Customer Case Study · AI Photography QC
The retailer's studios shoot thousands of product images for e-commerce, and every set must follow strict style guides on sequence, framing, cropping and background. HMC built a dual-tier vision AI pipeline that turns the style guides into JSON rulebooks, filters junk images cheaply, and judges every image set with a clear, explainable QC report.
Figures indicative; validated metrics available on request.
Studio shoot, manual review against the style guides, then the bottleneck: slow, subjective and error-prone checks before new SKUs could go live.
Every season, every category, from several studios and agencies.
Sequence, shot types, framing, creases and backgrounds differ by category and season.
QA teams compared each image to the guide by eye, set by set.
Different reviewers, different answers; reshoots found after the set was struck.
A pipeline that spends expensive vision AI only where it adds value.
Select a step to see what happened.
Sat with QA reviewers and photographers.
Deliverable: Rejection taxonomy and pipeline design.
A machine-readable rulebook per category.
Deliverable: Rulebook schema and rule registry.
End to end in one app.
Deliverable: Working QC pipeline.
Feedback before the set is struck.
Deliverable: Pilot results and studio SOP.
Global alignment across studios.
Deliverable: Roll-out plan.
Turn a style guide into rules, push a batch through the gatekeepers, and judge image sets. Rearrange or break the images and watch the verdict change. Product images here are drawn illustrations; with an OpenAI key you can judge your own photos.
Optional. Without a key, rulebooks come from the registry and the judge applies the rules to the sample sets. Add your OpenAI key to extract a rulebook from a style-guide screenshot and to judge your own product photos.
Each category's visual style guide is read and turned into a strict, machine-readable rulebook: the required image sequence and the quality standards. Pick a category, then extract.
A batch of 50 SKUs arrives from the studio and an agency. The router matches each SKU to its rulebook; cheap gatekeepers check that each image decodes, meets resolution and file size, and actually shows the right garment. Only clean image sets go to the Tier 2 judge.
Run the batch to see the logs.
Synthetic batch; garment images are illustrations. Thresholds: 2000 × 2667 px minimum, 3 MB maximum, apparel classifier confidence 0.80.
Pick a submitted set. Drag images to reorder, or use the menu under each image to break it (crop an edge, add creases, swap in a model shot). The judge re-checks sequence, shot type, cropping, creases, negative space and background, with evidence and crop coordinates.
Rulebooks marked "AI draft" are not reviewed by the retailer; for demonstration only.
The batch rolls up into one report: pass rate, the rules that fail most, and a reshoot list by studio. QA managers review it and make the final call; photographers get a plain note on what to fix.
| SKU | Category | Studio | Failed rules |
|---|
QA hours per thousand images fell by about 90%, and the studio clears far more SKUs a day for the site.
Donut: where the review time goes
Spiral heatmap: one ring per week, one cell per day, week 1 at the centre
Figures indicative; validated metrics available on request.
A JSON rulebook is something every studio, agency and model can follow the same way.
Basic checks and small classifiers stop junk for a fraction of the cost of a vision model.
A score, an observation and the exact image and crop make the AI's call easy to trust and to act on.