Calabi Labs · Guide · 2026-06-19
Ai image translator
If someone searches "AI image translator," they might be looking for a tool to translate text inside an image — but that's not what Calabi does. Calabi is a one-pass web tool that strips the invisible AI detection signals from any AI-generated image or video and injects authentic phone-capture identity, so platforms read it as a normal recording. If you need language translation inside an image, use Google Translate's camera mode. If you need to post an AI image without it getting flagged as AI-generated, that's what Calabi handles. This page covers the second problem.
What actually gets your AI image flagged — and it's not the visible content
When you upload an AI-generated image to Instagram, Reddit, or TikTok, the platform doesn't flag it because of how the picture looks. It flags it because of invisible metadata signals embedded in the file. These signals are the real problem.
The most consequential is C2PA / Content Credentials — a cryptographic manifest stored as JUMBF atoms that literally declares "this image was generated by AI." A single AI export can contain 18 or more of these JUMBF atoms and 16 C2PA references that point back to the generation provenance. The C2PA standard is built into Adobe, Microsoft, and most major AI image tools. Platforms that support Content Credentials read this manifest automatically.
Then there's the XMP AI flag — a metadata tag formally called DigitalSourceType: trainedAlgorithmicMedia. This is a single field that explicitly tells any scanner "this came from an AI model trained on data." Add to that encoder fingerprints like Lavc (FFmpeg's encoder) or x264 SEI messages in video bitstreams — telltale signs that a file was processed by a computer, not a camera. An unprocessed AI export carries these fingerprints at the byte level.
Finally, a real phone photo has GPS coordinates, a capture timestamp, a real device Make/Model, and genuine EXIF lens data. An AI export has none of that — just generation timestamps and software metadata. The absence of authentic camera identity is itself a signal.
Why cropping, screenshotting, and re-uploading don't work
These are the most common workarounds people try, and they all fail at the metadata layer:
- Cropping removes visible content but leaves the metadata intact. The JUMBF atoms, XMP tags, and encoder fingerprints survive any pixel-level crop because they're stored in the file header, not the image pixels themselves. If your AI image has a visible watermark in the corner, cropping removes that visible mark — but the invisible detection layer remains, and that's what gets you flagged.
- Screenshotting re-encodes the image through your display pipeline, which strips some metadata — but not C2PA or XMP AI flags if they're embedded deeply enough. You also inherit your display's color profile and a new capture timestamp with no GPS data, which looks suspicious to forensic scanners.
- Re-uploading through a second platform (e.g., uploading to Twitter, then downloading and posting to Instagram) strips some metadata, but platforms like Instagram re-apply Content Credential scanning on upload. You also lose quality from each generation cycle.
None of these approaches touch the core detection signals: C2PA manifests, the trainedAlgorithmicMedia flag, or encoder fingerprints. They treat the symptom — visible pixels — not the disease.
How Calabi actually cleans an AI image
Calabi runs an automatic three-stage pipeline on every upload. There's no manual editing, no sliders, no region selection.
- Strip: Calabi removes every detection signal from the file — all 18 JUMBF atoms, all 16 C2PA references, the
DigitalSourceType: trainedAlgorithmicMediaXMP flag, Lavc and x264 encoder fingerprints, and generation software tags. A raw AI export that carries 144 metadata tags comes out with roughly 94 neutral structural tags — no AI identity. - Inject: Calabi writes authentic phone-capture identity into the file — a real device profile (iPhone 15 Pro, Pixel 8 Pro, Galaxy S24 Ultra), real Make/Model/Software fields, GPS coordinates, and a capture timestamp. The file now looks exactly like it came from that phone's camera.
- Verify: Before download, Calabi generates a forensic proof card — the same ExifTool scan that platforms and newsrooms use — showing exactly what was stripped and what was injected. You see the clean report before you see the clean file.
You upload, wait for the automatic pipeline, review the proof card, and download. One pass. No decisions required on your end.
FAQ
Does Calabi remove visible watermarks like the Midjourney sparkle or Sora logo?
Calabi does not edit pixels or remove visible marks. If your AI image has a visible watermark, cropping removes it — but cropping alone doesn't remove the metadata layer that survives the crop. Calabi handles the invisible detection signals (C2PA, XMP flags, encoder fingerprints) that cropping can't touch.
Will Instagram or TikTok still detect my AI image after cleaning?
No tool can guarantee a platform won't flag you — results vary by platform and source model. Calabi removes the documented metadata signals (C2PA Content Credentials, the trainedAlgorithmicMedia flag, encoder fingerprints) that automated scanners specifically look for. A cleaned file reads as an authentic phone recording at the file level.
What device profiles can Calabi inject?
Calabi supports real device profiles including iPhone 15 Pro, iPhone 16 Pro, Pixel 8 Pro, and Galaxy S24 Ultra, with matching software versions and encoder names for each. The injected profile is what shows up in the forensic proof card and what platform scanners see when the file is read.
Try Calabi free at calabilabs.com — 10 cleans, no card.