PassRealPricingApp

How to remove AI metadata from a PNG

An AI-generated PNG can hide data in several places: text chunks that hold the prompt, an XMP packet, an EXIF chunk and a C2PA manifest chunk. List them all with `exiftool -a -G1 file.png`, delete them, then list again. Saving the pixels to a fresh file removes every chunk at once.

Last updated: 2026-10-08

Why a PNG is a special case

A JPEG stores metadata in a handful of well-known segments. A PNG is built from chunks, each with a four-byte type, and the format lets anyone add a chunk. That flexibility is why generation tools use it: they can write whatever they need, and decoders skip chunks they do not know.

If you want the three text chunk types explained, read how to strip text chunks from a PNG. This page covers the other half of the question: all the places an AI image can keep something, and how to be sure you found them.

A map of where things hide

A tool that reports only text chunks will miss the last three. That is the usual way a cleaned file still gets labelled.

Removing everything

ExifTool does the job in one pass. The command exiftool -all= file.png deletes writable metadata, and the C2PA group may need -jumbf:all= if it is still listed. ExifTool leaves the old file with an _original suffix by default. Keep it until you have checked the result, and delete it afterwards if you do not need it. Its manual warns that the overwrite option should be used only with separate backups.

An alternative is to build a new file from the pixels. Decode the image to an array, create a new image from that array and save it without copying the info dictionary. Because the new file never had the chunks, nothing has to be found and deleted. That is the approach PassReal uses, and its report counts what the original carried: on a Gemini export, 174 markers before and 0 after.

Verification you should not skip

The failure mode is quiet. A tool says done, and a chunk is still there.

What this does not fix

Removing metadata changes what a file says. It does not change what the picture contains. If the generator embedded a watermark in the pixels, as SynthID does for Google tools, it stays. If a detector reads pixel statistics, it reads the same pixels. In our tests, ZeroGPT still gave scores of 0.96 to 0.97 on images that PassReal had processed, which is why this site does not promise detector results.

An empty file is also its own tell. Cameras write EXIF, so a PNG or JPEG with no metadata at all is unusual. If that matters for your use, see how to add EXIF data to an AI image.

Choosing a method

For one or two files, ExifTool is exact and free. For many, a folder-wide command or a batch service saves time. In both cases, verify the output once by hand before you trust the process.

Questions

Does converting the PNG to JPEG remove the metadata?

Usually the text chunks do not carry over, but it depends on the tool, and some converters copy EXIF or XMP across. List the groups on the output file instead of assuming.

Why does ComfyUI add so much to a PNG?

It stores the prompt and the workflow graph as text so that dropping the image back into the interface restores the whole setup. That is useful for you and revealing for anyone else who opens the file.

Will removing the chunks make the PNG smaller?

Yes, by the size of the removed data. A workflow graph can be large, but the saving is small next to the pixel data.

Can I keep the prompt for myself?

Save a copy of the original before cleaning, or paste the prompt into your own notes. Once the chunk is gone from the only copy, the settings are gone too.

Does PassReal accept PNG?

Yes, with PNG, JPEG and WebP up to 20 MB and 8192 pixels on the longest side. The result is encoded as a JPEG with a consistent camera profile.

Limits

Naturalization operates on the pixels and on the file, not on the content of the image. A subject that is implausible - six fingers, inconsistent reflections, text that does not read - stays implausible after processing, and a human reviewer will notice it. PassReal changes what a statistical classifier measures, not what a person sees.

Results vary by generator, by subject and by detector, and detectors are retrained. No pass rate is published on this page because none has been measured in a way that would still hold next month.

Process three images free

Three images, no card, no expiry.