Your China-based production partner sends a folder of edited catalog images. Ten look good in the review call. By launch morning, several files have been replaced. Which images did your approval cover?
For overseas marketing teams working with Chinese manufacturers, the practical answer should live in a release manifest: a fixed list of final files with their review records. ONCE recommends separating checks of product facts from agreed aesthetic sampling, then approving replacements against the work they actually change. A folder name is too loose to carry that decision.
Automated QA does not define the scope of your approval
Photoroom’s September 14, 2026 guide, How to automate product image quality assurance at scale with Visual QA and Fidelity Raters, distinguishes visual quality from product fidelity and describes retries and human review. The category-specific fidelity raters listed there cover food and fashion. This is evidence of one company’s product design, not proof that unattended approval works across every catalog.
Ask a supplier to connect one final output to its reference, checks and results. Then ask whether the delivered file is that same version. A score on an earlier image cannot approve its replacement.
A clean sample can miss a known problem
Consider a hypothetical batch of 100 images containing exactly 10 defective images. Select 10 different files uniformly at random, without replacement. Assume reviewers detect the defect whenever they encounter it. The probability of selecting only acceptable images is:
C(90,10) / C(100,10) ≈ 33.0%. C denotes combinations. Even with ten defective files present, this sample misses every one about a third of the time. This is a calculation under stated assumptions, not a measured AI error rate or a recommended sample size. A supplier’s selection of its best images does not meet the random-selection assumption.
The NIST explanation of acceptance sampling separates deciding whether a lot is likely acceptable from estimating its quality. That distinction is useful here; it does not turn a manufacturing sampling method into certification for ecommerce imagery. Agree sample size and release conditions against actual use, consequences and available history. A stratified visual review may cover important variations, but it does not inherit the probability calculation above.
Make the release manifest part of the purchase
Freeze the final file list when submitting a batch for approval. Record the file ID and version, SKU and sale configuration, approved reference version, processing workflow ID, review status and approver. A checksum can help detect replacement under the same filename. The workflow ID should lead to the model version, prompt or template, crop and export settings actually used, wherever those records are available.
Define two review scopes. ONCE recommends checking buyer-relevant product facts in every final file: garment construction, included accessories, ports, defining text and similar attributes. Tools may assist; unresolved details go to the product owner. Another image passing cannot resolve an unknown detail. For aesthetic checks explicitly eligible for sampling, cover the different product versions, materials and processing routes.
Price generation, factual review, aesthetic sampling, repairs and replacement review as identifiable work. Name who supplies accurate product references and who approves each scope. A limited sample review should not be signed as a claim that every detail was checked. Even a complete review is not a promise of perfect detection.
A replacement needs a new decision
Imagine an apparel brand discovers an extra button in a generated image. Hold that file and trace other outputs sharing its reference, editing workflow or template. Shared inputs identify files to investigate; they do not prove those files are defective. If the connections cannot be established, expand the review instead of assuming only the sampled image is affected.
A wrong reference calls for correcting that reference and reviewing affected outputs. A local retouch calls for checking the change and its consequences. Give replacements new versions and complete their required checks. Unchanged files confirmed to be unaffected can retain their earlier approval records. The handoff should identify both newly approved versions and approvals being carried forward.
Before expanding production, ask for a frozen file list and a clear account of what was checked on every image, what was sampled and what happens after replacement. For related production planning, explore ONCE’s video production articles. The reusable asset is the connection between product references, file versions and defect causes. Any labor or margin benefit needs to be demonstrated in actual project records.
Scope: Sources checked November 8, 2026. The NIST page has no publication date displayed. The manifest is ONCE’s procurement recommendation; the apparel scenario and arithmetic are hypothetical. This article reports no client results or software test. The cover is an original AI-generated concept image.
