Most supplier photos show up looking like they came from five different shops. One SKU sits on a grey card. Another was shot on somebody’s workbench, next to a coffee mug. A third is cropped so tight the handle’s gone.
Reshooting all of it costs a lot, so sellers lean on AI photo editing for supplier images to tidy things up. With Pollo AI you type what you want (“remove the cables, leave the label alone”) and it can run the same edit across a whole batch. That’s a big time saver. It’s also where things can quietly go sideways.
The edit that looks perfect and ships wrong
A person retouching with a clone brush copies pixels that already exist. A generative editor works differently. Ask it to remove an object or widen the frame, and it fills the space with whatever seems most likely to belong there.
For a beach photo, that’s harmless. For a product you’re selling, “most likely” and “accurate” can be two different things.
Here’s a made-up but very typical case. A supplier sends a crossbody bag shot with the strap cut off at the edge. You extend the canvas to a square, and the AI finishes the strap for you, complete with a small silver buckle. It looks great. The real bag has no buckle, and you find out when the buyer opens a return.
Returns like that were climbing before generative tools went mainstream. Back in 2019, Klarna surveyed 2,000 UK shoppers and found 19% had sent something back because it looked different online, up from 8% in 2017. Eighty-one percent said better photos and descriptions would help. It’s old data, and it says nothing about AI specifically. It does show how little room there is for an image that drifts from reality.
Start before the editor opens
1. Check whether the photo is worth saving
If the product is blurry, partly blocked, shown in the wrong variant, or carrying packaging you can’t read, stop there. An editor can’t recover detail the camera missed. It’ll invent replacement detail, and that detail will look crisp and believable.
2. Write down what the product actually is
Colour, shape, materials, what’s in the box, where the logo sits, the exact wording on the label. Call it a product-truth checklist. It takes about five minutes per SKU and gives you something firmer to judge against than “looks fine”, which is exactly the bar a good generative model is designed to clear.
Getting the frame right for each marketplace
3. Sort the canvas out early
Tight supplier crops rarely suit every slot you need: a square listing, a wide banner, a sponsored placement. So expand product images to those ratios before you add badges or copy.
Then look hard at the new outer strip. Pollo’s own FAQ says its uncrop tool currently decides on its own how the missing area gets rebuilt, and you can’t steer it yet. Pay most attention wherever the product touched the original crop line. That’s where straps, cords and handles get “finished” for you.
4. Pick one background rule per category
Amazon makes the main-image decision for you: pure white at RGB 255, 255, 255, with the product filling roughly 85% of the frame and accurately showing the item you’re selling. Off-white fails the automated check, even when it looks white to you. Lifestyle thumbnails are more flexible. A soft neutral works, provided every product in the range gets the same one.
Cleaning up without changing what you sell
5. Only delete things that aren’t part of the order
Hands, stray packaging, loose cables and random props can go. Anything that ships in the box stays in the shot. Take the charging lead out of a headphone photo and you’re advertising a different bundle.
6. Light first, colour second
Brighten a dark subject and calm the hot spots before you touch colour at all. Get the order wrong and you can end up “fixing” a shadow into a shade the product doesn’t come in.
7. Colour-match to something you trust
That might be an approved sample, a manufacturer’s colour code or an in-house photo you’ve already checked. The temptation is to push saturation because brighter clicks better. A navy jumper nudged towards royal blue gets clicks and then gets returned.
Typing “don’t change the logo” helps less than you’d hope. Models drift, and if you’ve ever wondered why AI often ignores instructions about what to leave alone, the same thing applies here. Be specific about what’s protected, and check it afterwards anyway.
8. Watch edges and shadows
A cutout with no contact shadow looks like it’s floating, and shoppers pick up on that even if they couldn’t explain it. Zoom in on handles, mesh, fur, glass, chrome and the tiny gaps between straps or chair legs. Halos and smeared reflections tend to show up there first.
One approved image, then the variations
9. Sign off a master before you resize
Edit once, check it against your product-truth list, approve it, and only then crop out the square, portrait, and landscape versions. Editing the product again for each channel is how slight differences creep in.
Batch editing makes this sharper. If one setting is off, it’s off on all 40 files. It behaves like other silent failures in automation: a SKU recoloured a touch too warm won’t throw an error. It just sits in the catalogue until someone complains.
10. Look small, then look big
At thumbnail size, ask whether the product reads instantly on a phone screen. At full size, hunt for what AI tends to mangle: tiny label text, logos, product edges, repeating textures like knitwear or wood grain, and any patch of rebuilt background.
A quick reference
| Problem in the supplier photo | What AI can safely do | What you still check |
| Cluttered setting | Strip props, simplify the scene | Outline and contact shadow survive |
| Flat or patchy light | Relight, balance exposure | Texture and real colour hold |
| Wrong shape for the slot | Extend or recrop | No invented parts, scale still comparable |
| Dust or a small mark | Spot-retouch it | Surface finish still looks real |
| Label hidden, feature missing | Nothing; get another photo | Don’t let AI guess product detail |
Keeping the whole catalogue in step
Put each edited image next to the ones around it, rather than only checking it against its own original. A short written standard per category settles a lot of debates later. It should cover background tone and margins, colour temperature and contrast, camera angle and apparent size, and the boring export details: ratios, dimensions, compression, file names. Skip that last part and someone’s quick re-export will replace your approved masters within a month or so.
Also check your plan. Pollo gives commercial-use rights on paid tiers, so confirm yours before edited images go live on a marketplace.
When to ask for a new photo instead
If the photo can’t show what the buyer is actually getting, editing won’t fix it. That includes blurred or partly hidden products, wrong colours, missing parts and old packaging. It also includes shots where you can’t judge texture, finish or size with any confidence.
AI will polish an honest photo nicely. Ask it to finish a product it can’t see and it’ll do that just as smoothly, which is the problem.
Want a low-risk first try? Take one accurate supplier photo with a single obvious flaw, write a prompt that names what must not change, and compare the result line by line with your checklist. Note anything that changed without being asked.
Related: Scenario AI Review (2026): Pricing, Features, and Real Credit Costs
