Calabi Labs · Guide · 2026-06-14

Remove text from image

Can You Actually Remove Text from an AI Image — and What Gets You Flagged?

When you export an image from Midjourney, Sora, DALL-E, or any other AI generator, the visible text you typed is part of the pixel output — it's baked into the image and cannot be removed by any metadata tool. However, what actually gets your upload flagged, labeled "Made with AI," or suppressed by platforms like Instagram and TikTok is an invisible detection layer embedded in the file itself. That layer — the metadata, cryptographic signatures, and encoder fingerprints — can be stripped, and that's exactly what Calabi handles.

What Actually Gets Your AI Image Flagged

Platforms don't read the pixels to decide an image is AI-generated. They scan the invisible metadata attached to your file. Here's what's actually in a typical AI export:

C2PA / Content Credentials: A cryptographic manifest stored as JUMBF atoms. This is the "Made With AI" manifest — it lists the model that generated the image, the software version, and a cryptographic signature trail. Instagram, TikTok, YouTube, and Reddit all read this. A fresh AI export can contain 18 or more C2PA atoms declaring its origin.
XMP AI flags: Tags like DigitalSourceType: trainedAlgorithmicMedia sit directly in the image header. This is a machine-readable declaration that the file came from a trained AI model. It's not subtle — it's an explicit label.
Generator and encoder fingerprints: Export settings like Lavc (FFmpeg's encoder library) or x264 SEI NAL units in the video equivalent leave recognizable encoder signatures. These tell forensic tools "this came from a render farm, not a phone."
Missing capture metadata: Real phone photos include Make, Model, Software version, GPS coordinates, and a capture timestamp. AI exports typically have none of this. That absence itself is a signal platforms flag.

A raw AI export can carry 144 metadata tags. Platforms don't need to look at the pixels — they just read the header.

Why Cropping, Screenshotting, and Re-Uploading Don't Work

Most people try the obvious fixes first. Here's why each fails:

Cropping the image: Removes the visible pixels you don't want. Does nothing to the metadata layer — the C2PA manifest, XMP flags, and encoder fingerprints survive a crop because they're stored in the file header, not the pixel grid. Crop the image 90% smaller and the detection signals are still there.
Screenshotting: Taking a screenshot of an AI image and posting that is a workaround for the visible output, but screenshots carry their own problems: screen capture adds display metadata, re-encodes through a display pipeline, and still may carry the original file's metadata if the capture is of a browser rendering. More importantly, screenshot quality is ruined for print, ads, or any professional use.
Re-exporting from a photo editor: Opening the AI image in Photoshop and saving it as a new JPEG strips some metadata, but photo editors preserve C2PA manifests if the format supports it, and they don't inject authentic device identity. You're removing some signals but not the critical ones — and you're adding no legitimate phone identity to replace what was stripped.
Re-uploading through a second platform: Downloading from Midjourney and re-uploading to Imgur before posting to Reddit doesn't remove C2PA — platforms can re-read the original generation metadata even after a single re-encoding pass. The detection signals are persistent.

How to Actually Clean an AI Image File for Upload

Calabi runs a one-pass pipeline that handles the detection layer specifically. Here's what it does:

Strip detection signals: Calabi removes every C2PA / Content Credentials manifest, reduces JUMBF atoms from 18+ to zero, strips the DigitalSourceType: trainedAlgorithmicMedia XMP flag, removes generator and tool tags, and eliminates encoder fingerprints like Lavc and x264 SEI entries. After the pipeline, a raw AI export's 144 metadata tags are reduced to roughly 94 neutral structural tags — no AI declaration, no generation trail.
Inject authentic phone identity: Calabi writes real device metadata in its place: iPhone 15 Pro, Pixel 8 Pro, or Galaxy S24 Ultra profiles including Make, Model, Software version, GPS coordinates, and a capture timestamp. This is what a genuine phone photo looks like in the file header.
Return a forensic proof card: Before you download, Calabi shows you the ExifTool scan — the same forensic tool newsrooms and platform trust-and-safety teams use. You'll see exactly what was stripped (C2PA atoms, AI flags, encoder entries) and what was injected (phone device profile). This is your verification that the file now reads as a phone capture, not an AI export.

FAQ

I want to remove visible text from my AI image — does Calabi do that?

No. Calabi does not edit pixels, remove objects, erase logos, or reconstruct any region of an image. If you have a visible watermark, logo, or text overlay in the corner of your image, cropping removes it cleanly — Calabi handles the invisible metadata layer that survives cropping and that cropping alone cannot touch.

Will screenshotting bypass AI detection?

Partially and unreliably. Screenshotting re-encodes through a display pipeline which can disrupt some metadata signals, but platform forensic tools are getting more sophisticated about detecting screen-capture artifacts. The quality loss from screenshotting makes it unsuitable for anything beyond social sharing. Calabi cleans the file at the metadata level while preserving the original image quality — and it re-injects the phone identity that screenshotting strips out.

What happens after a platform re-encodes my upload?

Most platforms re-encode your file during upload (Instagram compresses JPEG to ~80% quality, TikTok transcodes to H.264). This can disrupt some metadata — but C2PA manifests are designed to survive re-encoding, and encoder fingerprints persist through many transcodes. More importantly, the phone identity Calabi injects survives standard platform re-encoding, reinforcing the "this was a phone capture" narrative that platforms read during the scan window.

Clean the File, Not Just the Pixels

If you're posting AI-generated visuals on social media, the platforms aren't scanning for visual artifacts — they're reading the file header. A visually clean image can still carry 18 C2PA atoms and a trainedAlgorithmicMedia flag that reads "AI" in two seconds of automated scanning. Calabi strips that layer, injects authentic phone identity, and shows you the proof before you download.

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

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

Try free →