Automate product image editing for marketplaces

You shoot a new collection, and then the real work starts. Someone sits down and lifts each product onto white by hand, crops it square for one channel and portrait for the next, compresses, renames. Thirty articles is an afternoon. Three hundred is a week nobody planned for, and by the end the pictures are the reason an article still is not online. There is a way to automate product image editing, and it is not a better button in your photo editor.

Why a product photo is never just one photo

Every sales channel writes its own rules. Amazon asks for a pure white background on the main image and a minimum share of the frame that the product has to fill. Other marketplaces prescribe aspect ratios and minimum edge lengths. Your own shop prescribes nothing at all, but it has a loading budget, and a 4 MB file straight out of the camera eats the whole thing on its own.

So if you sell on four channels, you edit the same photo four times. The usual tools each solve one slice of that: one cuts out the background, another compresses, a third resizes. The product description still gets typed by a person at the end.

Automate product image editing as one chain, not five tools

The way out is not a better single function. It is one run that goes all the way through: upload, cut out, resize, compress, describe. A whole batch at once, with the settings for each channel stored once and picked from a list afterwards.

That is what we built PicClean for, our own platform for product photography. Five AI models work together inside it: one separates product from background, one enlarges small shots without visible loss, one describes what can be seen, one turns that into copy for the article page. Amazon, eBay, Etsy, Shopify, Zalando and Otto come with their requirements already set up.

The part that mattered to us: it runs on hardware you control. No billing per image, no queue at somebody else's service, and the shots of a collection that has not launched yet stay in the building.

• Define the preset per channel once, then only pick it
• Upload a batch instead of opening one file after another
• Check the results where it counts, not everywhere

What that adds up to across a full range

Do the sum on your own catalogue. A retailer with 300 new articles a year, four photos per article and four channels ends up at roughly 4,800 image variants. At two minutes of handwork each, that is about 160 hours. Four working weeks, spread across the year, usually in the evenings before a launch.

The expensive part is not the time, though. It is the delay. An article whose pictures are missing does not sell. In seasonal ranges, garden in March, heating in autumn, one lost month decides half a season.

And then there is the quiet damage: things stop matching. When three people cut out products over two years, one sits the item in the middle, another near the edge, one uses pure white, the next a hint of grey. On the category page you see it immediately, even if nobody can say what exactly is off. A preset that applies to everybody fixes that on the side.

Where automation starts and where it stops

This pays off when a lot of images arrive in a short time, when the channels want different things, or when the shots are confidential until launch day. It does not pay off for a handful of articles a quarter. Then an afternoon of handwork is simply cheaper than any setup.

The honest limit runs along the ambition of the single image. A hero shot that a photographer staged stays handwork. The other 4,799 variants are assembly line work, and that is exactly the part a machine should be doing.

Common questions

Does this replace the photographer?

No, it replaces the rework. Light, setup and the idea behind the picture stay craft. The cutting out, cropping and compressing that follow are dull enough for a machine, and dull work is where machines are good.

What about difficult subjects like glass, fur or shiny metal?

No model gets those right first time, every time. That is why a visual check is part of the run: the batch goes through, a person fixes the outliers. In a typical product range that is a minority of the images, not the rule.

Does it have to run on our own hardware?

It does not. That is the route we recommend when the volume is high or the shots are confidential before launch. Where a cloud service is the better fit, we say so just as plainly. The question is which one is faster and cheaper in your case.

How we approach it

PicClean came out of a customer problem, not a product idea: too many articles, too little time for pictures. The machinery behind it, models, queue, scaling under load, is the same one we use to build shops and the systems behind them, and the same thinking we apply when we run AI inside a company.

If images are your bottleneck, tell us how many articles arrive per month and how many channels they run on. That is enough to work out in one conversation whether automating it pays for itself. Talk to us.