Glossary

PNG text chunk: where prompts and XMP live in a PNG

A PNG text chunk is an ancillary chunk of type tEXt, zTXt or iTXt that stores a keyword and a text value inside a PNG file, used by AI tools for records such as parameters, prompt, workflow and invokeai_metadata and by editors for XMP under the keyword XML:com.adobe.xmp.

Code-drawn illustration for PNG text chunk

Illustration · PNG file

  1. Signature + IHDR
  2. iTXt XML:com.adobe.xmprecord
  3. tEXt parameters
  4. tEXt prompt / workflow
  5. tEXt invokeai_metadata
  6. eXIf
  7. caBX C2PA manifest
  8. IDAT pixels
  9. IEND

AI source type declared

The file labels itself as AI-made.

IPTC DigitalSourceType such as trainedAlgorithmicMedia is read from XMP and shown with its exact value.

Check an image →
AI source type declared: where this record sits inside a PNG file and what the isGenAI report says about it. The picture is a code-drawn illustration.

How the chunks work

The PNG specification defines tEXt for Latin-1 text, zTXt for zlib-compressed text and iTXt for UTF-8 text with optional compression and language tags. Each chunk has a length, a type, data, and a CRC-32 checksum. Keywords are 1 to 79 bytes. Because the chunks are ancillary, decoders may ignore them and encoders may drop them, which is why re-saving a PNG usually loses saved prompts.

AI tools chose text chunks because they are simple and human-readable; nothing about them is signed.

How isGenAI reads them

The checker and the browser provenance reader walk the chunk list, verify every CRC, inflate compressed text within a bounded size, reject duplicate keys and malformed international text, and stop at IEND. Only the keywords parameters, prompt, workflow and invokeai_metadata are read as records; other text is ignored so that a casual comment is never mistaken for a prompt. The PNG text records guide includes downloadable examples of each chunk type.

Removing or comparing chunks

The metadata remover rebuilds a PNG from only the essential chunks and lists which text chunks it removed. The image metadata compare tool shows which chunks differ between two copies.

Chunk order is flexible: text chunks may appear before or after the image data, as long as they come before IEND. Readers that stop early at the first image data can miss a prompt saved near the end, which is why isGenAI walks the whole file.

How to use this term when checking a file

  1. Upload the original file to the AI image detector or the tool named on the page.
  2. Find the finding that uses this term under “How we checked” and read the source field it names.
  3. Weigh it by kind: a signed record, a declared field, a saved prompt and a matching copy are different strengths of evidence.
  4. Remember that absence of a record is not evidence of a human author.

Read the PNG text records guide

Frequently asked questions

What is the difference between tEXt, zTXt and iTXt?

tEXt is plain Latin-1, zTXt is compressed, iTXt is UTF-8 with optional compression and language. AI tools use all three.

Which keywords do AI tools use?

parameters (AUTOMATIC1111), prompt and workflow (ComfyUI), invokeai_metadata (InvokeAI), Comment (NovelAI) and XML:com.adobe.xmp for XMP.

Why did my PNG lose its text chunks?

Ancillary chunks are dropped by most editors, converters and platforms on save. Keep the original.

Can a text chunk be faked?

Yes. It is unsigned. isGenAI reports it as a declaration of moderate strength and checks the CRC to catch corruption.

Sources

Definitions follow the published standards linked above; the isGenAI behaviour described is the deployed checker as of October 2026.