isGenAI · Open examples
See what an image file actually stores.
The open reference for image metadata behavior: 7 controlled cases, 10 downloadable PNGs, exact SHA-256 hashes and recorded outputs, all CC0. Pick a question, inspect the records and reproduce the result yourself.
Choose a question
PNG records
One saved prompt in three PNG text encodings
tEXt, zTXt and iTXt files return the same prompt and seed, despite having different file bytes.
Missing records
The same pixels with and without a saved prompt
A PNG without the attached record produces no supported saved prompt, although its pixels match the reference image.
Conflicting records
A PNG with conflicting saved generation settings
Two parameters records disagree on seed and step count. The reader does not select an arbitrary original prompt.
Record comparison
Identical pixels, different saved seed and steps
The saved seed changes from 42 to 84 and steps from 20 to 30, while the decoded image remains identical.
Export behavior
A byte copy and a fresh PNG export retain different records
The byte copy retains the reference hash and prompt. The tested PNG re-encode preserves pixels and loses the supported prompt record.
Workflow records
Two ComfyUI graph records in one PNG
One PNG contains a two-node editor workflow and a separate two-node API graph. Neither is executed during inspection.
Saved settings
An InvokeAI record with seed zero
The reader preserves a saved seed of zero and a CFG scale of 4.5 instead of treating zero as a missing value.
What these examples measure
Every case starts from a picture drawn by code. We attached known text and workflow records, then inspected the files with the isGenAI readers and recorded the observations on . The catalog makes those fixtures easy to explore case by case; it is a controlled reference, not a detection benchmark.
The examples measure record reading, file equality and pixel equality with exact hashes. They measure neither whether a generator made a picture nor real-world detector accuracy, and every case states that scope so a citation carries it.
Compare with the separate 18-case metadata export experiment, or follow the worked image and API guides.
Download, reproduce and cite
Fixtures and recorded observations are released under CC0-1.0: use them anywhere, no permission needed. Each case links to its original files and SHA-256 hashes.
- All observations, hashes and dependency versions (JSON)
- Standalone reproduction script (Node.js and sharp)
- Fixtures on Hugging Face
Suggested citation: isGenAI, Controlled image metadata examples, observations dated 2026-09-19, catalog published 23 September 2026. Cite the individual case URL when describing a result and retain its stated limits.
Automate inspection with the workflow and metadata comparison API guide, including runnable Node.js examples.