AIMETADATAFILE STRUCTURES / 2026READ A FILE ↓
TECHNICAL GUIDE / FILE ANATOMY

What AI-generated
files can reveal.

Pixels, container structures, ordinary metadata, generation workflows and signed provenance are different layers. Each survives different edits and answers a different question.

Pixels are not metadata

The rendered image is one layer. Its container organises chunks, segments or boxes around those pixels. Metadata describes capture, creation, rights, software or generation settings. A field is informative, but it is not automatically trustworthy unless a cryptographic system binds it to the file.

What generation settings can be stored?

Common fields include a positive prompt, negative prompt, seed, sampler, scheduler, steps, CFG scale, checkpoint or model hash, dimensions, denoise strength, LoRA and VAE names. A seed guides randomness; it does not reproduce the same image without the same model, software, workflow and deterministic operations.

PNG text chunks

PNG is a sequence of typed chunks. Applications can store text in tEXt, compressed zTXt or international iTXt chunks. ExifTool documents these locations. Keys are application conventions rather than universal AI fields.

Source: ExifTool PNG Tags

ComfyUI workflow and prompt JSON

ComfyUI PNG files can carry a full editable workflow and a separate prompt graph. The project documents that generated images can be loaded back into the interface to restore a workflow. Current writers use keys such as prompt and additional PNG information. A workflow graph and an API prompt graph are related, but not identical representations.

AUTOMATIC1111 parameters

AUTOMATIC1111 commonly writes a human-readable parameters block into PNG text. Its PNG Info interface can read the saved prompt and settings. Export options, conversion and metadata stripping determine whether those fields survive.

EXIF, XMP and IPTC

EXIF covers imaging and capture fields such as dimensions, timestamps, software and camera settings. XMP is an extensible XML/RDF data model defined by ISO 16684-1 and used by creative tools. The IPTC Photo Metadata Standard defines descriptive, rights and editorial fields. These values are usually editable.

Sources: Adobe XMP specifications · IPTC Photo Metadata Standard

JPEG, WebP and ISO BMFF

JPEG and WebP typically place EXIF or XMP in format-specific segments or chunks rather than PNG text. AVIF, HEIC, MP4 and MOV use ISO Base Media File Format boxes. Conversion between formats may rewrite or discard metadata even when the visible image looks unchanged.

Audio and video containers

Audio and video have container-level tags and stream-level technical fields. A title or artist tag belongs to a different layer than codec, frame rate, dimensions or sample rate. ffprobe exposes format tags and stream data separately, which is why this site reports them separately.

Source: FFmpeg ffprobe documentation

C2PA is not ordinary metadata

A C2PA manifest can contain actions, ingredients, thumbnails and descriptive assertions, but a signature and content binding make later changes detectable. It may be embedded or externally referenced. The C2PA specification treats provenance as a validation problem, not a free-form text field.

Source: C2PA Technical Specification 2.4

Which editors and readers can open these records?

ToolBest used forScope
ComfyUIIts embedded workflow and prompt graphsComfyUI conventions
AUTOMATIC1111 PNG InfoSaved prompt and generation parametersAUTOMATIC1111-style blocks
ExifToolBroad EXIF, XMP, IPTC and container tagsMetadata, not an AI verdict
Adobe Bridge / PhotoshopCommon XMP and IPTC editingFields supported by the application
ffprobeAudio/video formats, streams and tagsTechnical container inspection
C2PA validatorManifest, signature and asset-binding validationSupported Content Credentials

Why prompts and workflows disappear

A screenshot, messenger, social platform, “strip metadata” option, re-encode or editor export can remove these records. A generator can also be configured not to save them. Missing prompt or workflow data is not evidence that AI was not used.

Why this site reads supported structures dependably

  1. Identify the file container before interpreting fields.
  2. Inspect multiple standard and application-specific storage locations.
  3. Keep raw bounded entries separate from normalised prompt and parameter fields.
  4. Parse known ComfyUI and AUTOMATIC1111 conventions without inventing absent values.
  5. Validate C2PA separately from editable EXIF, XMP, IPTC and PNG text.
  6. Report unsupported or missing layers explicitly.
Dependable parsing has a defined scope. No metadata reader can recover fields that were never saved or were removed from this copy.

Frequently asked questions

Can a prompt always be recovered?

No. It must have been stored and preserved in a supported structure.

Does a seed reproduce the exact image?

Not by itself. The model, software, workflow and other settings also matter.

Can metadata prove AI generation?

An explicit workflow or generator field can be strong evidence for that file, but ordinary metadata is editable. Signed provenance has a different validation model.

Read the file you have.

Read metadataUse the free API