One photo, ten SKUs: the unglamorous work that gets a product catalogue online

Most conversation about AI creative production is about the glamorous end: hero films, campaign concepts, localisation across twelve markets. Then a real brief lands, and it looks like this: a client who sells industrial fasteners online is refreshing their product photography. There will be roughly 500 raw camera files. Out of those, around 3,000 finished images need to exist on the website — sized, labelled, watermarked, optimised, named and tagged. Turnaround: as soon as possible, because the site launch is waiting on the images.

Nobody puts this work in a portfolio. It decides whether the site ships on time.

The anatomy of a bulk product-image job

Briefs like this are more interesting than they look, because the multiplication is doing all the work:

• One photo, many SKUs. A hex bolt comes in ten sizes. Nobody shoots ten near-identical bolts — you shoot one, and produce ten images with the size rendered on the image: M8, M10, M12. One photo becomes ten catalogue entries. This is where 500 shots become 3,000 images.
• Brand furniture. Every image carries the company logo as a watermark — same position, same opacity, every time.
• Platform spec. Exact pixel dimensions, exported as both JPEG and WebP, under a file-size budget (say, 100KB) without visible quality loss — because page speed is a conversion and ranking factor, and product pages live or die by it.
• Findability. Filenames follow a naming convention from a spreadsheet. Alt text goes into the metadata, also from a spreadsheet — which is simultaneously an accessibility requirement and free search visibility most catalogues never claim.
• RAW input. The photographer hands over CR3s, not JPEGs. Everything above sits on top of consistent RAW processing — exposure, white balance, colour — so image 2,847 matches image 12.

None of these steps is difficult. That's the trap.

Why it goes wrong when humans do it by hand

Three thousand images times seven operations is over twenty thousand operations, every one of them trivial and none of them allowed to be wrong. Manual production fails in predictable ways: the watermark drifts a few pixels between operator one and operator three; a batch gets exported at the old dimensions after the spec changes; filenames diverge from the spreadsheet somewhere in row 900 and surface as a mess at upload time; alt text gets quietly skipped when the deadline tightens, and the SEO benefit evaporates before anyone notices it was gone.

And then the spec changes — it always changes. "Actually, can we do 1200px, and move the size label to the top-left?" With hand production, that sentence costs weeks. It should cost one overnight re-run.

Treat it as a workflow, not a task list

The fix is to stop thinking of this as 3,000 editing tasks and set it up as one workflow that runs 3,000 times:

1. The spreadsheet is the brief. SKUs, size variants, filenames, alt text — the client's own spreadsheet becomes the machine-readable spec the workflow executes against. If a row is wrong, you fix the row and re-run, not the image.
2. RAW processing as a profile, applied uniformly, so consistency is a property of the system rather than of whoever was working that day.
3. Variant expansion from the size matrix — one master image per product, every size variant generated with the label typeset to brand rules, not stamped on by hand.
4. Watermark, resize, export as deterministic stages: both formats, compression tuned to the budget, dimensions from the spec.
5. Naming and metadata written straight from the spreadsheet — alt text embedded, filenames exact.
6. Human QA on samples, not on everything. Reviewers check a statistical sample per batch and anything the system flags as anomalous; people supply judgement, the workflow supplies repetition.

This is the same shape as generating fifty ad variants from one brief: inputs × a variant matrix × brand rules, executed by a system, checked by people. The subject matter is less glamorous; the economics are identical. AI earns its keep in the judgement-adjacent corners — consistent RAW-to-clean processing, background standardisation, smart cropping, drafting alt text when the spreadsheet doesn't exist yet — and disciplined automation does the rest. What the client experiences is simpler: the spec goes in, the finished catalogue comes out, and a re-run costs almost nothing.

The boring stuff compounds

A catalogue produced this way is faster (WebP under budget on every product page), more findable (complete, consistent alt text and naming), and more trustworthy (every image visibly from the same family). And when the product line grows or the photography refreshes next year, the workflow is sitting there, ready to run again — which is the real difference between buying an editing job and building a production capability.

If you're writing a brief like this, the checklist above is the brief: input format, variant logic, labelling spec, watermark, dimensions, output formats and file-size budget, naming convention, alt text source. Get those eight things on one page and you're most of the way there. Running it at volume, on deadline, without drift — that's the part we do every day.