See and remove AI metadata from your images

What AI metadata is

An image file holds more than pixels. It can also hold text and other data, called metadata, that apps write next to the picture. Cameras use it for the date and location. AI image tools use it to record how an image was generated.

For tools that run Stable Diffusion and similar models, that record is often very detailed. It can include the full prompt, the negative prompt, the seed, the sampler and number of steps, the guidance scale, the image size, and the name or hash of the model. ComfyUI can go further and save the complete node graph you built. The data is there so you, or the tool, can read the settings back later and make the same image again.

The catch is that anyone else can read it too. When you send the original file to a client, post it on a site that keeps metadata, or share it in a chat that sends files unchanged, the recipient gets your recipe along with the picture. Hosted generators may also attach Content Credentials, a signed provenance record, which is a different kind of AI metadata covered further down.

What each generator stores

AUTOMATIC1111 and Forge

The AUTOMATIC1111 Stable Diffusion web UI, and Forge, which is built on it, write a block of text that the code calls the infotext. In a PNG it goes into a text chunk named parameters. The first lines are your prompt. A line that starts with Negative prompt: holds the negative prompt. The last line lists the settings, such as Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 965400086, Size: 512x512, Model hash: 45dee52b, followed by the model name and any extras. In the web UI the options to write this text, and to add the model name and hash, are all on by default.

For JPEG and WebP images the same text goes into the EXIF UserComment field instead. The web UI has a PNG Info tab that reads it back and can send the settings straight to a new generation, which shows how easy it is to reuse.

ComfyUI

The standard Save Image node in ComfyUI writes two PNG text chunks. prompt holds the graph that was run, as JSON: every node, its settings and the text of your prompt nodes. workflow holds the full editor graph, including node positions and links. The official ComfyUI examples say their images can be loaded into ComfyUI, or dragged onto the window, to get the full workflow that made them. Anyone you share a PNG with can do the same with yours. ComfyUI has a --disable-metadata start-up option that turns this off for future images.

To read a prompt and copy it, or save a workflow.json, without removing anything, use the Stable Diffusion metadata viewer.

The checker counts the nodes in the workflow and shows the prompt text it finds in text-encode nodes, sorted into positive and negative where the sampler wiring makes that clear. It also lists the sampler settings and the model file names.

NovelAI

NovelAI images carry PNG text chunks such as Title, Description, Software, Source and Comment. The comment is JSON with the generation settings. The checker reads it when Software says NovelAI. NovelAI also says it stores the information in the image's alpha channel, in a format it calls stealth pnginfo. That copy is part of the pixels, so a metadata remover that leaves the pixels untouched, like this one, cannot remove it.

Content Credentials from hosted generators

Some hosted services attach Content Credentials, based on the open C2PA standard, instead of prompt text. Adobe says it automatically applies Content Credentials to assets where all the pixels were generated with Adobe Firefly. The record names Adobe as the issuer and lists the date, the app, the AI tool used and the actions taken. OpenAI says images generated with ChatGPT, Codex or the OpenAI API include Content Credentials and a SynthID watermark, and that its provenance signals do not include details about the user, the organisation or the prompt.

Other services and other tools may save different data, or none. The checker reports what is actually in your file. It does not guess where an image came from.

Why remove it

• Prompt privacy. A good prompt can take hours of testing. With the prompt, seed, model and settings, someone can often get very close to your image, or reuse your style on their own work, without asking.
• Workflow IP. A ComfyUI workflow is more than a prompt. It is the whole pipeline: upscalers, ControlNets, LoRAs and custom nodes. If you sell images, run a service or teach, that graph may be the valuable part.
• Personal data. Prompts sometimes contain names, places or private notes. Workflows can list model and LoRA file names and the names of images you loaded into them, which may show what is on your computer. Regular photo metadata, such as GPS location, can also be present if you edited a real photo.

Keeping your prompt private is a normal creative choice. It is not the same as hiding that an image was made with AI, which is covered below.

