Since late September, every image a ProductAI subscriber generates in the studio gets a second look. A vision model is shown the photo the seller uploaded next to the finished render and asked one question: is this still the same product? The scene, the light and the angle are supposed to change. The product is not.
In the first 18 days that check looked at 495 renders and flagged 111 of them, about one in five. We read every one of those 111 notes by hand. This is what they say about where AI product photos go wrong, and what you can do before you press Generate.
What went wrong, counted
| What changed on the product | Flags | Share |
|---|---|---|
| Label text and numbers | 39 | 35% |
| Colour or finish | 19 | 17% |
| Shape | 16 | 14% |
| Wrong, extra or absent product | 11 | 10% |
| A part missing | 9 | 8% |
| Other, mostly false alarms | 17 | 15% |
One problem is bigger than the next two combined: text.
1. Small text is where the models break
More than a third of all flags were about writing on the product. The failures came in three kinds.
- Numbers that quietly change. A perfume bottle labelled 50 ML came back as 30 ML. A tube marked 118 mL came back as 110 mL, and on another run as 116 mL. A spray whose label states an active ingredient at 9.6% was rendered at 5.6%.
- Words that almost survive. "MADE WITHOUT PARABENS" became "MADE WITNONT PARABENS". "Silicone-based" lost a letter. A two-word engraving on a pendant changed one letter, twice, in two different ways.
- Labels that go blank. A pharmacy bottle with printed dosage instructions came back with a clean white label on three different models.
The first kind is the dangerous one. A garbled word looks wrong at a glance. "110 mL" looks perfectly fine, and it is a false statement about your product on your own listing.
What helps: keep the product large in the frame. Many of the cases we opened were products shown small, in a hand or in a wide lifestyle scene, where the label is a few dozen pixels tall. If the label has to be readable, shoot it close.
2. Colours drift toward their neighbours
Nineteen flags were colour or finish. Navy was rendered purple on two different models, for the same product. Silver jewellery came back gold on one run and rose gold on another. A burgundy bag turned black. A white bottle became transparent.
What helps: say the colour and the finish in the prompt. "Navy blue, matte" or "polished silver" costs four words and gives the model something to hold on to.
3. What is not in the photo tends to disappear
Nine flags were a missing part, and the parts were consistent: the chain of a necklace, the battery pack of a power tool, a lid, a charm on a keyring. Sixteen more were shape changes, where an unusual outline was "corrected" into a familiar one. A teardrop-shaped compact came back round. A cropped sweatshirt came back as a bodysuit.
What helps: upload a photo that shows the whole product with every part attached. If the chain is out of frame in your reference, expect it to be out of the render too.
4. Several products in, the wrong one out
Eleven flags were the wrong product altogether. In about half, the product was simply absent: the render showed an empty pedestal, or only the gift box the product comes in. Four more had one cause: several different products uploaded as references for one image. The model picked the sibling product from the same brand, or added an extra one that was never asked for.
What helps: one product per generation, unless a group shot is really what you want.
The checker was wrong too, and that was the useful part
About one flag in seven was a false alarm, and they were not random. Our own instructions to the checker had asked for most of them.
- It flagged a jar because the lid was off in an in-use shot.
- It flagged a product for lying in a model's hand instead of standing upright.
- It flagged a pendant because the water droplets from the reference photo were gone.
- It flagged a street scene, uploaded as a mood reference, because the parked cars were missing.
The most instructive one was a pill bottle flagged because its label had changed to a different medication. That looked like the most serious catch in the whole set, until we read the prompt: the seller had asked for exactly that label swap. The checker never saw the prompt, because of how the studio stores it. A checker that does not know what was asked for cannot tell a mistake from a request.
We are testing a revised checker now. It ignores a product being open, held, tilted or styled, it reads the prompt, and it names the kind of change it found instead of raising a generic warning.
Does the model matter?
In this sample, yes, with a large caveat.
| Model | Renders checked | Flagged | Rate |
|---|---|---|---|
| Nano Banana Pro | 336 | 81 | 24% |
| GPT Image 2 | 64 | 7 | 11% |
| Seedream 5 Pro | 51 | 13 | 25% |
| Nano Banana | 37 | 9 | 24% |
GPT Image 2 was flagged about half as often as the others. That fits what we saw in two client catalogues, where it was the model that kept label text correct. The caveat: all of this comes from 13 accounts, and different sellers used different models for different products. Treat it as a direction, not a ranking.
What we changed in the studio
A flag after the fact is useful. Not needing one is better. So the three most common mistakes now show up as a one-line tip above the prompt box, and only when your prompt touches that subject:
- Mention a label, a logo or text, and it reminds you to keep the product large in the frame.
- Mention a chain, a cap or a battery, and it reminds you that parts missing from your photo tend to go missing from the render.
- Mention a colour or a finish, and it reminds you to name it exactly.
One tip at a time, and you can dismiss each one for good.
A checklist before you generate
- Label must be readable? Keep the product large in the frame. Use a close shot for the listing image and save the wide lifestyle scene for mood.
- Check every number. Volume, weight, strength, count. These change silently and look plausible.
- Name the colour and the finish in the prompt.
- Show every part in the photo you upload.
- One product per generation.
How we measured this
- The sample is small. 495 checks from 13 accounts between 23 September and 11 October 2026: 10 paying subscribers and 3 of our own team, who made 19 of the checks. A few sellers account for most of the volume, so product categories are uneven: personal care and jewellery are over-represented.
- The judge is a vision model with a deliberately conservative prompt that answers "ok" when in doubt. It misses things, and it raises the false alarms described above.
- The categories are ours. We read all 111 notes and sorted each into one group by its main complaint. Three checks failed on oversized images and are not in the per-model table.
- The false-alarm rate (about 14%) is our judgement from looking at the images, not a second automated check.
