Why AI Changes Your Shirt: Prompt Rules for Better Garment Accuracy

One of the strangest things about AI fashion photography is that the image can look excellent while the product is wrong.

The shirt is blue, the model looks realistic and the lighting is beautiful. But the collar has changed. The pocket moved. The checks became larger. A button disappeared.

If you are creating content around a real garment, visual quality is only half the job. Product identity has to survive the generation.

Why AI changes clothing details

A generative image model is not a traditional photo editor. It reconstructs an image from patterns it has learned and the instructions you provide. Small construction details can be treated as flexible unless you explicitly make them important.

That is why “put this shirt on a model” can produce a convincing shirt that is not quite your shirt.

Rule 1: Write a preservation list

Instead of describing only what should change, state what must remain unchanged.

Preserve the exact shirt color, collar shape, button placement, chest pocket, sleeve construction, print scale, hem shape and fabric texture. Do not add or remove design details.

Think of this as a product lock.

Rule 2: Separate product instructions from scene instructions

A messy prompt might say: “Make a luxury Italian street photo with a handsome man in this premium shirt, cinematic, editorial, expensive, beautiful…”

The model may prioritize the mood and reinterpret the garment to suit it.

A clearer prompt has two blocks:

Product: exact garment details that must not change.
Scene: model, pose, trousers, background, lighting and camera feel.

Rule 3: Name the collar precisely

Do not assume the AI understands that the reference collar must remain. Say spread collar, point collar, band collar or whatever is actually present. If there are no collar buttons, say do not add button-down collar buttons.

Rule 4: Treat patterns as geometry

Checks, stripes and repeating prints are common failure points. Ask to preserve:

• pattern scale
• line thickness
• spacing
• direction
• color relationship

For embroidery, specify the exact area where the motif appears and ask the model not to extend it across plain sections.

Rule 5: Do not let “luxury” rewrite the garment

Words such as luxury, designer and couture can push an image model toward shinier fabrics, more elaborate construction or decorative details. Describe the photography as premium, not the product as something it is not.

For example: premium editorial photography of the exact reference cotton-blend shirt.

Rule 6: Ask for natural fabric behavior

Real clothes wrinkle, pull, fold and cast small shadows. If everything looks vacuum-sealed to the body, the picture feels artificial.

Ask for natural drape, believable sleeve folds, realistic tension around elbows and waist, and visible but not exaggerated fabric texture.

Rule 7: Use a negative instruction for common failures

A short “do not” list is useful:

Do not change the garment color. Do not create a second pocket. Do not add logos. Do not alter the print. Do not add back stitching that is absent from the reference.

Keep it specific rather than writing a giant list of generic negatives.

Rule 8: Check the generated image like a cataloger

Before using the result, compare it against the real product. Inspect collar, placket, cuff, pocket, hem, pattern, color and back construction. Zoom in.

For a deeper checklist, see our 12 prompt details most fashion brands miss.

Frequently asked questions

Why does AI keep changing shirt buttons?

Small repeated details can be reconstructed inconsistently. Explicitly state the button count or placement when it matters and review the final image closely.

Can a generated image be 100% product accurate?

Accuracy can be improved, but every generated image should still be checked against the real product before commercial use.

Should I describe the garment if I already uploaded a reference?

Yes. The reference supplies visual information, while the text tells the model which details are non-negotiable.

In fashion AI, the best prompt is not the most poetic one. It is the one that tells the model exactly where creativity is allowed—and where it is not.