- Your phone merges several frames into one image by default, which the standards vocabulary already calls computational capture rather than a plain photograph.
- The IPTC vocabulary separates algorithmic enhancement, which does not change the main content, from editing with a generative model, which does.
- Background removal and exposure correction sit on the safe side of that line. Generative fill, scene replacement and recolouring do not.
- Google requires a unique image for every variant and only one variant per image, which is the rule most small catalogues quietly break.
- The minimum image size for Merchant Center rises to 500 by 500 pixels for all products from 31 January 2027.
You photograph forty products on a kitchen table with a window on the left and a phone in your hand. Then you open an app that removes the background, brightens the shadows and straightens the lot in one pass. Somewhere in there, the question of whether you still have a photograph of your product has an answer, and it is more precise than you might expect.
It is worth knowing where that line is, because marketplaces and search engines have started to ask, and because product photography is one of the few things a small seller controls completely.
What does your phone do before you touch anything?
It merges several exposures into one image and makes a large number of decisions about noise, sharpening and tone on your behalf. The output of a modern phone camera is not a single capture, and the standards people have a name for it.
The IPTC digital source type vocabulary, which is the controlled list the major platforms draw on, distinguishes digital capture sampled from real life from multi frame computational capture, the latter defined as the result of capturing multiple frames from a real life source and then automatically merging them. Night mode, high dynamic range and portrait mode all land in the second category.
This is not a problem, and nothing requires you to declare it. It is useful context, because it means the distinction people reach for in conversation, between a real photo and an edited one, was already blurred before any AI tool opened. The line that actually matters is drawn somewhere else.
Where is the line that matters?
Between changing how your product is depicted and changing what is depicted. The same vocabulary draws it precisely, and the wording repays reading.
Algorithmically altered media is defined as modification or correction by algorithm without changing the main content, initiated or configured by a human, with sharpening and noise reduction given as examples. Editing using generative AI is defined as augmentation, correction or enhancement using a generative AI model, such as with inpainting or outpainting operations. The first keeps your product; the second begins to invent parts of it.
Two further entries cover what most shop editing produces. A composite of captured elements is a mix of several elements that are all captures of real life. A composite including generative AI elements is a mix where at least one element came from a model. Drop your cut out kettle onto a photographed marble surface and you have the first. Drop it onto a generated kitchen and you have the second.
| What you did in the app | Where it lands | Product still depicted truthfully? |
|---|---|---|
| Brightened, corrected colour, sharpened | Algorithmically altered | Yes, if the colour is now more accurate rather than less |
| Cut out and placed on plain white | Human edit or composite | Yes |
| Placed on a photographed surface | Composite of captured elements | Yes, if the scale is honest |
| Placed in a generated room | Composite with generative elements | The product yes, the context no |
| Generative fill to extend a cropped edge | Edited using generative AI | Depends entirely on what got filled in |
| Recoloured to show another variant | Edited using generative AI | No, and this is the one that causes refunds |
The last row is the one worth tattooing somewhere. Generating a colourway you did not photograph is cheap, convincing and the fastest route to a return, because the customer receives a colour that does not match the image and is entirely right to say so.
What does the platform actually require?
Accuracy per variant, no promotional furniture, and a minimum size that is about to rise. Google's image requirements are the most explicitly documented of the major channels and they are a reasonable proxy for the rest.
On variants, the guidance for the image link attribute asks for a unique image representing the distinguishing details of each variant, states that you should not reuse one general photo across different variations, and adds that only one variant should appear per image. For a small maker selling one object in six colours, that is six photographs, and it is the requirement small catalogues most often fail.
On overlays, product images must offer an unobstructed view of the item, with post production text, marketing callouts, badges, barcodes, watermarks and superimposed logos all listed as grounds for disapproval. Logos are permitted only where they are an inherent physical part of the product. Placeholders are separately prohibited: no coming soon graphics, no company logo standing in for a photograph, no shared stand in image reused across products.
On size, the page carries a dated change worth diarising. The minimum rises to at least 500 by 500 pixels for all products beginning 31 January 2027, with 1500 by 1500 or above recommended for best performance across listing formats. There are ceilings too, with no image larger than 64 megapixels and no file larger than 16MB.
There is also an automatic repair mechanism that is useful to know about and unwise to rely on. Merchant Center can attempt to fix promotional overlays for you by removing detected text, watermarks and logos, but the documentation is candid that it may not resolve everything, naming complex multi panel collages, heavy text overlapping the product and dense watermarks as cases where the original image is kept instead.
Requirements and dates here are as published at the time of writing, October 2026, and apply to one channel. Marketplaces set their own image rules and several are stricter about generated imagery. Check the policy for each place you list before you reshoot a catalogue.
What about the metadata?
Where you have used a generative tool, the tag that says so is expected to survive your editing pipeline. Google's image guidance states that images created using generative AI must carry metadata indicating this, gives the IPTC digital source type property as the example, and asks that such tags are not removed.
The practical risk is not refusal to comply, it is loss in transit. Resizing scripts, format conversions and bulk upload tools strip metadata routinely and without telling anyone, so an image can arrive stripped despite nobody deciding to remove anything. We went through that failure mode, and what the EU AI Act adds on top, in our piece on how an AI product photo is allowed but a missing tag is not, which is the one to read if the labelling side is your concern rather than the production side.
A workflow that keeps you on the right side
Shoot more than you think you need, edit in one pass, and never let a generative step touch the product itself. The rest is logistics.
