Calabi Labs · Guide · 2026-06-14
```html
No tool can pixel-paint over a TikTok logo and make it disappear—but that's not what gets you flagged anyway. TikTok's automated systems scan the invisible metadata layer of your video file: the encoder fingerprints, the AI-generation tags buried in the file structure, and the cryptographic Content Credentials that travel with AI exports. Calabi strips those signals and injects the metadata signature of a real phone recording, so your upload reads as authentic capture rather than AI-generated content. The visible watermark is a separate problem that cropping solves; the metadata detection layer is what Calabi handles.
When you export a video from Sora, Runway, Kling, or any AI generator, the file carries a forensic trail that platform scanners detect within seconds of upload. This has nothing to do with whether the video looks real—it lives in the file's metadata and encoding structure.
The primary signal is C2PA / Content Credentials, stored as JUMBF atoms in the file. This is a cryptographic manifest that explicitly declares a video was machine-generated, listing the model used, the generation parameters, and a digital signature. TikTok, Instagram, and YouTube all check for this. A single AI-exported clip can contain 18 or more of these JUMBF atoms pointing to AI provenance.
Beyond C2PA, there are XMP metadata tags—specifically fields like DigitalSourceType: trainedAlgorithmicMedia that flag the content as AI output. These aren't visible anywhere in the video itself. There are also encoder fingerprints: AI generation tools like FFmpeg (Lavc) and x264 embed SEI (Supplemental Enhancement Information) messages in the bitstream that identify them as machine-generated encoding. Finally, authentic phone recordings carry signals AI exports don't have: GPS coordinates, a real capture timestamp, and device-specific encoder names like iPhone 15 Pro or Pixel 8 Pro. The absence of those authentic signals is itself a red flag.
You can crop the TikTok logo out of a frame, and that removes the visible mark—but the file's metadata trail survives untouched. The C2PA manifest, the XMP AI flags, and the Lavc/x264 encoder fingerprints are embedded in the file structure, not in the pixels of the video itself. Cropping changes the visual content; it doesn't rewrite the file's invisible forensic signature.
Screenshotting a video and re-uploading the screenshot has the same problem in reverse. Your screenshot now carries the metadata of whatever device took the screenshot, but the platform scanners aren't primarily checking device metadata on individual frames. They're scanning for the specific AI-generation signals in video bitstreams. And if you're uploading an AI video that originally came from a generator, cropping the visible logo doesn't remove the AI metadata that platforms actually detect.
A simple re-encode (uploading to another platform and downloading) disrupts some encoder fingerprints, but it doesn't touch C2PA manifests or XMP AI flags. Platform scanners have gotten sophisticated enough to catch re-encoded AI content because the structural metadata survives most casual re-encodes.
Calabi works on the file level, not the pixel level. It doesn't erase, paint over, or reconstruct any region of your video. Instead, it runs a three-stage pipeline that rewrites the file's forensic identity.
Stage 1: Strip. Calabi removes all AI-generation signals from the file. This means zeroing out every C2PA / Content Credential JUMBF atom (18 down to 0 in tested AI exports), stripping the DigitalSourceType: trainedAlgorithmicMedia XMP flag, removing generator and tool tags, and clearing Lavc / x264 SEI encoder fingerprints that identify the video as AI-encoded. The result is a file that no longer carries the machine-generation declaration in its metadata.
Stage 2: Inject. Calabi writes authentic phone-capture identity into the file. It injects a real device profile—iPhone 15 Pro, Pixel 8 Pro, or Galaxy S24 Ultra—with matching Make, Model, Software version, GPS coordinates, and a capture timestamp. It also adds a genuine phone encoder name to replace the removed AI encoder fingerprints. The file now looks, at the metadata level, like something recorded on a physical device.
Stage 3: Verify. Before you download, Calabi generates a forensic proof card—a scan using ExifTool, the same tool newsrooms and forensic investigators use. It shows exactly what was stripped (the C2PA atoms, the trainedAlgorithmicMedia flag, the encoder fingerprints) and what was injected (device profile, GPS, timestamp). You see the before-and-after metadata state that platforms will read when you upload.
Does Calabi remove the TikTok logo from my video?
Calabi doesn't edit pixels or remove visible marks. Cropping removes the visible TikTok logo from your video frame. What Calabi handles is the invisible detection layer—C2PA manifests, XMP AI flags, and encoder fingerprints—that survives cropping and gets your video flagged on upload even when the logo isn't visible.
Can I use Calabi on videos I've already posted?
No—platforms scan uploads, not already-posted content. Calabi prepares a file before upload so it passes platform scanning as authentic capture rather than AI-generated content.
Will this guarantee my video won't get flagged on TikTok?
No tool can guarantee that. Platform detection systems vary and evolve. Calabi removes the specific metadata signals—C2PA, XMP AI flags, encoder fingerprints—that automated scanners currently check. Results vary by platform and by the source model used to generate the video.
Try Calabi free at calabilabs.com — 10 cleans, no card.