Calabi Labs · Guide · 2026-06-19

Detect an invisible watermark

Can You Detect an Invisible Watermark in AI Images? Here's What Actually Works

Yes — invisible watermarks in AI-generated images can be detected, but the methods are specialized, and no tool guarantees complete removal. The detection layer that matters most for platform posting isn't the invisible pixel pattern itself, it's the metadata and encoder signatures baked into every AI export. That's where Calabi operates: stripping the cryptographic and metadata signals that platforms actually scan for, not attempting to decode steganographic pixel patterns.

What "Invisible Watermark" Actually Means

The second — and the one that actually gets you flagged on Instagram or TikTok — is the metadata and encoder detection layer. When you export from Midjourney, Sora, Runway, or DALL-E, the file carries an invisible trail: C2PA Content Credentials stored as JUMBF atoms, XMP tags like DigitalSourceType: trainedAlgorithmicMedia, and encoder fingerprints in the video bitstream (Lavc, x264 SEI messages). Platforms scan for these signals automatically, often within seconds of upload. This is the detection layer Calabi removes.

Why You Can't Just "See" These Watermarks

Invisible pixel watermarks are imperceptible by design. You cannot detect them by zooming in, examining metadata tables, or screenshotting the image. They require either:

Frequency analysis tools that expose the modulated patterns, which most creators don't have access to.

This is why screenshotting or cropping an AI image doesn't remove the invisible watermark — you're not touching the pixel-level signal. And re-exporting through a photo editor may disrupt some patterns through recompression, but it doesn't strip the metadata layer that platforms actually key on.

What Calabi Actually Handles

Calabi works on the detection layer that survives cropping and most re-encodes: the invisible metadata and encoder signatures. The pipeline runs in three stages:

Strip — Removes C2PA / Content Credentials JUMBF atoms, XMP AI flags (including DigitalSourceType: trainedAlgorithmicMedia), generator/tool tags, and encoder fingerprints like Lavc and x264 SEI from video bitstreams.
Inject — Adds authentic phone-capture identity: Make, Model, Software version, GPS coordinates, capture timestamp, and a real-phone encoder profile (iPhone 15/16 Pro, Pixel 8 Pro, Galaxy S24 Ultra).
Verify — Returns a forensic proof card — the same ExifTool scan platforms use — showing exactly what was stripped and what was injected, before download.

The re-encode step during processing disrupts some invisible pixel watermark patterns, but results vary by source model and watermark type. Calabi makes no guarantee about steganographic patterns it cannot directly inspect. What it does guarantee: the metadata and encoder signals that automated platform scanners rely on are removed to zero.

The Honest Edge Cases

A few things to know plainly:

Visible watermarks (corner logos, Sora's sparkle) — cropping removes the visible mark. Calabi removes the invisible metadata layer that survives cropping. These are separate problems with separate solutions.
Invisible pixel watermarks — a lossy re-encode disrupts some patterns, but no tool can guarantee removal. Focus on the metadata/encoder signals Calabi fully strips.
Platform detection — no tool can guarantee a platform won't flag you. Results vary by platform, source model, and the detection methods in use at any given time.

How to Check If an Image Has Invisible Watermark Signals

If you want to inspect an image yourself before running it through Calabi:

Open the file in ExifTool (the same forensic tool newsrooms and platform trust teams use).
Check for C2PA, Content-Credential, or JUMBF atoms in the metadata structure.
Look for XMP fields: DigitalSourceType, GenAI, or any generator/tool name in the XMP packet.
For video: scan the bitstream for Lavc, x264, or SEI encoder signatures.

If any of these are present, an automated scanner can flag the file as AI-generated — regardless of whether a human could perceive a watermark.

FAQ

Can I remove an invisible watermark by taking a screenshot?

No. Screenshotting captures the visual output, not the metadata layer or the steganographic pixel pattern. Some invisible watermarks are designed to survive screenshotting and recompression. Calabi strips the metadata signals that survive cropping and re-encoding, but a screenshot doesn't remove encoder fingerprints or C2PA metadata from the original file.

Does re-exporting through Photoshop remove AI metadata?

Partially. Re-saving through a photo editor recompresses the image and may disrupt some invisible pixel patterns, but it typically doesn't remove C2PA/JUMBF atoms, XMP AI flags, or encoder fingerprints from the metadata structure. Platforms scan metadata, not just pixel patterns. Calabi explicitly strips all three layers.

How do platforms detect AI images in 2026?

Instagram, TikTok, YouTube, and Reddit all run automated scanners that check for C2PA Content Credentials, XMP DigitalSourceType: trainedAlgorithmicMedia flags, encoder fingerprints (Lavc, x264 SEI), missing GPS/timestamp, and perceptual hash databases. Some also run neural classifiers on pixel patterns. Calabi eliminates the metadata and encoder signature layer; the pixel-pattern classifiers are a separate, less standardized detection method.

Try Calabi free at calabilabs.com — 10 cleans, no card.

10 free cleans. See the forensic proof before you download.

Try free →