AI product photo editing: change the background, not the product
Review AI-edited product photos by setting permitted changes, checking masks and comparing approved features with the original.
Article language: English

In this article
The instruction sounds harmless: “Change the background to a bright workshop; leave the product alone.” After editing, the ceramic jug sits in beautiful light, but the opening in its handle is narrower, the rim has straightened and small irregularities in the glaze have vanished. Marketing has received a better photograph and a slightly different jug. The second one is not waiting in the warehouse.
Take this as a hypothetical image to review. The goal is a new setting, not a redesigned vessel. Before choosing the most attractive result, record what may change and identify the parts of the product that must match the approved photograph. “It should look the same” is too broad a criterion to settle a disagreement about the handle.
Reviewing an image edit differs from judging its overall appearance. A photograph can be cleaner, brighter and more elegant while misrepresenting a feature a buyer needs. In that case, the background change is only partly successful. Separate the two results instead of averaging them into “very nice”.
Agree the change on a specific image
Discuss scope with the selected file in front of you, not just the product model name. The same jug has different occlusions, reflections and visible parts in different photographs. One frame shows the entire handle; in another, a hand hides half of it. Replacing the hand with the surroundings requires filling in the obscured section. That is a greater intervention than changing a visible wall in the background.
The product owner confirms that the photograph shows the right version and may be used as the reference. If the spout shape has changed, images from an earlier shoot should not define the current offer’s appearance. When the source is ambiguous, establish the correct reference first.
Check whether the source file allows the feature to be assessed at all. If a printed mark consists of a few blurred dots in the original, a newly legible letter is not recovered evidence of what it says. It may be an invention. A sharper photograph or an approved artwork will be needed. Enlarging an image gives a bigger view; it does not supply information that could not be recognised in the source.
For our jug, the permitted change might cover the wall, tabletop and background visible through the handle opening. Preserved features include the vessel’s outline, handle proportions, spout and distinctive glaze marks. Record whether the direction of light and the contact shadow—the shadow where product meets surface—may change. Those elements sit at the boundary between object and setting and are easily missed in a short brief.
Establish what the photograph is for. A mood image can show the atmosphere of a collection; a product-page photograph should help someone recognise the variant being sold. A scene with an extra lid suggests that the lid is included, even if it was added only to “finish the composition”. Scope therefore needs to relate to what the customer understands, not just the number of altered pixels.
These decisions begin when choosing brand photography for a website (in Polish). Once you know why the image is being made, it is easier to decide whether a new decoration helps or changes the offer being shown. You can approve a bold setting while keeping the handle’s exact shape.
Do not call removing a product feature “cleaning up the photograph”. A crumb beside the jug and a small irregularity in its glaze belong to different categories. Someone who knows the object decides whether a mark is accidental dirt, a defect of that specimen or a feature of the collection. The editing tool should not make that decision for them.
For text on a product, supply the approved wording and the correct mark variant. A decorative stroke in a logo can look like an accidental mark that a model decides to tidy up. A logo package with the correct variants (in Polish) gives the team a reference instead of asking it to reconstruct lettering from memory.
A mask identifies an area; the result still needs comparison
A mask is an image that marks the area to edit. OpenAI’s GPT Image documentation says the model uses it as guidance and may not follow its shape with complete precision. That specific limitation is described in the image-editing guide. Using a mask therefore does not confirm that every point of the product stayed unchanged.
The Gemini guide recommends specifying which details to preserve alongside the edit request. This advice on preserving image details helps prepare the instruction. Acceptance still requires looking at the result. Writing “keep the handle opening” and actually keeping it are different parts of the work.
In our example, marking the background must include the gap beneath the handle. Leaving part of the old wall there creates a visible setting error. Painting the entire opening the colour of the ceramic changes the construction. The reviewer should inspect the opening’s outline and interior, not just the vessel’s outer silhouette.
Some materials complicate the division. A transparent wall shows the setting through the product; a mirror surface shows it in reflection. A rule that “everything inside the outline stays” may preserve the old background. A rule to “fix everything reflecting the room” may change the surface’s shape. Agree the permitted scope and the material needed to resolve doubtful areas.
Begin comparison with the full frame. Is it the same variant, with the proportions and all visible parts retained? Only then move to enlargements. Several flawless details cannot repair an overall shortening of the vessel. Equally, a similar silhouette does not prove that a cable opening or the correct lettering survived.
