Calabi Labs · Guide · 2026-06-19
Remove object from video
When you search "remove object from video" today, you might be looking for two very different things: you either want to erase something visually — a person walking through your shot, a logo in the corner, a timestamp overlay — or you're trying to strip the invisible signals that tell platforms your content is AI-generated. Calabi handles the second one. It doesn't paint over pixels or reconstruct frames. What it does is remove the cryptographic and metadata layer that gets you flagged, and it does it in a single pass.
What Actually Gets Your Video Flagged
Platforms like Instagram, TikTok, YouTube, and Reddit aren't scanning your video frame by frame looking for a Midjourney-style grid or a Sora sparkle watermark. They're reading the metadata layer underneath — the stuff most editors never see. The main signals are:
- C2PA / Content Credentials: A cryptographic manifest stored as JUMBF atoms that explicitly declares "this was generated by an AI model." A raw Runway or Sora export can carry 18 or more of these atoms. That's what gets you flagged the first time you upload.
- XMP AI flags: Fields like
DigitalSourceType: trainedAlgorithmicMediaembedded directly in the file's XMP metadata. This is a machine-readable declaration that the content came from a trained model. - Encoder fingerprints: When AI video gets exported, the encoder leaves traces — Lavc (LibavCodec), x264 SEI (Supplemental Enhancement Information) units, and similar fingerprints that don't appear in normal phone recordings. Platform detectors recognize these patterns.
- Missing capture metadata: Real phone recordings have Make, Model, Software, GPS coordinates, and capture timestamps. AI exports have none of these, or they have obviously synthetic values. That absence is itself a signal.
Why Cropping, Screenshotting, and Re-Uploading Don't Work
If you've already tried to remove an object from video by cropping it down, taking a screenshot, or re-exporting it through a different app — you already know it didn't work. Here's why:
Cropping removes the visible frame area, but the metadata survives. A cropped AI video still carries C2PA atoms, XMP flags, and encoder fingerprints inside the remaining pixels. Platforms read the file's metadata, not what the thumbnail looks like.
Screenshotting strips some metadata but adds new problems. Your screenshot becomes a compressed image sequence or a phone screenshot with its own capture metadata — which looks different from a video file and still doesn't have the full phone-capture identity a platform expects.
Re-exporting through DaVinci Resolve, HandBrake, or FFmpeg removes some visible encoder fingerprints but leaves the C2PA layer largely intact. And it adds a new encoder fingerprint — the one from your re-export tool — which is also recognizable as non-phone.
None of these approaches inject the phone-capture identity that platforms are actually looking for.
How to Actually Clean an AI Video File — The Calabi Way
Calabi works on the metadata layer, not the pixels. Upload your AI-generated video and an automatic three-stage pipeline runs:
- Strip: Remove every detection signal in one pass — all JUMBF/C2PA atoms, all XMP AI flags including
DigitalSourceType: trainedAlgorithmicMedia, encoder fingerprints like Lavc and x264 SEI, generator/tool tags, and the raw AI metadata fields that platforms scan for. - Inject: Write authentic phone-capture identity into the file — a real device profile (iPhone 15 Pro, Pixel 8 Pro, Galaxy S24 Ultra), realistic capture timestamp, GPS coordinates, and a proper encoder name. The file now looks structurally identical to something recorded on that device.
- Verify: Before you download, Calabi shows you a forensic proof card — the same ExifTool scan that newsrooms and platform trust-and-safety teams use. You see exactly what was stripped (18 JUMBF atoms → 0, C2PA references → 0,
trainedAlgorithmicMedia→ removed) and what was injected.
The result is a file that passes the checks platforms run — not because it looks different, but because its invisible identity is now consistent with a normal phone recording.
FAQ
Does Calabi remove visible watermarks like the Sora sparkle or Runway logo?
No. Calabi removes the invisible detection layer — the C2PA manifests, XMP flags, and encoder fingerprints. If your source has a visible watermark, cropping removes it since the watermark sits in the pixel data. Calabi handles the metadata that survives cropping.
Can I remove a specific object from my video with Calabi?
No. Calabi is not a video editor. It doesn't select, reconstruct, or paint over any region of the video. It's a forensic metadata tool for creators who need their AI-generated content to read as phone-captured at the file level.
What if a platform uses perceptual hashes instead of metadata?
Some platforms also scan perceptual hashes (pHash), which are derived from visual patterns in the actual pixels. A re-encode through Calabi's pipeline disrupts some perceptual patterns, but results vary by platform and source model. The metadata layer is what Calabi fully controls and guarantees — forensic proof shows exactly what was stripped and injected at that level.
Try Calabi free at calabilabs.com — 10 cleans, no card.