The wipe test: what changed when you asked for one edit

You asked for one edit and got a new photograph

The instruction was to remove a cable, or clean a reflection, or replace a background. The result that comes back looks like the same photograph with that one thing gone, which is the source of the problem: it is a new frame that happens to resemble the old one, re-rendered around your instruction.

That regeneration is invisible when you look at the result alone. You check the cable is gone, it is gone, and you approve. What you have not checked is everything the model also decided to change: the chair that moved four inches, the label on the box that now reads something almost right, the second light source that appeared in a window.

The wipe makes the drift visible by putting both versions in the same coordinates and revealing one over the other. Differences appear as a flicker at the divider, so nothing has to be remembered, listed or found. The change announces itself.

For product photography this is the single highest-value check in the workflow, because the product is usually the smallest part of the frame and the part most likely to be quietly re-rendered.

Align the frames before you judge them

Compare exports, not screenshots. Two files at the same crop and aspect ratio make the divider honest, because any movement at the boundary is a real difference. Compare a screenshot against an export and you will spend the session chasing differences in compression rather than content.

If the edit came back at a different size, crop the larger frame to the smaller one and compare again. A tool that crops both to the frame will hide some of the mismatch, but it cannot invent detail, so a genuine upscale will still show as softness on one side of the divider.

For video, export stills from the same moment in each clip and wipe those. The divider is a stills instrument; a moving comparison tells you about timing, which is a different question from what changed inside the frame.

And wipe slowly near the subject. Movement at the edges of a frame is expected and usually harmless; movement across a product's silhouette is a finding.

Four zones, checked in order

Behind the subject first. Background objects, their positions and the shadows they cast. Shadows are the tell that survives longest, because a model that removes an object often leaves the shadow it was standing in, or produces a shadow that no longer matches the light direction of anything in the scene.

Then surfaces. Glass, screens, polished metal and water all reflect what the edited scene contains, and a reflection of a plant you deleted is the classic failure. It is also the easiest to miss, because reflections are read peripherally rather than examined.

Then the packaging and the print. Text on a box, an embossed logo, a care label, the fine print on a label. Models re-render type as a shape resembling type, and the wipe shows the letterform drift as a shimmer at the divider even when both versions look plausible on their own.

Then the subject itself: colour, gloss, silhouette, and the geometry of something whose proportions a buyer will notice. A product that has gained a slightly narrower neck is a returns problem, not a design preference.

• Background objects, positions and cast shadows.
• Reflections in glass, screens and polished surfaces.
• Packaging text, logos and fine print.
• Product colour, gloss, proportions and silhouette.

Deciding what counts as a defect

Not every difference is a failure. A cleaner background, a removed distraction, a slightly warmer grade: those are edits, and some of them are the reason you ran the job. The question is whether each change was asked for, and whether it is defensible if a customer compares the image against the object.

Write down the ones you accept. An accepted change is a decision, and decisions are what you want on record when a client asks in three weeks why the hero shot no longer matches the packaging artwork.

This matters most for the changes nobody would notice but everybody could: a slightly different shade of blue on a product that comes in two blues, a logo that has been nudged, a background object that implies the product ships with something it does not.

The wipe is also a fast way to catch an edit that failed silently. If the cable is still there, or the reflection is untouched, you will see it immediately rather than discovering it after the asset is scheduled.

Catalogue imagery and the disclosure question

For e-commerce the wipe is a commercial control rather than an aesthetic one. A product that gained a slightly different colour in an edit is a return, a review and a customer service conversation, and none of those show up in a marketing dashboard until the month is over.

Watch for additions as closely as removals. A generative edit asked to clean a background will frequently add objects: a plant, a stylus, a cable, a second cup. Those additions change what the image implies is included in the box, and no amount of visual quality fixes a mismatch between the picture and the parcel.

Decide the disclosure position before the asset is approved rather than after it is published. For a reconstructed lifestyle shot, a caption is increasingly expected and costs nothing. For a product image that a buyer will use to judge a physical object, a reconstructed image is a different category of thing and should be treated as such.

Keep the original file. In any dispute the wipe is only possible if both versions exist, and the version that has been overwritten in a stock library is the one you will need.

Editing text inside an image

Signage, packaging and screen content are where an edit is most likely to be rejected and hardest to spot. Models render type as a shape, and while the shape is plausible at page size, a wipe across a sign shows the letterforms sliding even when both versions look like words.

This matters most for localisation, where an edit is asked to swap the language on a label. A regenerated label rarely produces correct characters in a second writing system, and a customer who reads that language will see it instantly. For anything in-market, retouch the type rather than regenerating the frame.

Legal and price text is the highest-risk category in the whole workflow. An edited image that alters a disclaimer, a dosage line, a currency symbol or a numeric claim is not a design decision, and it is usually invisible on the result alone because the replacement text looks like the text it replaced.

One practical habit: wipe directly along every line of type in the image, slowly. Letterform drift is a flicker at the divider, and it is the only reliable way to catch a change you did not ask for and would not notice.

When the edit is good and the brief was wrong

Some failed edits are technically perfect. The background is cleaner, the product is untouched, the lighting matches, and the image is still unusable because the request itself was wrong: the shot no longer shows something a customer needs to see.

That is a briefing problem and it is worth naming separately from a defect. If the instruction was remove the clutter and the clutter was the thing that demonstrated the product in use, the wipe will show a faithful execution of a bad idea.

The check that catches it is deliberately non-visual. Before you wipe anything, write down what the image must communicate and what a buyer must be able to see. Then wipe and ask whether the edit preserved both.

Two frames, one honest comparison

The comparison only works if both images are the same thing. An original shot on a phone and an edit rendered from a description are not a before and after; they are two different pictures, and the wipe will show differences that belong to the camera rather than to the edit.

Where the source was itself a render, say so in the record. Comparing two generations is still useful for checking that the instruction was followed, and calling that a before-and-after invites a claim about authenticity that the picture cannot support.

The divider as a tool for agreement

Wipes end arguments faster than descriptions. When a client is unsure whether an edit kept their product intact, dragging a divider across it answers the question in three seconds and without anybody defending their judgement.

That makes it worth keeping the two files even when the edit is approved. A later question about the original framing is a wipe away from an answer, rather than a request to the retoucher and a wait.

Where a wipe stops

It shows what changed, not why, and not whether the result is fair. An image of a person can be edited into something misleading with no drift at all — the same pose, the same lighting, a different context, and the wipe shows nothing because the frame is otherwise identical.

It also says nothing about consent, licensing or disclosure. Those are decisions made before and after the edit, and they are usually the ones that matter legally. A perfectly executed edit of a stock photograph can still be unusable because the licence does not cover generated derivatives.

And it cannot tell you whether the image is a fair depiction of the product. If the original was already a render, comparing an edit against it compares two fictions, and the wipe will report a clean result.

So treat the wipe as the fastest way to answer one question — did this edit do only what I asked — and route the rest to a person who can answer the others.