ORA Signals: 3 in 4 Rejected AI Product Images Fail Because the Product Changed
By the ORA Lab team · Updated 18 November 2026 · 6 min read
Key takeaways
- In ORA's client work from 3 August to 28 October 2026, clients reviewed 6,571 AI product images and approved 64.3% of them.
- Of the 2,348 rejected images, 1,739 (74.1%) were rejected because the product changed: its colour, shape, detail, stones, pattern, logo or text, scale, or parts.
- Scene or lighting problems explain only 6.4% of rejections, model or pose 4.4%, and technical faults such as resolution or cropping 4.0%.
- The single most common note clients typed was simply "product changed", at 17.4% of all rejections.
- The practical lesson: check the product before you judge the scene, and choose tools by how often the product survives, not by how good the backgrounds look.
ORA Signals is a short series where we publish what our own production data says about AI product photography. This first edition asks a simple question: when a brand looks at an AI product image and says no, why? The common assumption is that AI images get rejected because they look fake, the lighting is off, or the room feels wrong. In ORA's client work, that is not what happens. Most rejections are about the product.
The numbers below cover 7,350 AI generations made for 36 client accounts between 3 August and 28 October 2026. Clients reviewed 6,571 of them, and every rejection came with a short typed reason. All figures are as of 28 October 2026.
The headline: two in three approved, and most rejections are about the product
| Measure | Result |
|---|---|
| AI generations in scope | 7,350 |
| Images reviewed by clients | 6,571 |
| Approved | 4,223 |
| Rejected | 2,348 |
| Still pending review (excluded from rates) | 779 |
| Approval rate of reviewed images | 64.3% |
A 64.3% approval rate means roughly two in three reviewed images went straight into use. The interesting part is the other third. When we grouped every typed rejection reason by keyword, one group dwarfed the rest.
| Reason group | What it covers | Rejections | Share of rejections |
|---|---|---|---|
| Product changed | Colour, shape, detail, stones, pattern, logo or text, scale, missing or extra parts | 1,739 | 74.1% |
| Other or unspecific | Notes such as "wrong", "regenerate" or "test" | 261 | 11.1% |
| Scene or lighting | Background, setting, light | 151 | 6.4% |
| Model or pose | The person wearing or holding the product | 104 | 4.4% |
| Technical | Resolution, artefacts, cropping, duplicates | 93 | 4.0% |
What clients actually wrote
Clients are not writing essays. Most notes are a few words, and the same few words come up again and again. The most common exact reasons tell the same story as the groups.
| Note the client typed | Share of all rejections |
|---|---|
| "product changed" | 17.4% |
| "wrong product" | 5.7% |
| Background issues ("bg issue", "bg blurred", "blurred bg") | about 2.9% combined |
| "design changed" | 1.5% |
| "model changed" | 0.8% |
"Product changed" on its own accounts for 17.4% of all rejections, and "wrong product" adds another 5.7%. The background notes that people often assume dominate AI complaints add up to about 2.9%. Brands will forgive a slightly soft background. They will not forgive a necklace with a different stone, a sofa with the wrong arm, or a print that has shifted.
Why the product is what breaks
Image models are very good at inventing plausible scenes, because a plausible scene has many acceptable answers. A product has only one acceptable answer: the real one. Every small liberty the model takes with a stone setting, a stitch line, a logo or a proportion is invisible in a quick glance and obvious to the brand that made it. We cover the mechanics in why AI product photos look fake, and the short version is that models tend to smooth, simplify and "improve" exactly the details that make a product recognisable.
This is also why approval rates differ by category. Products with dense, precise detail fail more often. Jewellery, with its stones, settings and fine metalwork, is the clearest example, and we will look at it category by category in the next ORA Signals edition. For a sense of how much detail a single piece can carry, see our guide to kundan and polki jewellery photography.
What this means for how you check images
- 1.Check the product first, the scene second. Put the AI image next to the real product photo and compare colour, shape, pattern, logo and parts before you look at the room or the light.
- 2.Zoom in where the detail lives: stones and settings, prints and borders, stitching, hardware, text on labels. Most product changes are small and local.
- 3.Check scale against something in the scene. A product can be perfect in every detail and still be the wrong size.
- 4.Count parts. Missing legs, extra handles, a lost earring back or a doubled strap are among the easiest changes to miss.
- 5.Only then judge the scene, light and model. Those are the easier fixes.
Our AI product photo fidelity checklist turns this into a repeatable routine your team can run on every batch.
What this means for choosing tools
Most tool comparisons judge sample galleries: how good do the scenes look? This data suggests that is the wrong test. In ORA's client work, scene quality is rarely the reason an image is turned down. The better question is how often the product survives the generation unchanged. Run any tool on your three hardest products, compare each output against the real thing, and count usable images rather than total images. A tool that makes beautiful rooms around a slightly different product will cost you more review time, not less. Read more in AI product photo accuracy.
How we measured
- Scope: 7,350 AI generations from 36 client accounts in ORA's client work, made between 3 August and 28 October 2026. Internal and test accounts were excluded, as were images clients uploaded themselves and jobs that failed.
- Approval: an image counts as approved when the client approved it in ORA's review screen, and as rejected when the client rejected it there. The 779 images still waiting for review are left out of all rates.
- Reasons: every rejection carries a short note the client typed. We grouped these notes by keyword into the five groups above. About 11% of notes were too unspecific to classify and sit in "Other or unspecific".
- Concentration: a small number of large accounts produce most of the volume, so the totals lean towards their products and standards.
- Coverage: there was no beauty or watch work in this period, so these results say nothing about those categories.
- Timing: all numbers are as of 28 October 2026. Pending reviews will change them, and we will say so when we update.
If you want AI images where the product is checked against the real thing before you ever see it, .
Frequently asked questions
- Why do brands reject AI product images?
- In ORA's client work from 3 August to 28 October 2026, 74.1% of rejected AI product images were rejected because the product changed: colour, shape, detail, stones, pattern, logo or text, scale, or missing or extra parts. Scene or lighting problems explained only 6.4%.
- What share of AI product images do clients approve?
- Clients approved 64.3% of the 6,571 AI product images they reviewed in ORA's client work between 3 August and 28 October 2026, as of 28 October 2026. Images still pending review are excluded.
- What is the most common reason clients give for rejecting an AI image?
- The exact note "product changed" was the most common, at 17.4% of all rejections. "Wrong product" followed at 5.7%, while background issues added up to about 2.9%.
- How should I check AI product images before using them?
- Compare each image against a real photo of the product before judging the scene. Check colour, shape, pattern, logo and text, scale, and count the parts. Then check the background, light and model, which are easier to fix.