Guide · product images

AI product images: what AI does well and where it fails

AI makes product images possible without a studio. But image models have typical weaknesses. Here is, from my own work, what works reliably, where it gets stuck and how I catch it.

In short

  • Good: setting, light, mood, lifestyle scenes and many variants in the same look.
  • Weak: small print, logos, fine technical details, true proportions and hands.
  • Rule: the product comes 1:1 from your photos, only the scene is free. Every shot is checked against the reference and labelled.

What works reliably

Image models are good at placing a product in a fitting setting: stone and towels for a serum, an old-building window for a blouse, warm evening light for an autumn ad. Once the look is approved, more shots follow in the same style without a new date. That is the real advantage over a shoot.

One example is the showcase Atelier 026: a launch set for a serum from a made-up brand, with hero, detail, texture and ad visuals in one look.

Typical mistakes from real projects

Invented details

For a sample shot of a sportswear set, the prompt asked for ribbed fabric and a side pocket. The real set is smooth and seamless. The client’s design team saw it instantly, and the shot was useless. Since then my prompts only name details that are clearly visible in the photos.

Missing small parts

With technical products, models like to leave out small but typical parts, such as a hand crank or a latch. Laypeople hardly notice, experts notice at once. I ask for such parts explicitly in the prompt and check them zoomed in.

Text and logos

Small print on labels tends to change: letters blur, words get invented. I check logos and labels character by character against the original.

Wrong size

Without a reference point in the frame, models often render objects too small or too large. The fix is a clear anchor: a hand that covers only one corner, or an object of known size next to it.

Hands

Hands holding a product often take several attempts. In detail shots I only use them when they really add something.

How I check every shot

  1. Photos of the product from every side, plus logo and label.
  2. Record the construction: smooth or ribbed, seams, pockets, closures, prints.
  3. The prompt names only confirmed details, plus the instruction that the product must look exactly as in the photos.
  4. Side-by-side comparison with the photo, zoomed in on text and small parts.
  5. If something is off, it gets corrected or the shot is dropped.
  6. Images that look real get the “AI” label.

Labelling

A product image showing your real product in an AI setting usually looks real. In the EU it has to be labelled since August 2026. What that means in practice is in the guide labelling AI videos and images.

FAQ

Frequently asked questions

Will my product still look like my product in AI images?

Yes, if every shot is checked against your photos. Without checks, models like to invent details or leave out small parts.

Which photos do I need as a basis?

Phone photos from every side in good light, plus logo and label. No studio needed.

Can AI render small print on labels correctly?

Not reliably. That is why I check text and logos character by character and correct them where needed.

Do AI product images need a label?

In the EU, yes, if they look real. I put a small “AI” label directly into the image.