XMP vs EXIF vs IPTC vs C2PA: A Visual Comparison of Image Metadata

These four formats all live inside an image file. They don't replace each other; they layer. Each has its own era, its own standards body, and its own typical contents. Knowing which is which makes the difference between an "AI metadata scan" you can cite and one that misses half the trail.

EXIF (1994)

• Purpose: Camera settings. Built by digital cameras, for digital cameras.
• Format: TIFF tags inside an APP1 JPEG marker. Binary, fixed schema.
• Common contents: ISO, shutter, aperture, focal length, GPS, camera serial.
• Typical size: 1–4 KB on a DSLR shot.
• Limit: Designed for hardware, not software. EXIF has a Software field, but it's free-text — any tool can write anything, and nothing is signed.

IPTC (1970s, modernized early 2000s)

• Purpose: Newsroom photo credits and rights.
• Format: Binary IIM (Information Interchange Model) inside an APP13 marker.
• Common contents: Photographer, copyright, headline, caption, location keywords.
• Typical size: 0.5–2 KB.
• Limit: Static schema, hard to extend. Largely superseded by XMP for new fields, but Adobe and stock-photo workflows still write it.

XMP (2001)

• Purpose: A general-purpose, extensible metadata format that both Adobe apps and modern AI tools can write.
• Format: RDF/XML, often embedded inside an APP1 segment (alongside EXIF) or a PNG iTXt chunk.
• Common contents: Edit history, ICC profiles, AI-generator provenance (parameters, dalle, openai), software version, keywords.
• Typical size: 0.5–50 KB depending on what's stamped.
• Limit: Plain XML inside a binary container is awkward, and like EXIF it's unauthenticated — anyone can write any XMP block.

C2PA (2021)

• Purpose: Tamper-evident, cryptographically signed content provenance.
• Format: JUMBF (ISOBMFF) boxes, typically as APP11 in JPEG and as a chunk in PNG.
• Common contents: Signed manifest — issuer certificate, asserted actions (c2pa.created, c2pa.actions, stitching, trainedAlgorithmicMedia), and a hash chain linking every prior version.
• Typical size: 3–10 KB per assertion; grows with edit count.
• Limit: Adoption is partial; most editing tools strip credentials when they resave; almost no social platform renders them.

At a glance

| Format | Era | Authenticated | Common AI use | Stripped on social upload | | ------ | ----- | ------------- | --------------------------------- | ------------------------- | | EXIF | 1994 | No | Camera, GPS, sometimes Software | Usually stripped | | IPTC | 1970s | No | Caption, credit, copyright | Often stripped | | XMP | 2001 | No | AI parameters, edit history | Often stripped | | C2PA | 2021 | Yes (X.509) | Provenance, signed actions | Always stripped |

How they layer inside a single file

A modern JPEG from Adobe Lightroom might carry all four:

• EXIF in APP1 (camera settings from the original shot)
• XMP in APP1 (edit history, AI claims, color treatment)
• IPTC in APP13 (caption, photographer credit)
• C2PA in APP11 (signed provenance)

A scanner that walks the file segment by segment picks the ones it knows and reports them. AI Metadata Ninja's tool reports all four in one pass — useful, because a buyer running the same file through a competing scan might only see C2PA and miss an XMP dalle field from the same image.

Cleaning strategy by format

For most creators the question is "clean AI-related fields, keep the rest". A byte-level cleaner applies this rule per format:

• EXIF — keep camera settings if you want them, strip GPS for privacy, strip AI-software tags.
• IPTC — keep credits and copyright. There's almost never anything AI-trailing here.
• XMP — strip parameters, dalle, openai fields, AI-software tags. Keep Adobe editing history if useful.
• C2PA — strip by default. Keep it only if you want third parties to verify the chain.

A byte-level cleaner walks the file, removes only the AI-trailing segments, and leaves the rest. Pixel data and ICC profile are untouched.

The takeaway

These four formats are not interchangeable. EXIF is hardware, IPTC is newsroom credit, XMP is general-purpose, C2PA is signed provenance. Knowing which is which tells you where to look for AI generators (XMP, C2PA), what to keep (IPTC credits), and what to strip when you publish.