Reviewing generated imagery

Generated imagery fails rarely and badly. Reviewing it one image at a time is how the bad ones get through, because a wrong image looks perfectly fine until it is next to a right one.

Review in grids

Put twenty outputs from the same category on screen together. Inconsistencies in light, crop, shadow and scale become obvious immediately and are close to invisible in a single-image review. This one change catches more problems than any checklist.

The order to look in

1. Is it the right product? Not similar. The same one, with the same features, in the same variant.
2. Is anything invented? Text on packaging, a logo, a port, a button, a seam that does not exist.
3. Would it pass Merchant Center? No overlays, no watermarks, no text, product fills the frame.
4. Does it match the set? Light, background, crop, shadow, scale.
5. Is it better than what it replaces? Sometimes the original photograph was fine.

Reject outright

• Any invented text or logo, however small.
• A variant mismatch, such as the blue one rendered in red.
• Anatomically wrong hands or faces, if people appear at all.
• Products that have changed proportion or lost a defining feature.
• Anything you could not defend to a customer who received the item and disagreed.

Sample, do not inspect everything

At four thousand products, reviewing every image is not a plan. Review a stratified sample by category, measure the failure rate per category, and only go exhaustive where the rate is high. A category with a two percent failure rate needs a spot check; one with twenty percent needs a different template, not more reviewing.

Keep the rejections

The rejected outputs are the most useful data you have about where the template is weak. A category whose failures cluster on the same feature is telling you what to change in the prompt, which is a fix that prevents the next four hundred failures rather than correcting four.