Calabi Labs · Guide · 2026-06-14
Ai blemish remover
A blemish remover for AI-generated images doesn't work the way most people expect—cropping, screenshotting, or re-uploading won't strip the invisible detection signals baked into your file's metadata. What actually works is removing the hidden AI signatures at the file level: the C2PA manifests, XMP tags, and encoder fingerprints that platforms like Instagram and TikTok scan automatically. Calabi handles this in one pass, then gives you a forensic proof card showing exactly what was removed and what phone identity was injected in its place.
What actually gets flagged
When you export an image from Midjourney, DALL-E 3, Flux, or any AI image tool, the file doesn't just contain pixels. It carries an invisible metadata layer that forensic tools read to determine whether content was machine-generated. This layer is what gets you flagged, not the visual content itself.
The primary signal is C2PA / Content Credentials—a cryptographic manifest stored as JUMBF atoms inside your file. This manifest lists the AI model used, generation parameters, and a cryptographic signature. ExifTool, the same tool newsrooms and platforms use, reads these manifests to flag content. A raw AI export can contain 18 or more JUMBF atoms, plus 16 C2PA references that trace back to the generation pipeline.
Beyond C2PA, there's the XMP AI flag: the DigitalSourceType field gets set to trainedAlgorithmicMedia—a direct declaration that the image came from an AI model. Generator-specific tags identify the tool. Encoder fingerprints embedded during export—think Lavc (libavcodec) or x264 SEI NAL units in video—mark the file as machine-generated at the codec level. Missing fields that a real phone capture would always have—GPS coordinates, precise capture timestamp, Make/Model of the device—create a negative signal that automated scanners also flag.
None of this is visible in the image. But platforms scan for it automatically, often within seconds of upload.
Why the obvious fixes fail
Cropping removes the visible composition but leaves the metadata intact. The C2PA manifest, XMP tags, and encoder fingerprints survive because they're embedded in the file structure, not the pixel area. A crop tool doesn't touch them.
Screenshotting re-encodes the image through your display driver and OS, which strips some metadata—but not C2PA manifests that were burned into the file at generation. You also introduce screen artifacts and lose resolution. The remaining metadata still often flags as AI-generated.
Re-uploading through a platform like Twitter or Discord re-encodes again, which may strip some metadata, but platforms also run their own analysis on the file structure and codec fingerprints. You're trading one problem for another without solving the underlying detection layer.
The fundamental issue: these methods treat the visible image as the problem. The visible image isn't what's being scanned. It's the invisible metadata layer underneath.
How Calabi actually cleans AI-generated images
Calabi works on the invisible layer. Upload your AI-generated image or video and a single automatic pipeline runs three stages:
Strip: Calabi removes every detection signal in one pass. All JUMBF atoms and C2PA references are stripped to zero. The DigitalSourceType: trainedAlgorithmicMedia XMP flag is deleted. Generator tool tags and encoder fingerprints—Lavc, x264 SEI, and similar codec-level markers—are removed from the file structure. A raw AI export's 144 metadata tags are reduced to roughly 94 neutral structural tags.
Inject: Calabi writes authentic phone-capture identity into the file. It embeds Make, Model, Software version, GPS coordinates matching a real location, a capture timestamp, and a real-phone encoder name. Device profiles include iPhone 15 Pro, Pixel 8 Pro, and Galaxy S24 Ultra. The file now looks, at the metadata level, exactly like something captured on an actual phone.
Verify: Before you download, Calabi generates a forensic proof card—the same ExifTool scan that platforms use. You see exactly what was stripped (18 JUMBF atoms → 0, 16 C2PA references → 0, the trainedAlgorithmicMedia flag) and what was injected (phone device profile, GPS, timestamp, encoder). This is what gets you cleared, and this is what you can show if challenged.
If your AI-generated image has a visible watermark—a corner logo, a sparkle marker, a model name burned into the frame—cropping removes the visible mark. What Calabi removes is the detection layer that survives cropping: the metadata and encoder signatures that persist even after you've cut out the visible watermark. Both steps together give you a clean result.
Step-by-step: upload your AI-generated image or video to calabilabs.com. The automatic pipeline strips every detection signal, injects authentic phone identity, and generates your forensic proof card. Review what was removed and what was injected. Download the cleaned file. Post it.
If you're a creator using AI image tools and posting to platforms that scan uploads automatically, this is the layer that matters.
Frequently asked questions
Does Calabi edit pixels or remove objects from my image?
No. Calabi doesn't touch pixels, doesn't use inpainting or content-aware fill, and doesn't remove or alter any visible element. It works exclusively on the invisible metadata layer—stripping AI detection signals and injecting phone-capture identity. If you need to remove a visible watermark or object, use a crop tool or a photo editor separately; Calabi handles the metadata layer those methods leave behind.
Will this guarantee my post won't get flagged?
No tool can guarantee a platform won't flag content. Platform detection systems vary and evolve. What Calabi removes is the metadata and encoder fingerprint layer—the specific signals that ExifTool and similar forensic tools detect. After cleaning, your file looks like a phone capture at the metadata level, which addresses the primary automated scanning vector on most major platforms.
What does "forensic proof card" mean?
The proof card is an ExifTool readout of your file's metadata—before and after cleaning. It shows exactly which fields were stripped (C2PA manifests, DigitalSourceType flags, encoder fingerprints) and which were injected (Make, Model, GPS, timestamp, encoder name). This is the same forensic scan that platforms and newsrooms use. You can review it before downloading to confirm the clean.
Try Calabi free at calabilabs.com — 10 cleans, no card.