Skip to main content
Architecture & Principles

Built for Verifiable Privacy & Bit-Identical Media Quality

Strip AI Label was engineered to solve two widespread problems for photographers, designers, and video editors: false-positive social platform “Made with AI” flags on retouched camera work, and invasive cloud metadata strippers that re-compress your files on remote servers.

1. Surgical Binary Splicing Instead of Canvas Re-encoding

Most web-based metadata cleaners draw your image onto an HTML5 <canvas> and re-export it via toBlob(). That shortcut introduces generational JPEG/WebP compression artifacts, discards 10-bit/16-bit precision, strips Display P3 ICC profiles, and fails completely on MP4 or MOV video files.

By contrast, Strip AI Label parses container structures directly at the byte level—walking JPEG markers (`APP11`, `APP1`, `APP13`, `COM`), PNG chunks (`caBX`, `tEXt`, `iTXt`, `zTXt`, `eXIf`), WebP RIFF subchunks (`C2PA`, `XMP `, `EXIF`), and ISO-BMFF boxes (`uuid`, `moov.udta`, `stco`/`co64`). Only ancillary metadata segments are excised; your compressed DCT scan, `IDAT` deflate stream, or `mdat` video track remains 100% bit-identical.

“C2PA manifests are stored inside format-specific JUMBF superboxes—APP11 in JPEG, caBX in PNG, and uuid boxes in ISO-BMFF—separate from the compressed media bitstream.”

2. Honest Technical Boundaries (Container Tags vs. Pixel Signals)

We believe forensic tools should state clearly what they do and do not modify. Container-level provenance (`C2PA Content Credentials`, `IPTC DigitalSourceType`, `XMP CreatorTool`, prompt chunks, and `EXIF/GPS`) lives in distinct file boxes and is stripped 100% losslessly. Frequency-domain steganographic signals such as Google SynthID alter pixel values directly; because our engine never alters or inpaints pixels, pixel-embedded signals remain untouched.