Skip to content

AI watermark API for images, video and audio

Mark every file your model generates before it leaves your platform. Original Pictures adds an invisible watermark and signs each file with C2PA Content Credentials, so a copy can still be traced back to you after its metadata is gone.

  • Images, video and audio through one API
  • No change to your model, prompts or training
  • Watermark, fingerprint and C2PA signature in one request
Line-screen illustration of a marble bust with a laurel wreath, seen from behindLine-screen illustration of a marble bust with a laurel wreath, seen from behind
CR · Content Credentials
MANIFEST
Embedded · valid
WATERMARK
In the pixels or samples
RESULT
Declared origin readable

Recovery is a lead back to your signed record, not cryptographic proof. Picture: AI-generated illustration.

Mark AI-generated content before it is published

Generated files rarely stay where you published them. Within a few days, many of them have been downloaded, cropped, screenshotted or posted again on another platform. Any label stored as metadata tends to disappear somewhere along that path, and with it the evidence that your system made the file.

Regulators now expect more than that. Article 50(2) of the EU AI Act has applied since 2 August 2026, and it asks providers to mark AI output in a way machines can read. Systems that were already on the EU market before that date have until 2 December 2026. In California, the AI Transparency Act asks for a hidden disclosure in AI-generated images, video and audio. Neither law picks a technology, but the voluntary EU Code of Practice sets a clear default of two layers: signed, time-stamped metadata plus an invisible watermark.

The easiest place to add those layers is between generation and publication. That is where Original Pictures sits. Your model creates a file, our API marks and signs it, and your platform publishes the signed copy. If every output follows that route, nothing reaches your users without a mark.

Line-screen illustration of a glass jar with a blank labelLine-screen illustration of a glass jar with a blank label
AI-generated illustration

How it works

  1. Generate

    Your model produces an image, a video clip or an audio file, the same way it does today.

  2. Mark and sign

    You send the finished file to our API. We embed an invisible watermark, store a fingerprint of the file on our servers and sign it with a C2PA manifest.

  3. Publish

    You download the signed copy and publish it in place of the raw output.

  4. Check later

    When a copy shows up somewhere else, you send it back to us and get a clear result: still signed, traced through its watermark or fingerprint, or not found.

What is an AI watermark?

An AI watermark is a small signal hidden inside the content itself, in the pixels of an image or video frame or in the samples of an audio track. Viewers and listeners should not notice it, but detection software can.

This is what separates it from the other two common labels. A visible logo can be cropped out in a second, and a C2PA manifest lives in the file's metadata, which many platforms strip on upload. A watermark is part of the content, so it has a far better chance of staying with the file after compression or sharing.

Line-screen illustration of a strip of film negativesLine-screen illustration of a strip of film negatives
AI-generated illustration

All three defaults are built on open models from Meta's research team, and the methods are described in published papers (Pixel Seal, which covers images and video, and AudioSeal). That matters when your engineers review a vendor: they can read how the watermark works instead of trusting a black box. For images, you can also choose one of the Adobe TrustMark variants or Meta VideoSeal. The watermark documentation lists every option.

How embedding and detection work together

Embedding is the writing step. The watermark model makes tiny changes to pixel or sample values, and together those changes carry a short message. A person should see or hear no difference.

Checking a file is where most confusion starts, because "detecting a watermark" can mean several things. A detector can tell you that a mark is present. Reading the mark goes one step further and recovers the message inside it. That message, or a fingerprint of the file, is then matched to a record in your account, and only that match tells you which signed original the copy came from.

When you check a file with us, you get one of three plain results:

  • Valid. The C2PA manifest is still attached and matches the file. This is the only result that proves origin. Validators that use the C2PA Trust List show our signer as recognized, because our certificate is on it.
  • Recovered. The manifest is gone, but the watermark or fingerprint matched one of your signed records. Treat it as a strong lead.
  • Not found. We found no mark from our system. This does not mean a person made the file.

If a copy could match more than one record, the check says so instead of picking one. And since we only search the records in your own account, we cannot confirm marks placed by other providers.

Line-screen illustration of a satellite dish on a mastLine-screen illustration of a satellite dish on a mast
AI-generated illustration

Choose how many layers you need

Every file goes through one of three modes. Pick the one that matches what a copy should keep after someone strips its metadata.

ModeWhat it addsIf a copy loses its manifestRequest flags
SignA C2PA manifest onlyNothing can lead back to your original"watermark": false, "fingerprint": false
Sign + watermarkAn invisible watermark, then the manifestThe watermark can lead back to your record"watermark": true
Sign + watermark + fingerprintA watermark, a server-side fingerprint, then the manifestThe watermark or the fingerprint can lead back, even for some edited copies"watermark": true, "fingerprint": true

The full mode gives you two independent routes back to the original. You choose the mode on each request with the watermark and fingerprint flags, so different products or teams can run different setups.

Line-screen illustration of layers of torn paperLine-screen illustration of layers of torn paper
AI-generated illustration

Balance quality and watermark strength

A watermark always involves a trade-off. A stronger mark survives more editing, but it changes the file more, and at some point people may notice. A lighter mark stays hidden more easily, and it is also easier to lose. You choose between three strength levels:

  • Subtle

    Keeps the mark as light as the model allows. If it cannot be detected at that level, the job stops instead of making the mark stronger.

