Cropping in on the detail that gives the generation away

Artifacts live in the small parts

A generated image is judged as a whole and generated in patches, which is why the failures sit at the joints: fingers where a hand meets an object, ears where hair meets skin, lettering where a label wraps around a curve, the edge where a subject meets a background it was never photographed against.

At full frame those regions are a few hundred pixels and the eye completes them. A viewer's visual system is extremely good at inferring a hand from a plausible blob, which is why the same image that looks convincing in a feed looks wrong when someone opens it on a larger screen.

Cropping and magnifying replaces inference with inspection. You are not looking at the image any more; you are looking at the part of the image that a model had the least evidence about.

The magnification is not for drama. It is the difference between yes, that is a hand and no, that is five fingers arranged around a sixth stump, and only the second is a note somebody can act on.

Where to look first, in order

Hands, always. Fingers, knuckles, the way a grip wraps an object, whether a thumb emerges from a plausible place. Hands are the most reliable tell because they are complex, articulated and almost never the subject of the prompt, so the model has less instruction to follow.

Then text. Any sign, label, screen or piece of packaging. Magnified lettering from a generative model tends to be near-words: the right number of letters, the right general shape, and no actual language. A crop makes that unmistakable.

Then ears, teeth and eyes, especially at three-quarter angles. Teeth are frequently merged into a single ridge; ear cartilage frequently resolves into a swirl. Both are invisible at thumbnail size and comic at 4x.

Then the seams: hair against a background, the boundary of a shadow, a reflection meeting the surface it reflects, straps and collars where two materials meet. Seams are where generation shows its patchwork, and they are the failures that survive the longest because nothing looks obviously wrong in isolation.

• Hands: finger count, grip, thumb placement.
• Text: sign, label, screen and packaging lettering.
• Faces at angles: ears, teeth, eye direction.
• Seams: hair, shadows, reflections, straps, collars.

Crop, magnify, then decide

Draw the crop tight. A region containing the whole torso tells you nothing; a region containing one hand at 4x tells you everything. The mistake is to crop for context when what you need is resolution.

Look at the crop twice: once as a picture, once as a set of shapes. The first pass asks whether it reads as a hand; the second asks whether the parts are where parts go. Generative artifacts survive the first question and fail the second.

Then classify the finding rather than describing it. Minor, meaning a viewer will not notice at the intended size. Visible, meaning anyone looking closely will see it. Disqualifying, meaning the image cannot be published without embarrassment or a correction.

That classification is what makes the check portable. A hero image at 2400 pixels wide has a higher bar than a thumbnail in a list, and the same artifact can be disqualifying in one placement and irrelevant in another.

Crops as briefs for regeneration

A crop is not only a review artefact. Sent back to whoever generates the image, a tight crop of one hand is a far better brief than a sentence describing the problem, because it removes the argument about what looks wrong.

That changes the economics of rework. A vendor receiving a description will produce something similar and hope; a vendor receiving a crop of four fused fingers knows exactly where to look and can fix it with a smaller change, or mask that region and regenerate it.

Keep the crop with the finding. Six crops from one image, each with a one-line note, is a work list that can be closed, and it is also the record that shows what was fixed and what was waived.

The same trick works with clients. When somebody says the product looks off, a crop next to the specification photograph turns a disagreement into a measurement, and disagreements that can be measured get resolved in one round instead of three.

Compression artefacts versus generation

Not every soft edge is a model's mistake. A JPEG saved at low quality produces blocking around high-contrast edges, resizing produces ringing, and a screenshot of a scaled image produces softness everywhere. Reading those as generation artefacts wastes a regeneration round.

Compression artefacts have a signature: they repeat across the image in blocks, they cluster around edges rather than inside objects, and they get worse the further the file has travelled. Generation artefacts do the opposite: they are local to complex regions, they are absent from flat areas, and they do not change if you re-save the file at a higher quality.

The quick test is to compare the crop against the original export. If the softness is in the export, it is a generation problem. If it appeared after the file was optimised, size-reduced or screenshotted, it is a pipeline problem and the fix is in the asset pipeline rather than in the prompt.

This distinction matters commercially: one requires another generation, the other requires not exporting at 40 percent quality.

What a reviewer's eye skips

Human attention moves to faces and to the centre of the frame. That is useful in a photograph and unhelpful in an inspection, because artifacts cluster in exactly the places attention does not go: hands at the edge, feet under a table, the fourth object in a row, the background above a head.

The fix is to inspect by region rather than by looking at the picture. Divide the frame into nine and examine each one at magnification in order, whether or not anything there interests you. The boring regions are where the finding is.

Repetition is the second blind spot. A generated crowd will contain the same face twice, a shelf will contain the same label repeated, and the eye forgives it because repetition is normal in the world. Count instead of looking: three reflections where there should be two, the same logo on two different products.

And symmetry is the third. A generated building, face or object is often almost symmetrical, with one side mirroring the other imperfectly. The imperfection is a few pixels and it is a reliable signal, if you think to look at the half of the image nobody photographs.

Crop the same detail in the whole set

For a campaign with twenty images from one generator, the useful unit of inspection is the detail rather than the image. Crop the same region across all twenty and look at them side by side: the same hand failing in the same way is a prompt problem, and one hand failing in one image is a luck problem.

That changes the fix. Twenty failures on the same feature means the model has no reference for it and the prompt needs an example or the workflow needs a different tool. One failure means regenerate and move on.

It also makes the review affordable. Twelve crops take a few minutes and produce a pattern; twelve full-image inspections take an hour and produce twelve opinions.

Keep the contact sheet with the findings. A grid of crops is the most persuasive artefact available when somebody asks why a set should be regenerated rather than patched.

What a clean crop proves

That the region you examined holds up under magnification. It is real evidence, and it is the kind of evidence that stops a bad asset reaching a homepage. It is also strictly local.

It says nothing about the rest of the frame, about the composition, or about whether the image is a fair depiction of the thing it claims to depict. A perfect hand on a fabricated product is still a fabricated product.

It says nothing about rights, either. An image can be technically flawless and still be unusable because its subject did not consent or because the generation imitates a living artist, a real person or a trademark. Those are questions about permission, and they are answered by a contract rather than a crop.

The workflow that holds up is the boring one: inspect the artifact, check the bytes, check the claims, then decide about disclosure. Each step is small, and skipping any of them is the reason image QA is the discipline people discover after an embarrassing launch.