Photo SignalsAI Watermark Remover

Gemini watermark remover — the sparkle comes off, SynthID is measured, not removed

A picture made or edited in the Gemini app, including by Nano Banana, leaves with three separate marks. One is paint, one is a record in the file, and one is woven into the pixels with a detector only Google holds. This page says which is which before you press anything.

  • The sparkle badge
  • Content Credentials
  • SynthID

Open an image and paint over the mark

  • Nothing is uploaded
  • No account
  • Free, no limit
What Gemini and Nano Banana actually write

Three marks, three different answers

  • 1 · The visible sparkle

    Removed, and you can see it is gone

    Google's own words: images created or edited in the Gemini app carry a visible watermark — unless, since the setting below arrived, you switched it off. It is the sparkle in the corner, and it is paint on the picture — so you paint over it here and it fills in from the surrounding image. On a flat background it disappears; over hair, foliage or text it may leave a patch, and the tool tells you to compare before you keep it.

    Cheaper still: Google lets you switch the visible mark off before generating. In the Gemini app it is Settings → Media Watermark. If you still have the chat, turning that off and regenerating beats any remover, because nothing has to be filled in.

  • 2 · Content Credentials (C2PA)

    Removed, and the removal is proven

    Gemini attaches a signed C2PA record saying what made the image. It lives in the file's container, not in the pixels, so the rebuild simply never copies it across, and the result is read back with the same parser to prove it is gone. Note Google's warning on the setting above: switching the visible watermark off does not touch this record or SynthID. Only a rebuild does.

  • 3 · SynthID

    Reduced and measured — never claimed removed

    SynthID is Google's invisible watermark: embedded in the pixels the moment the image is created, designed to survive cropping, filters and lossy compression, and readable only by Google's own detector. The deepest pass here runs a learned model that weakens the high-frequency residual such marks live in, and reports how much it changed on your image, in decibels. What it cannot do is tell you whether Google's detector still reads it. Nobody outside Google can.

The honest test is Google's, not ours. After you download, open Gemini, upload the result and ask: "Was this image created or edited by Google AI?" That is the only detector that counts, and it is free. If it still says yes, the mark survived, whatever any remover's page told you.

Four steps

Paint the sparkle, then clean the container

  1. Open the image

    Drop it above. It is read into this tab; there is no upload endpoint on this site.

  2. Paint over the sparkle, or let the model suggest it

    Suggest the mark for me proposes a mask; nothing is removed until you have edited it and pressed Fill. The filled result is written as a fresh PNG, so it carries no record from the original.

  3. Remove the Content Credentials with proof

    Press Remove metadata, C2PA and hidden files instead on the same file. Pick Container only to leave every pixel byte-identical, or the deepest level to run the learned pass over SynthID's territory as well.

  4. Read the numbers before you trust the file

    The result is re-inspected and the page reports what was removed, what was reduced by how much, and what it cannot verify.

What is measured afterwards

Every result is read back by the same engine that produced it. These are the readings you will see, and what each one is allowed to mean.

  • Checked again on the result

    The container, re-read

    The cleaned file is parsed a second time. If a second pass would still change bytes, the page says so instead of calling the file clean. This is the proof that the C2PA record is gone.

  • Perceptual fingerprint moved N of 64 bits

    At the re-encode level

    How far the re-encode moved the picture's perceptual hash. Usually very little, and the page says so: nudging pixels does not defeat a platform's lookup, and this number is there so you are not told otherwise.

  • High-frequency band reduced by X dB

    At the deepest level

    What the learned pass did to your image, as a measurement. On our benchmark the model reduces an injected residual by 8.5 dB while keeping the picture at 38.9 dB PSNR — a synthetic residual, not SynthID, and the page labels it that way.

What this page will not say. It will not tell you SynthID is gone, because no bytes we can read prove it. Our deeper levels are known to damage marks of this family, and no honest party outside can tell you whether a specific image still trips a specific private detector. The word for that class is "not verifiable", and it stays on the findings list where you can see it.

Sources: Google's watermark settings for Gemini Apps, Google DeepMind on SynthID, how to verify an image in Gemini. Other generators: Sora, every AI mark, the limits.