# How to remove an AI watermark from an image

Source: https://passreal.com/guides/remove-ai-watermark

"AI watermark" can mean three different things, and each one has a different answer. A visible logo drawn on the picture can be cropped or retouched. Provenance data stored in the file, such as C2PA, IPTC and XMP, can be deleted. A signal hidden in the pixels, such as SynthID, is built to survive edits, and nobody serious promises to remove it. PassReal handles only the second kind.

## Three things people call an AI watermark

The phrase covers three separate mechanisms. They live in different places,
they are made by different tools, and they are removed in different ways, or
not at all. Mixing them up is the main reason people try a method that cannot
work. Work out which one you have before you pick a fix.

## Kind one: a visible mark drawn on the image

This is the easy one to recognise. A small logo sits in a corner, or a line of
text is laid over the picture. Some image tools add a sparkle or a name tag,
and some video tools place a moving logo on the frame. It is part of the
picture, in the same way a stamp is part of a printed photo.

Because it is drawn into the pixels you see, the fixes are the ones for any
drawn mark:

- Crop the image so the corner falls outside the frame.
- Retouch the area with a clone or healing tool in GIMP or Photoshop.
- Generate the image again at a size or plan that does not add the mark.

Check the terms of the tool that made the image before you do this. Some
services ask you to keep the mark on free output. PassReal does not remove a
visible watermark, and we would rather say so here than let you upload an image
and find out afterwards.

## Kind two: provenance data inside the file

Most AI image tools write a record into the file itself. It does not show on
screen. It sits in blocks next to the image data: a C2PA manifest, an IPTC
field that names the digital source type, XMP tags that list the software, and
text chunks in a PNG. Platforms read these fields to decide whether to show an
AI label. Meta says it uses the standard indicators in this way.

Because the record is separate from the pixels, you can delete it and the
picture does not change. You can do it yourself:

- Run `exiftool -all= image.jpg` to clear the common tags. Read the file again
  afterwards, because a C2PA block can be missed by a quick pass.
- On Windows, open Properties, choose Details, and use the option to remove
  properties and personal information.
- Re-save the image from a graphics editor that writes a fresh file.

PassReal does this in one upload, together with the other steps below. On an
export from Gemini, 174 C2PA and IPTC markers went to 0, and the report shows
you the count before and after. The longer walk-through is in
[how to remove AI metadata](/guides/remove-ai-metadata).

## Kind three: a signal inside the pixels

The third kind is a watermark written into the pixel values. You cannot see it
and you cannot find it with a metadata reader, because there is nothing in the
file structure to read. The best-known example is SynthID from Google, which
the company says is designed to stand up to cropping, filters and lossy
compression.

Deleting metadata leaves this kind untouched. Be wary of any site that
promises to wipe it with a button. Google does not publish a way to remove it,
we have no independent proof that anything does, and PassReal does not claim
to. [How SynthID works](/guides/synthid) explains what is known.

## How to check which one you have

1. Open the image at full size and look at the corners and edges. A visible
   logo or text means kind one.
2. Read the metadata with ExifTool. Any C2PA, IPTC digital source type, XMP
   software or PNG text field means kind two.
3. If the picture came from a Google tool and the file is clean, kind three may
   still be present. Google's own checks in Gemini, Search or Chrome can tell
   you, but a chat answer is a guess and the dedicated detector is the better
   test.
4. If you cannot tell, treat the image as having more than one kind. They
   often appear together.

## What PassReal does, in plain terms

One upload runs four stages. It removes every family of provenance metadata,
simulates a camera sensor with grain and a lens colour fringe, smooths the high
frequencies with a Fourier filter, and re-encodes the picture as a JPEG with
the quantization tables and EXIF fields of a real phone. SSIM and PSNR are
measured for each image and shown to you.

The promise is narrow. If a public AI image detector flags an image you paid
to process, email us within 14 days with the result and the payment is refunded
in full. The terms are in [section 7](/terms#guarantee). That guarantee covers
public detectors. It does not cover SynthID, and it does not cover a visible
logo. The free plan includes 3 images, enough to check the metadata count on
your own file before you decide anything.

## A note on disclosure

Removing a marker does not remove the duty to say an image was made with AI.
Platform rules and local law can require it, and that duty stays with the
person who posts the image.

## Questions

### Does PassReal remove the visible Gemini or Sora logo?

No. PassReal does not remove a visible watermark, meaning a logo or text drawn on the picture. It removes the provenance data stored in the file. For a drawn logo, crop the image or retouch it.

### How do I know which kind of watermark my image has?

Look at the picture first. If you can see a logo or text, it is the visible kind. If you cannot, read the file with ExifTool for C2PA, IPTC and XMP fields. If both are clean and the image came from a Google tool, a pixel watermark may remain.

### Can anyone remove SynthID?

Google says SynthID is designed to survive cropping, filters and compression. We have no proof that anything removes it, so PassReal does not promise it. Be careful with any tool that does.

### Is it legal to remove an AI watermark?

That depends on the platform rules and the law where you live, and it is your responsibility to check. Many platforms and some local laws require you to say when content is AI-generated, and removing a marker does not cancel that duty.

### Will the AI info label on Instagram go away if I remove the watermark?

Meta sets the label from standard indicators in the file, such as C2PA and IPTC, and from what the user declares. Removing those fields takes away the file signal. It does not change what you declare when you post.

Try it: https://passreal.com/app (three images free)