Remove metadata from ComfyUI images

What is in a ComfyUI PNG

In ComfyUI’s nodes.py, the Save Image node builds PNG info with one text entry named prompt (the API graph) and one entry for each key in the extra PNG info, which is where workflow comes from. Files are written as ComfyUI_00001_.png, counting up. The values are plain JSON: the checkpoint file name, the positive and negative prompt text, the seed, steps, sampler, and the node layout. That is a map of your setup, and it travels with every copy of the image.

The tests below use a PNG built with those same two chunk names and a seven-node graph (checkpoint loader, two text encoders, empty latent, sampler, decode, save). It was written with Pillow, not rendered by ComfyUI, so the graph is small; a real workflow chunk is larger and holds whatever custom nodes you use.

Why “AI only” is not enough here

The tool has three modes. For a ComfyUI PNG the difference matters, and we measured it:

Before (what the tool listed)	After (re-scan of the download)
AI-related text (prompt): prompt Text chunk (workflow): tEXt	Text chunk (workflow): tEXt

• Download: PNG, 1024x1024, 2.15 MB.
• Second opinion (ExifTool 13.25): 2 metadata tags before, 1 after (PNG Workflow).
• Pixels: decoded RGB values identical to the original (compared in Python).
• Time in the tool: 225 ms in headless Chrome on a server CPU. A phone will be slower.

AI only recognises the prompt key. The workflow chunk, with the same prompt text in its widget values, stayed.

Before (what the tool listed)	After (re-scan of the download)
AI-related text (prompt): prompt Text chunk (workflow): tEXt	Nothing found. Banner: “Cleaned. 2 fields removed — your picture is unchanged.”

• Download: PNG, 1024x1024, 2.14 MB.
• Second opinion (ExifTool 13.25): 2 metadata tags before, 0 after.
• Pixels: decoded RGB values identical to the original (compared in Python).
• Time in the tool: 234 ms in headless Chrome on a server CPU. A phone will be slower.

Original 2197.4 KB, cleaned 2195.5 KB: the saving is just the two text chunks, since the image data is copied untouched. The cleaned PNG decodes to the same pixels as the original, so there is nothing to compare by eye. Use All metadata for ComfyUI files; the SD page covers the Automatic1111 parameters key in Stable Diffusion PNG metadata.

Going forward vs the folder you already have

ComfyUI itself has a server option, --disable-metadata, which its argument parser describes as “Disable saving prompt metadata in files.” Use it if you never want the chunks in new renders. Two reasons you may still need to clean files: renders you made before you set it, and renders you want to keep re-openable. Keep the original PNG in your archive folder, copy the ones you will publish into a separate folder, and clean that folder.

The folder workflow

1. Copy the PNGs you will publish out of the output folder.
2. Open this page, set All metadata, and drop one file to read the report. One image is free.
3. For the rest, open the file picker inside the folder, select all the files (Ctrl or Cmd+A), or drag them in together. With Pro the tool cleans each file in your browser and returns one ZIP; a file it cannot read is skipped and the rest still finish.
4. Post from the ZIP. Dragging a cleaned PNG into ComfyUI no longer restores a graph, which is the point.

Command line instead

If you prefer a script, exiftool -all= removed both chunks from the test PNG too. What ExifTool does and does not do to a file is on ExifTool remove all metadata.

The general PNG chunk reference is on remove metadata from PNG.

What stays

• Pixels. A watermark or a pattern in the image itself is not metadata and is not touched.
• Any copy you already posted. Cleaning the file does not edit a page that already hosts the old one.
• Sites that read generation details from the chunks will show none for the cleaned copy.

Questions

What exactly does ComfyUI put in a PNG?

The Save Image node writes a PNG text chunk named prompt (the API-format graph with every input value) and, when the UI sends it, a chunk named workflow (the editor graph). That is what the ComfyUI source does in nodes.py; the tests on this page use a PNG built with those two chunk names.

Does AI only mode clear a ComfyUI image?

Not completely. In our test it removed the prompt chunk and left the workflow chunk, and ExifTool still listed PNG Workflow afterwards. All metadata removed both. Always read the After column.

Can I stop ComfyUI from writing this in the first place?

ComfyUI has a launch option, --disable-metadata, described in its source as “Disable saving prompt metadata in files”. It only affects images saved after you set it. Files you already rendered still carry the chunks, which is what this page cleans.

Will I lose the ability to reload the workflow from the image?

From the cleaned copy, yes: dragging it into ComfyUI will no longer restore the graph. Keep the original PNG, or save the workflow as a JSON file, and publish only the cleaned copy.

Does cleaning change the image?

In our test the decoded pixels of the cleaned PNG were identical to the original and the dimensions were unchanged. Only the text chunks were removed.

Can I clean the whole output folder at once?

One image is free. Pro is a one-time unlock for unlimited images and a single ZIP, so a folder of ComfyUI_00001_.png files is one run. Everything stays in your browser.

Related guides

• Clean Stable Diffusion PNG metadata (Automatic1111 and ComfyUI keys)
• ExifTool remove all metadata, tested
• AI metadata remover
• All metadata-removal guides by file type and device
• Open the Remove AI metadata tool on the homepage if you want the general cleaner.
• Short Instagram upload workflow if you only need the four-step version.