How it is made

The render is downstream of the catalogue

Nobody types a paragraph and hopes. A design is a set of choices made against described vocabulary, and the prompt is composed from those choices — which is why it can be checked before it is spent on, and audited after.

None of what follows is a model we trained. It is rules, notes and counts: the sort of thing a house can argue with.

The order it happens in

  1. Step 1

    The brief is composed, not typed

    A designer chooses from the catalogue. The prompt is then written out of those choices — each chosen word replaced by the description somebody authored for it.

  2. Step 2

    It is checked by rule, before anything is spent

    A craft with no description, a motif with no placement, an unset region on an ethnic brief, a word still awaiting review — each is a fault with one specific remedy.

  3. Step 3

    The render is told what it must not become

    For the garments that have actually been got wrong, the prompt carries the specific drift to avoid — and where the cloth is, not only where it is not.

  4. Step 4

    The result is read back and held against the record

    Through the same vision instrument used to catalogue somebody else’s photograph, with the claim withheld from it.

What the prompt is allowed to say

Composition adds no ingredients. The designer picks the crafts, the cloth and the motifs; the brief describes the ones they picked and nothing else. A term with no authored description contributes no description at all rather than a guessed one — the same refusal to fabricate that governs the rest of the platform.

Length is rationed, because a prompt is read front to back with diminishing attention and a sentence that adds nothing quietly deletes whatever falls off the end. Three terms per dimension: a prompt naming eight crafts describes none of them.

The garment’s own description gets a longer allowance than the terms hung off it, and that was learned expensively. A shweshwe dress lost the clause naming its cut to a mid-word truncation, the region’s rules filled the vacuum, and it came back as an Indian sari in indigo. A craft is an adjunct; the garment is the subject of the sentence.

Where the work sits, not just what it is

A craft description says what mirror work looks like. It does not say that the mirrors belong massed on the choli yoke and scattered over the odhni. Both are carried now — and for every chosen craft, not only the garment’s famous one, because a customer who picks three crafts and gets one placed gets one rendered.

Descriptions, not provenance

Many notes were written for a human reader and carry a second sentence about where a thing comes from. That half is useful to a person and noise to an image model, which will happily render a map of Gujarat into a border. The visual sentence is the one that travels.

Naming what it must not become

For the garments that have actually come back wrong — and only those — the record carries three things a description cannot: the structural facts that must be visible, where this garment’s own crafts sit on it, and the specific drift to avoid. Naming what a render must not become is the single most effective instruction these models take.

A chaniya choli must not become a bridal lehenga, a smooth skirt with one wide gold border and no bands, a saree, or one continuous length of cloth from hem to shoulder. Every item on that list is a render that actually happened.

And the counter-lesson, which cost two more renders to learn: a negative gives a diffusion model nothing to draw. “Not a saree pallu” was in the prompt and the odhni went across the torso anyway. It is now told where the cloth is — pinned at one shoulder and falling beside the arm, clear of the chest. An earlier attempt said “straight down the back”, and the model obliged by leaving it out of the picture altogether.

These rows are added only for garments that are got wrong. A plain kurta renders as a kurta, and a rule for it would be words spent to say nothing.

Before the money: rules, and a ruling on the cloth

Faults are structural, so they are free to find

Most of the ways a design goes wrong can be detected by rule at no cost: the brief named a craft nobody has described, so the prompt said its name and the model guessed; a motif had no placement, so it went wherever; an ethnic brief carried no region, so it carried no regional idiom at all.

Each has one remedy — write that description, place that motif — and once the remedy is done, that fault cannot recur for that word. An observed failure becomes a named authoring task, and finishing the task permanently removes a class of error.

Can this cloth carry this work?

Asked against what has already been ruled, then against what the market has actually been seen doing, and only then asked properly and remembered forever. The answer is not a refusal — it is what the platform says before it builds, along with what it would offer instead: georgette will not carry karchobi, and on that cloth the request is kamdani.

A person who insists still gets what they asked for. And a ruling that came from research rather than from practice is stored saying so, awaiting review, rather than passing itself off as fact.

Afterwards: looking at what we actually made

Checking a brief by rule is good and structurally blind — it has never seen a picture. So four chaniya cholis could be tagged with mirror work, rendered without a mirror on any of them, priced, matched to makers on that mirror work and filed in the library, and nothing in the platform noticed.

The finished render is now read back through the same vision instrument already trusted to turn a photograph of a garment into catalogue terms, and what it saw is diffed against what the record claims. It is not a new kind of intelligence; it is the instrument we already own, pointed at our own output instead of somebody else’s.

The reader is not told the answer. The vocabulary is sent — the reply has to come back in terms the catalogue uses — but the claim itself is withheld, because a check that whispers “it’s a chaniya choli” to the model it is about to ask “what is this?” is theatre.

Silhouette and craft are hard checks: a wrong garment is the worst thing this can find. Colour, motif, cloth, occasion, region and cut are soft — a reader that misses an idiom is reporting something worth knowing, not proving the design wrong. Price band is not checked at all, because nothing in a photograph distinguishes a premium tier from a mid one.

What it finds is counted per word. An element that has historically failed to appear is pushed to the front of the next prompt on the first render, rather than being corrected afterwards at the cost of a second one. It is a count and not a score, so the reason a design was changed can be explained to the person whose design it was.

What a house is handed

A spec sheet, with its gaps showing

Three controlled renders, one variable each, established that shape renders and sub-detail does not: a band collar draws clearly, a button-down collar comes back with the collar’s shape and without the buttons that make it a button-down. So canvassing, buttoned collar points, pick stitching and lapel roll have few or no pixels and can only ever live on a written sheet.

Every field on it is a value read from the record or blank, and blank travels all the way to the page so the sheet can say the record does not hold it. Nothing falls back, defaults, or infers.

A cost band built from labour

The silhouette’s own making cost, plus how much labour each chosen craft adds. Those multipliers are what stop a supplementary-weft handweave from being priced like machine sequinning — which is exactly what happened while they were missing and a single default was standing in for them.

They are estimates and are held as estimates: editable by whoever actually knows the costing, without a deploy. A silhouette with no basis gets no band, and the screen says so rather than inventing a figure.

The limit

None of this makes a render correct. It makes a wrong one findable, attributable to a named word, and cheaper to fix than to re-do blind — and it means the platform can tell you it got something wrong before you do.

Some things remain outside what a picture can settle at all. Cloth weight and hand, fibre content, and how a garment moves are judged from a sample, not from a screen.

Bring the garment you think it will get wrong.

It is the only useful test. Name the garment, the region and the craft, and we will show you what the catalogue holds for it — including where it holds nothing.