Generated images need more than a quick beauty check. Reviewers should know what the asset claims, which details carry the greatest risk and where the final file will appear. A structured pass catches factual, visual and production problems before they reach a client.
Create a risk map from the brief
Mark the details that affect identity, safety, product truth, legal clearance or brand trust. Typical hotspots include faces, hands, labels, reflections, packaging, uniforms, landmarks, cultural symbols and background text. The more a detail supports a claim, the more evidence it needs.
Record what is synthetic and what comes from approved photography. NIST recommends risk management across the lifecycle of generative AI systems; for a studio, that translates into defined owners, review criteria and documentation before release.NIST generative AI risk profile ↗
Use distinct review passes
First review composition and message without zooming in. Next inspect the file at intended use size. Finally check critical regions at 100 percent or higher. Switching between these views matters because microscopic artifacts can be irrelevant in a small card, while a subtle shape error can undermine a large poster.
Review colour and contrast in the destination template, not only on a neutral canvas. Confirm crop-safe areas, overlays and responsive behaviour. For meaningful images on the web, write alternative text based on purpose and context rather than turning the filename into a description.W3C image guidance ↗
Challenge plausible details
Ask whether a reviewer knows a detail is correct or merely finds it believable. Compare products, locations, people and marks with approved references. Verify quoted words separately. Never infer permission, endorsement or factual accuracy from visual polish.
If the image history matters, preserve prompts, settings, source references and edit notes. C2PA content credentials can help communicate provenance assertions when the tools and delivery path support them, but missing credentials do not prove misconduct and present credentials do not replace review.C2PA technical specification ↗
Make approval specific
Use named statuses such as needs correction, approved for concept only, approved for specified placements and final master. Record the approved crop and version. A general thumbs-up in a chat thread is too ambiguous for a multi-format delivery.
After export, reopen the actual files and sample them in their destinations. Check embedded metadata, colour, dimensions, transparency and compression. Then give the client a short alteration note and any use limitations that remain.
Pair creative and factual reviewers
A creative reviewer can assess message, rhythm and brand fit, but may not recognise a changed product port or incorrect uniform. Add a reviewer who knows the subject when the image includes consequential specialist detail. Give each reviewer a defined question so approval is not duplicated or assumed.
Use an annotated proof that circles high-risk regions and links them to references. Comments such as looks good are too broad. Better decisions state that the composition is approved, the product geometry matches reference B and the background sign was removed because its lettering could not be verified.
Add an adversarial review pass
After the normal review, ask someone to search for reasons the image should not ship. They should inspect peripheral figures, reflected text, repeated objects, hidden marks and contextual implications. Generated defects often survive because everyone focuses on the intended hero subject and stops looking at the edges.
Temporarily change the viewing method: flip the image, desaturate it, inspect channels, hide overlays or compare with a difference blend. These techniques disrupt visual familiarity and make duplicated texture, inconsistent lighting and repaired seams easier to notice. Use them selectively on risk areas rather than as theatrical process.
Prepare a client proof that supports a decision
Show the proposed asset at the approved placement and include close views only for consequential details. Label the source, output version and meaningful alterations. If more than one route is offered, explain the tradeoff rather than presenting nearly identical options with no recommendation.
Ask for approval against named criteria and record the response in the release register. If feedback changes the crop, text, product or person, issue a new proof. The client should never have to remember whether a later attachment contains the change they approved.
A practical decision table
| Pass | Focus | Failure example |
|---|---|---|
| Message | Composition and intended claim | Scene implies the wrong use |
| Use size | Crop, legibility and hierarchy | Label fails in the final placement |
| Detail | Anatomy, texture, text and seams | Invented mark or repeated pattern |
Release checklist
- Read the approved brief
- Mark high-risk details
- Compare approved references
- Review the whole composition
- Check at intended size
- Inspect critical regions closely
- Verify text independently
- Test destination crops
- Record version-specific approval
- Reopen final exports
Common questions
Does pixel-level review guarantee a good image?
No. It must be combined with message, factual, rights and in-context review.
Who should approve generated details?
A named decision owner, with subject specialists involved when identity, products, places or regulated claims matter.