Back to Guides
Metadata Standards9 min read

AI Image Metadata Explained: C2PA, XMP, EXIF and PNG Data

Published: Last updated:

A comprehensive overview of every key metadata architecture utilized in digital images and how they interact in the modern generative AI era.

Digital images are complex container files that encapsulate both visual raster arrays and multi-layered diagnostic metadata. Understanding how these systems differ is essential for privacy, technical auditing, and content authenticity.

1. C2PA (Content Credentials)

The newest metadata standard, C2PA packages cryptographic assertions signed by digital certificates. In JPEG files, it lives in APP11 markers; in PNG files, it lives in a caBX chunk. Unlike legacy tags that can be edited with basic hex editors without notice, C2PA binds a cryptographic hash of the image pixels to the manifest. If any pixels are altered, the signature breaks.

2. XMP (Extensible Metadata Platform)

XMP is an XML data packet stored inside image containers (APP1 in JPEG, iTXt chunks in PNG). It is designed to describe document properties, copyright licenses, and complex editing histories using Dublin Core and custom namespaces. Generative tools like Adobe Firefly store prompt and model information here.

3. EXIF (Exchangeable Image File Format)

Established in 1995 for digital cameras, EXIF records hardware parameters: camera body manufacturer, lens serial, shutter speed, aperture, and capture timestamp. It also includes the Software tag, which indicates what program last saved or edited the file.

4. PNG Text Chunks (tEXt, zTXt, iTXt)

The PNG specification allows key-value text chunks. The open-source AI community (Stable Diffusion, Automatic1111, ComfyUI) standardized on writing 'parameters', 'prompt', and 'workflow' chunks containing complete JSON graphs and sampler recipes directly into these blocks.

  • tEXt: Key-value pairs in Latin-1 text, for example parameters holding the prompt, steps, sampler, CFG scale and seed (Automatic1111 and Forge).
  • zTXt: The same idea, compressed with zlib.
  • iTXt: UTF-8 text, optionally compressed. ComfyUI stores its full node graph here as prompt and workflow JSON, which is why dragging a ComfyUI PNG back into ComfyUI can reload the entire workflow.

5. How Generators Store Parameters in JPEG and WebP

JPEG and WebP have no PNG-style text chunks, so generators use the standard metadata containers instead:

  • EXIF UserComment: The UserComment tag (ID 0x9286) often holds the same parameter text a PNG would carry in parameters.
  • XMP packets: XML that names the creating tool and, for some services, describes the generation.
  • C2PA: Signed Content Credentials in APP11 (JPEG) or a C2PA chunk (WebP).

Because all of these sit around the image data rather than inside it, a metadata remover can take them out without re-encoding the picture.

Audit Your Image Metadata

Inspect C2PA, XMP, EXIF, and PNG chunks with our in-browser checker.

Open AI Image Checker

Which container is in which format, and who reads it?

AI metadata containers by image format
ContainerJPEGPNGWebPCommonly read by
C2PAAPP11 segmentcaBX chunkC2PA chunkPlatforms, C2PA viewers
XMP (incl. digital source type)APP1 segmentiTXt chunkXMP chunkPlatforms, editors
EXIFAPP1 segmenteXIf chunkEXIF chunkViewers, editors
Text (prompts, workflows)Comment / UserCommenttEXt, zTXt, iTXtRareAI tools, viewers

Frequently asked questions

Which AI metadata triggers platform labels?

Meta has said it relies on C2PA and IPTC metadata. Prompts in PNG text are not part of those standards.

Is the PNG C2PA chunk called caDX?

No. The C2PA specification uses a chunk named caBX for PNG.

Can I keep my prompt but remove the AI label?

Yes, with a custom selection or the Metadata Editor, but platform and legal disclosure rules may still apply.

Sources

  1. C2PA Technical Specification 2.3 (C2PA)
  2. Digital Source Type vocabulary (IPTC)
  3. Portable Network Graphics (PNG) Specification, Third Edition (W3C)
  4. Meta changes its label from “Made with AI” to “AI info” (TechCrunch, July 1, 2024)

More guides on this topic