AI Watermark Remover

AI Watermark Remover
Finds the sparkle badge Gemini stamps into a corner and rebuilds that area from the picture around it. Drop in as many images as you like — every one of them is processed.
Fills the badge in from its surroundings. Same dimensions, same everything else.
Matches the sparkle against the spot Gemini always stamps it. That spot is repaired either way.
Margin
4px
Every pixel outside the repair stays exactly as it was. A photo saved as PNG grows.
Drop images here, click to browse, or paste with Ctrl+V
PNG, JPEG, WebP · nothing is uploaded
This removes the visible badge only. Gemini images also carry SynthID, an invisible watermark spread across the whole picture, which is untouched here and still identifies the image as AI-generated. Use this on pictures you made yourself.

Images from Gemini — the model everyone calls Nano Banana — come out with a small sparkle stamped into one corner. It is composited over the finished picture rather than drawn into it, which is what makes it removable: the badge is a lid, and the photograph underneath it is still a photograph. This page finds where that lid sits, discards those few thousand pixels, and rebuilds the area from the picture surrounding it. Everything happens in this browser, and every image you drop in is processed, not just the first.

A badge that is composited onto a finished image is a different problem from one that is drawn into it. Gemini writes its sparkle over the picture as a partly transparent layer, so the pixels underneath were once a real photograph and the question is only how to get them back. Knowing the glyph exactly would let you invert the blend and recover the original arithmetically; not knowing it means treating the badge as a hole and rebuilding it from its surroundings instead. The second approach is less exact and far more robust, because it does not depend on which version of the glyph was used, what size it was rendered at, or whether the file was re-encoded afterwards and the original alpha no longer holds.

Finding the badge is the harder half, and a fixed coordinate does not do it. The inset has moved between model versions and the newer ones calibrate it per aspect ratio, so a page that assumes a position will confidently repair empty sky on a portrait crop. What identifies a glyph reliably is that it is bright and thin at once: a morphological top-hat keeps small bright structures and discards large ones, which separates a stamped mark from an overexposed sky that a brightness threshold would flag just as loudly. Judging what survives on how far it stands out from the rest of the corner, rather than on how bright it is, is also what lets a white badge on a pale background be found at all — its absolute lift is barely ten levels, and against a backdrop that varies by almost nothing that is still unmistakable.

Doing this in the page rather than on a server changes what the tool is for. Generated images are often drafts nobody intends to publish — a mock-up for a client, a character sheet, something with a face in it — and a free web service that takes an upload has a retention policy most people have not read. Here the files never leave the tab, which also sets the real limit: a batch is bounded by the memory of the device rather than by an upload quota, and images are decoded one at a time because decoding a dozen large ones at once is the quickest way to have a browser tab killed outright.

How to use AI Watermark Remover

  1. Drop in your images — as many as you like. Clicking the box opens a file picker that takes a multiple selection, and Ctrl+V pastes whatever is on the clipboard, which is the fastest route from a chat window.
  2. Leave the position on "Find it automatically" unless you already know it is wrong. Gemini stamps the badge in the same place on every picture it makes — across 56 measured here, fifty carried it at exactly the same offset from the bottom-right corner and the rest within six pixels — so the page matches the badge's own four-pointed shape against that spot instead of hunting the corner for anything bright.
  3. Press Remove from all. Images are processed one after another with a running count, and Stop halts the run without losing what is already finished.
  4. Check each card. Under the thumbnail are two magnified squares showing the same patch before and after — that is where an imperfect repair shows up, and the thumbnail is far too small to judge it.
  5. Hold the eye button on any thumbnail to flick back to the original. Releasing it returns to the repaired version, which makes the comparison a single gesture rather than two downloads.
  6. Save one image with its own button, or Download ZIP to take the whole batch at once. Names keep their original stem with -clean appended.

Tips & gotchas

  • If a card says the badge was not found, the picture behind it was probably too bright for a white glyph to stand out — a half-transparent sparkle over a spray of water has almost no contrast left to match. The repair still goes to the measured position, which is usually right, so read the two magnified squares before changing anything. Set the corner and the sliders by hand when the picture came from somewhere other than Gemini; the sliders are in percent, so one setting fits every resolution in the batch.
  • The Margin slider widens the repaired area by a few pixels on every side. Raise it when a faint outline of the badge survives — JPEG compression smears a glyph a pixel or two past its own edge, and that spill is what is left behind.
  • Crop the strip off is the honest fallback for a badge that sits on text, a face, or a hard architectural line. Rebuilding those convincingly is not possible from the surroundings alone, and a slightly shorter image with nothing wrong in it beats a full-size one with a smear.
  • Save as PNG when the source was a PNG: every pixel outside the repaired box comes back bit for bit identical. A photograph saved as PNG grows a lot, so for a JPEG original, JPEG at 95 is the sensible trade.
  • Settings are remembered in this browser, so a corner and a size you had to set by hand once are still there next time.
  • Nothing is uploaded. The images never leave the tab, which also means a very large batch is limited by your own memory rather than by an upload quota.

