The challenge
A mid-size jewelry retailer with online and offline sales works with many suppliers and keeps its catalogue in an ERP system. Supplier photos arrive in inconsistent sizes, resolutions, orientations, and crops, so the online catalogue looks chaotic on mobile, where around 70% of shoppers browse.
The requirements made this harder than a batch resize:
- Each product category has its own display standard: rings in a set orientation, earrings shown face and profile in one frame, pendants framed wider to show the chain
- The existing catalogue held 7,000 to 8,000 photos that needed a one-time pass to a consistent look
- New items arrive at around 100 per month and must be processed without holding up publishing
- The brand’s bright, festive visual style had to survive the processing instead of being flattened into a generic look
The approach
The service slots into the current publishing process and keeps originals untouched: source images live in one folder, processed images land in another, and the filename (the SKU) is the join key back to the ERP.
Category metadata comes from the ERP, where a merchandiser has already classified each item, so the AI never guesses the category from the image. A pre-filter passes already-acceptable photos straight through, and the AI budget goes only to images that genuinely need rework.
Delivery is staged. First a one-time batch pass over the existing catalogue, with before-and-after exports so the client can validate each category. The automated pipeline for new arrivals starts only after the client accepts that pass.
Architecture
- Orchestrator: a Python backend that runs the full pass: download, category sorting, AI processing, AI quality control, upload
- AI: image editing and generation models suited to reorientation and recomposition, benchmarked per category during the batch pass
- Pre-filter: a vision model plus rule-based checks that skip the AI call when an image already meets the category criteria, keeping the cost down
- Configuration: per-category rules (target orientation, scale, crop, prompt) stored as editable config, so tuning needs no redeploy
- Transport: SFTP exchange with the client’s server, originals in a source folder and processed images in a target folder keyed by SKU
Results
Shoppers see one consistent catalogue on mobile instead of a patchwork of supplier photos. The pre-filter keeps the cost predictable: photos that already meet the standard never touch the AI, and per-category configs let the retailer tighten a standard without a redeploy.