write down what you expect the model to learn from them. Then write down what it cannot learn from pictures: why you made each choice. Keep both lists for afterwards.

Prove it

Explain what a fine-tune can change in a model and what it cannot, and make the case for when your own body of work is worth training on and when prompting is enough.

How it works

Fine-tuning adjusts the weights on new data. LoRA freezes the base model and trains small low-rank matrices inside selected layers beside it, which is cheaper and keeps the base intact. It is a general adaptation method — used for a language model's behaviour, a domain, a task or a format as much as for an image style; this card takes the image case because that is usually a studio's first need. Either method shifts what the model treats as likely. Neither adds judgment, and neither carries a signature over to a typology the training set never held. EVOLVING: LoRA is today's standard lightweight method and the tooling changes yearly. The reasoning about when to train is the durable part.

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