Calabi Labs · Guide · 2026-06-14
Remove watermark from image
Calabi does not erase visible watermarks pixel-by-pixel — that's a photo editor's job, and it requires a very different tool. What Calabi actually does is strip the invisible detection layer attached to your AI-generated image file: the metadata flags, Content Credentials, and encoder fingerprints that platforms like Instagram and TikTok scan to flag content as AI-generated in the first place.
What actually gets your AI image flagged
When you export an image from Midjourney, DALL-E, Firefly, Stable Diffusion, or Sora, the file carries an invisible forensic trail. Platforms don't just look at how an image looks — they scan the file's metadata layer for signals that give it away.
The most damning flag is C2PA / Content Credentials, stored as JUMBF atoms in your image file. This is a cryptographic manifest that cryptographically proves the image was generated by an AI tool. It lists the AI model, generation parameters, and every edit pass — and it's designed to survive re-saves and re-uploads. Then there's the XMP AI flag: fields like DigitalSourceType: trainedAlgorithmicMedia that explicitly label the content. Your file also carries encoder fingerprints — Lavc (LibAV encoder) markers and x264 SEI data in video exports that are dead giveaways of AI generation. Add to that missing fields: no GPS coordinates, no EXIF capture timestamp, no real device metadata. A clean phone recording has all of that. Your AI export has none of it.
That's the combination that gets you flagged. Not the content itself — the file's invisible signature.
Why cropping or screenshotting doesn't solve it
If you've searched for "remove watermark from image," you've probably tried the obvious moves: cropping out the visible logo, taking a screenshot to get a "fresh" copy, or re-saving in a different tool. Here's the honest truth: those methods do remove the visible mark, but they do almost nothing to the invisible detection layer.
A screenshot strips some metadata, but it preserves the perceptual hashes and encoder artifacts that forensic tools still recognize. Cropping removes the visible corner logo, but the C2PA manifest, XMP flags, and encoder fingerprints are embedded in the entire file structure — they're not in the corner pixels. They survive cropping. Re-saving in Photoshop or another editor removes some tags but almost never all of them, and it leaves the structural fingerprints that ExifTool and platform scanners still catch.
What you need is a complete metadata and signal overhaul at the file level — not a visual fix.
How Calabi actually cleans your file
Calabi runs a three-stage pipeline on your AI-generated image or video in a single pass. You upload the file, and Calabi handles everything automatically.
- Strip the AI detection signals. Calabi removes every C2PA / Content Credential manifest — all 18 JUMBF atoms and 16 C2PA references go to zero. It strips the
DigitalSourceType: trainedAlgorithmicMediaXMP flag and every generator/tool tag. Encoder fingerprints like Lavc and x264 SEI markers are removed from the bitstream. - Inject authentic phone-capture identity. Calabi writes real device metadata: a confirmed phone make and model (iPhone 15 Pro, Pixel 8 Pro, Galaxy S24 Ultra), a real encoder name, GPS coordinates, and a capture timestamp. The file now looks, at the metadata level, exactly like a photo your phone took.
- Verify with a forensic proof card. Before you download, Calabi shows you the same ExifTool scan that newsrooms and platform moderators use. You see exactly what was stripped (all those AI flags) and exactly what was injected (the phone identity). It's the same forensic scan — so you know it will pass.
The result: your AI-generated image reads as a normal phone photo at the file level. The visible logo or mark — if you cropped it or it's not there — is irrelevant. What platforms actually scan is this metadata layer, and that's what Calabi cleans.
FAQ
Does Calabi remove visible logos or text watermarks?
No. Calabi does not edit pixels, use inpainting, or clone-stamp tools. If your image has a visible logo or text overlay in the corner, you need to crop it out or use a photo editor first. What Calabi removes is the invisible detection layer — the metadata and encoder signals — that survives cropping and re-saving.
Can platform scanners still detect my image if I've cropped the visible mark?
Yes. Cropping removes the visible logo, but the C2PA manifest, XMP AI flags, and encoder fingerprints embedded in the file structure are not in the corner pixels. They survive cropping. Calabi strips those signals so the metadata no longer flags the image as AI-generated.
What does the forensic proof card show?
The proof card is an ExifTool readout — the same forensic scan that platform moderation systems and newsrooms use. It shows your file's metadata before and after cleaning: all 18 C2PA atoms and 16 references reduced to 0, the trainedAlgorithmicMedia flag removed, and 144 structural tags trimmed to about 94 neutral ones. You see exactly what was stripped and what phone identity was injected.
Try Calabi free at calabilabs.com — 10 cleans, no card.