Product Image Moderation for E-commerce

Why marketplaces need automated image screening

On any marketplace where sellers upload their own product photos, a fraction of those images will be a problem: explicit content disguised as a legitimate listing, weapons and regulated goods, graphic violence, or counterfeit products passed off with a copied brand logo. The volume makes manual pre-screening impossible — a trust-and-safety team cannot review every photo on a platform that ingests tens of thousands of new listings a day, and buyers see the bad ones long before a human does.

Automated screening flips the default. Instead of publishing first and reacting to reports, every image is scored by an API as it is uploaded. Pixicular's image analysis API gives you two services that matter here: content moderation for unsafe imagery and label detection for what is actually in the photo. The safe majority of listings go live with no human in the loop; reviewers spend their time only on the ambiguous and the dangerous.

How does the moderation pipeline work?

A typical product image moderation pipeline has four stages: a seller uploads a listing photo, the photo is sent for automated screening, the API returns label and moderation flags with confidence scores, and your service makes an approve / queue / reject decision against thresholds your policy owns.

The decision step is yours, not the API's. The API returns scored signals; your trust-and-safety policy decides what confidence on which category means auto-reject, what means queue-for-review, and what passes. That separation keeps moderation behaviour under your control and auditable, which matters for the regulatory obligations covered below.

What the content moderation and label services return

The detect-moderation service returns a confidence score per unsafe category. The detect-labels service returns the objects, brands and scene concepts present in the image. Together they cover the categories an e-commerce trust-and-safety team screens for.

Note: Pixicular's AI-generated image detection service is currently disabled while accuracy is being improved. The content-moderation and label-detection services above are live and are what e-commerce platforms use for product image screening today.

Code: the moderation API call

One request asks for both services. The image goes up once; the response is a single JSON document with one key per requested service.

curl -X POST https://api.pixicular.com/detect \
  -H "Authorization: Bearer $PIXICULAR_API_KEY" \
  -F "image=@./listing-photo.jpg" \
  -F "services=detect-moderation,detect-labels"

// Screen a seller-uploaded product image before the listing goes live.
async function screenListingImage(file: Blob) {
  const body = new FormData();
  body.append("image", file);
  body.append("services", "detect-moderation,detect-labels");

  const res = await fetch("https://api.pixicular.com/detect", {
    method: "POST",
    headers: { Authorization: `Bearer ${process.env.PIXICULAR_API_KEY}` },
    body,
  });
  const result = await res.json();

  const moderation = result["detect-moderation"]?.flags ?? [];
  const labels = result["detect-labels"]?.labels ?? [];

  // Per-category thresholds owned by your trust & safety policy.
  const blocked = moderation.some(
    (f: { name: string; confidence: number }) =>
      (f.name === "Explicit Nudity" && f.confidence >= 0.9) ||
      (f.name === "Violence" && f.confidence >= 0.85),
  );
  const counterfeitSignal = labels.some(
    (l: { name: string; confidence: number }) =>
      l.name === "Designer Logo" && l.confidence >= 0.7,
  );

  if (blocked) return "reject";
  if (counterfeitSignal) return "queue_for_review";
  return "approve";
}

Regulatory pressure: DSA Article 16 notice-and-action

The EU Digital Services Act puts marketplace image moderation firmly inside a compliance perimeter. Article 16 requires hosting providers and online platforms, including e-commerce marketplaces, to operate a notice-and-action mechanism: any individual or entity must be able to flag content they consider illegal, and the platform must process those notices in a timely, diligent, non-arbitrary and objective way, then notify the reporter of its decision and the available redress.

Automated image screening does not satisfy Article 16 on its own — the notice-and-action mechanism is a separate, mandatory channel. But it materially changes the obligation in three ways. First, proactive detection reduces the volume of illegal imagery that ever reaches a buyer, shrinking the notice queue. Second, every screened listing carries stored moderation and label scores, which is the evidence trail your team needs to action a notice quickly and produce the statement of reasons Article 17 requires. Third, consistent automated thresholds make decisions non-arbitrary and reproducible, which is exactly the standard regulators expect when they audit moderation behaviour. Pair the API with a human review queue for borderline cases and you have a defensible pipeline rather than a backlog. This is general information, not legal advice — confirm your obligations with counsel.

When to auto-block, queue, or approve

• Auto-reject on high-confidence unsafe categories — explicit nudity, violence and weapons above a strict threshold should never reach a buyer.
• Queue for review on borderline confidence or counterfeit-indicator labels — a human reviewer decides, with the scored evidence attached.
• Approve when no category exceeds threshold and labels match the expected product type — the listing publishes with no human in the loop.
• This same model extends to UGC image moderation for social platforms and other user-upload surfaces — see what image content moderation is for the broader picture.

Frequently asked questions