How to automate the retouching of 30,000 photos

Stock that does not exist for customers

The request did not arrive as "automate our retouching". It arrived as "part of our storage does not exist for our customers".

That nuance decides everything, because it sets the measure of success. It was not an aesthetic problem: it was invisible stock, and therefore revenue that never happened.

So the goal was never the most beautiful possible image, but a complete catalogue, then one moving at the pace of acquisitions. A photo taken during the day had to be live the next morning.

Twelve days were enough, end to end. That covers building the pipeline and processing the full 30,000 existing references, not just switching it on. It now absorbs around 12,000 images a year, produced with no studio, between two shelves of the storeroom.

The file that lies about its format

The first failure at scale had nothing to do with cutout, or with artificial intelligence. Part of the catalogue consisted of phone photos whose original format had simply been renamed during transfer.

To a human, those files open normally. To the pipeline reading them, they are unreadable: the extension says one thing, the content says another. Processing stopped dead, with a message that said nothing useful about the real cause.

The fix took one line of configuration. Finding it took several days, because the error presented itself as a damaged image rather than a format problem.

That is the project's main lesson, and it is not technical: before writing a processing pipeline, inventory what actually goes into it. Not what you believe goes into it. The photos in a real catalogue never resemble the test set prepared for the demo.

What fails never fails once

That format problem occurred hundreds of times. In the pipeline's logs it produced hundreds of error lines identical but for one detail, each carrying a different filename.

As long as you read raw logs, you see a river and conclude nothing. You fix the last error you saw, not the most frequent one, and start again the next day.

So we grouped occurrences by fingerprint: the message is normalised, identifiers and filenames replaced by markers, and everything sharing the same signature becomes one problem with a counter.

That change of viewpoint is the most profitable of the whole project, and it produces no images at all. A list of ten problems ranked by frequency is handled in order, in a morning. A river of thousands of lines is not handled at all.

What the machine does well, and what stays with the eye

On an opaque, properly lit object, automatic cutout produces a result a trained eye struggles to tell from a careful manual cutout.

But the real gain is not the quality of a single image: it is consistency. A catalogue processed by hand varies by person and by hour. A catalogue processed by a pipeline has the same margin, the same background and the same scale across its 30,000 records, and that uniformity shows immediately on a results page.

What stays with the human is judgement: deciding that a shade should be faithful rather than flattering, that a piece deserves an extra view, that a particular object would be better served by a reshoot. Those are catalogue decisions, not production gestures.

In the catalogues we handle, the share of images escalated to a human stays below 1%. It is nonetheless the ability to flag them that gives the whole thing its value: an automation that processes everything and never escalates forces you to check everything, and has therefore automated nothing.

The materials that need preparation

Four families of subjects are handled differently, and it is better to identify them during framing than in production: transparency, reflection, open-work, and monochrome on a close background.

What they have in common is instructive. In all four cases the answer lies at the shoot, not in the processing. A background chosen for the material solves upstream what would be expensive to correct afterwards.

A translucent bottle raises a question that is editorial rather than technical: should the record's white background show through the glass, should a tint be kept so the material stays readable, or should a neutral backdrop be reconstructed? None of those answers is superior. The choice belongs to the catalogue manager.

That is why a pipeline always begins by building a quality reference: a few hundred already-approved records that embody the company's intentions. Without them, "this is better" remains an opinion and quality discussions never conclude.

The volume at which it earns its place

The threshold is a one-line calculation: the preparation time for one image multiplied by your annual flow.

From a few thousand images a year, the question no longer arises. Below a thousand, a retouching provider is often still the right answer, and saying so is part of the framing work.

On the calendar, count one to two weeks for a catalogue of ordinary size, framing, build and initial processing included.

The staged rollout is worth defending to those who find it slow. Process one storage area, compare against historical records, correct, then move on: that is what lets you settle colour handling with the catalogue manager rather than discovering it live.

Building starts between 3,000 and 5,000 euros for a catalogue of ordinary size, then follows the number of formats, marketplaces and special cases. On top comes a usage cost, billed by volume processed, counted in cents per image. That second line decides profitability at three years, and it is almost always the one left out of the quote.

The only indicator that counts

The result is not measured by the number of images processed. It is measured by the delay between the shot and going live, because that is what the end customer feels and what turns into revenue.

The other effects follow from that one. The catalogue is uniform, so results pages read better. Records are complete, so clarification requests drop. New acquisitions are online immediately, so they get rented.

If you are considering a project like this, start with the cases to exclude rather than the ideal case. Gather a hundred photos representative of your catalogue, deliberately including the most awkward ones, the ones your retoucher saves for last. Those are what will tell you what is genuinely achievable.