kennmarq

Mark AI pictures, songs and clips for the EU AI Act and California's AI Transparency Act in one call. Each gets an invisible watermark (TrustMark for pictures, AudioSeal for sound, Video Seal for frames), and signed, time-stamped C2PA Content Credentials that say it was made with AI. Kennmarq.com helps you meet the Act's marking duty (Art. 50(2)) and the Code of Practice's two-layer measure; you stay responsible for your system.

pip install kennmarq                 # the cloud API: no dependencies
pip install "kennmarq[local]"        # local mode: files never leave your machine
pip install "kennmarq[safety]"       # the safety check on your machine too

A free key comes with 1,000 marks a month and every feature: kennmarq.com. The SDK reads it from KENNMARQ_API_KEY.

Cloud

from kennmarq import Kennmarq

k = Kennmarq()
result = k.mark("render.png", "render.marked.png", source="created", model="FLUX.1 schnell")
print(result.mark_id, result.watermark_id, result.marks_used)

The picture goes to Kennmarq.com's servers in Zurich, which mark it in memory and keep nothing.

Add label=True to draw the EU's visible AI label ("AI GENERATED", or "AI MODIFIED" for edits) into a picture or every frame of a clip, before the watermark; deepfake=True turns it on by itself, and label_language, label_position, label_size and label_style shape it (docs). Labels are drawn in the cloud.

Songs and clips

song = k.mark("track.mp3", "track.marked.mp3", model="Song Model 2")
print(song.marks_counted, song.audio_watermark)   # five marks per started minute; "claimed"

clip = k.mark("clip.mp4", "clip.marked.mp4", check="ai-act")   # waits for the job, then downloads
print(clip.watermark_id, clip.marks_counted)      # five marks per started second

Songs (MP3, WAV, M4A, up to 15 minutes) carry Meta's AudioSeal in the sound, shaped so no one hears it; clips (MP4 at up to 1080p, 4,500 frames: 2.5 minutes at 30 frames a second) carry Video Seal in every frame and AudioSeal in their sound. Both are marked in the cloud. A clip is a job: mark() waits for it and downloads the marked clip; mark(..., wait=False) returns at once, and wait_for(mark_id, dest) finishes later.

Local mode

result = k.mark_file("render.png", "render.marked.png", source="created", model="FLUX.1 schnell")

The watermark and the Content Credentials are made on your machine. Only the manifest goes to Kennmarq.com: about 13 KB of claim and assertions, never pixels. Kennmarq.com signs and time-stamps it, and the SDK puts it back in your file. Before writing anything, the SDK checks the file reads as signed by Kennmarq.com, with a validated time-stamp and the watermark in place; if not, it raises InvalidSignedFile and writes nothing. Each file counts one mark.

• Credentials your file already carries (from an upstream model's API, say) are kept as an ingredient, as the Code of Practice asks. They travel with the manifest, with any thumbnail their signer put in them; keep_input_credentials=False keeps and sends none.
• The watermark's model (TrustMark Q, 65 MB) downloads once, to ~/.cache/kennmarq/models or KENNMARQ_MODELS_DIR, and is checked against its pinned SHA-256.
• Where c2pa-python has no build for your system (Windows, for now), mark_file makes the watermark on your machine and sends the watermarked picture to the cloud for its manifest (mode == "hybrid").
• Without a key, watermark_file() adds the watermark and the IPTC AI label alone, with no Content Credentials.

The safety check

answer = k.check("render.png", policy="ai-act")              # cloud: two marks
answer = k.check("render.png", policy="ai-act", local=True)  # here: nothing sent, nothing counted
if answer["verdict"] == "block":
    ...

It screens a picture for what the AI Act's new bans are about (Art. 5(1)(ba) and (bb)): sexual content with someone who appears to be a child, and with a realistic face, which needs the person's consent. ai-act blocks the first and flags the second; safe-only also blocks anything explicit. It returns a verdict and a reason, never an age or an identity. It's a risk signal with published error rates (docs), not a guarantee: you decide, and stay responsible. Locally its models download once (about 2.3 GB). mark() and mark_file() take check= too: a block raises ContentBlocked and nothing is marked.

Check a file

k.verify("render.marked.png")    # free, no key needed

What its Content Credentials and its watermark say, and the credential the watermark leads to when the metadata was stripped.

Errors

API errors raise KennmarqError with the API's code, status and request_id (docs). ContentBlocked carries the check's answer in safety.

MIT-licensed. The paid part is Kennmarq.com's signature, the time-stamps and the stored credentials. The face model and the prompt embeddings bundled for the safety check carry their own licences: THIRD_PARTY_NOTICES.md.