Calabi Labs · Guide · 2026-06-14
Add a watermark
The search results show a crowded field of tools that let you overlay a visible logo or text stamp onto an image or video. That's a legitimate need — but it's not what Calabi does. Calabi does the opposite: it strips the invisible detection signals that make platforms flag your AI-generated file in the first place. Here's why the distinction matters.
What Actually Gets Your AI Video or Image Flagged
When you upload an AI-generated video or image to Instagram, TikTok, YouTube, or Reddit, the platform isn't scanning for a visible logo or your editing software's name in the corner of the frame. It's reading a hidden metadata layer that gets baked into every AI export by default. That layer contains signals that say "this was made by an AI model," and platforms use it to apply labels, reduce reach, or reject the upload outright.
The most common detection signals are:
- C2PA / Content Credentials — a cryptographic manifest stored as JUMBF atoms that attaches a provenance chain directly to the file, linking it to the AI model that generated it. A raw AI export can carry 18 or more of these atoms.
- XMP AI flags — specifically DigitalSourceType: trainedAlgorithmicMedia, an XML metadata tag that explicitly labels content as algorithmically generated. This is separate from any visible watermark.
- Encoder fingerprints — tools like Lavc (FFmpeg's encoder) and x264 SEI messages leave distinct bitstream signatures in video files. Platforms maintain lists of these fingerprints as AI indicators.
- Missing capture metadata — real phone recordings include Make, Model, GPS coordinates, and capture timestamp. AI exports have none of these. That absence itself is a signal.
None of these are visible in the image or video itself. You can't see them without a forensic metadata tool like ExifTool. But platforms read them automatically, often within seconds of upload.
Why Visible Watermarking Tools Won't Help Here
If you're searching for how to "add a watermark" to avoid being flagged, you might be thinking of adding a visible logo, a text overlay, or your own stamp in the corner of the frame. That approach targets the wrong layer entirely. Here's what it actually does:
- Adding a visible logo or text stamp — Tools like Canva, Clideo, or iLoveIMG let you overlay your brand mark or name onto your content. This is useful for branding and copyright notice. It does nothing to the hidden metadata layer that platforms actually scan.
- Screenshotting or re-recording — Capturing your AI video with a phone camera or screen recording software removes some metadata, but encoder fingerprints in the bitstream survive re-encoding, and the pixel patterns that perceptual hash systems detect are largely preserved.
- Cropping — Cropping can remove a visible corner logo (like Sora's sparkle or Runway's watermark). But the C2PA manifest, XMP flags, and encoder fingerprints inside the file are completely unaffected by cropping — the file's metadata layer survives.
- Metadata strippers — Some tools remove basic EXIF fields. But they don't touch C2PA/JUMBF atoms, XMP DigitalSourceType tags, or bitstream-level encoder fingerprints — the signals that actually get you flagged.
The visible watermark problem and the invisible detection signal problem are two completely separate issues. The tools in the search results above solve the first one. Calabi solves the second.
How to Actually Remove the Signals That Get You Flagged
Calabi works on the invisible layer. It runs an automatic three-stage pipeline on your AI-generated video or image file:
- Strip — Remove every detection signal in one pass: all C2PA / Content Credentials JUMBF atoms (reduced from 18+ to 0), all DigitalSourceType: trainedAlgorithmicMedia XMP flags, all generator/tool tags, and all Lavc / x264 SEI encoder fingerprints from the video bitstream.
- Inject — Write authentic phone-capture identity into the file: a real device profile (iPhone 15 Pro, Pixel 8 Pro, Galaxy S24 Ultra), matching Make, Model, Software version, GPS coordinates, and a capture timestamp. The file now looks, at the metadata level, like it came from an actual phone.
- Verify — Get a forensic proof card — the same ExifTool scan that platforms use — showing exactly what was stripped and what was injected, so you can confirm the result before downloading.
After Calabi processes your file, a raw AI export that carried 144 metadata tags (including all the AI flags and encoder fingerprints) ends up with roughly 94 neutral structural tags — and zero C2PA references, zero trainedAlgorithmicMedia flags, and zero Lavc/x264 fingerprints. The platform sees a phone recording. Nothing more.
FAQ
Q: I have a visible logo on my AI video — can Calabi remove it?
No. Calabi does not edit pixels, remove logos, or touch anything visible in the image or video. If you have a visible watermark or logo baked into the frame, cropping or a photo editor is the right tool. Calabi handles the invisible metadata layer that survives cropping.
Q: Does Calabi work on video, or just images?
Both. The pipeline strips C2PA atoms and x264 SEI encoder fingerprints from video bitstreams, and handles XMP AI flags and C2PA manifests for both image and video files.
Q: If I screenshot my AI video, do I still need Calabi?
Screen recording or screenshotting removes some metadata, but encoder fingerprints and perceptual hash patterns survive re-encoding. If you're posting to platforms that use bitstream analysis or pixel-pattern detection, a screenshot may still carry enough signals to trigger a label. Calabi gives you a clean metadata layer to stack on top of whatever visual approach you choose.
Try Calabi free at calabilabs.com — 10 cleans, no card.