  • Balanced

    The default. It starts light and increases strength only when the mark would not be detectable in the output.

  • Robust

    Allows the full strength range of the model.

Today, balanced and robust use the same range for every model we offer. No setting makes a watermark impossible to remove, and we would be wary of any vendor that claims otherwise.

Published research gives a fair picture of where the limits are. In Meta's own tests (Pixel Seal paper), PixelSeal still recovered about 96% of its message bits after heavy JPEG compression, while a crop that kept only about a third of the image brought results close to random guessing. AudioSeal handled MP3 and AAC compression well in its paper but became weaker after filters that cut low frequencies. These figures come from the model authors, not from our own tests. Independent studies also show that a determined attacker can remove a watermark or copy it onto another file, which is exactly why we never rely on the watermark alone.

Line-screen illustration of a volume knob on an amplifierLine-screen illustration of a volume knob on an amplifier
AI-generated illustration

Connect watermarking with C2PA and fingerprinting

Each layer covers a weakness in the other two.

LayerWhat it doesWeak pointProves origin?
C2PA Content CredentialsA signed record of origin, tied to the exact bytes of the fileScreenshots, re-encoding and some CDNs strip itYes, when valid
Invisible watermarkA hidden signal that can lead back to the signed recordHeavy edits and targeted attacks can damage itNo, it gives a lead
FingerprintA summary of how the content looks or sounds, kept on our serversDepends on our service, and similar files can match by mistakeNo, it gives a lead

C2PA calls a watermark or fingerprint used to find a manifest again a soft binding, and our design follows the same idea. The signed manifest stays the source of proof. The watermark and the fingerprint are how you find it once it has been separated from the file.

Deploy within your content pipeline

Integration adds one step to your publishing flow: after generation, before the file goes live. Here is the signing request your backend sends once a file has been uploaded.

# After the upload is complete (see the quickstart):
curl -sS -X POST "$OP_API/v1/sign" \
  -H "Authorization: Bearer $OP_API_KEY" \
  -H "Idempotency-Key: $(uuidgen)" \
  -H "Content-Type: application/json" \
  -d '{"asset_id": "ast_...",
       "watermark": true,
       "fingerprint": true}'
# 202 returns a job: poll GET /v1/jobs/{job_id}
# or wait for the webhook.

The job runs in the background, and you can poll for the result or receive a webhook when the signed file is ready. For common image formats with watermark and fingerprint on, there is also a faster mode that can return the finished job and a download link in the same request, which suits apps where a user is waiting on screen. Uploads, polling and verification are covered step by step in the quickstart, and batch requests in the signing guide.

Book an AI watermarking demo

Evaluate coverage, limits and pricing

StatusMedia
Watermark, fingerprint and signJPEG, PNG, WebP, TIFF, AVIF, WAV, MP3, MP4 (H.264)
Sign only, no watermarkAAC audio (which can still take a fingerprint) and several other image, audio and video formats
Signed manifest beside the fileText, documents, fonts and ML model files
Not availableText watermarking, MP4 with HEVC or HDR video, watermark removal

A few points to keep in mind when you plan:

  • Text is not watermarked. Article 50(2) also covers text, so a platform that only generates text needs a different solution.
  • A signed file shows where it came from and whether it changed. It cannot prove that the scene or event in it really happened.
  • Watermark and fingerprint matching runs on our records, so it needs our service. The C2PA manifest itself can be read by any C2PA-aware tool without us.

Pricing. You pay a $99 monthly platform fee, then per operation. Signing has a fixed price per file, and watermarking and fingerprinting are each priced per image, whatever its size, or per started second of audio or video. Failed jobs are not charged. Rates are listed on the pricing page and in the pricing documentation.

The best way to judge the watermark is on your own outputs. In a demo, we run a few of your files through the full flow and show you the results before and after a screenshot or re-encode.

AI watermarking questions

Does watermarking require retraining our AI model?

No. The watermark is added to the finished file after your model has made it. Your model, weights and prompts stay as they are, and it works with any generator.

Does Original Pictures support text watermarking?

No. A reliable text watermark has to be built into the language model while it writes, and paraphrasing or translation weakens it. For text and HTML, we provide a signed C2PA manifest stored beside the file.

Can Original Pictures detect watermarks from other providers?

No. We only search the records in your own account, so we cannot confirm that another company marked a file.

Does an absent watermark mean the content is human-created?

No. A "not found" result only means we found none of our marks. The file may come from another tool, or the mark may have been damaged. The honest answer in that case is "unknown".

How does watermarking differ from an AI classifier?

A classifier looks at a file and estimates how likely it is that a model made it, and that estimate can be wrong for any single file. A watermark is added on purpose before publication, so checking for it means looking for a known signal, not guessing from style.

Does a watermark survive everything?

No. It survives a lot of normal sharing and compression, but heavy editing, very tight crops, strong video compression and deliberate attacks can break it. That is why it works as one of three layers.

Does Original Pictures cover our legal obligations?

No product can promise that. We support your marking and detection process, but compliance also depends on your disclosures, testing and governance. This page is not legal advice.

How is watermarking billed?

One unit per image, or one unit per started second of audio or video. Units are graduated per month, from $0.015 down to $0.0025, and billed separately from signing.

See it on your own files

Bring a handful of images, clips or audio files from your model. We will mark and sign them live, then show you what a check returns for the original and for an edited copy.