Solution · Content moderation

Catch repeated harmful media before it spreads

MediaLayer matches incoming uploads against known-bad media and recently-actioned content so moderation queues see clusters, not duplicates. Image, video, and audio similarity behind one JSON envelope — built to survive the obfuscation users actually try.

Talk to MediaLayer AI Labs View public API on RapidAPI

The problem

Where this hurts in production

Re-uploaded harmful media
Once content is removed, near-duplicate copies reappear with re-encoding, cropping, audio swaps, or watermarks. Pixel hashes miss them; perceptual matching does not.
Duplicate uploads under different accounts
The same video is uploaded across hundreds of accounts within hours. Without similarity grouping, every copy gets reviewed independently and decisions drift across reviewers.
Known-bad lists go stale fast
A static hash list catches the original asset and almost nothing else. Adversaries iterate; the matching layer needs to too.
Cross-modal evasion
Bad audio gets re-uploaded inside a different video; a still frame reappears as a meme image. Matching has to span media types — not assume image is image and video is video.

How MediaLayer fits

Same APIs. Same JSON envelope. Targeted at this workflow.

MediaLayer wraps image, video, and audio similarity in one JSON request shape. POST two URLs to /image/match, /video/match, or /audio/match and the response carries a similarity score, a confidence label, and (for video and audio) aligned matched segments showing exactly where two assets overlap.

Matching is built for the things adversaries actually do: re-encoding, resizing, watermarking, mirroring, audio transcoding, pitch-shifting, and re-uploading clips trimmed from longer originals. Reviewers see clusters of near-duplicates instead of having to action 40 separate copies of the same thing.

For platforms with very large reference sets — known-bad libraries, prior-removed content, NCMEC / GIFCT-aligned hash matching, or industry-shared databases — Enterprise media search ingests the catalog into a similarity index and runs one-to-many lookups on every new upload. Talk to MediaLayer AI Labs for direct API access, private deployment, and bulk ingestion pipelines.

Operationally, T&S workflows call the API server-to-server with their own x-rapidapi-key (or, on direct-API enterprise deployments, with private credentials inside a VPC). URL validation rejects private, loopback, and cloud-metadata addresses, which is the right default when uploads include internal preview URLs that should never be fetched. Per-request size and duration caps prevent runaway processing on adversarial uploads, and the uniform JSON envelope across image, video, and audio keeps moderator-action audit logs consistent regardless of which media type triggered the match.

Workflow example

From media in to match decision out

1
Capture the upload
Trust & safety pipeline picks up the new image, video, or audio URL from the upload queue.
2
Match against reference set
POST source_url + target_url to the matching endpoint for that media type.
3
Score the response
similarity_score + confidence drive action: auto-remove, route to review, or pass.
4
Group before review
Cluster near-duplicates so reviewers action one decision against a group, not against each copy.
5
Promote at scale
Move from pairwise calls to one-to-many Enterprise media search as the reference catalog grows.

RESPONSE · RE-UPLOAD MATCH

{
  "match": true,
  "confidence": "high",
  "similarity_score": 0.96,
  "processing_time_ms": 1280,
  "media_type": "video",
  "matched_segments": [
    { "source_start": 0.0, "source_end": 22.4, "target_start": 4.6, "target_end": 26.9, "score": 0.97 }
  ]
}

Relevant API endpoints

Drop these into your pipeline

POST /image/match

Detect re-uploaded harmful images even after re-encoding, cropping, and watermarking.

Learn more →

POST /video/match

Compare videos and surface aligned matched segments — built for partial reuse and frame-level overlap.

Learn more →

POST /audio/match

Match audio across re-encoded uploads so audio-only evasion (audio swap, dub) doesn't slip through.

Learn more →

Real-world examples

Patterns we see in this space

Removed-content recurrence
Hash recently-removed media and match every new upload against it. Catch re-uploads in minutes instead of waiting for a second user report.
Coordinated-upload bursts
When the same media gets uploaded by hundreds of accounts in one window, similarity grouping turns the burst into a single cluster decision.
Cross-platform reference matching
Match incoming uploads against industry-shared reference hashes (CSAM, terror content, etc.) using one-to-many Enterprise search.