Skip to content
SokuPhoto
Photo notes
AI image quality control

Did AI change your product? A free pre-publish photo checklist

Put the generated candidate beside an approved real photo and compare specific product features, not overall attractiveness. Mark hidden details as unverified. Reject confirmed changes to the item, and keep the original when the candidate cannot be checked reliably.

AI-generated cleaning-product scene used as an illustration, not a passed product-accuracy test
AI-generated layout example. Not an original product photograph or a measured test result.

Choose a reference that can answer the question

Use the current product, packaging and variant. A back view cannot verify a front label; a cropped source cannot verify a missing handle. If the candidate shows a new angle, take a real photo from a comparable angle before approving the newly visible details.

Do not use another generated image as the sole reference for an item you actually sell. It may repeat the same mistake. Keep the original filename and the candidate filename together so someone else can understand what was checked.

Review five product facts in a fixed order

Work through the same order for every candidate. This prevents attractive lighting from distracting you from a changed closure or extra unit. The following is a manual editorial checklist, not an automated certification system.

CheckCompareExample of a rejection
ShapeSilhouette, openings and closuresPump becomes a dropper
LabelReadable names, numbers and marksA different volume or invented badge
QuantityUnits and included accessoriesA one-pack becomes two
VariantApproved color and finishMatte black becomes glossy gray
Physical detailSeams, supports and connectionsA handle joins the wrong surface

Use pass, review and reject—not a vague quality score

Pass means the observable product details agree and the scene is suitable for the intended use. Review means you cannot verify a detail: a hidden label or a strong color cast is a reason for another check, not evidence that the model preserved it. Reject means a confirmed mismatch or unsupported product claim.

Not every pixel difference is a product change. Lighting naturally changes highlights and shadows. Compare the corresponding material and ask whether a customer would infer a different variant or finish. If the answer is uncertain, return to a neutral real photo rather than forcing a pass.

  1. Compare the full silhouette and count first.
  2. Zoom into the label, logo and moving or connecting parts.
  3. Inspect the image as a small listing thumbnail for misleading emphasis.
  4. Record pass, review or reject with the specific reason and save the reference.

Check the final export and the destination

A correct candidate can still become misleading after you add text, crop it or attach it to the wrong variant. Inspect the exported file and the actual listing, including the mobile version. Confirm that promotional wording does not imply accessories or results not supported by the product.

Google's Merchant Center requirements include accurate depiction and relevant AI-source metadata. Those are additional destination checks, not a promise that this checklist guarantees approval. SokuPhoto does not automatically certify product accuracy. If regeneration still changes the same critical feature, use the real photograph and keep editing limited to changes you can verify.

Reference: Google Merchant Center: Product image requirements and AI metadata

Before you publish

  • Reference is real, current and comparable.
  • All five product checks have a recorded decision.
  • Hidden details remain unverified rather than assumed correct.
  • The final export and selected listing variant were inspected.
Review and edit a candidate

Sources & editorial notes

Original workflows and examples by SokuPhoto, informed by the linked sources. Prompts are examples, not guarantees. Platform features and requirements can change; sources checked on September 16, 2026.