AI Image Metadata Viewer and Remover
Read the prompt, settings and workflow hidden in a generated image — or strip them out.
Loading the viewer…
About the AI Image Metadata Viewer and Remover converter
Stable Diffusion tools write far more into an image than most people realise. Automatic1111 stores the prompt, negative prompt, seed, sampler and model; ComfyUI stores the entire node graph as JSON; and the whole lot travels with the file when you send it to somebody.
This reads all of it back — from PNG, JPEG and WebP — and removes it when you would rather not hand it out. The file never leaves your device, which matters when the thing you are inspecting is unreleased work.
What each generator leaves behind
Automatic1111, Forge and SD.Next write a single "parameters" block: your prompt, then the negative prompt, then a line of settings — steps, sampler, CFG scale, seed, size and model hash. In a PNG that sits in a text chunk; save as JPEG or WebP and the same block goes into the EXIF UserComment tag instead. Both are read here, and the settings line is parsed into labelled fields rather than shown as one wall of text.
ComfyUI writes two chunks, "prompt" and "workflow". The second is the whole node graph including positions and widget values, which on a complicated generation runs to hundreds of kilobytes. ComfyUI only does this for PNG, so a JPEG from it will usually be empty.
InvokeAI writes invokeai_metadata, Fooocus uses its own scheme, and NovelAI puts prompt and settings in a comment chunk. All of them are read here, along with any XMP a WebP happens to carry.
The part nobody expects
A ComfyUI workflow contains the filenames of every checkpoint, LoRA and custom node it touched — and very often the absolute path to them. On Windows that means C:\Users\yourname, so an anonymous post can carry a real first name and surname in it.
Anything like that found in the file is called out at the top, before you share the image rather than after.
Midjourney, DALL·E and Firefly show nothing
These are the ones people are surprised by. Midjourney delivers finished images through Discord and the web, DALL·E through ChatGPT, and both arrive re-encoded with no prompt block in them. Adobe Firefly is the same. Whatever they knew about how the picture was made did not travel with the file.
Some of them do embed C2PA Content Credentials instead — a signed provenance manifest saying which tool made the image and when. That is a different format from the text chunks read here, and it is not something this page decodes.
So a blank result is not evidence of anything. It means the file has no readable generation data, which is equally true of a Midjourney render, a photograph, a screenshot, and a Stable Diffusion image that was saved as JPEG by an editor along the way. If you want to know whether something is AI generated, metadata is a hint at best.
Removing it without touching the picture
The clean copy is written by rebuilding the PNG without its text chunks and copying the compressed image data across byte for byte. Nothing is redrawn through a canvas, so the picture that comes out is identical to the one that went in.
That is worth insisting on: the usual way to strip metadata is to re-encode the image, which quietly costs quality every time somebody does it.
Frequently asked questions
No metadata was found. Does that mean it is not AI generated?
No, and this is worth being clear about. Converting to JPEG or WebP drops these chunks, as does posting to most social platforms, screenshotting, and re-saving in most editors. Generators can also be configured not to write metadata at all. Absence proves nothing either way.
Why does my Midjourney or DALL·E image show nothing?
Because they do not write a prompt block into the file. Midjourney and DALL·E hand you a re-encoded image with no generation data in it, and Adobe Firefly is the same — some of them embed C2PA Content Credentials instead, which is a different format this page does not read. Stable Diffusion tools like Automatic1111 and ComfyUI are the ones that leave a readable trail.
Does it work on JPEG and WebP?
Yes. There are no text chunks in those formats, so Automatic1111 puts the same parameters block into the EXIF UserComment tag instead, and that is read here too. ComfyUI only writes its workflow into PNG, so a JPEG will usually have the prompt and settings but not the node graph.
Which formats can it clean?
All three. PNG loses its text chunks, JPEG loses the EXIF and comment segments, and WebP is rebuilt without its EXIF and XMP chunks. In every case the compressed image data is copied across rather than re-encoded.
Is the image uploaded anywhere?
No. The file is read as bytes in your browser and the cleaned copy is built there too. Nothing is sent anywhere, which is the point when the thing you are checking is unpublished work.
Does removing the metadata reduce the quality?
No. The compressed image data is copied across untouched and only the text chunks are dropped, so the picture is bit-for-bit what it was. The file just gets smaller.
Can I recover my settings from an old image?
Yes, that is one of the main uses. If the PNG came straight out of the generator, the seed, sampler, model and prompt are all still in it, and can be copied straight back.