All articles

Privacy

Can everyone see your prompt? How image generators store text inside files

Where PNG and JPEG files may expose prompts, negative prompts, seeds, models and entire workflows — and why the answer varies by generator and export path.

A finished coastline print partly covering handwritten process notes and thumbnail sketches
Privacy

Some generator exports keep reusable workflow details inside the file. Whether they are present depends on the tool and export path.

This is not steganography

A prompt is usually not hidden in pixels. The generator writes a text chunk or EXIF/XMP field so the creator can restore settings later. Any compatible metadata viewer can read it when an upload platform has not removed it.

AUTOMATIC1111 may record positive and negative prompts, seed, sampler, steps, CFG, dimensions and model. Its project documentation describes PNG Info as a text chunk, and its settings can disable metadata infotext; other output formats should be checked separately.

ComfyUI can reveal the whole workflow

Native ComfyUI PNGs commonly carry prompt and workflow data. That convenience lets a creator drag an image back into the interface to restore the graph. It can also expose model and LoRA names, node parameters and instructions. Conversion or social uploads may remove the data, but you should verify rather than assume.

  • Look for PNG text chunks named parameters, prompt or workflow.
  • In JPEG, check EXIF UserComment, ImageDescription and XMP.
  • Review the complete field list, not only a friendly summary.

What about Midjourney?

Midjourney’s job interface lets a creator copy a prompt, Job ID and seed. Creation visibility also depends on account and privacy settings. That does not mean every downloaded Midjourney JPEG is guaranteed to carry the complete prompt in EXIF. A public job history and embedded file metadata are separate disclosure channels.

The accurate question is whether parameters are available from this particular file or a linked public page. An image interrogator can sometimes guess a description, but it cannot reliably recover the creator’s exact original prompt.

Checklist before client delivery

Create a separate publication copy and inspect it locally. Remove only fields you do not intend to disclose, while keeping the generation original in a private archive. Blind stripping can destroy useful process or licensing evidence.

  • Check prompt, negative prompt, seed, model, workflow and local file paths.
  • Make sure comments contain no client names, tokens, internal URLs or draft instructions.
  • Inspect again after conversion because a new container may preserve some XMP.

Primary sources

  1. AUTOMATIC1111: generation parameter storagegithub.com
  2. AUTOMATIC1111: metadata infotext settinggithub.com
  3. ComfyUI: workflow metadata in generated imagesdocs.comfy.org
  4. Midjourney: prompt, Job ID and seed controlsdocs.midjourney.com

LINKED / TOOL CHAIN

Tools for the formats and platforms discussed above.

Stable Diffusion metadata ComfyUI workflow metadata Midjourney metadata JPG / PNG metadata

LINKED / RESEARCH

More guides on file history, privacy and platform rules.

TechnologyWhat is C2PA, and why ordinary EXIF cleaners do not remove itTechnology / watermarkingSynthID vs C2PA: which AI labels are metadata, and which live in the media itself?