How to Check AI Product Photos for SKU Mistakes Before You Publish

The dangerous AI product image is not the obviously bad one. It is the polished image that looks convincing until you notice the clasp has changed, the label contains a different word, or a ring has gained an extra stone.

You should still inspect generated images yourself, but a vision model can act as a useful second pair of eyes. Instead of asking it whether an image “looks good,” give it the real product photo and the AI-generated version together and ask it to look specifically for SKU differences.

This tutorial works with multimodal tools that can analyse more than one image. Google’s current Gemini image-understanding documentation explicitly shows a two-image prompt asking what is different between the images. ChatGPT also supports image inputs on free and paid plans, subject to account and usage limits.

Step 1: Keep one real image as the source of truth

Start with a genuine photograph of the exact SKU. Choose the sharpest view that shows the details customers would rely on. For jewellery, that may include the clasp, chain construction, stone count, engraving, pendant shape and metal finish. For packaged products, prioritise the label, logo, cap, quantity and colour. For electronics, check ports, buttons and proportions.

If one photograph does not show an important feature, keep a second detail view ready. A vision model cannot verify a clasp hidden in the source photo any more reliably than a human can.

Step 2: Upload the real image and the generated image together

Add the real product photograph first, then the AI-generated derivative. Tell the model which is which. Do not simply ask, “Are these the same product?” That invites a broad judgement and can miss small commercial differences.

A more useful prompt is: “Image 1 is the real product and is the source of truth. Image 2 is an AI-generated marketing image. Compare Image 2 against Image 1 only for product accuracy. Ignore changes to the background, model, lighting and camera framing unless they change how the product itself is represented.”

Step 3: Give the model a product-specific checklist

Generic comparison prompts produce generic answers. Tell the model exactly which details matter for your category.

For a bracelet, ask it to compare the clasp type, number and path of chain links, engraving, width, visible thickness, connector shape, stone count and metal finish. For a skincare bottle, compare label wording, logo, bottle shape, dispenser, cap, colour and pack size. For clothing, compare seams, buttons, pockets, print placement and visible hardware.

Then ask for a simple output such as: “For each feature, return Match, Possible difference, Clear difference, or Not visible enough to verify. Do not guess when the source image is unclear.”

Step 4: Treat “not sure” as useful information

One of the most valuable results is not a detected error. It is a feature the model cannot confidently verify. That tells you where your source material is weak.

If the real engraving is too small to read, add a macro photograph and repeat the check. If the clasp is hidden, provide another angle. This is better than asking the model to infer what the missing detail probably looks like.

Step 5: Zoom in yourself on every flagged area

AI vision is a screening tool, not a product-certification system. OpenAI’s current image-input guidance explicitly warns that image analysis can be inaccurate, that precise spatial localisation can be difficult, that images may be resized before analysis, and that object counts can be approximate.

That matters for ecommerce. A model may miss one altered chain link or incorrectly report a difference caused only by perspective. When it flags an area, open both images at full size and decide yourself. If the feature is commercially important and still uncertain, use the genuine photograph instead.

A reusable prompt for product-image QA

You can save this structure and change the checklist for each product category:

“Image 1 is the genuine product and the source of truth. Image 2 is an AI-generated derivative. Ignore intentional changes to the background, lighting, model and composition. Compare only the product. Check: overall shape and proportions; count and placement of components; logos or engraving; clasps, connectors and hardware; colour and material finish; visible dimensions or thickness; and any extra, missing or redesigned feature. Return a short table with Feature, Status, Evidence and What I should inspect manually. Use only Match, Possible difference, Clear difference or Not visible enough to verify. Do not assume hidden details are correct.”

Where this fits in a real seller workflow

Run the check after you have chosen the generated image you actually want to keep. There is little value in auditing every rough generation. First choose the best scene, pose or advert. Then compare that finalist with the real SKU before it reaches Amazon, Etsy, Shopify, TikTok Shop or an ad campaign.

For teams producing many product images, the same checklist can become a repeatable approval step. Lustra Studio is designed to reduce product drift by working from jewellery references, but generated outputs should still be checked before commercial use. A vision-model comparison can make that final review more systematic without pretending to replace it.

The takeaway

AI can help create product images, and it can also help review them. The useful trick is to give the vision model a real source-of-truth photograph, a tightly defined checklist and permission to say when something cannot be verified.

Use the result as a second opinion, then perform the final human check yourself. The goal is not to prove that an AI image is perfect. It is to catch more quiet product changes before a customer sees them.

Sources: Google AI for Developers, Image understanding, accessed 5 September 2026; OpenAI Help Center, ChatGPT Image Inputs FAQ, updated September 2026.