Calabi Labs · Guide · 2026-06-16
Pdf watermark
When someone searches "PDF watermark," they're usually trying to solve one of two problems: removing a visible logo or stamp someone else placed on a document, or stripping the invisible AI-detection metadata baked into a file they created. Calabi handles the second one — the hidden metadata layer that platforms like Google Merchant Center, Instagram, and content moderation systems actually scan for.
What Actually Gets Flagged in an AI PDF
Most people assume a "watermark" means a visible logo or stamp. But the detection layer platforms use in 2026 is almost entirely invisible metadata — and it lives in the file's structure, not on the page.
The primary signal is C2PA / Content Credentials: a cryptographic manifest embedded in the file that says, in machine-readable form, "this was generated by AI." This uses the JUMBF (JPEG Universal Metadata Box Format) structure and stores information about the generative model, software version, and creation timestamp. Google Merchant Center's AI content policy explicitly requires images with AI-generated metadata to carry an IPTC DigitalSourceType value of trainedAlgorithmicMedia — that's the specific XMP tag platforms check.
Beyond the AI-specific flags, the metadata also reveals the encoder fingerprint: software like Adobe Firefly, Midjourney, or DALL-E stamp the file with their own tool identifiers, encoder names, and generation parameters. A raw AI export can carry over 140 metadata tags. A platform scanning for "AI-generated" content doesn't need to analyze the pixels — it just reads the XMP and IPTC headers. Missing fields matter too: a real phone-captured PDF has GPS coordinates, a capture timestamp in a specific EXIF format, and a real device Make/Model. AI exports typically lack all of these.
Why the Obvious Fixes Don't Work
If you've tried removing a visible watermark from a PDF before, you know the frustration of screenshotting a document only to have the quality degrade. But even if you don't mind the quality loss, screenshotting and re-uploading doesn't solve the metadata problem — because the detection signals aren't in the pixels, they're in the file structure.
Re-exporting a PDF from preview software, printing to PDF, or converting formats strips some visible artifacts but leaves the C2PA atoms, XMP AI tags, and encoder fingerprints largely intact. Platforms like Google and Instagram run automated scanners that read metadata headers before they ever render the file. A PDF that looks "clean" visually still broadcasts "AI-generated" in its metadata layer — and that metadata survives most conventional transformations.
For visible watermarks — a corner logo, a "SAMPLE" stamp, a text overlay — the honest answer is that Calabi doesn't erase logos pixel-by-pixel. That's a photo-editor job, not a sanitizer. But cropping or using a PDF editor to remove the visible mark does work, and critically, cropping doesn't remove the invisible metadata layer. That's where Calabi comes in after you've handled the visible elements.
How Calabi Cleans an AI PDF
Calabi works on the file-level signals, not the visual content. It runs a three-stage pipeline on whatever file type you upload — PDF, PNG, JPEG, MP4, or WebP.
Stage 1 — Strip: Calabi removes every detection signal from the metadata layer. That means JUMBF / C2PA atoms are zeroed out, DigitalSourceType: trainedAlgorithmicMedia XMP flags are deleted, IPTC AI tool tags are stripped, and encoder fingerprints like Lavc SEI data are removed. In testing, 18 JUMBF atoms and 16 C2PA references drop to 0.
Stage 2 — Inject: Calabi replaces the stripped signals with authentic phone-capture identity. It writes real device profiles — iPhone 15 Pro, Pixel 8 Pro, Galaxy S24 Ultra — including Make, Model, Software version, GPS coordinates, and a capture timestamp. The encoder fingerprint changes from "Adobe Firefly 3.0" to a real phone encoder name.
Stage 3 — Verify: Before you download, Calabi generates a forensic proof card — the same ExifTool scan that platforms use — showing exactly what was stripped and what was injected. You see the before-and-after metadata count: a raw AI export's 144 tags reduced to about 94 neutral structural tags.
The process is one-pass and automatic. Upload, wait for the pipeline, download the cleaned file with its proof card.
FAQ
Will Calabi remove a visible watermark from my PDF?
No. Calabi removes the invisible metadata layer — C2PA, XMP AI flags, encoder fingerprints — not pixel-level content like logos or text overlays. If your PDF has a visible logo or stamp you need gone, use a PDF editor to crop or redact it first, then run Calabi to clean the metadata layer underneath.
Does re-saving my PDF in Adobe Acrobat remove AI metadata?
Not reliably. Standard "save" and even "export as" operations often preserve the C2PA atoms and XMP headers that AI generation tools write. Some settings in Acrobat can strip metadata broadly, but that removes everything — including legitimate author and copyright info — without replacing the detection signals with authentic phone-capture identity.
How do I know what metadata my PDF has right now?
You can run an ExifTool scan on any file to see its full metadata payload. Look for C2PA / JUMBF boxes, DigitalSourceType in the XMP namespace, and tool identifiers like "Adobe Firefly" or "Midjourney" in the metadata fields. Platforms like Google Merchant Center check exactly these fields when determining whether content carries an AI-generated disclosure requirement.
Try Calabi free at calabilabs.com — 10 cleans, no card.