If you have not generated the image yet, you do not need a remover at all. Since August 2026 Gemini has a setting called Media watermark that stops the visible mark being added in the first place. Turn it off, generate, and the file comes out clean.
If the image already exists, that switch cannot help you — it only affects new generations. For those, you need to take the mark off the finished file, which is a different job. Both routes are below.
Route 1: the official setting (for images you have not made yet)
This is the cheapest, cleanest answer and it comes from Google itself. The control lives under Settings > Media watermark in Gemini. Switch it off before you generate, and the visible sparkle logo is not drawn onto the image at all.
It is worth understanding what the switch changed. Before it existed, the visible mark was a fixed part of every output — you could not opt out, and removal meant editing the finished picture. When Google made the mark optional on 14 August 2026, the visible layer became a preference rather than a guarantee.
Which models the setting covers: Google described it as applying to Nano Banana, Omni and Lyria, rolled out across Gemini and the video editor Flow.
We're striking a balance here between creative control and safety: while the visible watermarks are now optional, invisible SynthID watermarks and C2PA metadata are still being used for transparency.
What the official switch does not do
Three limits catch people out, and all three are worth knowing before you rely on it:
- It is not retroactive. Turning it off changes future generations only. Every image you already have keeps its mark.
- It does not remove SynthID. The invisible watermark Google embeds in the pixels — and the C2PA metadata attached to the file — are unaffected. The setting is about the visible layer specifically.
- It is a preference, not a guarantee. Settings can be reset, accounts are shared, and a generation made before you changed the setting will still carry the mark.
That last point is the reason remover tools exist at all. If the switch were the whole story, nobody would need one.
Route 2: removing the mark from an image you already have
This is the case the switch cannot fix. The file exists, the mark is in the pixels, and the only way to get rid of it is to rewrite those pixels. There are three families of approach, and they are not equally good:
1. Crop the corner
The mark sits at the bottom-right, so cropping it away is tempting. It works, in the narrow sense that the logo is gone — but you lose part of the picture, the aspect ratio changes, and the crop is obvious to anyone who knows where the mark normally sits. Fine for a thumbnail, poor for anything else.
2. Paint over it
Clone-stamping or content-aware fill produces a plausible patch. It also produces invented pixels: the area under the logo is replaced with something the model or the brush guessed. It looks right at a glance and falls apart under scrutiny, because the texture no longer matches its surroundings.
3. Reverse the blend
This is the approach our tool takes. Gemini applies its logo by ordinary alpha compositing, which is a known
equation. If watermarked = alpha × logo + (1 − alpha) × original, then the
original can be recovered by solving it backwards. Nothing is invented: the pixels outside the watermark box
are copied straight from your file, and the pixels inside are computed rather than guessed.
The method page sets out the arithmetic and the two watermark geometries we calibrated. The short version is that the recovery is exact when the template matches the image, which is why the tool scores both generations against your picture and keeps whichever lines up better.
Which route should you actually use?
It depends entirely on one question — do you have the image already, or not?
- Not yet generated → use the official setting. It costs nothing, it is lossless, and there is no reason to edit a file that could have been clean from the start.
- Already generated, want to keep the framing → reverse the blend. You keep the full composition and nothing is invented.
- Already generated, mark is in a dead corner you were going to trim → cropping is legitimate, and it is the simplest thing that works.
The one route worth avoiding is painting over the mark on an image whose detail matters. You will be looking at fabricated pixels and calling them your own.
Doing it without uploading anything
Most remover tools require the image to be uploaded to a server, which means a copy of your file exists somewhere you do not control. Our tool runs entirely in the browser tab you already have open: your own device reads the file, the arithmetic happens locally, and the finished image is written back by your browser. There is no account, no sign-up, and no per-day limit — the tool itself explains the boundary on the page, and it deliberately does not remove SynthID.
What none of these routes change
Removing the visible mark is not the same as making an image untraceable, and it is not a licence to misrepresent what you are showing. SynthID and C2PA survive all three routes above. If you are publishing somewhere that asks you to disclose AI generation, the disclosure obligation is unchanged by anything on this page — and the rules around that are a separate question from the mechanics.
Sources
- Josh Woodward, VP for Gemini, announcement post on X, 14 August 2026 — quoted above; the source for the Media watermark setting and the models it covers.
- Google, Generative AI Prohibited Use Policy — policies.google.com/terms/generative-ai/use-policy (last modified 17 December 2024).
- Google DeepMind, SynthID documentation — ai.google.dev/responsible/docs/safeguards/synthid.
- Our own tool's behaviour, measured on this site: only the pixels inside the watermark box are rewritten.