Supplier photography, made to look like your shop.

Every product arrives with somebody else's photographs. The pipeline takes them apart — duplicates, background, composition — and rebuilds them in one house style that holds.

An image is the one part of a listing a buyer looks at before reading anything. It is also the part a supplier gets wrong most consistently, because their photograph was made for their page, not yours. The pipeline closes that gap mechanically — and stops before it does damage, which is the harder half of the problem.

Five stages, in order.

01. Deduplicate — The same photo arriving twice takes only one slot and one processing run, however often the supplier includes it.
02. Classify — A small model sorts the image into a product category. Optional, and only used where it changes the outcome.
03. Cut out — The background comes off — but only if the guard below decides it is safe to do so. A bad cutout is never published quietly.
04. Compose — The result is placed onto your template: canvas size, layers, a watermark, one to three badge lines, rotation where you want it.
05. Serve — WebP, at full size and four thumbnail widths, through links that carry the file's own fingerprint.

The guard, and why it refuses.

Background removal fails in a particular way: it takes the product with it. A dark object on a dark background, a soft edge, a subject that blends into the set — the algorithm returns something confidently wrong. A listing with a half-eaten product photo is worse than one with the supplier's original.

So the pipeline measures before it writes. If the cutout would remove more than half of the product pixels, or leave under three percent of the frame, the guard refuses the step and the original is kept — and you see that it was skipped, and why. Larger inputs are resized rather than refused.

It is a small rule, and it is the one that makes the rest trustworthy: a step that knows when it is out of its depth can be left running.

What a template can hold.

Canvas — Between 64 and 4096 pixels. One size per template, so a catalogue rendered from one template comes out uniform.
Layers — Up to twelve per template, in three kinds: an image, a watermark and a badge — which is what carries the style.
Anchors — Nine anchor positions, so a layer sits in the same place regardless of the source image's shape.
Badges — One to three lines of text, rotated where a corner ribbon reads better than a flat one.
Watermarks — Applied as a layer, and separately embeddable as an ownership mark that can be verified.

Questions, answered

Does the background removal run on my own computer?

No. It runs on our servers, on a resized copy capped at fifteen hundred pixels on the long side. Your originals are never modified — the processed result is a new file.

What happens to an image the guard refuses to cut out?

It is kept as it arrived and the step is reported as skipped. The listing still gets composed from the original; what it does not get is a broken cutout pretending to be finished work.

Where are the finished images stored?

Either on the service's own storage or on an image CDN — a deployment setting. Either way the public URLs carry the file's hash, which both defeats caching of an old version and stops the URLs from being guessed.

Can different product categories get different styles?

Yes. Several templates can be kept and chosen per run, so a category that needs a badge can have one without forcing the badge onto everything else.

Go deeper

Product preparation — The whole pass: images, data and the listing draft.
AI listing drafts — What the draft says about an image: alt text, title, and where it goes.
Trust and security — Ownership marks, and how an image can be proven to be yours.

Back to product-preparation

Take one supplier image through all five stages.

It takes a minute and it shows exactly what the pipeline does and where it stops.

Get Started