Calabi Labs · Guide · 2026-06-19
Ai image colorizer
When you run a black-and-white photo through an AI colorizer, the output file carries invisible signals—C2PA manifests, XMP tags, and encoder fingerprints—that platforms like Instagram and Reddit scan for automatically. Calabi doesn't colorize images; it strips those detection signals and injects authentic phone-capture identity so your colorized photo posts without triggering an AI flag. Here's exactly what goes on inside the file and what actually works to clean it.
What actually gets flagged inside a colorized image
AI colorizers—from free web tools to Photoshop's Neural Filters to dedicated apps—embed specific metadata that forensic scanners detect before your post ever goes live. The three layers that matter:
- C2PA / Content Credentials: The JPEG or HEIC file carries a JUMBF (JPEG Universal Metadata Box Format) manifest with cryptographic proof that the image was AI-generated. This is the same structure Adobe, Microsoft, and Google are building into Creative Cloud and Bing Image Creator. It's not visible in the image—it's stored in the file header as a chain of signed claims about the image's origin.
- XMP AI flags: The metadata namespace includes fields like
DigitalSourceType: trainedAlgorithmicMediaandGeneratorAItags. ExifTool reads these flags specifically. This is separate from the C2PA layer and sits in the standard XMP packet that any image editor reads. - Encoder fingerprints: Most AI colorization tools render output using FFmpeg under the hood, which stamps the video bitstream with Lavc (libavcodec) or x264 SEI NAL units. Image exports use library names like
LibRawordcrawin the metadata. Platforms maintain blocklists of these encoder identifiers because they're strong signals of automated generation.
Beyond these three layers, platforms also flag on the absence of expected phone-capture metadata: no GPS coordinates, no EXIF Make/Model, no capture timestamp in the standard fields, and no lens provenance. A photo with AI-generation signals and zero device identity is a red flag regardless of how good it looks.
Why cropping, screenshotting, and re-saving don't work
These are the moves most creators try first. None of them fully remove the metadata layer:
Cropping removes the visible pixel area but preserves the file header, XMP packet, and C2PA manifest intact. The JUMBF box sits in the metadata region outside the image bitmap—you're cutting pixels, not metadata. A forensic scan of a cropped AI image still shows the C2PA chain and XMP flags.
Screenshotting (taking a photo of your monitor) does strip metadata because you're capturing the displayed pixels through a screen recording or camera, which writes fresh device metadata. However, you're also introducing compression artifacts, losing resolution, and potentially adding screen moiré patterns. For print-quality restoration work this is never a acceptable workflow. And for video exports, there's no equivalent "screenshot" workaround.
Re-saving in Photoshop or Preview writes a new file but does not strip C2PA manifests or XMP AI flags by default. Both macOS Preview and Adobe Photoshop preserve XMP data unless you explicitly check "Reset Metadata" in the Save dialog—or in Photoshop's File Info panel. Even then, embedded C2PA manifests require explicit removal tooling. A casual re-save without these steps leaves the detection layer fully intact.
How to actually clean a colorized image for posting
Calabi runs a one-pass pipeline that strips the three detection layers and injects authentic phone-capture identity in a single step. Here's the process:
- Upload your AI colorized file—JPG, PNG, HEIC, or WebP. No format conversion needed from your end.
- Automatic strip: Calabi removes all JUMBF/C2PA atoms (18 down to 0), strips the
DigitalSourceType: trainedAlgorithmicMediaXMP flag and any generator tags, and clears encoder fingerprints like Lavc and x264 SEI references. A raw AI export's 144 metadata tags compress to approximately 94 neutral structural tags. - Automatic inject: Calabi writes fresh device identity—Make, Model, Software version, GPS coordinates, and capture timestamp—from a real phone profile. You can select from iPhone 15 Pro, Pixel 8 Pro, Galaxy S24 Ultra profiles. This gives the file the metadata fingerprint of a normal phone capture.
- Forensic proof card: Before download, you see exactly what was stripped and what was injected, verified with the same ExifTool command that newsrooms and platform trust-and-safety teams use to audit images.
- Download the cleaned file and post directly. The proof card is your audit trail if anyone questions the image's origin.
Note on visible watermarks: if your AI colorizer outputs an image with a visible corner logo or sparkle mark, cropping removes the visible element. Calabi does not erase visible marks pixel-by-pixel. What Calabi removes is the invisible detection and metadata layer—the C2PA manifest, XMP flags, and encoder fingerprints—that survives cropping and that platforms actually scan for.
FAQ
Will Instagram or Reddit know my colorized photo was AI-generated?
Platforms scan for C2PA manifests, XMP AI flags, and encoder fingerprints. A colorized image posted without cleaning carries these signals in the file header. After Calabi's pipeline, those signals are removed and replaced with phone-capture metadata. Results vary by platform and source model, but the metadata layer that automated scanners rely on is fully stripped.
I colorized my grandfather's WWII photo—why does metadata matter for that?
Genealogists and historians often don't realize that many free AI colorization web tools embed creator flags and AI-generation claims in the XMP metadata. If you're posting historical photos to a family group or heritage site that uses moderation tools, these flags can get your post flagged or removed as "AI-generated content" regardless of the photo's actual subject matter. Cleaning the metadata lets the image stand on its own content.
Can't I just use a metadata stripper tool instead?
Generic metadata strippers remove standard EXIF fields—camera make, lens, GPS—but they don't touch C2PA/JUMBF manifests or XMP AI namespace flags. ExifTool's -overwrite_original flag strips EXIF but leaves C2PA atoms and the trainedAlgorithmicMedia tag intact. You'd need to know exactly which fields to target and run multiple commands. Calabi handles the full detection-layer strip in one pass and then rebuilds the device identity, which is what platforms actually check.
Try Calabi free at calabilabs.com — 10 cleans, no card.