You can check a generated product scene in ten minutes — if you know where the render stops matching the photo. The expensive mistakes are rarely obvious at thumbnail size. They show up when the buyer unboxes the product and finds a handle, a seam, a colour or a fitting that was never there.
A generated product scene is a presentation image built from a real product photo, with the room, lighting and camera angle supplied by software. The scene is allowed to change the context. It is not allowed to change the product. That one line settles most of the arguments your team will have about a render.
The render can change the room; it cannot change the product.
How to check a generated product scene in 10 minutes
Do the comparison before anyone comments on how good the image looks. Put the source photo and the render side by side at 100% zoom on the same screen. Toggle between them, then compare in this order:
- Identity first. Does the silhouette, part count and hardware match the exact SKU?
- Fidelity second. Do colour, finish, markings and texture still belong to the real product?
- Context last. Does the room, prop or background imply a size, material or use the product does not have?
That order matters. A beautiful scene with the wrong hinge is worse than a plain scene with the right one, because the wrong hinge survives every review that starts with "does it look real?"
Keep the source photo open for the whole review. A render reviewed on its own has no reference, and the eye fills in the gaps with whatever looks plausible.
The 7 errors an AI product photo fidelity check catches
An AI product photo fidelity check compares the generated image against the source photo and the product record, not against other renders. These are the seven failures that still get through a quick glance.
1. Silhouette drift
The model straightens a curved backrest, deepens a cabinet by a few centimetres, or narrows a seat cushion. At full size the change looks like styling. At delivery it looks like the wrong product.
Open both files at 100% and trace the outer edge with a temporary line. Compare the widest point, the height, and any straight edge that should stay straight. If the edge moves, stop and mask the product back to the source.
This is not a cosmetic preference. eBay's image policy asks sellers to use images that are truthful, accurate and accurately represent the item; blurry, undefined or distorted images are listed as not allowed. A stretched silhouette is a distorted image even when it is sharp.
2. Invented hardware and parts
Generation fills gaps. Give it a low-resolution photo of a cabinet and it may add a handle, a hinge, a drawer line or a screw head that the product does not have. The extra detail makes the image look more expensive and the product less accurate.
The fix is boring: start from a sharper source, keep the real hardware list open, and remove anything the model invented. If the product has three visible screws, the render gets three. If the handle is on the left, it stays on the left.
3. Colour and finish drift
Walnut turns orange, matte powder coat turns glossy, and a brushed finish becomes mirror-polished. Buyers notice finish changes because they compare the render with the sample they already have, or with the product that arrives.
Use a colour-managed workflow and keep the ICC profile attached through export. Compare the render with a physical sample under neutral daylight, roughly 5000–6500 K, rather than under office lighting or a warm phone screen. If the sample is RAL 7016, the render is RAL 7016, not "close enough".
4. Markings and labels redrawn
Generated text is where renders become dangerous. A CE mark can gain the wrong spacing, a thread grade can change from A2-70 to A4-80, and a size label can turn into convincing nonsense.
Keep every marking that carries a legal, safety or specification claim from the original photo. Crop it at 100%, place it back on the render, and check that the letters, numbers and spacing match. Never let a model redraw a certification mark, a capacity rating or a traceability code. If the source photo does not show the marking clearly, the render does not get to invent one.
5. Scale cues that lie
A chair looks generous beside a tiny plant and cramped beside an oversized sofa. Props are not neutral decoration; they are a ruler the buyer reads without thinking.
Google Merchant Center asks for the product to fill no less than 75%, and no more than 90%, of the image. That rule keeps the product readable, but it does not fix a lying prop. Use props with known dimensions, or remove them. For a marketplace main image, follow the platform's background rule rather than the mood board.
6. Shadows and contact errors
A product that floats above the floor, casts two shadows from different directions, or reflects a window that is not in the scene reads as fake. Most buyers cannot explain why, but they trust the image less.
Match the light direction to the source photo, add a contact shadow where the product meets the floor, and check the reflection at the base. If the source has soft window light from the left, the render does not get hard sunlight from the right.
7. Handoff and metadata errors
The scene can be perfect and still fail at the handoff: the file is named for the wrong SKU, the AI metadata was stripped during export, or the image is below the marketplace's minimum size.
