Calabi Labs · Guide · 2026-06-13
Remove glare from photo
Output:
```html
How to Actually "Remove Glare" from an AI-Generated Photo Before Posting
When people search "remove glare from photo," they usually mean the visible light reflection on glass, water, or a screen. But if you're posting an AI-generated image and worrying about getting flagged, labeled, or suppressed — that's a different kind of glare entirely. The "glare" platforms actually see isn't on the pixels; it's embedded in the file's metadata and signal structure. Calabi strips those invisible flags in one pass so your AI image reads as a normal phone recording at the file level.
What Actually Gets Your AI Image Flagged
Platforms like Instagram, TikTok, YouTube, and Reddit don't flag images by looking at them — they scan the file's metadata layer first. Three specific signal families trigger detection:
C2PA / Content Credentials (JUMBF atoms). When you export from Midjourney, DALL-E, Sora, Flux, or Stable Diffusion, the image carries a cryptographic manifest baked into the file as JUMBF (JPEG Universal Metadata Box Format) atoms. This manifest — part of the C2PA standard — explicitly lists the generating tool, model version, and a digital signature stating "this was AI-created." Instagram and Facebook read this on upload and attach an "AI Info" label automatically. A single Midjourney export can contain 18 or more of these JUMBF atoms.
XMP metadata flags. The DigitalSourceType XMP property is set to trainedAlgorithmicMedia in most AI export pipelines. Combined with Generator, Software, and CreatorTool tags, this tells any forensic reader "not a camera capture." Raw AI exports routinely carry 140+ metadata tags — roughly 50 of which are explicit AI fingerprints.
Encoder fingerprints. Specific video and image encoders leave detectable traces. The Lavc (FFmpeg/libavcodec) and x264 SEI (H.264 supplemental enhancement information) markers in exported files are recognized encoder signatures. A photo that claims to be an iPhone capture but carries Lavc metadata in its bitstream fails the basic authenticity check.
Platforms also look at the absence of expected signals: no GPS coordinates, no capture timestamp, no camera Make/Model, no lens metadata. A file with all the AI fingerprints but none of the phone-capture identity looks suspicious by omission.
Why the Obvious Fixes Don't Work
Screenshotting — Taking a screen capture of an AI image removes some metadata, but the underlying pixel data still carries the generation artifacts (oversaturated regions, unnatural texture patterns) that perceptual hash detectors are trained on. You also lose resolution and introduce compression artifacts.
Cropping — Cropping removes the visible canvas, but the metadata block and C2PA manifest survive the crop unchanged. If you export a 512×512 SD image and crop it to 400×400, the JUMBF atoms and XMP tags are still embedded in the new file. This is the most common misconception: cropping makes the visible watermark disappear but does nothing to the invisible detection layer.
Re-encoding / saving as a new file — Re-encoding disrupts some perceptual hashes but doesn't remove C2PA manifests or XMP metadata unless you use a specialized stripping tool. Many AI export tools re-embed the C2PA block on save anyway. And re-encoding degrades image quality with each pass.
None of these approaches touch the core signals that automated detection systems are actually scanning.
How to Actually Clean an AI Image Before Posting
Calabi runs a three-stage pipeline on upload — no manual editing, no quality loss:
- Strip the AI signals. Calabi removes every C2PA / JUMBF atom (verified down to 0), the
DigitalSourceType: trainedAlgorithmicMediaXMP flag, all generator/tool tags, and encoder fingerprints like Lavc and x264 SEI. A raw AI export with 144 metadata tags is reduced to roughly 94 neutral structural tags. - Inject authentic phone-capture identity. Calabi writes Make, Model, Software version, GPS coordinates, capture timestamp, and a real-phone encoder name. You can choose a device profile — iPhone 15 Pro, Pixel 8 Pro, Galaxy S24 Ultra — and the metadata reflects exactly what that phone would have written.
- Verify with a forensic proof card. Before download, Calabi returns the same ExifTool scan that platforms use — showing exactly what was stripped and what was injected. You see the before/after so nothing is guesswork.
The result is a file that looks, to automated scanners, like it was captured on a real phone — because at the metadata level, it is.
Frequently Asked Questions
Does this work if my AI image has a visible watermark or logo in the corner?
Calabi does not edit pixels, so it cannot erase a visible logo or text overlay. Cropping the image removes the visible mark, and Calabi then cleans the metadata layer that survives the crop — the C2PA manifest and AI flags that would otherwise get detected even after cropping. This is the combination most creators use.
Will platforms still detect my image as AI-generated after cleaning?
Calabi removes the metadata and encoder signals that automated systems scan for. Results vary by platform and by the source model — some detection also relies on perceptual hashes that analyze pixel patterns, which re-encoding partially disrupts. Calabi handles the metadata layer completely; the perceptual layer depends on the specific platform and image content.
Does Calabi change how my image looks?
No. Calabi works entirely at the file/metadata level. Zero pixels are modified, painted, or reconstructed. The image appearance is identical to what you uploaded.
Try Calabi free at calabilabs.com — 10 cleans, no card.
```