Remove Stable Diffusion and ComfyUI metadata from PNG files.
Strip the prompt, negative prompt, seed, sampler, model name and full ComfyUI workflow that AI image tools save inside PNG files. The pixels stay identical and nothing is uploaded.
What AI image tools hide in a PNG
A PNG file is a series of chunks. Most describe the image, but text chunks (tEXt, zTXt and iTXt) can hold anything, and local AI image tools use them to save how the image was made. Anyone who opens the file in a metadata viewer, or drags it back into the same tool, can read it.
| Tool | Text field | What it contains |
|---|---|---|
| AUTOMATIC1111 and Forge-style web UIs | parameters | Prompt on the first line, then the negative prompt, steps, sampler, CFG scale, seed, size and model name or hash |
| ComfyUI | prompt | The executed node graph as JSON, including every input value such as seed and checkpoint file name |
| ComfyUI | workflow | The full editor workflow, which can be loaded again by dropping the image into ComfyUI |
| Other front-ends | Their own field names | Usually the same kind of settings in JSON or plain text |
Libraries that read these files, such as the open-source sd-parsers project, document the same field names. This tool does not depend on the names: it removes every text chunk, whatever it is called, and flags the file as containing AI generation settings when it recognises them.
Why remove it
- Your prompt is your work. A detailed prompt and its settings can take hours to refine. Sharing the PNG shares the recipe.
- The seed makes it reproducible. With the prompt, seed, sampler and model, someone can regenerate a near-identical image.
- Model and LoRA file names say a lot. A ComfyUI workflow lists the checkpoint and add-on files you used, by file name.
- It travels further than you think. Email attachments, cloud drives and file transfers pass the original file along unchanged, text chunks included.
What we tested
We built PNG files that store settings the way these tools do: an AUTOMATIC1111-style parameters field in a plain tEXt chunk, the same field compressed in a zTXt chunk, and ComfyUI-style prompt and workflow fields. In every case the report listed the text fields and marked them as AI generation settings, the cleaned file contained no text chunks at all, and the image data chunks were copied byte for byte.
How to check a PNG for prompts yourself
- Here: drop the file in. If it holds generation settings, the report shows a line such as "2 text fields: prompt, workflow · includes AI generation settings" before anything is removed.
- In AUTOMATIC1111: the PNG Info tab reads the
parametersfield and shows the full prompt and settings. - In ComfyUI: dropping the image onto the canvas loads the embedded workflow. If a workflow appears, the file is carrying it.
- On the command line: ExifTool lists PNG text fields along with everything else, for example
exiftool -a -G1 image.png.
Checking a cleaned copy the same way should show no prompt, no settings and no workflow.
Cleaning a whole folder of renders
Generation sessions produce dozens or hundreds of files. You can select them all at once or drag the whole set onto the page. Two files are processed at a time, each gets its own report, and when they are done a single button downloads everything as one .zip, named with a -clean suffix so the cleaned copies never overwrite your originals. There is no limit on the number of files or their size.
What stays in the PNG
Only chunks that hold metadata are removed. Everything needed to display the image exactly as before is kept:
- The image data itself (
IDAT), copied without re-compression - Transparency (
tRNS) and the palette for indexed images (PLTE) - The color profile (
iCCP,sRGB,gAMA,cHRM) and HDR information (cICP,mDCv,cLLi) - Pixel density (
pHYs), so print size does not change - Animation frames in animated PNGs (
acTL,fcTL,fdAT)
Because nothing is re-encoded, a 4K render comes out at exactly the same size and quality minus the text, which is usually a few kilobytes but can be much more for large ComfyUI workflows.
Does this affect AI labels?
Probably not on its own. Platforms that explain how they label content describe reading C2PA Content Credentials and the IPTC "AI generated" field (Meta), not PNG text chunks. A PNG straight out of a local Stable Diffusion setup normally has no C2PA manifest at all. Removing the settings is about keeping your prompt and workflow private. If you also want to know whether a file carries C2PA or IPTC flags, drop it into the AI metadata checker.
Other ways to do it
AUTOMATIC1111 has a setting to stop saving generation parameters as PNG text chunks, and ComfyUI can be started with its --disable-metadata option. Those are good options for new images. For images you have already made, or a folder of mixed files, dropping them here clears everything at once and lets you download a single .zip.
Frequently asked questions
Will removing the metadata change how the image looks?
No. Only text chunks are removed. The image data, transparency and color profile are kept exactly as they were.
Can someone still load my ComfyUI workflow from the cleaned image?
No. The workflow lives in the workflow and prompt text chunks, and the cleaned file has no text chunks left.
Does it work with JPG and WebP exports from AI tools?
Yes. Those formats store settings in EXIF comments or XMP instead of PNG text chunks, and both are removed as well.
Can I keep the prompt for myself?
Yes. Your original file is never modified. Keep it as your record and share the cleaned copy.