Take the Gemini watermark off.

Drop an image in. It is solved away in your browser with the same alpha math that applied it, so the pixels underneath come back.

Drop images here

or click to browse, or paste from the clipboard

PNG, JPEG or WebP up to 50 MB each

Corner of a Gemini image after the watermark was removed
The same corner with the sparkle watermark still on it
Watermarked Cleaned

Drag the handle. The leather grain survives.

This is real output from the engine on this page, not a mockup. Nothing is blurred, cloned or repainted, so texture that sat under the mark stays exactly where it was.

watermark 42 x 42 px at 900, 644

decision validated-match, one pass

output lossless PNG

How to remove a Gemini watermark

Three steps, no account, and the file never leaves the tab. The same flow works for a single picture or a folder of them.

  1. Drop the image in

    Drag a PNG, JPEG or WebP onto the box at the top of this page, click it to browse, or paste from the clipboard. Select as many as you like at once.

  2. Wait about a second

    The mark is located and solved away while you watch. There is no queue and no upload, so a normal Gemini image finishes in roughly the time your browser takes to decode it.

  3. Download the result

    Drag the handle on the result to compare before and after, then download the PNG. A batch comes down as a single ZIP.

Why the result is exact rather than a guess

Gemini does not burn its mark into the picture, whether it came from Gemini itself or the Nano Banana image model. It blends it, with a known logo and a known transparency map. That is an equation with one unknown, and the unknown is your original pixel.

what Gemini does watermarked = α × logo + (1 − α) × original
what this page does original = (watermarked − α × logo) ÷ (1 − α)

Find the mark first

The size and corner offset are matched against the catalogue of known Gemini output sizes, then refined by searching for the logo anchor nearby. If nothing matches, your file is returned untouched instead of damaged.

Solve, then verify

Every candidate removal is scored against the surrounding pixels. A result only ships if it actually reduced the mark, which is why some files come back flagged rather than silently altered.

No model, no guessing

Still images need no neural network at all, so a batch finishes in about the time it takes to decode the files. Video adds a small local denoise model for the frames where compression smeared the mark.

Video works the same way, frame by frame

Gemini and Veo clips are decoded with WebCodecs, cleaned in the corner where the mark sits, and re-encoded to MP4 in the tab. Your audio track is copied across untouched.

Open the video tool

What it handles

Files

  • PNG, JPEG and WebP images up to 50 MB, saved back as PNG
  • MP4, WebM and MOV clips up to 500 MB, exported as MP4
  • Whole folders at once, downloaded as a single ZIP
  • Paste straight from the clipboard

Marks

  • The 48 px and 96 px Gemini sparkle, and the scaled sizes between them
  • Newer outline variants on light and dark backgrounds
  • The Veo text mark, and the Gemini Notebook mark in NotebookLM video
  • Marks that moved inward from the corner in newer outputs

Browsers

  • Images work in any current Chrome, Edge, Firefox or Safari
  • Video needs WebCodecs H.264 encoding, so Chrome or Edge
  • Runs offline once the page has loaded
  • Nothing to install and no account to make

Questions

Does my file get uploaded anywhere?
No. Decoding, removal and encoding all happen inside this tab. The file is read from your disk into memory and never crosses the network. You can open your browser's network panel and watch: once the page and its code have loaded, no file data leaves. The video tool fetches its cleanup model from this site once, and that request carries nothing about your video.
How is this different from an AI eraser?
An AI eraser invents plausible pixels to cover the mark. This inverts the blend that Gemini applied, so for still images the pixels underneath are recovered rather than imagined. Video is the one place a small local model helps, because compression smears the mark before it reaches you.
Some images come back with a note instead of a clean result.
That is deliberate. If the detector cannot confirm a Gemini mark, or the removal scored badly against the surrounding pixels, the tool says so rather than quietly damaging the corner. Re-exporting the image from Gemini at a standard size usually resolves it.
Why is video slower than images?
Every frame is decoded, cleaned and re-encoded on your own hardware. A ten second clip runs roughly in real time on a normal laptop. The first run also fetches a 2.7 MB cleanup model from this site, once.
Is it free?
Free, with no account, no quota and no payment. The removal engine is the open source project by GargantuaX, MIT licensed, and it is bundled here unmodified.
Is it safe, and do I need to sign up?
No sign up, no account and no upload. Nothing to install either, so there is no app asking for permissions on your machine. The page is static and the processing code runs inside the tab, which is why it keeps working with your connection switched off.
Can it clean a whole batch at once?
Yes. Drop in as many images as you like, or a whole folder's worth. They queue up, each one gets its own before and after, and the finished set downloads as a single ZIP. Nothing is capped and nothing is metered.
Does it work on Nano Banana images and Veo 3 clips?
Yes to both. Nano Banana output carries the same Gemini mark, so the image tool treats it identically. Veo and Veo 3 clips are handled by the video tool, which also covers the Veo text mark rather than only the corner sparkle.
Am I allowed to remove these watermarks?
That depends on where the content came from and what you do with it. The mark is a provenance signal, and removing it from work that is not yours, or using it to pass AI output off as something else, is on you. See the terms.

Further reading

Longer write-ups on how the marks are built and what it takes to lift one off.

All guides