Google Merchant Center now requires images created with generative AI to keep metadata that identifies them as AI-generated, such as the IPTC DigitalSourceType tag. It also asks for at least 500 × 500 px, recommends around 1500 × 1500 px or above for performance, and caps files at 16 MB. The IPTC DigitalSourceType vocabulary defines the values used to mark trained algorithmic media, composite synthetic media and algorithmic media.
Stripping that metadata to "clean up" a file is the wrong move. Keep it, and add the SKU to the filename so the next person can match the image to the product without opening a chat thread.
| Error |
What the buyer sees |
Fix before sending |
| Silhouette drift |
A product that looks slightly wider, taller or straighter than the one delivered |
Overlay at 100%; mask the product back to the source |
| Invented hardware |
Handles, hinges or screws that do not exist |
Sharper source; compare against the hardware list |
| Colour drift |
A finish that does not match the sample |
ICC profile; compare under neutral daylight |
| Redrawn markings |
Wrong grade, spacing or certification mark |
Crop the real marking; never regenerate text |
| Bad scale cues |
A product that looks bigger or smaller than it is |
Known prop dimensions, or no props |
| Shadow errors |
Floating product, double shadows |
Match light direction; add contact shadow |
| Metadata errors |
Wrong SKU, missing AI tag, undersized file |
Keep IPTC metadata; name by SKU; check pixel size |
The product image QA checklist before you send
Use this as a gate, not a reminder. The reviewer signs off only when every box is true.
What evidence to keep from a generated product image review
A render without a record is hard to defend six months later, when the product changes and nobody remembers which image matched which version. Keep a small evidence pack with the scene.
| Evidence |
Why it matters |
Where to keep it |
| Source photo hash or file name |
Proves which photo the render came from |
Asset folder, next to the render |
| Model, prompt and date |
Explains how the scene was produced |
Job log or project note |
| Colour target and ICC profile |
Shows the finish was matched, not guessed |
Export settings |
| SKU, variant and BOM version |
Ties the image to the exact product record |
Product folder |
| Crops of real markings |
Defends a grade or certification claim |
Review folder |
| Reviewer name and date |
Makes the approval traceable |
Sign-off sheet |
| AI metadata flag |
Shows the file was not stripped |
File properties |
This pack is also useful when a buyer asks for a different angle. You can rebuild the scene from the same source and the same product facts instead of starting from a new photo and hoping the product stays consistent.
Where the product scene background check stops
A background check answers a narrow question: does the scene contradict the product? It does not answer whether the product is the right size, the right material or the right configuration. Those answers come from the measured record, not the render.
That is why a generated scene works best beside a measured spec diagram, not instead of one. A buyer can look at the room image to understand the look, then read the dimensions and clearances in the same handoff. Product visuals generated from your own photo keep that link between the source and the presentation asset; changing a product photo background without altering the product is the narrower version of the same discipline.
The boundary is worth stating plainly: AI-generated images cannot measure dimensions. A render can show a plausible room. It cannot certify a clearance, a load rating or a fit. If the buyer's decision depends on a number, put the number in the table and keep the render in its place.
FAQ
How do I check a generated product scene before publishing?
Start with the source photo at 100% zoom and compare identity before style. Check the silhouette, part count, hardware, colour, markings, scale cues and contact shadows, then confirm the file kept its AI metadata and meets the platform's size rules. A scene that passes the style check but fails the identity check is not ready.
Do AI-generated product images need metadata?
Yes on Google Merchant Center: images created with generative AI must keep metadata that identifies them as AI-generated, such as the IPTC DigitalSourceType tag. Do not strip embedded metadata during export. Check the rule for every other marketplace you use, and keep the flag even where it is not enforced yet.
How do I verify AI product visuals before a buyer sees them?
Verify AI product visuals against the real product record, not against another render. Compare the image with the source photo, the SKU sheet and a physical sample, then record who approved it and when. The record is what turns a good-looking image into a reviewable asset.
Can a generated product scene show the real dimensions?
No. A generated scene can show shape, context and finish, but it cannot establish a measurement. Dimensions come from a tape measure and a recorded annotation on the real product. If the buyer needs a clearance or a fit, send a measured spec diagram with the scene, not a number invented from the render.
Why does a generated product photo look fake to buyers?
Usually because the contact shadows, scale cues or silhouette drifted. Buyers read the floor line, the prop size and the product edge before they read the styling. Fix the physical logic of the scene first; the lighting and polish matter less than a product that sits correctly in the room.
Sources & References
