A vision model as a catalogue QA pass
A small script that checks the product images on your best sellers for blur, low resolution and wrong-language overlay text, and hands you a report instead of a fix.
A large electronics catalogue accumulates bad product images: a blurry supplier photo, a 400px thumbnail that got promoted to the main image, a picture with English marketing text baked in on a Ukrainian store. Nobody has time to audit thousands of products, so it never gets done.
Scope it to what sells
The script pulls the products sold in the last seven days, sorted by quantity, and only checks those. The long tail of the catalogue isn’t worth a vision API call; the fifty products actually getting views this week are.
For each one it sends up to five images to a vision model and asks two things: is this blurry or low-resolution, and does it contain overlay text in a language other than Ukrainian. It flags anything scoring poorly or under 800px on the short side.
Output a report, not an edit
The script writes an HTML page — one row per flagged product, with buttons: open it in the admin, run a reverse image search, generate a prompt for an AI replacement image, mark it done. It does not touch the store.
That’s deliberate. Image replacement is a judgement call — is this one bad enough to spend time on, is there a better source, is the overlay text actually a problem here — and a script that auto-swaps images would get some of those wrong silently. A human clears the list in an hour with the context to make the call.
Cache the analysis
Vision calls aren’t free and the best-seller list barely changes day to day. Results are cached by product, so a re-run only pays for what’s new.