TL;DR
If a generated face looks off,
or an item that was never in your garment photo shows up in the result,
the prompt usually isn’t the problem — the input is.
Crop out anything that doesn’t belong in the shot (a face, an extra colorway,
an item you’re not even selling), and those errors drop.
When the Result Doesn’t Match What You Uploaded
You upload a garment photo, generate a lookbook, and something’s off.
Maybe the face reads a little strange,
or inconsistent with the model you set up elsewhere.
Maybe an item shows up in the result that wasn’t part of your product at all.
The reflex is to reword the prompt and try again.
That’s often not where the problem sits.
It usually comes down to what else is packed into the garment photo alongside the item you’re actually trying to generate.
When an input carries more than one signal — a face, two colorways,
an unrelated piece of clothing — the model has to decide what to prioritize,
and that’s exactly where errors creep in.
Clean up the input, and a lot of that ambiguity disappears before generation even starts.
Here are three crops worth making a habit.
Crop the Face Out of On-Body Shots
If your garment photo is an on-body shot, it comes with a face in frame —
and that face has to compete with the model face you’ve already set elsewhere in the generation.
Two faces in one request means the model has to reconcile them,
and that’s often where odd or inconsistent faces come from.
Crop the face out before uploading, and only the garment information gets through.
Here’s that in practice.
This jacket photo has the face cropped out of frame before being used as input.
Nothing else changed about the input — just the crop —
and both the face and the jacket come through clean.
Upload Only the Garment You’re Generating Right Now
A second source of confusion is less obvious: two colorways of the same item in a single input.
Say a jacket comes in red and yellow, and you upload both photos at once,
hoping to generate a lookbook for each.
From the model’s side, that isn’t “the same jacket in two colors” —
it’s a mismatch it has to resolve somehow,
and the color it lands on may not match either shot cleanly.
One generation, one garment.
If you need both colorways, run them as separate generations.
Crop Out Anything That Isn’t Part of the Look
The same principle covers whatever else happens to be in frame.
Product shots sometimes catch more than the item you’re selling —
another piece folded in the corner, packaging still in shot,
an accessory that isn’t part of this particular look.
None of that belongs in the generation, but if it’s sitting in the input,
the model still has to account for it somehow.
Crop down to just the item you want, and nothing else.
This applies whether you’re shooting clothing or non-apparel pieces like bags and accessories.
Summary
- A face in your garment photo can compete with the model face and cause face errors — crop it out.
- Two colorways of the same item in one input creates ambiguity —
upload one garment per generation. - Anything else in frame, from other garments to packaging to unrelated accessories,
still gets factored in — crop it out too. - The goal for every input is the same: one item, cleanly framed, nothing extra.
For more on how your inputs shape a model’s proportions,
our guide to model proportions covers that side.
For the different ways to combine tops, bottoms, and accessories in one generation,
see our combo input guide.
And if your product photos are cutout shots rather than on-body ones,
our guide to using cutout garments walks through that path.