How to check and remove it with the tool

1. Choose or drop one or more PNG, JPEG or WebP images into the checker at the top of the page.
2. Read the report. A purple box lists any AI generation data: the tool that made it, the prompt, the negative prompt, the settings and, for ComfyUI, the number of workflow nodes. Long prompts are cut short, with a Show more button.
3. Check the other sections for location, device, date and software details, and for Content Credentials.
4. Leave Also remove Content Credentials unticked unless you have a reason to remove them. It is off by default.
5. Press Download clean copy. The file gets -clean added to its name, and your original is not changed.
6. If you like, open the clean copy in the checker again to confirm the prompt data is gone.

Everything happens in your browser. The images are not sent to SaveMingo, and closing the tab discards them.

What is removed and what is kept

Removed by default: all PNG text chunks, which is where AUTOMATIC1111, Forge, ComfyUI and NovelAI put the prompt, settings and workflow. Also EXIF, including the UserComment that holds A1111 data in JPEG and WebP files, plus XMP, IPTC, comments and any extra data after the end of the image.

Kept by default: the colour profile, so colours look the same, a tiny orientation tag if the image needs one, and Content Credentials. You can remove the colour profile or the Content Credentials with the boxes in the tool.

Never changed: the pixels. The tool copies the image data byte for byte and leaves out the metadata blocks, so there is no re-compression and no quality loss. That also means anything stored in the pixels themselves stays, such as the NovelAI alpha-channel data or an invisible watermark. For more on regular photo metadata, see the image metadata remover.

Content Credentials and disclosure

Content Credentials are different from prompt metadata. They are a signed record of where an image came from, not a copy of your recipe, and the ones OpenAI describes do not include the prompt. That is why the tool keeps them unless you tick the box. Removing them is your choice, but it does not make an AI image anything other than an AI image.

Some files also carry an IPTC digital source type. The checker shows trainedAlgorithmicMedia as created with generative AI and compositeWithTrainedAlgorithmicMedia as edited with generative AI, following the IPTC vocabulary. When it sits in XMP or IPTC metadata, the default clean copy removes it. A copy inside Content Credentials stays with them.

If a platform, a client contract or a law requires you to say that an image was made with AI, you still have to say so after cleaning the file. Removing metadata does not change what an image is, and it does not change how platforms detect AI content in other ways. OpenAI, for example, adds an invisible SynthID watermark alongside Content Credentials and says it may survive changes that remove metadata. OpenAI also says Content Credentials and watermarks do not replace visible labels or other notices that may be required.

If a real photo of yours was labelled as AI on Instagram, Facebook or Threads, read why real photos get the AI info label. Before you post or sell an image that uses someone else's work, see our copyright page.

Questions

Does this work for Stable Diffusion images? Yes, if the tool that saved them used one of the formats above. Images from the AUTOMATIC1111 web UI, Forge, ComfyUI and NovelAI are recognised. Other text chunks are shown under Other metadata with a short preview, and they are removed too.

Can I remove the ComfyUI workflow but keep the image quality? Yes. The workflow lives in PNG text chunks, and the clean copy leaves them out without touching the image data.

What about Midjourney or other services? The checker shows whatever text and provenance data a file contains, whichever service made it. We only describe formats we could check in a public source, so other services are not listed here.

Will this stop AI detection? No, and that is not what it is for. It removes the notes stored next to the pixels. It does not change the pixels, remove watermarks or change how any platform labels content.

Is my image uploaded? No. The browser reads the file and builds the clean copy on your device. See our privacy policy.

Sources

• AUTOMATIC1111 Stable Diffusion web UI source: modules/images.py (PNG parameters text, EXIF UserComment for JPEG and WebP), modules/processing.py (infotext format), modules/infotext_utils.py (how it is parsed) and modules/shared_options.py (defaults)
• Forge source: modules/images.py
• ComfyUI source: nodes.py (Save Image writes prompt and workflow) and comfy/cli_args.py (--disable-metadata), plus the ComfyUI examples
• NovelAI: novelai-image-metadata (alpha channel, stealth pnginfo) and the naipng reader (PNG text chunks)
• Adobe: Content Credentials overview (Firefly)
• OpenAI: Provenance signals in OpenAI-generated content
• IPTC: Digital Source Type vocabulary