Calabi Labs · Guide · 2026-06-13
Here is the page, written as clean HTML with the required structure and length:
```html
To comment a picture on a TikTok video, tap the speech bubble icon on the right side of any video, tap the camera or gallery icon in the comment bar, select your image, and post. But if that image was generated with AI — a meme template, a edited still, a Midjourney export — TikTok's detection systems may read the invisible metadata layer and flag it before your comment even lands. Here's what actually happens at the file level, and what actually fixes it.
When you upload any image to TikTok — including in comments — the platform runs it through automated scanners that don't look at what the image shows. They look at what the file claims about itself. Three invisible signal layers trigger AI content labels:
C2PA / Content Credentials. Since January 2025, TikTok uses the C2PA (Coalition for Content Provenance and Authenticity) standard to detect AI-generated content. When you export an image from Midjourney, DALL-E, GPT Image, Flux, Stable Diffusion, or almost any other AI image tool, it embeds a JUMBF (JPEG Universal Metadata Box Format) block containing a cryptographic manifest — essentially a "made by AI" certificate. TikTok's scanner reads this block and applies an AI label automatically. A single AI export can carry 18 or more of these JUMBF atoms. TikTok counts them down to zero as part of its detection pipeline.
XMP AI metadata tags. Outside the JUMBF layer, AI generators write fields like DigitalSourceType: trainedAlgorithmicMedia into the image's XMP metadata. This is a C2PA-defined property that explicitly flags the image as produced by a trained AI model. TikTok's scanner flags this the same way it reads JUMBF data.
Encoder fingerprints. When AI image tools render an export, they use specific software libraries — Lavc (from FFmpeg), certain x264/x265 encoding signatures, or internal renderer names — that appear in the file's structural metadata. These encoder fingerprints are consistent across outputs from the same tool. Platforms use them as a secondary detection signal alongside C2PA. A raw AI export typically carries 144 metadata tags; a clean phone photo carries fewer than 100.
The most common workarounds people try — and why they fail against TikTok's detection layer:
TikTok's scanner operates on the file's metadata, not its visual content. That means the fix lives at the file level too.
Calabi strips the detection layer from AI-generated images and injects authentic phone-capture identity so the file looks like a normal photo taken on a real device. Here's the actual process:
DigitalSourceType: trainedAlgorithmicMedia XMP property and all related AI flag tags.Will Calabi remove the "AI generated" label TikTok already applied to my post?
Calabi cleans the file you upload. If TikTok has already processed and labeled a post, that label is tied to the platform's internal record — you'd need to delete and repost the content with a cleaned file. Calabi works on the upload stage, not the platform's post-record.
Does Calabi erase visible watermarks like the Sora sparkle or Midjourney logo?
No. Calabi doesn't edit pixels, so visible logos or watermarks remain. If a visible watermark is present, cropping it out is the standard fix — and Calabi removes the invisible detection metadata that survives that crop, which is what actually gets you flagged after you've already trimmed the image.
Can I use Calabi on video clips I want to post as TikTok comments too?
Yes. Calabi handles both images and video files. The same pipeline strips C2PA/JUMBF atoms, removes the trainedAlgorithmicMedia XMP flag, clears Lavc and x264 SEI encoder fingerprints from the video bitstream, and injects phone-capture metadata for video as well.
Try Calabi free at calabilabs.com — 10 cleans, no card.
```