Calabi Labs · Guide · 2026-06-15

Metadata remover

```html

What a Metadata Remover Actually Does — and What It Doesn't

A metadata remover strips the invisible signals baked into your AI-generated video or image file — the cryptographic manifests, XMP tags, and encoder fingerprints that tell platforms like Instagram, TikTok, and YouTube that a human didn't press the shutter button. It does not erase pixels, remove logos, or edit what the image looks like. If you downloaded a clip from Sora, Runway, Midjourney, or any generative AI tool and want it to read as a normal phone recording at the file level, that's exactly what Calabi does in one pass.

What Actually Gets Your File Flagged

When you upload an AI-generated image or video, platforms don't flag you because of how it looks. They flag you because of what's embedded in the file itself — layers of metadata that survive cropping, compression, and re-saving. Here's what's actually being scanned:

C2PA / Content Credentials: This is the big one. C2PA embeds a cryptographic manifest — stored as JUMBF (JPEG Universal Metadata Box Format) atoms — directly into the file. This manifest says who made the tool, what model generated it, and when. It's the "Made by AI" label baked into the file structure, not visible on screen. Platforms like Adobe, Microsoft, and the C2PA coalition itself are building automatic scanning for this. A raw AI export can carry 18 or more of these atoms.
XMP AI Flags: Specifically Iptc4xmpExt:DigitalSourceType set to trainedAlgorithmicMedia. The IPTC published guidance in 2023 requiring AI tools to tag synthetic media this way. Google Merchant Center actually requires this tag on AI-generated product images. It's a machine-readable declaration: "this came from an AI model."
Generator and Tool Tags: Fields like Generator, Software, Producer, and CreatorTool in the EXIF and XMP headers name the exact AI tool and version that made the file. "MagicMedia" or "DALL-E 3" in your metadata is a direct flag.
Encoder Fingerprints: Video files generated by AI carry tell-tale encoder signatures. The Lavc (FFmpeg/libavcodec) encoder identifier and x264 SEI (Supplemental Enhancement Information) NAL units are common in AI video exports. Real phone recordings use hardware encoders — the difference is detectable.
Missing Phone-Capture Signals: A real phone photo has Make, Model, Software version, GPS coordinates, and a capture timestamp. An AI export has none of these, or has them set to generic values. That absence is itself a signal.

In total, a raw AI export can carry 144 metadata tags. A significant portion of those are the AI-signaling type described above.

Why the Obvious Fixes Fail

If you're thinking "I'll just crop out the watermark" or "I'll screenshot it and re-upload" — those approaches address the visible layer, not the metadata layer. Here's the problem:

Cropping doesn't remove C2PA. The cryptographic manifest is embedded in the file structure at the byte level. Cropping the visible image doesn't touch the JUMBF atoms buried in the file. Even at 50% crop, the manifest survives intact. Platforms scanning files don't look at your image — they parse the metadata.
Screenshots re-encode but don't cleanse. Taking a screenshot creates a new file with your display's encoder, which removes some metadata. But the new file still carries whatever the screenshot captured, and if your display software embeds any AI-generation flags, those come along. Screenshots also destroy resolution and introduce visual artifacts — a poor tradeoff.
EXIF strippers remove the obvious tags but miss the rest. Most free metadata strippers remove a layer of EXIF fields — camera make, GPS, date. They don't touch C2PA/JUMBF manifests, XMP AI flags, or encoder fingerprints. The file still carries enough signals for a forensic scan to flag it. Some strippers also remove legitimate phone-capture data, making the file look even less like a real photo.

The detection layer and the visual layer are two separate things. You need to fix the file-level signals, not the pixels.

How Calabi Actually Cleans Your File

Calabi runs a three-stage pipeline on every upload. No manual editing, no sliders, no region selection. Just upload, process, and download a file that reads as a normal phone capture.

Strip the AI signals. Calabi removes all C2PA / Content Credentials JUMBF atoms, XMP DigitalSourceType: trainedAlgorithmicMedia flags, generator/tool tags, and encoder fingerprints like Lavc and x264 SEI. A raw AI export's 144 metadata tags are reduced to roughly 94 neutral structural tags — the same number a standard image would carry. The AI "fingerprint" is gone from the metadata layer.
Inject authentic phone identity. Calabi writes real device profiles into the file: iPhone 15 Pro, Pixel 8 Pro, or Galaxy S24 Ultra. That means Make, Model, Software version, GPS coordinates, and a capture timestamp — all the signals that say "this was a real phone." It uses genuine hardware encoder names, not Lavc or FFmpeg.
Return a forensic proof card. Before you download, Calabi shows you an ExifTool-readable proof card — the same forensic scan that platforms and newsrooms use. It shows exactly what was stripped (all 18 JUMBF atoms, all 16 C2PA references, the trainedAlgorithmicMedia flag) and what was injected (the device profile). You see the clean result before you commit to it.

The result is a file that passes the same automated scans platforms run — without any visible editing, without losing resolution, and without the visual degradation that comes from screenshotting or re-encoding.

What Calabi Doesn't Do

Calabi works on invisible file-level signals, not on pixels. It does not perform inpainting, content-aware fill, object removal, logo erasing, or any form of pixel-level editing. If you have a visible Sora sparkle watermark or a corner logo in the frame, cropping removes that visible mark — Calabi removes the invisible metadata layer that would flag the cropped file anyway. Those are two separate problems, and Calabi solves the one that standard cropping and screenshotting can't touch.

Frequently Asked Questions

Will this guarantee my post won't get flagged? No tool can guarantee that. Platform detection systems evolve constantly, and they use multiple signals beyond metadata — including perceptual hashes and behavioral patterns. Calabi removes the metadata and encoder signals completely (verified by ExifTool), which addresses the automated scanning layer. Results vary by platform, source model, and how the file has been previously processed.

Does Calabi remove visible watermarks or logos? No. Calabi removes invisible detection signals and metadata — it doesn't edit pixels. If you need to remove a visible logo or watermark from the frame, you'd use a photo editor for that. But the metadata layer that survives after cropping — C2PA, XMP flags, encoder fingerprints — is what Calabi handles.

What file formats does Calabi support? Calabi processes both images and videos in a single pipeline. For images, it handles standard formats carrying EXIF, XMP, and C2PA metadata. For video, it strips Lavc/x264 SEI encoder fingerprints and injects hardware encoder identity. Both output a forensic proof card showing exactly what changed.

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

```

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

Try free →