The AI Image Quality Control Pass: What to Check Before You Publish

Generated images fail in specific, recognisable ways. That is good news: a defect with a known signature can be checked for deliberately in a couple of seconds, rather than hoped against.

What does not work is the informal glance. A batch of ninety images gets skimmed at thumbnail size, everything looks fine, and three weeks later someone points out that the technician on your main service page has six fingers. At thumbnail size it was invisible. At the size a visitor sees it, it is the only thing they see.

This is a structured pass instead. Per image it takes about fifteen seconds; the batch-level checks take a few minutes and catch the problems that matter more.

The per-image pass, in priority order

Check these in this order, because they are ordered by how quickly a viewer notices.

1. Hands and fingers. Still the most common failure and the most immediately jarring. Count fingers. Check that hands attach to arms at plausible angles and that fingers holding a tool actually wrap around it. Zoom in — a hand three hundred pixels across in a full-size image is a thumb-sized detail at review scale and a clear one on a laptop.

2. Text on any surface. Signage, labels, packaging, screens, vehicle livery, book spines. Generated text is usually approximate glyph-shapes rather than words, and readers spot it instantly because they try to read it. The reliable fix is to prompt for surfaces with no text at all — blank signboards, unlabelled equipment — rather than hoping the letters come out.

3. Object coherence. Follow each significant object to its edges. Tools that merge into benches, chair legs that pass through table tops, cables that begin and end nowhere. These read as "off" before a viewer can say why.

4. Lighting and shadows. Pick the light source, then check every shadow points away from it. Mixed shadow directions and objects with no shadow at all are the most common physics failures, and they are what makes an otherwise good image feel synthetic.

5. Background architecture. Backgrounds get less model attention than subjects. Look for windows at inconsistent heights, doorframes that taper, tiles whose perspective drifts, and repeating patterns that lose their rhythm partway across.

6. Framing and aspect ratio. Confirm the image is actually the ratio you asked for and has no black bars. Letterboxing is a specific, common artefact triggered by cinematic phrasing in prompts, and it is easy to miss on a dark image.

7. Subject accuracy. Does the image show what the page claims? A generic workshop where you specified a diagnostic bench is not a defect in the image; it is a defect in the pairing, and it is the one that most undermines the page.

Defect	Fastest check	Usual fix
Hand and finger errors	Zoom to 100% on every hand	Compose people out of the scene
Garbled text	Read every surface bearing marks	Prompt for blank, unlabelled surfaces
Merged or floating objects	Trace object edges	Regenerate with a simpler scene
Wrong shadow direction	Identify the light source, follow shadows	Specify a single named light source
Warped background	Check verticals and repeating patterns	Shallower depth of field, simpler background
Letterboxing	Look at the outer edges	Remove cinematic language; crop in post
Wrong subject	Compare against the page it will sit on	Rewrite the prompt around the subject

The batch-level checks that matter more

The individual pass catches embarrassments. The batch pass catches the thing that actually wastes the project.

Sameness. This is the important one. A set produced from one prompt with a variable swapped — a city name, an adjective — comes back as one image in forty slightly different colours. If the purpose was distinct imagery for pages that need to look distinct, that outcome is a complete failure even though every individual image is fine.

Test it properly: put every image in the batch on one screen at thumbnail size and look at the grid. Sameness is obvious at grid scale and invisible when reviewing one at a time. If the grid reads as one texture, the prompts varied the wrong things — vary subject and setting, not adjectives. Prompting for genuinely photorealistic, varied results is the upstream fix, and it is far cheaper than regenerating a batch.

This matters because the reason for producing distinct images at all is that reusing one image across many pages costs you the ability to differentiate them. Forty near-identical generated images are functionally the same problem with a larger bill.

Consistency where images appear together. The opposite requirement, and both can apply in one project. Images sharing a page or a carousel should agree on colour temperature, lighting style and framing. A warm, soft image next to a cool, hard one looks like a mistake even when both are good.

