Calabi Labs · Guide · 2026-06-19
Compress midjourney image
If you're trying to compress a Midjourney image to get it past platform detection or make it look less like AI-generated, you're solving the wrong problem. File size has nothing to do with it. What actually gets you flagged is the invisible metadata layer — the cryptographic signatures, XMP tags, and encoder fingerprints baked into every AI export. Reducing the kilobytes won't touch those.
What Actually Gets Your Midjourney Image Flagged
When you download a Midjourney export, it's carrying a full invisible dossier that screams "AI-made" to every platform scanner. The three layers that matter:
C2PA / Content Credentials (JUMBF atoms) — Midjourney embeds a cryptographic manifest using the JUMBF (JPEG Universal Metadata Box Format) standard. This stores who made the content, what model generated it, and when. Instagram, TikTok, YouTube, and Reddit all parse this automatically. A single Midjourney export can contain 18 or more JUMBF atoms declaring its AI origin.
XMP metadata with DigitalSourceType: trainedAlgorithmicMedia — This specific XMP tag is the smoking gun. It was added to the EXIF spec specifically to flag AI-generated images. Platforms check for this field directly. If it's present, you're flagged before a human even sees your post.
Encoder fingerprints — Midjourney exports carry Lavc (FFmpeg) or x264 SEI NAL units in the video bitstream for images saved as video files. These are the same fingerprints forensic tools like ExifTool use to identify AI-generated content. They survive re-encoding and are platform-readable.
A raw Midjourney export typically carries 140+ metadata tags. Platforms don't need to look at your image — they just scan the file header.
Why Standard Compression Doesn't Help
Tools like TinyPNG, Squoosh, or HandBrake reduce file size by recompressing the pixel data. That's a lossy or lossless operation on the visible layer. It does nothing to the metadata layer sitting underneath. Your 5MB Midjourney export becomes a 400KB Midjourney export — still carrying every JUMBF atom, every XMP flag, every Lavc fingerprint.
The same applies to:
- Cropping — Removes visible elements like Midjourney's corner watermark, but the metadata layer survives intact. Cropped AI images still trigger platform detection.
- Screenshotting — A screen capture may strip some metadata, but platforms have gotten sophisticated about detecting resampled AI imagery through perceptual hashing regardless.
- Re-saving in Photoshop — Even "Save for Web" preserves XMP metadata unless you deliberately strip it, and most users don't know how.
The visible image and the invisible metadata are two separate systems. Compression and editing tools operate on pixels, not on the cryptographic manifest embedded in the file structure.
How to Actually Clean a Midjourney Image (Or Any AI Export)
Calabi runs a one-pass pipeline that treats the problem at the actual layer where platforms check:
- Strip the signals — The pipeline removes all JUMBF / C2PA atoms, all XMP AI flags including DigitalSourceType: trainedAlgorithmicMedia, generator/tool tags, and encoder fingerprints like Lavc and x264 SEI. An export's 140+ metadata tags get reduced to roughly 94 neutral structural tags.
- Inject authentic phone identity — Calabi writes Make, Model, Software version, GPS coordinates, capture timestamp, and a real-phone encoder name into the metadata. You can choose device profiles including iPhone 15/16 Pro, Pixel 8 Pro, and Galaxy S24 Ultra.
- Verify with a forensic proof card — Before download, you see the exact same ExifTool scan platforms use, showing what was stripped and what was injected. It's the same forensic standard newsrooms and platform trust-and-safety teams rely on.
What you're left with is a file that reads as a normal phone recording at the file level — not a resized Midjourney export, but a file with authentic capture metadata from a real device.
FAQ
Can't I just use a metadata stripper tool? Some EXIF strippers remove basic camera data, but they don't touch JUMBF / C2PA atoms or the DigitalSourceType: trainedAlgorithmicMedia XMP flag. You'd need a forensic-grade tool and know exactly which fields to target. Calabi handles the full signal layer in one pass.
Does this work for other AI generators besides Midjourney? Yes — the C2PA / Content Credentials standard, XMP AI flags, and encoder fingerprints are used across AI image and video tools. Calabi's pipeline handles Sora, Runway, DALL-E, Leonardo, and others the same way it handles Midjourney exports.
Will this guarantee my post won't get flagged? No tool can guarantee that — platform detection varies by source model, image content, and each platform's evolving models. What Calabi removes is the metadata and cryptographic layer that automated scanners check within seconds of upload. Results vary, but you're not fighting the system blind anymore.
Try Calabi free at calabilabs.com — 10 cleans, no card.