Most tools that fix an AI-generated product image fix it by generating a new one over the top. For a single hero frame that is usually fine. It stops being fine when the pixels that ship have to be accountable, and when forty images have to match each other.
A good number of tools now exist for this, and the honest position is that for one image, quickly, they are faster than us and cost a fraction of what we do. We would rather say that plainly than pretend the category does not work.
What follows is the part that is harder to find written down: how the fix is actually made, and what that means for the image you put on a product page.
Speed, cost, and no minimum. You upload one frame, brush the wrong area, and get a result in seconds. There is no brief, no onboarding, no invoice to approve. For a single lifestyle image, a social post, or a test of a concept before anyone commits budget, that is the right tool and we will say so to a buyer who asks.
They are also improving quickly. Nothing below is an argument that they do not work.
There are two different operations hiding behind the same word.
The first is regeneration. The tool takes the area you marked, takes a reference you supply, and generates replacement pixels that resemble the reference. This is the usual mechanism, and several vendors describe it openly in their own documentation: the logo is rebuilt from your input. The result can be very close. It is still machine-made.
The second is reconstruction by hand. A retoucher opens the frame alongside the real product and rebuilds the wrong area from the actual artwork — the vector logo, the correct Pantone, the real weave — in layers, by hand. It is slower. The pixels that ship were put there by a person who can be named.
Both are reasonable. They are not interchangeable, and the difference only shows up in three situations.
A tool rebuilds from the reference you hand it. If that reference is low resolution, shot under different light, or quietly the wrong colourway, the tool will still produce something plausible. It has no way to know the reference was wrong, and a plausible result is the dangerous kind, because nobody queries it.
A retoucher looking at a brief with eight colourways in it notices when the reference supplied for one of them does not belong to it. That is not a clever feature. It is just someone paying attention, and it is the thing most often missing from an automated pass.
You may be asked — by a marketplace, a retail partner, a regulator, or a customer who thinks the product looked different online. The question is how the final image was produced.
If the correction was made by generating new pixels, then the part that shipped was machine-made, whatever the step was called in the interface. That is worth knowing in advance rather than discovering it in an email. We have written separately about where the disclosure line actually sits, which is narrower than most teams assume: does retouched product photography need an AI label?
This is the one that costs money, and it is invisible until volume.
A tool judges one image at a time. It has no memory of the other seven colourways of that garment and no opinion about whether all eight will sit together in a category grid. Corrected one by one, each result is defensible on its own and the set drifts — a little warmer here, a little more contrast there. The buyer does not notice any single image. They notice the row.
Holding a set together is a different job from fixing a frame. It needs one retoucher across the whole set, a reference that does not move, and someone checking the set as a set before it goes out. More on what that pass involves: AI image correction, and for generated models specifically, what still has to be done by hand.
| If this is true | Then |
|---|---|
| One image, fast, nothing much riding on it | Use a tool |
| You are testing an idea before committing budget | Use a tool |
| The logo, printed text or trim must match the real product exactly | Hand pass |
| A range has to stay consistent across every frame | Hand pass, one retoucher across the set |
| You need to be able to state how the pixels were made | Hand pass, in writing |
| Hundreds or thousands of frames a month to a fixed standard | Hand pass with QC |
You do not need to audit anyone. One question separates the two mechanisms:
Were the corrected pixels generated, or reconstructed by hand?
There are three answers you might get, and each tells you something:
“Generated from your reference.” Clear and usable. Now you know, and you can decide whether it matters for that image.
“Reconstructed by hand, and we will put that in writing.” Also clear. Ask for the written version anyway, because an answer you can forward is worth more than one you remember.
A description of the outcome rather than the method — “pixel-perfect”, “studio quality”, “AI-assisted”. That is not an answer to the question. Ask it again.
We correct and refine; we never generate. Every image is retouched by hand, and that includes frames a client generated themselves before sending them to us. Client files never pass through a generative tool in our workflow and are never used as training data for any model. We will confirm that in writing, and we will sign your DPA.
We are not the cheapest answer to “fix this AI image”. We are the answer when the corrected image has to survive being looked at closely, and has to match the other forty beside it. Rates are published rather than quoted on request: pricing. A free test is five images back within six hours, and live work runs 24 to 48 hours.
For one image, quickly and cheaply, a tool is often the right choice and we will say so. A hand pass is the right choice in two cases: when the corrected pixels must not themselves be machine-generated, and when a set of images has to stay consistent with each other.
If the correction was made by generating new pixels, then the part that shipped was machine-made, whatever the step was called. Whether that triggers any disclosure duty depends on what the image shows rather than on the retouching itself, and we have written about that line separately.
Because it evaluates one image at a time. Each correction is defensible alone, but without a single reference and one person across the whole set, the set drifts. The inconsistency shows up in a category grid rather than in any single frame.
No. We correct, retouch and refine images that already exist, including generated ones. We do not produce generated imagery, and that boundary is deliberate rather than a limitation of capacity.
We can usually work from a poor reference, but we will tell you it is poor rather than quietly inventing the missing detail. If the reference is wrong for the colourway, we will flag it before correcting forty images against it.
This article describes how image corrections are produced and is not legal advice. Check anything here that carries consequence with your own counsel.
Photographed or generated — send the frames and the garment reference, and judge the match.