Image moderation
Check a user-uploaded photo before it goes live — unsafe content, spam and whether it matches its caption — and block or escalate it, with the Image Moderation template.
Goal: people upload photos with their listings. Dcision looks at each photo before it goes live: unsafe images and spam are blocked, unrelated ones and photos that don't match their caption go to a moderator.
Create the decision
Templates → Image moderation → Create decision:
| Part | Content |
|---|---|
| State | photo (image, required), caption (string) |
unsafe | probability — nudity, violence, weapons, drugs or hateful symbols; minimum confidence 0.6 |
content | choice — product, spam, unrelated, other |
matches_caption | probability — the photo shows what the caption describes |
| Policies | unsafe ≥ 0.5 → block; content = spam → block; content = unrelated → escalate; matches_caption < 0.3 → escalate; fallback action escalate |
photo is an image field: PNG, JPEG, GIF or WebP. Run this decision on OpenAI Decisions for the engine to see the image itself; on Jev or Laya it reads a description written by OpenAI.
Call it
curl -X POST https://api.dcision.io/v1/decisions/image-moderation \
-H "Authorization: Bearer $DCISION_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"state": {
"photo": {"url":"https://example.com/listing-photo.jpg","filename":"listing-photo.jpg"},
"caption": "Used road bike, size M"
}
}'{
"schema": "image-moderation",
"result": { "unsafe": 0.02, "content": "product", "matches_caption": 0.93 },
"action": "continue",
"action_reason": { "type": "default" },
"metrics": { "engine": "openai", "input_files": 1 }
}Act on it
const decision = await dcisionDecide("image-moderation", { photo: fileFromBytes(upload, upload.name), caption });
if (decision.action === "continue") return publish(listing);
if (decision.action === "block") return reject(listing, decision.action_reason);
return queueForModeration(listing, decision);Tune it
- Describe your platform in
context: what a normal photo looks like there. - Add options to
contentfor what you don't allow (watermarks, photos of screens). - Raise the minimum confidence of
unsafefor public listings; unsure cases go to a moderator.