How to get AI ad creative that keeps the product exact

Say a customer sees your ad for a jar of face cream and orders it. The parcel arrives. The lid is narrower than the one in the picture, and the label has a line of small print the ad never showed.

Nobody planned that. An image model drew the ad, and a model can hand back a jar that looks like yours without being yours. The customer does not think about models. They see that the picture was wrong.

Two things help you get AI ad creative that keeps the product exact: a tool that starts from your photograph instead of a description of it, and a minute spent comparing the result with the real pack before the ad runs.

What AI product photo ads change besides the background

An AI tool that makes product photo ads can change more than the scene around the product. When Meta announced new generative AI features for advertisers in May 2024, it described image variations inspired by your original ad, and its coffee example included adjustments to the cup itself.

A variation inspired by your ad is a new picture, so the product in it needs checking like one. Meta's Ad Creative Generative AI Terms say whoever uses its ad tools is solely responsible for evaluating what comes out and for its accuracy.

The same goes outside Meta. A tool that takes your photo and returns an ad has made a picture of your product, and unless it says it placed your photo untouched, you cannot assume the pack in the ad is the pack you shot.

Why a model gets labels and lettering wrong

OpenAI's guide to its image model lists text as a limit: the model can still struggle with precise text placement and clarity, and with keeping brand elements the same across several generations.

Researchers have studied one cause of wrong lettering. A 2022 research paper on spelling in images blamed part of the problem on the popular models of the time having no input that carried a word's letters, which made it harder to draw a word as a row of letter shapes. Models trained with that input spelled better.

Newer models have moved on. Google's Gemini image documentation says its models are good at rendering text, and OpenAI calls its own text rendering significantly improved. Improved and exact are different standards, and a label has to meet the second.

Shape needs a test of your own. How often a lid, a bottle shoulder or the number of items in a pack comes out wrong depends on the tool and the product, so measure it on yours. Generate the same product ten times, lay the results beside the photograph, and count the misses.

How to spot a redrawn product

Put the real photograph and the ad side by side, at the same size, on the screen your buyers use most.

• Count what can be counted: items in the pack, buttons, stripes, rings. A count that is off by one still looks natural.
• Compare shape. How tall is it against how wide, and what do the lid, the cap, the shoulder of the bottle and the handle look like?
• Read the label, smallest line first, letter by letter. Lettering that is nearly right reads as right when you scan.
• Look for what you never printed: a seal, a claim, an ingredient, a symbol.
• Put the product itself next to the screen and compare colour. That is the comparison your buyer makes when the parcel arrives.
• Check the scale. A pack in a hand or beside a mug tells the buyer how big it is, so the size in the ad has to match the size in the box.

Ask the person who packs your orders to look as well. They know the product better than the photograph does. These checks are the product half of the five checks before an AI ad runs.

What keeps the product exact in AI ad creative

Start the tool from the photograph. Model makers offer this under different names. Google lists reference photos of objects that a model can carry into the final image with high fidelity, and OpenAI's cookbook describes, in an archived example, a high input fidelity setting for edits, which keeps faces more accurately than its standard mode and keeps logos true to the original assets.

Ask a tool two questions. Does it start from my photograph, or from a description of my product? Can it show me the photograph and the ad at the same size? A tool that cannot answer the first is asking you to trust a description of your own jar.

Change one thing at a time. If only the lid is wrong, ask for the lid. A new generation is a new picture, so the label that was right can come back wrong.

Words that must be exact, such as a price or a discount, belong in the text of the ad, typed by you, rather than in the picture.

Where this stops working

A check catches what you can see. A pack with tiny print, a transparent bottle or a pattern that repeats across the whole label gives you more to compare and more chances to miss one difference. A glossy bottle also changes with the light around it, so exact means the same product, not the same pixels.

The photograph sets the ceiling. If it shows last year's label, a perfect copy is the wrong label. And the documents named here describe models that change often, so read the current version before relying on a limit quoted from one.

An exact pack does not make an ad sell either. If the pack is right and the ad still looks like somebody else's brand, that is why AI ads come back off-brand, and it has a different fix.

What Nomi does with the product photo

Nomi starts from the photograph you attach. It copies the shape, the packaging, the label text, the colours and the logos, and changes the light, the background and the props, as its ad creative page describes.

When one spot is wrong, you point at it and say what should be different there, and the correction comes back as a new version beside the one you liked. You still hold the result up to your photograph before it runs.