A store adds a product. Someone uploads one good photo. Then that single image has to become a square for the catalog grid, a white-background main for Amazon, a vertical for TikTok Shop, a lifestyle frame for Etsy, and a specific ratio for the Google Shopping feed. Each one cropped so the product still looks right.
We make AI product images for brands every week, so I have watched this tax up close. Every marketplace wants different image specs, and a single product across Amazon, Shopify, Etsy, and TikTok Shop needs somewhere around 8 to 12 distinct images per channel. Multiply that across a catalog and you get a quiet, recurring job that never makes the roadmap because it is always almost handled.
Why the manual version stops scaling
The old answer is a studio reshoot. A single SKU runs $150 to $500 once you count lighting and post. A 100-SKU catalog selling across three marketplaces needs roughly 600 to 1,500 compliant images, which is $90,000 to $750,000 in studio time. That math breaks at any real catalog size.
The new answer is to fan one source image out to every format in software. The cost lands closer to a few dollars per finished listing instead of a few hundred. The blocker is not the idea, it is doing the resize without wrecking the photo.
Why naive resizing makes it worse
Resizing one source to each required size is where product photos go to die. Stretch a square shot into a wide banner and the product distorts. Hard-crop a tall frame from a centered photo and you slice the top off the bottle. Pad it with whitespace and the product shrinks into a stamp.
The reason is that a dumb resize treats every pixel as equal. It has no idea where the product is, so it cannot protect it. What you want is the opposite: keep the subject correct and in frame, and let the framing adapt around it. Content-aware resizing finds the subject first, then fits it to each target ratio without distorting it, so one source image becomes every placement and none of them look squished or cropped wrong.
The part that bites at scale
There is a failure that only shows up across a big catalog, and I have learned to respect it. AI image output is usually about 90 percent correct, and 90 percent reads as done when you are moving fast. A resized shot comes back looking like a finished listing, and nobody zooms in. Then you notice the crop clipped the product or the framing is off.
At one image, you catch it. Across three hundred products, you do not, and a bad crop ends up live on a page where the image is doing most of the selling. A shopper reads a distorted or clipped product photo as carelessness, at the exact moment they are deciding whether to trust you with a card. The fix is not only better resizing, it is a check that scores every output and holds back the ones that are wrong. The pipeline behind our customer BetterPic ranks every generated image and only lets the good ones reach the customer, across more than 35 million images over the past two and a half years. That scoring layer is what makes automation safe to run unattended.
Where to automate it
If your volume is low, a careful person and a good editor are fine. Once the catalog grows or the channels multiply, you want this in code. An image resizing API can take one source image and return every aspect ratio your channels need with the subject preserved, as a single step in your product-ingest flow. You upload the photo once, and the placements generate themselves.
The build-versus-buy line here is the usual one. Writing subject-aware cropping in-house is a real computer-vision project. Calling an endpoint that already does it is an afternoon. For most stores the image problem is worth solving but not worth building.
One product shot should not become a manual rendering job across five channels. Stop resizing blindly, start resizing around the subject, and put a check at the end so nothing distorted ships. Whether you do that by hand or by API is a question of volume. The standard, that no channel gets a broken product image, should not move.
