Calabi Labs · Guide · 2026-06-19
Ai image extender
AI image extenders — sometimes called outpainting tools — expand the borders of an image by generating new pixels that blend with the original scene. They're useful for fixing tight crops, adding context to a frame, or turning a square image into a 16:9 landscape. But here's what most creators discover the hard way: after you extend an image with Midjourney, DALL-E, Firefly, or Stable Diffusion, platforms like Instagram, TikTok, and Reddit often still detect it as AI-generated — even if the extension looks seamless to the human eye.
The reason isn't visible. It's buried in the file.
What Actually Gets Your Extended Image Flagged
When an AI model generates or extends an image, it doesn't just produce pixels. It embeds a forensic layer of metadata that tells platforms "this was made by artificial intelligence." That layer survives cropping, screenshotting, re-encoding, and re-uploading — because it's stored in the file's metadata structure, not in the visual content itself.
There are three main components platforms scan for:
- C2PA / Content Credentials — The most significant flag. C2PA embeds a cryptographic manifest (stored as JUMBF boxes) directly into the image file that cryptographically asserts the content was generated or modified by an AI model. This is the "Made with AI" badge you see on some Instagram posts. It includes a full provenance chain: what model was used, when, and by whom. Extended images from tools like Firefly, DALL-E, and newer Stable Diffusion exports carry this by default.
- XMP AI flags — Specifically the
DigitalSourceTypetag set totrainedAlgorithmicMedia. This is a plain-text metadata field — no cryptography required — that explicitly states the source was algorithmic. It's embedded in the XMP metadata block that ExifTool, Adobe, and platform scrapers can read in seconds. - Encoder fingerprints — AI generation pipelines use specific encoders — Lavc (FFmpeg's libavcodec), x264 SEI messages,梵谷 (and similar) — that leave distinct structural fingerprints in the file. A normal iPhone 16 Pro photo uses the HEIC encoder with Apple-specific atom structures. An AI-extended image often carries a different encoder signature, even after a "save as JPEG" round-trip.
A raw AI export can carry 140+ metadata tags. Many of those tags — the JUMBF manifest entries, the AI source tool name, the model version string, the generation timestamp — are invisible in any image viewer but trivially easy for a platform scanner to find.
Why Cropping, Screenshots, and Re-Saving Don't Fix It
These are the methods most creators try first, and they each fail for a specific reason:
Cropping removes pixels, not metadata. When you crop an AI-extended image, you're cutting the visible canvas — but the metadata block attached to the file is unchanged. Platforms that scan the file header, not the pixels, will still find the C2PA manifest and the XMP trainedAlgorithmicMedia flag. The visible evidence of extension is gone; the forensic evidence remains.
Screenshotting removes some metadata but adds new problems. Taking a screenshot of an AI image does strip the C2PA and XMP layer — because you're capturing the screen, not the file. However, you also introduce screen-capture artifacts (DPI and resolution anomalies, the UI elements at the edges, compressed color bands) that themselves can be detected as signs of screen capture, and platforms increasingly flag those too.
Re-saving in Photoshop or Preview strips some tags but not C2PA. Legacy metadata strippers can remove IPTC and some XMP fields, but C2PA manifests use cryptographic signing. Simply re-saving a JUMBF-embedded file in Photoshop doesn't remove the manifest — in fact, it may leave a broken or orphaned C2PA reference that forensic tools read as "AI-generated and then tampered with."
How to Actually Clean an AI-Extended Image
If you want an extended image to read as a normal phone capture at the file level — with no AI metadata, no Content Credentials, and the encoder identity of a real device — you need to handle all three layers: strip the forensic signals, inject authentic device identity, and verify the result.
That's exactly what Calabi does in one pass:
- Upload your AI-extended image file. No settings, no manual editing — one file at a time.
- The pipeline strips every forensic signal: 18 JUMBF/C2PA atoms reduced to 0, 16 C2PA references to 0, the
DigitalSourceType: trainedAlgorithmicMediaflag removed, generation tool names and model version strings purged, and the AI encoder fingerprint replaced with a clean structural profile. - The pipeline injects authentic phone-capture identity: a real device profile (iPhone 15 Pro, Pixel 8 Pro, Galaxy S24 Ultra, and others), with Make, Model, Software version, GPS coordinates, and capture timestamp — all written as standard EXIF and XMP fields.
- You see the forensic proof card — the same ExifTool scan platforms use — showing exactly what was stripped and what was injected, before you download.
- Download the cleaned file and post it as normal.
Note on visible watermarks: if your AI-extended image has a visible logo, sparkle icon, or corner badge (common on Firefly and Sora exports), cropping removes the visible mark. Calabi removes the invisible detection layer — the C2PA manifest and AI metadata — that survives cropping and is what platforms actually scan for.
FAQ
Does re-uploading to Instagram remove AI metadata?
No. Instagram and TikTok read the file's metadata on upload — they don't strip it. In fact, Instagram's AI detection often runs within seconds of upload and can trigger reduced reach or a "Made with AI" label even on images that look completely natural. Re-uploading a screenshot doesn't help because the platform scanner reads the uploaded file, not a reference to what you screenshotted.
Will Calabi change how my extended image looks?
No. Calabi does not edit, regenerate, or reconstruct any visible pixels. It works entirely on the invisible metadata layer — stripping forensic signals and writing device identity — so your extended image looks exactly the same when you download it. The change is in the file's digital fingerprint, not the visual content.
Which AI image extenders embed the most metadata?
Adobe Firefly embeds the heaviest C2PA/Content Credentials layer since it's built into Adobe's provenance pipeline. Midjourney exports carry generation tool tags and model identifiers in XMP. DALL-E and Stable Diffusion variants vary by version and export settings, but all newer versions (DALL-E 3, Stable Diffusion 3, Flux) have adopted C2PA or equivalent AI provenance standards. The safest approach is to assume any AI-generated or AI-extended image carries detectable metadata unless you've run it through a cleaner.
Try Calabi free at calabilabs.com — 10 cleans, no card.