AI image detector by file format

PNG AI image detector: every record a PNG can hold

PNG is the richest format for AI evidence because it stores saved prompts in tEXt, zTXt and iTXt text chunks, EXIF in an eXIf chunk, XMP in an iTXt chunk and C2PA manifests in a JUMBF chunk, and isGenAI reads every one of those from an original PNG.

Code-drawn illustration for PNG

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.

Where PNG keeps provenance

The PNG specification defines textual chunks, tEXt, zTXt and iTXt, each with a keyword and a value. AI tools use them freely: AUTOMATIC1111 writes a parameters chunk, ComfyUI writes prompt and workflow chunks, InvokeAI writes invokeai_metadata, and XMP travels in an iTXt chunk keyed XML:com.adobe.xmp. EXIF has its own eXIf chunk. C2PA Content Credentials are stored as a JUMBF box inside a dedicated chunk, as described in the C2PA specification.

All of these are ancillary chunks: a decoder can drop them without changing a single pixel, which is exactly what happens on most re-saves.

What isGenAI reads from a PNG

The AI image detector accepts PNG up to 20 MiB and validates the C2PA manifest, reads the IPTC DigitalSourceType from XMP, quotes AI statements, and reads the parameters, prompt and invokeai_metadata chunks after verifying each chunk’s CRC. The image prompt extractor and ComfyUI workflow viewer read the same chunks in detail. The EXIF viewer lists the readable fields in your browser, and the metadata remover rebuilds a PNG with only the essential chunks and tells you which ones it removed.

Animated PNG, damaged CRCs and oversized text records are refused rather than partially read.

Converting a PNG

Converting to JPEG or WebP with the image converter produces a new file written by your browser without the text chunks and without the manifest. That is fine for sharing and wrong for evidence. Keep the original PNG for checking, and use the converter on a copy.

To prove that two PNGs differ only in metadata, run the image pixel compare tool: identical pixels, different records.

How to check a file in this format

  1. Keep the original bytes. Do not open and re-save the file in an editor before checking it.
  2. Use the tool that accepts the format: the AI image detector for PNG, JPEG, WebP, AVIF and TIFF; the C2PA checker for those plus HEIC, PDF, MP4 and MOV.
  3. Read each finding with its source field. The report names the container record it came from.
  4. If you must convert, convert a copy and compare both files with the image metadata compare tool so you know what the conversion dropped.

Check an image now

Frequently asked questions

Does a PNG keep the AI prompt?

It can. AUTOMATIC1111, ComfyUI and InvokeAI write their records into PNG text chunks, and isGenAI reads them from the original file. A re-saved or screenshotted PNG has lost them.

Can a PNG carry C2PA Content Credentials?

Yes. The C2PA specification defines a JUMBF chunk for PNG, and both the AI image detector and the C2PA checker validate it.

Why does my PNG show no metadata?

Most editors, converters and social platforms drop ancillary chunks on save. Check the original export; the image metadata compare tool shows what a copy lost.

Can I remove the prompt from a PNG?

Yes. The metadata remover rebuilds the PNG with only the essential chunks in your browser and lists exactly which chunks it removed.

Sources

Container facts come from the published specifications. Tool support describes isGenAI as deployed in October 2026.