FLUX metadata check: the evidence sits in PNG text
FLUX is usually run through the same front-ends as the rest of the open-weights ecosystem, which decides where its evidence ends up: PNG text chunks holding a prompt or a serialised workflow, sometimes XMP tool fields, and — for video output — frame-level checks as a fallback. AICheck365 reads all of those in your browser and labels which front-end convention each chunk follows.
Short answer: read the PNG text chunks, not the EXIF
Open the original PNG and read its tEXt/iTXt/zTXt chunks. A parameters key follows the AUTOMATIC1111 convention; prompt and workflow keys indicate ComfyUI and carry the node graph as JSON, which is how most FLUX pipelines are driven. XMP fields may name a tool as well. None of this is signed, so it explains provenance well without proving it.
Sample status and evidence profile
| Bundled FLUX sample | Not bundled yet. This page does not claim a reproduced FLUX test until a licensed original can be published alongside the other samples. |
|---|---|
| Related bundled image | A ComfyUI / Stable Diffusion XL PNG is included and exercises exactly the same prompt and workflow chunk paths that FLUX pipelines write to. |
| Evidence level | Unsigned descriptive metadata: a strong explanation of how the image was produced, but not tamper-evident the way a C2PA signature is. |
Signals to Check
| PNG text chunks | tEXt, iTXt and zTXt carry the prompt, generation settings or a full workflow JSON depending on which front-end wrote the file. |
|---|---|
| Workflow JSON | When a ComfyUI-style graph is preserved, node structure, model names and sampling settings can be read and often re-imported. |
| XMP / EXIF | Tool and software fields sometimes name the generating stack. Useful corroboration, unsigned and editable. |
| Frame watermarks | For video output, sampled frames are compared against known patterns when the container yields nothing. Not a general AI classifier. |
Why the front-end matters more than the model
Which model produced an image and which application wrote the file are separate questions, and only the second one determines what you can read back. FLUX weights driven through ComfyUI leave a workflow chunk; the same weights driven through a hosted interface may leave nothing at all, because the writing convention belongs to the front-end rather than the model. So a FLUX image with no readable parameters is unremarkable — it usually means the pipeline did not write them, not that something was hidden.
What destroys these chunks
Text chunks survive copying but rarely re-encoding, and the failure modes are predictable: converting PNG to JPEG discards them outright since JPEG has no equivalent container, social platforms and messaging apps re-encode uploads and drop them along the way, and a screenshot produces a new file from pixels with none of the original text. Copying or moving the file, or uploading somewhere that stores it verbatim, preserves everything.
Limits
- PNG text chunks and XMP fields are unsigned: they can be edited, removed or copied onto an unrelated image.
- Converting to JPEG, re-encoding on social platforms and taking screenshots all destroy them.
- No bundled FLUX sample is published yet, so this page describes field coverage rather than a reproduced result.
- Finding no parameters does not mean the image was not generated — only that this copy carries none.
Use the PNG parameter extractor for prompts and workflows; run the full detector when you also want C2PA and EXIF/XMP checked.
FAQ
How do I read FLUX generation parameters?
Load the original PNG into the parameter extractor here. It reads the tEXt/iTXt/zTXt chunks in your browser, free and without uploading the file, and labels which front-end convention the chunk follows.
Why does my FLUX image have no metadata?
Most often because the front-end never wrote any, or because the file was re-encoded. Hosted interfaces frequently omit generation chunks entirely, and converting to JPEG or posting to social media removes whatever was there.
Is PNG metadata proof that an image is AI-generated?
It is a strong explanation but not proof. These chunks carry no signature, so they can be edited or copied between files — unlike a valid C2PA manifest, which is tamper-evident.
Are my images uploaded?
No. Parsing runs locally in your browser, so prompts and workflows stay on your device.
Sources and verification records
- AICheck365 public sample records — Which samples are bundled, their provenance, test dates and known limits.
- W3C PNG Third Edition — Standard reference for PNG chunks and the tEXt/iTXt/zTXt formats.
- Official ComfyUI examples — Source of the bundled SDXL PNG that exercises the same chunk paths.
- AICheck365 methodology — How unsigned metadata is weighted against signed provenance.