Use case · marketplaces & e-commerce
Product image compliance, automated honestly
Product image compliance is the automated checking of seller or catalog photos against listing rules — clean background, no watermarks or overlaid text, product fully in frame, adequate quality — before a listing goes live. The stack that usually fits: deterministic checks for the objective rules, a prompted vision-language model (VLM) returning a per-rule verdict and reason, and confidence thresholds that send uncertain photos to your existing review queue instead of guessing.
Who this is for
The team that owns the listing bar
Marketplaces and e-commerce products where sellers upload their own photos: the product team owns conversion and catalog quality, trust & safety owns the rules, and today enforcement is manual review, spot checks, or nothing. The volume is real — hundreds to hundreds of thousands of photos a day — and the pain is symmetric: bad photos hurt buyers, wrong rejections hurt sellers.
The decision maker is usually a product engineer or founder without a vision team, choosing between a moderation API that does not know their rules, a model provider’s demo, and an agency quote. What they actually need first is an architecture shaped by their rule set — and then evidence on their own photos.
The decision table
Match the approach to the rule, not the hype
Compliance is not one recognition problem — it is a bundle of rules with different shapes. The most common mistake is forcing every rule through one model.
| Rule type | Best-fit approach | Why |
|---|---|---|
| Resolution, aspect ratio, file properties | Deterministic checks — no model | Exact, free, and instant. Never spend model budget on rules code can decide. |
| Clean background, product fully in frame, framing quality | Prompted vision-language model (VLM) with a per-rule output schema | Compositional judgments expressed in language; rules evolve faster than retrained models. |
| Watermarks, overlaid text, logos | VLM first; dedicated detector if violations are frequent and adversarial | Casual violations are caught by prompting; determined evasion eventually justifies a specialist. |
| Prohibited content categories | Pretrained safety / moderation API, tuned thresholds | Mature providers cover common safety classes well; domain-specific bans still need the VLM rules. |
General fit conditions, not measurements — which approach clears your bar is exactly what a future proof run would test.
What to measure
Metrics that decide this use case
| Metric | Why it matters here |
|---|---|
| Per-rule precision | Of photos flagged for a rule, how many truly violate it — false flags erode seller trust fastest. |
| Per-rule recall | Of true violations, how many are caught — misses define the policy risk you still carry. |
| False-rejection cost | Every wrong rejection is a delayed listing and a support ticket; weight it explicitly. |
| Abstention rate | Share of photos routed to human review — the dial that trades queue load against error rates. |
| Latency at upload | The check sits in the seller’s flow; tail latency, not average, decides the experience. |
| Cost per thousand photos | Computed under your real volume assumptions, including retries and review overhead. |
The non-negotiable: abstention
Start small
The first experiment
Have your own moderators label 200–300 recent photos per rule — including hard cases: busy backgrounds that pass, subtle watermarks, borderline framing, and photos violating several rules at once. Run the VLM stack on that set, compare its per-rule verdicts to the labels, and read the failures by slice. That one afternoon of labeling converts the decision from opinion to evidence — and the labeled set becomes your regression suite.
A worked (fictional) version of this exact scenario is the example free report — same product shape, written out in full.
In short
For product image compliance, split the rules: deterministic checks for objective properties, a prompted vision-language model with per-rule structured verdicts for judgment rules, and confidence thresholds that route uncertain photos to human review — then prove the design on a few hundred labeled photos before trusting it.
Related: image classification & routing · how to test vision models · glossary