Frequently asked questions

Which watermark does this actually remove?

The visible one: the small sparkle badge that Gemini and Nano Banana composite into a corner of a generated image. The figure usually quoted is 48 pixels square inset 32 pixels on images up to 1024 across, and roughly double that above it — and no picture measured here matches it. Averaging 56 Gemini pictures over a window pinned to the bottom-right corner leaves the badge standing while everything else averages away, and what is left is a four-pointed star about 42 pixels across whose centre sits 98 pixels in from both edges of a 687x1024 original. Two pictures in the set at a different size carry it 103 pixels in, which is the same figure scaled by the size of the picture. That measured offset is what the page looks at, and it turned out to be far more dependable than searching the corner for anything bright and thin.

Does it remove SynthID?

No, and it does not try to. SynthID is a second, invisible watermark that Google spreads across the whole image in the frequency domain rather than placing in a corner. Nothing here touches it, so an image processed on this page is still identifiable as AI-generated by a detector that looks for it. Removing the visible badge and removing the provenance signal are different things, and only the first one is what this tool does.

How does the repair work — is it generating new content?

No. Two steps, both arithmetic. First the page looks for a patch elsewhere in the same image whose surroundings most closely match the surroundings of the badge, so a repair over brickwork is made of that brickwork and a repair over sky is made of that sky. Then it solves a Poisson equation across the hole: the copied patch supplies the texture, and the real pixels bordering the hole supply the colour, so the edge of the repair equals the pixels it abuts exactly. There is no seam to feather because there is no seam. Where nothing nearby is a good enough match it falls back to the smoothest surface that meets the boundary, which is right for a plain background and a soft blur on a busy one.

Does it change the rest of the picture?

Not one pixel, as long as you save as PNG. The repair writes only inside the box shown on the card, and PNG stores what it is given. Choosing JPEG or WebP re-encodes the whole image, which is a small loss everywhere rather than a large one anywhere — usually invisible at 95, and the right choice if the original was already a JPEG, since re-saving a photograph as PNG can multiply the file size several times over.

Can it do a whole folder at once?

Yes — that is what Remove from all does. Select any number of files in the picker, or drag them in together, and each one is detected and repaired in turn with a count showing progress. They are processed one at a time on purpose: decoding a dozen large images simultaneously is the fastest way to have a browser tab killed for running out of memory. Download ZIP then packages every finished image into one archive.

Why does it sometimes say the badge was not found?

Because it would rather say so than claim a find it cannot support. Detection matches a model of the glyph — a four-pointed star drawn from a formula, not a copy of Google's artwork — against the place the badge is measured to sit, then scores how well the shape fits and how far it lifts the whiteness of the picture. What defeats it is contrast rather than clutter: a half-transparent white sparkle over a bright spray of water or a near-white sky leaves almost nothing to match, and two of the 56 pictures measured came out exactly that way. The page still repairs the measured position and labels the card, so "not found" is an instruction to look at the two magnified squares rather than a refusal. Setting the corner by hand is what a picture from a different generator needs.

The repair leaves a faint smudge. What now?

Try three things in order. Raise the Margin slider by a few pixels, since a leftover outline is usually compression spill just outside the detected box. Then check what the badge is sitting on — a repair reads as a smudge mainly where it crosses a hard edge, so a badge over a doorframe or a horizon is the hard case. If it still shows, switch to Crop the strip off, which cannot go wrong in that way.

Is it all right to do this?

For your own generated images, generally yes — and it is worth reading the terms attached to whichever model produced them, because they differ and they change. What is not all right is running someone else’s work through it: a watermark on a stock photograph, a photographer’s signature or a studio’s mark is there to assert ownership, and removing it does not transfer any. The invisible provenance mark on an AI image is also untouched here, which is deliberate.

Related tools
Color Contrast Analyzer
Icon Generator
CSS Generator
Background Remover