isGenAI.com

PNG records · Controlled example

One saved prompt in three PNG text encodings

tEXt, zTXt and iTXt files return the same prompt and seed, despite having different file bytes.

By isGenAI · Published

Read a saved image promptDownload the test files

Does the PNG text encoding change the recovered prompt?

All three controlled files return the same saved prompt, negative prompt, model label and settings. Their SHA-256 hashes differ because the bytes used to store the records differ.

How the case was made

The same AUTOMATIC1111-style parameters string was attached to the same code-drawn PNG as tEXt, compressed zTXt and compressed iTXt. Each file was read by the saved-prompt parser and checked against the same expected object.

Code-drawn geometric test image used for the controlled metadata examples
The same code-drawn source picture is used across the fixtures. The saved demonstration prompts do not describe how these pixels were made.

Files and observed output

  • parameters-tEXt.png · 13,371 bytesSHA-256: e76d2af4e6c273e1ba8ca6cf2b5b7885ceda2321bce893f0bb5f9e64ace17c9f
  • parameters-zTXt.png · 13,346 bytesSHA-256: 2cbac7bbc872ec95d8979132ef871b3a1f816ffc55bc12e68ea10ab5ac2d7818
  • parameters-iTXt.png · 13,349 bytesSHA-256: 4123149f352daeb842ecf241f7c7e6ee0fc4c5a08175ecb37fc690e093484cbf
Recorded result
[
  {
    "chunk_type": "tEXt",
    "file": "parameters-tEXt.png",
    "result": {
      "format": "automatic1111",
      "prompt": "Controlled example: a paper boat on a blue table",
      "negative_prompt": "text, blur",
      "recorded_model": "demonstration-only",
      "settings": {
        "seed": "42",
        "steps": 20,
        "sampler": "Euler",
        "cfg_scale": 7,
        "width": 1024,
        "height": 768
      }
    }
  },
  {
    "chunk_type": "zTXt",
    "file": "parameters-zTXt.png",
    "result": {
      "format": "automatic1111",
      "prompt": "Controlled example: a paper boat on a blue table",
      "negative_prompt": "text, blur",
      "recorded_model": "demonstration-only",
      "settings": {
        "seed": "42",
        "steps": 20,
        "sampler": "Euler",
        "cfg_scale": 7,
        "width": 1024,
        "height": 768
      }
    }
  },
  {
    "chunk_type": "iTXt",
    "file": "parameters-iTXt.png",
    "result": {
      "format": "automatic1111",
      "prompt": "Controlled example: a paper boat on a blue table",
      "negative_prompt": "text, blur",
      "recorded_model": "demonstration-only",
      "settings": {
        "seed": "42",
        "steps": 20,
        "sampler": "Euler",
        "cfg_scale": 7,
        "width": 1024,
        "height": 768
      }
    }
  }
]

Observations recorded 2026-09-19 using Node v23.11.0 and sharp 0.35.4. JSON seed values are strings to retain integer precision. A null prompt means the reader did not return a supported saved prompt.

Reproduce this result

  1. Download the files above and verify their SHA-256 hashes.
  2. Open the linked isGenAI tool and choose the corresponding file, or the before/after pair for comparisons.
  3. Compare the returned fields with the recorded output. Different tool versions or altered files can produce different results.

The standalone reproduction script creates the fixture variants and compares bytes and pixels. The complete JSON observations also include parser results. Use the workflow and comparison API examples to inspect downloaded files programmatically.

Read the format guide and its primary sources.

What this result does not establish

This tests three supported encodings of a known record. It does not establish support for every PNG producer, establish authorship or measure AI-detection accuracy.

Files and observations: CC0-1.0. Cite this case URL and include its method and limits. Report a reproducibility issue.