Here is a failure every ethnic-wear seller who has tried AI photos knows: upload a thigh-length kurti, get back a beautiful image of a knee-length one. The model didn't render your garment; it rendered the category. Telling it "keep the length accurate" in a prompt barely helps — instructions lose to statistical priors.
For a seller, the wrong length is a wrong product photo, exactly like the wrong colour. It is also a returns machine: the buyer ordered what the picture showed.
Measure, then anchor — never just instruct
CatalogX's answer has three parts:
- Measure. The analysis pass measures comparative proportions from your own photo — the garment's length-to-width ratio, and its length against a companion bottom when one is in frame (the strongest signal on co-ord hanger shots). Vision models are unreliable at absolute measurement but decent at comparative ratios, so the bands are anchored to a known reference garment and calibrated on real products.
- Classify carefully. The measurement maps to a hem category — hip, upper-thigh, knee, longer — with a deliberate asymmetry: calls that shorten the garment act on medium confidence, calls that lengthen it require high confidence, because the failure mode only ever errs long.
- Anchor. The measured category selects the length control automatically and injects an evidence-quoting instruction into generation. Your explicit length choice, if you make one, always outranks the measurement.
Worn photos are read differently
If your photo shows the garment on a person, proportion ratios are meaningless — so the hem is read directly against body landmarks instead, and a worn bottom is never mistaken for a companion piece on the hanger.
Enforced at review time, not just requested
In Quality mode, the reviewer receives the expected hem: output at the wrong length fails the garment-match check and triggers the automatic, steered re-shoot. Length isn't a hope; it's a pass/fail criterion.