Check a video for our marks

Every video ClonifyLab makes carries two invisible marks that show it was made with AI by ClonifyLab. Here is how to check them yourself, for free.

1. Content Credentials (C2PA)

Our files are signed with C2PA Content Credentials. They state that the file contains content generated by AI (“Composite including generative AI elements”) and that ClonifyLab signed it, with a trusted time stamp.

  • Open the file in a free Content Credentials checker, such as contentcredentials.org/verify, or run the open-source c2patool.
  • Our signing certificate is issued by our own root, which isn’t on public trust lists, so a checker may call the signer “unknown”. Compare the certificate with ours below.
  • Social platforms often remove this metadata from posted videos. The watermark (step 2) stays.
CertificateValidSHA-256 fingerprint
ClonifyLab C2PA Root CA · download26 Oct 2026 – 23 Oct 20365f4a9477a368165e7712651099ae1119133647a96f14b8dee55c4929112b3bec
ClonifyLab C2PA Signer (render service) · download26 Oct 2026 – 25 Oct 2031217fc1de011c780248939677b2a3ea15105dc21fe7c31148ffa9997f52fb296b

2. The invisible watermark

The picture carries a Meta Video Seal 1.0 watermark (256 bits) and the sound a Meta AudioSeal 0.2 watermark (16 bits). Both use one public key: the SHA-256 digest of the text ClonifyLab AI-generated v1:

9c249661471a5e7fc3aef2b1714b4280b416d12cf5db18c87a9d8c6984fa36d6

The picture uses all 256 bits, the sound the first 16 (byte 0 first, most significant bit first).

Run the free check

  • Download our script: verify.py (the same script our own final check runs on every video).
  • You need Python 3.10 or newer and FFmpeg 5.1 or newer (ffmpeg and ffprobe on your PATH).
  • Download the two model files (MIT licence; the script checks their SHA-256 before loading them):
Model fileSHA-256
y_256b_img.jit (Video Seal)5c7a4581c36fc6090aafdcfb3999123bae5172a4847f22e2da4e7fd1a39d1e1b
detector_base.pth (AudioSeal)8a78e8a83584113523e161fc599fcab10fd0e94c04d2eb9d2fa1e9ec91ab69d9

Then install the libraries and run the check:

pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install numpy omegaconf einops
pip install --no-deps audioseal==0.2.0
python verify.py my-video.mp4 --video-model y_256b_img.jit --audio-model detector_base.pth

It prints, for example, AI-generated by ClonifyLab: yes (p = 4.1e-62), plus one line each for the picture and the sound. The p-value is the chance that a video without our watermark would match our key this well; the script says “yes” at one in a million or less.

What the marks can and can’t prove

  • The watermark is a label, not a signature: the models are open and the key is public, so someone could copy or remove it. The signed C2PA Content Credentials are the proof that a file came from us.
  • Heavy edits, such as cropping more than 10–20%, rotating, screen recording or re-making the video with AI, can remove the watermark.
  • A picture with almost no texture (one flat colour) can’t carry the picture watermark; our final check rejects such a video before it’s delivered.
  • Found our mark in a video that misuses someone’s face or voice? Report it, and we act within 48 hours.