Calabi Labs · Guide · 2026-06-14
Text remover
What Happens When You Search for a "Text Remover"
Most people who type "text remover" into Google are looking for a tool that physically erases visible words from an image or video — a caption, a watermark, a subtitle — and fills in the background behind it. That's what Pixelcut, CapCut, Canva, and dozens of other apps market under that name: select the text, inpaint the area, done. Calabi doesn't do that. If you need to visually erase words pixel-by-pixel, use one of those tools.
But if you're trying to stop a platform from detecting that your content was AI-generated — or recognizing it as a repost, a screen recording, or anything other than a fresh phone capture — then what you actually need is a metadata and signal-level cleaner. That's what Calabi does. The search term "text remover" gets you visual editors; what you likely need is something else entirely.
What Actually Gets Your File Flagged
Platforms like Instagram, TikTok, YouTube, and Reddit don't scan your video frame-by-frame looking for the word "AI-generated" burned into the image. They scan the invisible layer underneath: the metadata. When you export a video from Sora, Runway, Kling, or any other AI video tool, the file carries a specific set of forensic signals that automated systems are designed to catch on upload — often within seconds.
Here's what's actually in that file:
- C2PA / Content Credentials (JUMBF atoms): A cryptographic manifest embedded in the file that explicitly states when and how the content was generated by AI. This is the single most damning signal. A raw Sora export contains 18 JUMBF atoms and 16 C2PA references pointing directly to an AI model.
- XMP metadata tags: Fields like
DigitalSourceType: trainedAlgorithmicMedia— an Adobe-coined XMP property that explicitly flags AI-generated source material. This tag alone is enough for some detection pipelines. - Encoder fingerprints: Video exports from AI tools carry specific encoder signatures — Lavc (libavcodec), x264 SEI (Supplemental Enhancement Information) NAL units — that don't exist in footage from a real iPhone 16 Pro or Pixel 8 Pro.
- Missing authentic capture signals: A genuine phone recording has Make, Model, Software version, GPS coordinates, and capture timestamp. An AI export has none of these by default. That absence itself is a red flag.
A raw AI video export has roughly 144 metadata tags. A real phone recording has around 60–70. The gap is obvious to forensic tools, and platforms use those tools automatically.
Why Cropping, Screenshots, and Re-Encoding Don't Work
The intuitive fix is to physically remove the evidence: crop out the watermark, take a screenshot to break the file structure, re-encode the video to strip the metadata. These approaches fail for a specific reason — they remove what humans can see, not what automated systems are actually scanning.
When you crop a visible watermark, you remove the visual logo. The C2PA manifest, XMP tags, and encoder fingerprints are still embedded in the file. They survive cropping because they're stored in the file header, not in the pixel region you're cropping. The same logic applies to screenshots and re-encoding: the visible layer changes, but the metadata layer persists unless it's explicitly stripped. Re-encoding also introduces its own encoder fingerprint — often Lavc or FFmpeg — which is itself an AI-export signal rather than an authentic one.
Some people try to manually edit EXIF data with tools like ExifTool. That's closer to the right approach, but it requires knowing exactly which fields to remove and which to inject — and if you remove everything without replacing it with authentic device identity, the file looks like a stripped asset, which is also a flag.
How Calabi Actually Cleans Your File
Calabi runs a one-pass pipeline that simultaneously strips detection signals and injects authentic phone-capture identity. Here's what happens in that single pass:
- Upload your AI-generated video or image. Calabi accepts the raw export file — no manual selection or region marking needed.
- Calabi strips the detection layer. All 18 JUMBF / C2PA atoms are removed. All 16 C2PA references are zeroed. The
DigitalSourceType: trainedAlgorithmicMediaXMP tag is deleted. Encoder fingerprints like Lavc and x264 SEI NAL units are stripped. Your 144 metadata tags are reduced to approximately 94 neutral structural tags. - Calabi injects authentic phone identity. The file receives a real device profile — iPhone 15 Pro, iPhone 16 Pro, Pixel 8 Pro, or Galaxy S24 Ultra — including Make, Model, Software version, GPS coordinates, and a capture timestamp. The encoder fingerprint is replaced with one matching the device profile.
- You receive a forensic proof card. Before download, Calabi shows you the same ExifTool scan that platform detection systems use. It lists everything that was stripped (C2PA, XMP flags, encoder signatures) and everything that was injected (device profile, GPS, timestamp). You can verify it yourself before the file ever reaches a platform.
- Download the cleaned file. The file is ready to upload with the metadata profile of a genuine phone recording.
No visual editing. No pixel reconstruction. No inpainting. The file looks identical to a platform's scanning system because the scanning system reads metadata, not pixels.
FAQ
Does Calabi remove visible watermarks or logos from images?
No. Calabi doesn't modify pixels, select regions, or reconstruct any part of an image or video frame. If you have a visible logo or watermark you want gone — a Sora sparkle mark, a tool's corner badge — a visual editor like Pixelcut, Canva, or CapCut is the right tool for that. Calabi handles the metadata layer underneath: the invisible signals that survive after you've already cropped or edited the visible content.
I took a screenshot of my AI video and uploaded it. Why did it still get flagged?
Screenshots break the video file structure, which removes some metadata, but screenshot files (JPEG/PNG from a screen capture) carry their own signals — screen capture timestamps, display resolution metadata, and software fingerprints — that detection systems also flag. A screenshot of an AI image is still recognized as a screen recording of AI-generated content by many platform scanners in 2026. Calabi works on the original file, not a captured reproduction.
Can I use Calabi for content I've edited visually already?
Yes. If you've already cropped a watermark, run the file through Calabi before uploading — it will strip the residual AI metadata and injection signals that still exist in the edited file, then inject authentic phone identity. Many creators run their exports through Calabi as a final step regardless of whether they edited visually, because the metadata flags persist even after visible signs are removed.
Try Calabi free at calabilabs.com — 10 cleans, no card.