Producing product images with AI works well in most categories. It does not work on modest clothing, and the reason is technical.
What goes wrong
General-purpose image models resemble the data they were trained on. Modest clothing makes up a very small share of that data. The result is not that the model "does not know" modest wear — it pulls it towards what it has seen more often.
In practice that shows up as four errors:
The skirt gets shorter. The model drifts to the average it has seen, and a long skirt is pulled towards the knee.
The loose cut gets tighter. A cut that does not hug the body turns into the fitted silhouette the model sees more often.
The closed front opens. On closed-front garments such as abayas and evening wear, the neckline and front change.
The headscarf sits wrong. Tying style, hairline and neck coverage come out inconsistent.
Why it matters
When these images go into a catalogue, the customer receives something that does not look like what they ordered. That means returns, and for a small brand returns are the most expensive thing there is.
So this is not an aesthetic problem; it is a commercial one.
How I handle it
For modest clothing I do not use the general pipeline. I use a separate production flow that preserves silhouette and cut and takes the product's original proportions as the reference.
More importantly: I do not present the result as a claim. I put the generated image next to the original product and show both. Is the hem the same, is the cut preserved, did the front stay closed — these are things you can see by looking, and they need to be looked at.
One limit
Not every product gives the same result. Very finely detailed fabrics, lace and layered cuts are still hard. For a product like that we run a single test first; if it does not work, I say so.