Filenames and metadata. Confirm every file has a descriptive, distinct filename and that the metadata written is correct for that image rather than copied across the set. A batch where every file carries the description of image one is worse than a batch with no metadata, because it looks deliberate.

Subject distribution. Count what the set actually depicts. Batches drift — ask for varied workshop scenes and you often get thirty benches and two vehicles. Check the spread against what the pages need before accepting the batch.

Regenerate, edit, or discard

Three options, and picking correctly saves a surprising amount of time.

Regenerate with a changed prompt when the defect is structural — bad hands, garbled signage, warped background. Change the prompt, not just the seed. Regenerating identical wording and hoping is the most common time sink in this workflow; the model failed at that description once and will tend to fail again.

Edit when the image is otherwise exactly right and the flaw is small and localised — one stray object, a blemish, a small area of texture. Inpainting or a quick clone-stamp pass takes two minutes. Do not commit to editing a fundamentally wrong image; sunk cost dominates fast.

Discard when the concept itself is not working after two or three attempts. The subject may be one the model handles poorly — crowds, complex machinery in operation, specific branded objects, text-heavy scenes. Change the concept rather than fighting it. A photograph of the actual thing is sometimes the right answer, and recognising that quickly is a skill.

Where honesty enters the review

One check does not belong to craft.

Ask, for each image: is this pretending to be documentary evidence of something real? A generated image illustrating a service is ordinary marketing illustration. A generated image presented as your workshop, your team or your completed job is a claim about reality, and it is false.

The distinction matters commercially before it matters ethically — customers do notice, platforms remove content, and "before and after" images that were never before or after anything are the kind of thing that ends up screenshotted. Where disclosure and labelling obligations currently stand is worth reading if you publish at volume, because the requirements are tightening rather than settling.

The line that holds up: generated imagery illustrates; photography documents. Use photographs when you are making a claim about a specific real thing, and generated images when you are illustrating a general one. The wider comparison of stock, generated and commissioned photography sets out where each earns its place.

Rights are the other half of this. Confirm the licensing position for the images you are shipping — particularly if a client will own them — and be sure the answer is written down somewhere. What you can and cannot do commercially with generated images covers the current position, and it is a conversation that goes much better before delivery than after.

The checks that belong to the page, not the image

Two items are usually treated as separate work and are far cheaper to do while the image is still in front of you.

Write the alt text now. You are looking at the image, you know what page it is going on, and you know what it is meant to convey. That is the entire input needed, and it is exactly the context that has evaporated by the time someone works through a spreadsheet of missing alt attributes three months later. The rules for writing alt text that is actually useful take five minutes to internalise and then cost nothing per image.

Check the image is not carrying information as pixels. If a generated image has produced something resembling a label, a chart or a sign, decide whether the page depends on reading it — and if it does, that content needs to exist as real text on the page rather than only inside the file. This is the most common accessibility failure introduced by generated imagery, and the failures that alt text alone cannot fix covers the rest of that set.

While you are there, confirm the contrast of any text your design will overlay on the image. A dark headline planned for the top-left corner of a hero fails immediately if the model put a bright window there, and finding that at review costs nothing while finding it after the page is built costs a regeneration.

Making the pass survive contact with a deadline

Any QA process that depends on discipline decays. Two things keep this one alive.

Make it a gate, not a habit. Images do not leave the review folder until the pass is done. One person, one sitting, before upload. A pass that happens "when there is time" happens in month one and never again.

Keep a rejection log. One line per rejected image: what was wrong, what the prompt said. After two batches you will have a short list of the phrasings that reliably fail for your subject matter, and that list is worth more than any general prompting advice — it is specific to your use case and it stops the same defect being regenerated.

For an agency delivering to clients, this is not optional and the log becomes part of the deliverable — the agency workflow for producing client imagery at volume treats review as a named step with a named owner for exactly this reason. A defect that reaches a client costs the review time many times over, and it costs it in credibility rather than in minutes.

Fifteen seconds an image. The one on your highest-traffic service page is worth considerably more than that.