If You Only Describe the Shirt, the Model Picks the Trousers

Here is a prompt that looks completely reasonable, and that anyone selling apparel online would write on their first try:

Photo of a model wearing a blue linen shirt, studio background, natural light.

You will get a blue linen shirt. You will usually get grey or beige trousers with it, because that is the most common thing in the training data next to a blue linen shirt. And every so often you will get a person wearing a blue linen shirt and, below the hem, nothing you can put on a listing.

It is not a filter failing. Nothing in that prompt asked for trousers. The model completed what you specified and improvised everything else, which is exactly what it is built to do. Underspecification is not a neutral choice. It is a decision you handed to a system that has no idea you are going to publish the result on Myntra tomorrow.

Why this happens more with product photos than with anything else

If you have prompted image models for illustrations or landscapes, you may never have hit this. Product photography is unusually exposed to it, for three reasons that stack.

1. Your prompt is a product description, not a scene description. You are pulling text from a catalogue field. Catalogue fields describe one item, because that is the item for sale. That habit produces a prompt with exactly one garment in it.
2. Apparel training data is full of partial bodies. Crops at the waist, crops at the shoulder, flat-lays, mannequin torsos, detail shots of a cuff. A model trained on that has seen an enormous number of images where the lower half of a person simply is not in frame or is not clothed. It has no strong prior that a full-body shot must be fully dressed.
3. You are generating at volume. One image, you would have spotted it. Two hundred SKUs on a Sunday night, exported straight to a bulk upload, and you are relying on a spot check that nobody does properly at 11pm.

The third one is what turns an odd generation into a real problem. Nobody reviews the two-hundredth image as carefully as the first.

The cost is not embarrassment, it is your seller account

If you are selling on marketplaces, this is not a taste issue. Every major platform has an explicit image policy, and every one of them treats nudity or partial nudity as a listing-level violation rather than something to email you about. Amazon, Flipkart, Myntra, Meesho and Nykaa all reject on it, and repeated violations escalate to the account, not the SKU.

Meta is stricter still, because ad review is automated and unforgiving. A rejected creative is a wasted afternoon. A pattern of rejected creatives is a restricted ad account, and getting one of those reinstated is a genuinely bad week.

So the downside is asymmetric. The upside of a slightly shorter prompt is that you typed forty fewer characters. The downside is a compliance strike on the channel your revenue depends on. That is not a close call.

Write the whole person, not the hero garment

The fix is not clever prompt engineering. It is just being complete. Name every garment zone, even the ones nobody is buying.

Two things worth noticing about that table. First, the rewrites are longer but not smarter. There is no incantation, no magic token, no negative-prompt trick. You are simply removing the model's freedom to improvise in the places where improvisation costs you.

Second, the accessory row matters most and gets forgotten most. If you sell earrings, bags, watches or footwear, your prompt is about a small object and says nothing at all about the person wearing it. That is the most underspecified prompt of the lot, and the one people are most surprised by.

The other fields people leave blank

Clothing is the one that bites hardest, but it is the same failure in every direction. Anything you leave vague, the model resolves on your behalf, and it resolves toward whatever is most common in its training data rather than whatever is right for your brand.

• Age. Leave it out of an apparel prompt and you are letting a generator decide the apparent age of a person who is going to appear in a paid advertisement. Always state an adult age range explicitly. This is the one to be least relaxed about.
• Pose and framing. "Model wearing X" with no framing gives you a lottery of crops, and a waist-up crop is exactly the framing where the missing-garment problem hides.
• Background. Unspecified backgrounds drift toward busy lifestyle scenes, which most marketplaces reject for a primary image. Say "plain white studio background" if that is what the listing needs.
• Skin tone and body type. If you serve a specific market, say so. Otherwise the output will quietly converge on whatever the training set over-represents, which for most models is not your customer base.

Check before you publish, not after

Whatever tool you use, a thirty-second review habit catches nearly all of this. Sort your batch by thumbnail, scan the lower third of every full-body frame, and look specifically for waist-down gaps rather than reading each image as a whole. Your eye will skip a missing garment when you are assessing an image for colour accuracy. It will not skip it when that is the only thing you are looking for.

And do the scan before the bulk upload, not after the rejection email. Marketplace appeals cost days.

The general lesson

Image models are not being perverse when they do this. They are doing the only thing they can do with a prompt that describes one garment on a whole human being. The gap between what you meant and what you wrote is where every strange generation lives, and with product photography that gap has a specific, predictable, compliance-shaped consequence.

Describe the whole outfit. It is a duller prompt and a much better photo.

Published 27 August 2026. Written from our own experience running an on-model generation pipeline; the failure mode described is common to current image models generally, not specific to any one vendor.