If view and scale have stayed the same, switching between two images in the same position can help. An overlay or difference image can find changed areas, but does not determine their meaning. A lighting correction can change many pixels without altering shape. A missing small notch can affect few pixels and change how the object is used.

Keep viewing conditions comparable. If one image is smaller or more compressed, some differences may come from file preparation. First align versions at the same scale and inspect the area without extra display filters. Then assess the final thumbnail separately. The tests answer different questions: whether a feature changed, and whether it remained legible after preparation for publication.
If an edit rotates the product or creates a new perspective, a simple overlay is no longer the right test of conformity. You need a reference from the corresponding view or other confirmed material. Do not judge an unseen side of the jug by whether the generated image looks plausible. A convincing appearance is not enough to establish the shape of this particular object.
Record the result against each feature: retained, changed or unresolved. “Handle matches from the front; back not visible in source” is more honest than “product preserved”. A limited scope does not prevent use of the frame if the frame is meant to show what was actually checked.
Compare every version with the original too
Retouching often happens in several passes: first the background, then the shadow, then removal of a small artefact. Save each result as a separate file. After the next change, compare it with both the previous state and the approved original. Otherwise, a gradually altered handle can pass through five small corrections and end up as an entirely new part.
Record the instruction’s scope and a link to the reference used with the files. There is no need to reconstruct the entire conversation in a photograph’s filename. A recognisable version and a note identifying what changed and which part has already been approved are enough. The article on the provenance of AI images in company materials (in Polish) develops that audit trail.
If you adjust the shadow under the jug, inspect its base again. A local goal does not guarantee a local effect. The edge may become thicker, or a dark detail may merge with the shadow. In the acceptance record, identify the corrected area and adjacent features checked after the change. For this correction, inspect the vessel’s bottom, outline and point of contact with the table.
For a colour dispute, return to the material’s purpose. Atmospheric light may warm the whole frame, but if a customer chooses between light and dark glazes, the distinction between variants must remain understandable. An assessment on an arbitrarily configured screen is not a colour measurement. If colour needs precise confirmation, agree a separate control method with the appropriate person.
Before another generation, decide whether the correction really requires processing the whole image again. Sometimes conventional retouching, a deliberate background composite or a new detail photograph is enough. Choose the method for the required result. There is no prize for involving the most talkative tool in the simplest removal of a speck.
In a report, avoid “not very realistic” when you can identify a specific error. “After editing, the gap between handle and body was closed” leads to a concrete correction. Lighting credibility and aesthetics can be assessed separately. The designer then knows whether to change the mood or restore a part of the product.
Attach two corresponding crops with the area marked, plus a full view of both files, to the report. A crop alone may hide a change in perspective. The full frame alone may make a narrow opening difficult to find. “Version 3, upper-right part of the handle, opening reduced” lets the next correction be checked without re-establishing what the conversation concerns.
When the same image becomes the starting point for animation, retain the approved version’s identifier. The next stage should start with the right file. Continuity of an object and action in AI video (in Polish) needs an additional check for changes over time; approval of a photograph does not automatically cover later movement.
The approved image must reach the right offer
After choosing a result, check the files prepared for publication. A crop can remove part of the handle, compression can blur the lettering and a new thumbnail can show different equipment from the main image. Accepting the large source file does not replace looking at the version a customer will see on a phone.
Assign the approved image to the specific catalogue page and variant. A file called only “jug_final” is easy to swap between sizes. At handover, identify the model name, glaze version and purpose of the frame. If an image was prepared only as a general presentation of the collection, it should not automatically replace photographs of every variant sold. Correct retouching alone does not prevent a wrong assignment.
Check the description too. If the photograph shows a scene with extras, the text should not describe them as part of the set. The problem can also run in reverse: an automatic description may add a function based on a generated detail. For AI-written product descriptions (in Polish), the basis should remain confirmed features, not an interpretation of retouching.
In a product gallery showing details (in Polish), pair the styled image with a view explaining the construction. Both must show the correct variant. A correct detail does not excuse an incorrect silhouette in the main photograph, though it can add information the mood image simply does not show.
Approval should identify the specific file, intended use and person confirming product conformity. Keep unresolved features separate. If the project does not cover assessment of material, colour or construction properties, do not turn an image review into that expertise. You can, however, decide whether the file is fit for the agreed presentation.
A review of the tool, material workflow and approval rules can be discussed through Aeneia Audit. The result should be a clear decision about this photograph, not blanket approval for every later generation. The bright workshop can stay. The new, uncommissioned shape of the jug needs another pass.







