Reference

Image metadata glossary

Every term here is one we had to implement or deliberately decide not to. The definitions describe how these structures actually behave in files, not how a specification summarises them.

A

AI Info: Meta’s label for AI provenance; The notice Instagram, Facebook and Threads apply when an upload carries C2PA or XMP declaring AI involvement. It reflects what the file says about itself, not an analysis of the pixels.
APP segment: JPEG metadata block; A length-prefixed block in a JPEG, numbered APP0 to APP15. EXIF and XMP live in APP1, ICC profiles in APP2, C2PA in APP11 and Photoshop/IPTC data in APP13. They sit before the compressed image data, which is why they can be removed without re-encoding.
AVIF: AV1-based image format; A modern format built on the ISOBMFF box structure rather than JPEG segments or PNG chunks. Metadata lives in a meta box. Not yet supported here, because it needs its own parser and we would rather reject a format than convert it silently.

C

C2PA: Signed provenance manifest; The Coalition for Content Provenance and Authenticity standard. A cryptographically signed record of how a file was created and edited, written by OpenAI, Adobe, Google, Microsoft and some camera makers. The marker platforms read most reliably.
caBX: C2PA chunk in PNG; The PNG chunk type that carries a C2PA manifest, equivalent to the APP11 JUMBF box in JPEG.
Chunk: PNG structural unit; A PNG is a list of chunks, each with length, four-character type, data and a CRC. Critical chunks (IHDR, PLTE, IDAT, IEND) are required; ancillary chunks like tEXt are optional. Deleting a whole chunk needs no CRC recalculation, because every surviving chunk keeps its own.
Client-side processing: Work done in your browser; Computation that runs on your own machine rather than a server. For a metadata tool it is the difference between a promise not to store your files and never receiving them at all. Verifiable by disconnecting from the network.
Content Credentials: Consumer name for C2PA; Adobe’s branding for C2PA manifests. When a platform says it is showing content credentials, it is reading a C2PA manifest.

D

DigitalSourceType: IPTC field declaring origin; An IPTC field whose value states how content was made. The value trainedAlgorithmicMedia is a formal machine-readable declaration of AI generation, and platforms can read it without parsing a C2PA manifest.

E

Entropy-coded data: The actual JPEG picture; The compressed image data following the Start of Scan marker. Lossless metadata removal means copying these bytes untouched; re-encoding means decoding and recompressing them, which costs quality.
EXIF: Camera and device metadata; Exchangeable Image File Format. Camera make, model and serial number, lens, exposure settings, timestamps and GPS coordinates, stored as a TIFF-structured block. Also carries a Software field that often names the editor, including AI editors.

G

GPS coordinates: Where the photo was taken; Latitude and longitude stored in the EXIF GPS IFD, often with altitude and bearing. Accurate enough to identify a specific building. Note that AI-markers-only mode keeps EXIF, and therefore keeps GPS.

H

HEIC: Apple’s photo container; High Efficiency Image Container, the default on iPhone. Like AVIF it uses the ISOBMFF box structure. Not yet supported here.

I

ICC profile: Colour rendering data; Describes how colour values should be rendered. Technically metadata, but removing it visibly shifts colour, which is why it is kept by default and removing it is an explicit choice.
IDAT: PNG image data chunk; The chunks holding the compressed pixel data. Lossless stripping copies every IDAT byte unchanged; our verifier hashes them before and after to prove it.
IFD: Image File Directory; The TIFF structure EXIF is built from. IFD0 holds general fields, the Exif IFD holds camera specifics, and the GPS IFD holds location. Pointers between them are why a naive parser can miss GPS entirely.
IPTC: Publishing metadata; International Press Telecommunications Council fields: creator, credit, copyright, caption and usage terms. Legitimate and often worth preserving, which is why AI-only mode leaves attribution alone.

J

JUMBF: Box format carrying C2PA; JPEG Universal Metadata Box Format, the container C2PA manifests travel in. In JPEG it appears inside an APP11 segment.

L

Lossless: No change to pixel data; Here it has a strict meaning: the decoded pixel data is byte-for-byte identical before and after. Not "visually similar", not "high PSNR". We verify it on every run rather than asserting it.

M

Metadata: Data about the image; Everything in the file that is not the picture: EXIF, XMP, IPTC, C2PA, PNG text chunks and colour profiles. Usually a tiny fraction of file size and disproportionately revealing.

P

PNG text chunk: tEXt, zTXt and iTXt; Plain or compressed text stored in a PNG. Stable Diffusion writes full generation parameters here and ComfyUI writes an entire node graph as JSON, readable by opening the file in a text editor.
PSNR: Peak signal-to-noise ratio; A measure of difference between two images, in decibels. Higher is closer; infinite means identical. A tool claiming "same pixels" while measuring 39.78 dB is re-encoding, which is exactly what we measured on a leading competitor.

Q

Quantization table: JPEG compression settings; The DQT tables defining how aggressively each frequency band is compressed. If these change between input and output, the image was re-encoded, a reliable way to catch a tool that claims otherwise.

R

RIFF: WebP container format; The chunked container WebP uses. Removing an EXIF or XMP chunk also requires fixing the RIFF size field and the VP8X feature flags, or decoders may reject the file.

S

Seed: Reproducibility number; The value that makes a generation repeatable. Combined with prompt, sampler and model, a published seed lets anyone reproduce your image almost exactly.
SynthID: Google’s pixel watermark; An invisible watermark embedded in the pixels of Google-generated images, designed to survive screenshots, cropping, resizing and compression. Because it is not metadata, no metadata tool can remove it, including this one.

T

trainedAlgorithmicMedia: IPTC value meaning AI-generated; The standard DigitalSourceType value declaring that content was produced by an algorithm trained on data. A formal, machine-readable statement that an image is AI-generated.

V

VP8X: WebP extended header; The chunk advertising which optional features a WebP uses, including whether EXIF, XMP or ICC data is present. Removing those chunks without clearing the corresponding flags produces a file some decoders reject.

W

Web Worker: Background browser thread; A separate thread in the browser. Running the parser there keeps a 30-file batch from freezing the page, and keeps all processing on your machine.

X

XMP: Adobe’s XML metadata block; Extensible Metadata Platform. The usual home for AI generation parameters: Midjourney prompts and seeds, Adobe creator tools, and IPTC source-type declarations.