Artifacts are what the image could not resolve.

An artifact is a visible consequence of the generator lacking information. It resolved the subject well and the periphery badly, or resolved the lighting globally and the contact shadows not at all.

That makes artifacts predictable rather than random. They cluster in the places where a photograph would contain data and a description would not: the far side of an object, the space behind a subject, the texture of something incidental.

They also cluster where structure is expensive: repeated identical objects, precise geometry, and anything that must be consistent with something outside the frame, such as a reflection or a shadow.

So inspection is not a hunt for anything odd. It is a short list of places to look, in an order that finds the common failures fastest.

These are the errors that survive scrutiny longest, because they look fine at a glance and absurd on inspection.

Fine texture is generated statistically, which produces two failures: uniform repetition and smearing.

Repetition shows as identical bricks, identical hair strands, identical fabric folds at an identical interval. Look at a patch of wall and move your eye along it; real texture varies in a way that is hard to describe and easy to spot.

Smearing shows as a blur that follows no optical logic, often in the corners or around hair. It reads as a low-quality JPEG artefact but does not respond to sharpening, because there is no detail underneath.

Both are more visible in print and on large screens than at card size, which is why the publication size decides whether they matter.

A generated image usually has coherent global lighting and incoherent local shading. The scene is lit from one direction, and an individual object is lit from another.

Check the direction first: every cast shadow should agree, and the highlights on curved surfaces should be on the same side as the light. Then check softness, which should match across objects at similar distances.

Contact shadows are the most frequent omission. An object sitting on a surface with no shadow beneath it looks pasted, and this is one of the clearest signs of a composite or a generation.

Reflections are the hardest case. A mirror or a window should show the reverse of what faces it, and almost never does, which is worth checking whenever the image contains glass, water or polished metal.

Every legible-looking string of characters is a test. If it is a word, read it. If it is a label, ask whether a label would be there.

The characteristic failure is pseudo-text: characters with the right density and shape that form no word, often at the exact size where the eye would not resolve them in the source but will in the export.

Text also reveals provenance because it is expensive to generate and trivial to add afterwards. A poster with perfect typography is more likely to have been designed than generated in one pass.

If the image is a mockup or a product shot, text errors are the most costly defects on the list, because a client will read the label.

Hair fails at the boundary: strands that end in mid-air, a hairline that changes direction, or locks that pass through a shoulder.

Skin fails in texture consistency: patches that are poreless next to patches that are grainy, or a shine that follows no light source.

Eyes fail in the details a viewer checks unconsciously: pupils that differ in size, catchlights in conflicting positions, eyelashes that merge into the lid.

These are the regions where human attention is strongest, so they are worth checking even when the overall image reads well.

Look at the frame boundary. Generated images often become less resolved towards the edges, and objects are cut in ways that no photographer would choose.

Check what the composition is for. An artifact-free image that leaves no room for a headline is useless for its placement, and a slightly flawed image with clean negative space is the better asset.

Also check for accidental remnants: a signature-like smudge, a watermark ghost, a logo fragment. These are rare and expensive when they appear on commercial work.

Composition errors are the ones that survive downscaling, so review them last, at the size they will ship.

Small artifacts are often cheaper to inpaint than to regenerate: a bad patch of wall, a duplicated object, a bent line in the background.

Anything structural is not. If the geometry is wrong, inpainting a corner will not fix the room, and the effort produces a seam in a different place.

Text should be added after generation, in a design tool, with a real typeface. Regenerating to fix a label usually breaks something else.

For commercial work, budget for a retoucher. An hour of professional cleanup is cheaper than a day of regeneration roulette and produces a predictable result.

Artifacts are the most accessible evidence of generation, and the least reliable proof. A visible bend in a door frame strongly suggests a generated image; its absence proves nothing.

Photographs also contain variants of the same features: lens distortion bends lines, compression smears texture, heavy retouching removes contact shadows.

So use artifacts to decide whether to look closer, and use credentials when the answer has consequences. Content credentials, camera records and a documented edit history are evidence about production; a crooked table is not.

Publishing the distinction protects you twice: you catch your own defects, and you avoid accusing a photographer of something an artifact only hints at.

One-off review works until you publish thirty assets a week, at which point the check quietly stops happening. The fix is to make the check cheaper rather than optional.

Build a contact sheet of four crops per asset: top left, top right, bottom left, bottom right, at full resolution. Ten assets on one page, evaluated in a single pass, is a different task from opening ten files.

Then put the assets in two tiers. Anything that ships at hero size, in print or beside a product name gets full review. Decorative backgrounds at card size get the contact sheet only.

Record the outcome per asset, even when it passes. A log of what you checked is what makes the process real, and it is also the only way to notice that one prompt is producing the same defect every time.

Quadrants at 100% zoom, looking for smearing and repeated texture. Then straight lines, corner to corner. Then contact points and shadows. Then any text.

That order catches the most frequent and most expensive defects in under a minute, which is short enough to become routine.

Then step back. The last question is whether the image works at the size it ships and in the layout it ships in, which is a different question from whether it is technically clean.

Use the sweep on this page to keep it consistent, and record what you found so a brief that produced three failures can be rewritten.