Last updated
Open a freshly generated PNG in a plain text editor and you may find your own prompt in it, word for word. Another file from another tool might hold nothing readable at all but a signed certificate saying which company's model drew it. Both are metadata, both leave with the picture when you upload it, and neither shows when you look at the image. Below: where the records sit, what each generator leaves, and how to delete the lot safely.
Start here: what did your generator write?
| You used | Look for | Gone after cleaning? |
|---|---|---|
| AUTOMATIC1111, Forge and similar web UIs | A PNG text entry called parameters holding prompt, negative prompt, seed, sampler and model hash; the same text in an EXIF comment for JPEG and WebP | Yes |
| ComfyUI | Two PNG text entries: prompt (graph to execute) and workflow (layout of nodes) | Yes |
| InvokeAI and other interfaces | Own PNG text entries, named differently in each program | Yes |
| ChatGPT image generation, Firefly, Photoshop generative tools, Microsoft design tools | A Content Credentials manifest, often together with XMP that names the tool | Yes, the copy inside the file |
| Google tools | An invisible SynthID watermark, sometimes IPTC source-type fields | Fields yes, watermark no |
| Midjourney and others | Varies with the tool and with how you downloaded the image | Whatever the file holds |
A prompt in one string
The parameters entry of a Stable Diffusion web UI is ordinary text. With invented values it looks like this:
a lighthouse at dusk, oil painting, warm light
Negative prompt: blurry, text, watermark
Steps: 28, Sampler: Euler a, CFG scale: 7, Seed: 1234567890, Size: 768x1024, Model hash: 1a2b3c4d5e, Model: sdxl-base-1.0, Lora hashes: "style-oil: 9f8e7d6c"Share that and you share your craft: a competitor can regenerate your image, and the model names may expose private or licensed checkpoints. ComfyUI goes further. Its workflow entry holds the whole graph, including file names of models, input images, custom nodes and sometimes folder paths from your computer. Dropping such a PNG onto the ComfyUI canvas rebuilds the project in one move.
Content Credentials: a signed note, not a lock
The C2PA standard lets a tool attach a signed statement to a file: who or what made it, which edits were applied, and a hash that ties the statement to the exact pixels. Edit the pixels and the tie breaks. Delete the statement and you hold a normal image. It sits in a caBX chunk (PNG), APP11 segments (JPEG), a C2PA chunk (WebP) or a box (MP4, MOV).
- Deleting it from your file does not delete the provider's own copy, if they keep one.
- Some providers can still recognise an image through an invisible watermark or a fingerprint.
- You can read a file's credentials at the verify page; after cleaning it should find none.
Next to it sits the IPTC field Digital Source Type, a one-line AI label stored in XMP. trainedAlgorithmicMedia means generated, compositeWithTrainedAlgorithmicMedia partly generated or edited, digitalCapture camera-made. Cleaning removes the XMP packet whole, so the label goes too.
Platforms read these records
Meta, TikTok and LinkedIn have stated that they use Content Credentials or IPTC fields to attach AI labels, and YouTube wants creators to declare realistic synthetic material at upload. So a label may appear because of metadata alone, even on a photograph with a single AI-assisted edit. It may also appear for other reasons: platforms add watermark detection, classifiers and your own declaration. Cleaning a file does not end an obligation to disclose when a platform requires it.
What no metadata cleaner can touch
- Invisible watermarks such as SynthID. They are written into pixel values, and we do not claim to remove them.
- Settings encoded in the pixels, which a metadata tool cannot tell from a normal picture.
- Visible marks: logos and labels drawn on the image.
- What the provider remembers about your account and generation history.
Why clean at all
- Prompts and workflows are your intellectual work and can name clients.
- Model, LoRA and folder names describe your setup.
- A ComfyUI PNG can weigh tens of kilobytes more than it needs to.
- Clients and marketplaces sometimes ask for files without embedded data.
How it works for you
- Add PNG, JPEG, WebP or AVIF files. One or many.
- Press Clean. Records are listed under plain names: “Generation parameters (prompt, seed, model)”, “ComfyUI workflow”, “C2PA Content Credentials”, “XMP packet”, “JPEG comment”.
- Download. Pixels are identical to the original. A second scan checks that nothing remains before the file earns the verified clean label.
Want to see all fields first, without changing anything? Use the metadata viewer, and afterwards the compare tool to put original and clean copy side by side.
Save the recipe, clean the copy
Records deleted from a copy are not recoverable from that copy. So copy your prompt and settings into a notes file, export ComfyUI workflows as JSON, and keep the original image in a private folder. Clean only the copies you publish.
For an output folder that keeps filling up, let an assistant clean it through the MCP server: it processes the PNGs and writes each clean copy next to the original. The API offers the same for scripts. Both leave originals untouched.
Check it yourself
exiftool -a -G1 -s cleaned.pngResponsible use
Provenance records exist so people can judge what they see. Protecting prompts and privacy is a good reason to clean. Presenting generated images as photographs where authenticity matters (news, evidence, insurance, contests that exclude AI) is not, and many places require AI-made or AI-edited content to be labelled whatever the file contains. Rules vary by country and change quickly, and this is not legal advice. If your audience would expect to be told, tell them. Everyday photos are covered by the image tool.
Questions
Are Stable Diffusion prompts removed?
Yes. The parameters entry and similar ones are PNG text chunks and all text chunks are dropped. The same text in EXIF or XMP fields of JPEG and WebP files is removed too.
Is a ComfyUI workflow removed?
Yes. Both prompt and workflow entries go, so the image cannot rebuild the graph in ComfyUI any more.
Are C2PA credentials removed?
From the file, yes: the PNG caBX chunk, JPEG APP11 segments and the WebP C2PA chunk. A provider's own online copy is out of reach.
Does it work for ChatGPT, Gemini, Firefly or Midjourney pictures?
It removes whatever metadata the file holds, whichever tool made it. Invisible watermarks such as SynthID are not metadata and remain.
Could a detector still call the picture AI?
Yes, possibly. Detectors read pixels, and watermarks survive cleaning.
Is image quality reduced?
No. PNG, JPEG and WebP keep their pixel data unchanged. AVIF files are redrawn and saved again, so they are not bit-identical.
Can I recover the prompt afterwards?
Not from the cleaned file. Store prompt and settings before you clean.
Why do platforms still label my image?
They combine file records, watermark detection, classifiers and your declaration, and may also remember an earlier upload. Cleaning removes the file's records, not theirs.
Are my images sent to a server?
No. The cleaning runs in your browser; the network panel shows that nothing is uploaded.
Is it free? Can I do a folder?
Guests can clean 50 images a day, a free account 500. Starter, Pro and Business have no daily limit and bigger batches. For folders use the API or the MCP server.