Fix the light before you fix anything in software. A window and a sheet of white card as a reflector beats any amount of shadow recovery, because recovered shadows carry colour casts that then make your product the wrong colour. This is the step that most affects returns and it costs nothing.
Shoot every variant you intend to list. Given the one image per variant rule, this is not optional, and it is far quicker in one session than discovering it a month later. Lay them out, shoot them in the same position with the same light, and you have a consistent set for free.
Include the shots that answer questions. The scale reference, the underside, the texture close up, the thing in a hand. These are the images that reduce support contact, and they are the ones a seller skips when running late. Depth of field is the usual obstacle on a phone at close range, and the standard technique for it is focus stacking, combining several frames shot at different focus distances into one image that is sharp throughout. Several phones now do this without being asked.
Do the boring edits in bulk, the risky ones never. Exposure, white balance, straightening, cropping to a square and background removal all run safely across forty images at once. Generative fill does not, because it produces a different invention each time and nobody reviews forty of them properly.
Check colour on a second screen. Your phone screen is bright and saturated. Looking at the batch on a laptop, ideally next to the actual object, catches the one that went green before a customer does.
Is generated lifestyle imagery ever worth it?
Sometimes, and the test is whether the generated part makes a claim. A kettle photographed accurately and placed in a generated kitchen is making no claim about the kitchen, and nobody buys the kitchen.
The problems start when the generated context implies something about the product. Scale is the usual culprit: a bag generated onto a model's shoulder communicates a size, and if that size is wrong you have made a representation you cannot support. Materials behave the same way, because generative tools render fabric and metal with a uniform plausibility that flattens the difference between your linen and a cheaper blend.
So the usable rule is that generated backgrounds are a presentation choice while generated anything touching the product is a factual claim. Keep the main image a photograph against plain background, which is what the platform rules push you toward anyway, and confine atmosphere to the secondary images where the buyer has already seen the object plainly.
Why does an honest photo still cause returns?
Because colour and scale are the two things a photograph communicates worst, and neither is fixed by better equipment. Both are judgement calls your editing pass makes on your behalf, usually in the direction of making the product look nicer.
Colour is the larger problem and the easier one to improve. A customer's screen is not calibrated, your screen is not calibrated, and the light you shot in has a cast you stopped noticing an hour in. Nothing makes this perfect, but three habits remove most of the gap. Shoot everything under the same light rather than across a day as the sun moves. Include one genuinely white object in a reference frame so you can correct against something known. And describe the colour in words in the listing, because "warm mustard, closer to ochre than to yellow" survives any screen and an image does not.
Scale is simpler to fix and more often ignored. One image in every gallery should contain something whose size everyone knows, held or placed next to the product. A hand works for small objects, a standard mug for homewares, a doorway for furniture. Dimensions in the description are necessary and insufficient, because almost nobody converts centimetres into a mental picture accurately, and the ones who try get it wrong in the optimistic direction.
There is a commercial reason to care beyond the refund itself. A return for "not as pictured" is the most expensive category you can generate: you pay the outbound shipping, often the return shipping, the handling, and the item comes back having lost its newness, all for a sale that was never going to stick. Everything in this section costs an afternoon once and then applies to every product you ever list.
Should you rephotograph an old catalogue?
Only where the images are failing a rule or losing you money, and in that order. Reshooting everything is a project that stalls at product fifteen, and a half reshot catalogue looks worse than a consistently mediocre one.
The sequence that works is to fix disapprovals first, because those products are currently invisible. Then fix the best sellers, where a better image has the most traffic to act on. Then fix anything whose return rate is above your average, since that is where the pictures are most likely to be misleading. Everything else can wait for the next time you handle the stock anyway, which for most small shops is a natural rhythm rather than an extra task.
The images you are producing are also data
Every one of these files acquires a filename, a size, a position in a gallery and a description, and all of those are read by software long before a person sees them. Treating the shoot as a data exercise rather than an art exercise is what makes a catalogue maintainable.
Consistency is the whole of it: same aspect ratio, same background, same order of shots across every product, predictable filenames that contain the SKU. A catalogue built that way can be re exported, re cropped or migrated in an afternoon. One built as forty individual decisions cannot, and the cost arrives the day you change platform or channel.
That consistency is also what makes the surrounding fields cheap to fill. The alt text on each image, which is both an accessibility requirement and a thing search engines read, is far easier to write when image three is always the scale reference, a point we worked through in how to sort shop images by role before generating alt text. The same structure feeds your shopping feed, where the image rules above are enforced automatically and silently, as we covered in what a product feed is judged on.
If you are setting up the shop rather than maintaining one, deciding the image contract early is one of the few choices that stays cheap forever: how many images per product, in what order, at what ratio, named how. It is the kind of structure a well built store builder should be imposing on you rather than leaving open, because the freedom to do it differently on every product is not a freedom anybody wants in year two.
What to do with the forty photos on your phone right now
Check whether every listed variant has its own image. That is the most likely rule you are currently breaking and the one with an immediate commercial cost, since a disapproved product shows nowhere.
Check the shortest edge of your smallest image against 500 pixels, with the January 2027 date in mind, and note that the recommendation sits three times higher than the floor.
Then look at your main images and ask, for each one, whether anything in the frame was invented. If the answer is yes for the product itself, reshoot it. A photograph of the thing you are actually selling, taken badly on a phone next to a window, outperforms a beautiful generated image of something slightly different every time a parcel is opened.