# AI Product Images Get Scale Wrong, and Shoppers Notice Before You Do

Source: https://aicommercewatch.com/blog/2026-11-10-ai-generated-product-images-scale-errors.html

November 10, 2026

*The product image looks great. The bag is on a table next to a coffee cup, the lighting is perfect, the stitching is sharp. The only problem is the bag is the size of the table, and the coffee cup is the size of a bucket, and nobody noticed until the first customer opened the box and laughed.*

## Scale is the thing AI cannot see

Generative image tools are good at texture, lighting, and composition, and they are bad at scale, because scale is a fact about the physical world and the tool has no physical world to check against. A bag next to a chair, a lamp next to a sofa, a bottle next to a hand: the tool renders each object beautifully and sizes them by visual balance, not by measurement. The result is a scene where every object is individually plausible and collectively impossible, and the shopper who buys on that scene buys a size that does not exist.

The tell is the reference object. A coffee cup, a hand, a door frame, a person: anything with a known size. When the reference object is wrong, everything next to it is wrong, and the shopper has no way to correct for it, because the listing gives them no other scale cue. A dimensions table in the description does not fix a misleading image, because the shopper trusts the image more than the table. They bought what they saw.

## Where the errors actually cost money

The cost shows up in returns, and it shows up as a specific complaint: smaller than expected, larger than expected, not as pictured. Those are the three most common size-related return reasons in ecommerce, and AI-generated lifestyle images make all three worse, because the image is the shopper's only scale reference and it is wrong. The return costs the shipping both ways, the restocking, and the review that says the product is not as pictured, which then costs the next sale too.

The fix is not to stop using generated images. It is to anchor them. Every generated lifestyle image needs a real scale cue: the product next to a standard object, or a dimension overlay, or a flat lay with a ruler in the frame for the products where size is the buying decision. And every generated image needs a human review against the spec sheet before it ships, checking the one question the tool cannot answer: is this the size the product actually is. For the description version of the same failure, see [AI Product Descriptions Invent Features, and Returns Pay for It](https://aicommercewatch.com/blog/2026-11-09-ai-product-descriptions-invent-features-returns.md), and for the size chart version, [AI-Generated Size Charts Measure Wrong](https://aicommercewatch.com/blog/2026-11-07-ai-generated-size-charts-measure-wrong.md).

## The rule for generated imagery

Treat a generated image like a claim, not a photo. A photo of a real product has a real size, and the camera recorded it. A generated image has no size at all until a human checks it against the spec sheet and anchors it with a reference object. The check takes two minutes per image. The return it prevents costs twenty dollars and a review.

## Frequently asked questions

### Why do AI-generated product images get scale wrong?

Generative tools render by visual balance, not by measurement. They have no physical world to check against, so a bag next to a chair can come out the size of the chair while looking individually plausible.

### What is the tell for a scale error?

The reference object: a coffee cup, a hand, a door frame, a person. When the known-size object is wrong, everything next to it is wrong too, and the shopper has no way to correct for it.

### How do scale errors show up in returns?

As smaller than expected, larger than expected, and not as pictured, the three most common size-related return reasons. The image is the shopper's only scale reference, and when it is wrong the return costs shipping both ways plus a review.

### Should we stop using AI-generated product images?

No. Anchor them instead. Put the product next to a standard object, add a dimension overlay, or shoot a flat lay with a ruler, and have a human check every image against the spec sheet before it ships.