Gemini AI Watermark Remover
Free & Browser-Based

Drop the original file onto the upload area at the top of this page. The tool works out the watermark's size and position from the image dimensions, solves for the pixels underneath, and shows you the result. Download it as a PNG. There is no selection step and no account. Use the file Gemini gave you rather than a screenshot.

Use the Videos tab, which opens a separate beta tool. Pick a short MP4, press process, and watch the frame counter. Be aware the video calibration is still provisional, so quality is unverified and varies between clips. It needs a desktop Chrome, Edge or Opera browser, and audio handling has not been tested yet.

Yes. Select as many files as you like and the tool switches to a batch list. It works through them three at a time rather than all at once, showing a running count and a per-image status. When the batch is done you can download everything as a single ZIP, or grab individual images separately.

No. There is nothing to install, no extension, no sign-up and no email address. The tool is a web page; opening it is the whole setup. Because processing happens in your browser rather than on a server, there is also no account needed to track usage or quota.

Yes, with no usage limit. There is no paid tier, no credits and no watermark on the output. This is possible because your browser does all the work — there are no per-image server costs to recover. The page carries standard analytics, which is unrelated to your images.

No, outside the watermark square nothing is touched at all — those pixels are copied through unchanged, and the PNG export is lossless. Inside the square the pixels are reconstructed by reversing the blend, which is accurate where the logo was semi-transparent. In the most opaque parts of the logo, very little original signal survived to recover.

It is solving the compositing equation backwards. The watermark was applied by mixing the logo into your image at a known opacity, which is a reversible operation wherever that opacity is below 1. Given the opacity at each pixel, the original value can be calculated directly rather than guessed. Nothing is generated or copied from elsewhere in the picture.

Because they cover the watermark instead of undoing it. Blur averages the logo's bright pixels into their neighbours, leaving a pale haze and a soft rectangle against sharp surroundings. Eraser and clone tools copy pixels from elsewhere, which duplicates visible features or lands at the wrong brightness. Neither knows what was actually underneath.

JPG, PNG and WebP. Other formats are ignored when you select them. Output is always PNG regardless of what you put in. If your file is in another format, convert it first — but be aware that converting re-encodes the pixels, which can slightly degrade the result.

20 MB per image. Files above that are skipped rather than processed. There is no separate limit on resolution — a large PNG well under 20 MB is fine — though very large images are ultimately bounded by how much your browser can hold in a canvas. In a batch, the limit applies to each file individually.

Because JPG would re-compress the pixels just recovered. JPEG is lossy and works in 8×8 blocks, adding artefacts around high-contrast edges — precisely the corner just worked on. PNG stores the values exactly. If you need JPG for size reasons, convert afterwards, knowing that conversion introduces its own loss.

Two: 48 × 48 px and 96 × 96 px. The 96 px version is used only when both width and height are above 1024 pixels; everything else uses 48 px. Margins differ too — 64 px from the right and bottom for the large mark, 32 px for the small one. The preview panel shows which was chosen.

Then this tool cannot remove it. The position is calculated arithmetically from the image dimensions, always anchored to the bottom-right. There is no search step and no manual selection, so a mark anywhere else is outside what the tool handles. A common cause is an image that has been cropped or rotated since generation.

Almost always because the file is not the untouched original, or the mark is not where the size rule expects. Screenshots, re-saved copies, cropped or resized images and non-Gemini sources all break one of those assumptions. Anti-fingerprinting browser extensions can also interfere by adding noise to canvas pixel reads. The troubleshooting section above covers each case.

Yes, when the output carries the standard Gemini corner watermark. Nano Banana is a nickname for one of Google's Gemini image models, and its outputs are marked the same way. What matters is not the model name but whether the visible logo is 48 or 96 pixels in the bottom-right corner of the original file.

Not reliably yet. The video tool will attempt any MP4, but its calibration comes from the still-image captures because no video-specific capture has been made. That means quality on Veo output is unverified and varies. Treat it as an experiment, check every result yourself, and do not rely on it for anything that matters.

Only if the image carries the same Gemini visible watermark at the same size and position. The opacity maps were captured from that specific mark. A different watermark has a different shape and opacity profile, so subtracting this one produces a wrong result in the corner rather than a clean image.

No, or at least not well. A screenshot has been scaled to your display, colour-managed and re-encoded, so the watermark is no longer exactly 48 or 96 pixels and no longer sits at the expected offset. The calculation then subtracts the wrong values from the wrong place. Download the original file from Gemini instead.

Because the arithmetic assumes the pixel values are the ones Gemini wrote. Screenshots rescale and re-encode; re-saving as JPG shifts values, most around high-contrast edges like a logo. The inverse blend divides by a small number, which magnifies any error already present. Small deviations in the input become visible residue in the output.

No. This tool only removes the visible logo from the corner. SynthID is a separate, imperceptible signal embedded across the pixels of the image, and nothing here touches, weakens or detects it. It is designed to survive re-saving, compression, resizing and cropping, so ordinary editing will not remove it either.

Yes. Removing the visible logo changes nothing about SynthID or any other embedded signal, and a detector with access to SynthID verification can still identify the image. If your purpose depends on the image not being identifiable as AI-generated, this tool does not achieve that, and you should not assume otherwise.

A visible watermark is a logo composited into the picture where you can see it — bounded to one region, and reversible because the original pixels were mixed rather than discarded. An invisible watermark such as SynthID is a pattern spread across the whole image below the threshold of perception, designed to be machine-readable and to survive editing.

No. Your file is decoded, modified and re-encoded entirely inside the browser tab; there is no upload step in the code. You can verify this yourself: open developer tools, watch the Network tab while processing an image, and you will see no request carrying your file.

No, because your files never reach us in the first place. There is no server-side storage, no processing queue and no log of images, since nothing is transmitted. The page does load standard web analytics that record anonymous page usage — visits and clicks — which is separate from your images and never includes them.

Yes, for images, once the page has finished loading. All processing happens in code already running in the tab, so disconnecting the internet and then processing an image still works — a useful way to prove nothing is uploaded. The page's fonts and styling load from external CDNs, so it may look plain while offline.

It depends on the image and what you do next, and we are not lawyers. Removing a mark from an image you generated yourself for your own use is a different matter from removing one to misrepresent authorship, or from stripping marks off someone else's work. Local law, Google's terms and the context all matter.

Possibly — read Google's current terms rather than trusting a summary here. Terms for Gemini and related products change over time, and what they say about watermarks and generated output is the authoritative answer for your account and region. If you are unsure, or the use is commercial, check directly or take advice.

That depends on your rights to the image, not on this tool. Whether Google's terms permit commercial use of your generations, and under what conditions, is the deciding factor — removing the logo does not grant rights you did not already have. Where disclosure of AI generation is expected, that expectation still applies.

Often yes, and if you generate a lot it is the cleanest route — no post-processing at all. What each paid tier includes changes over time, so check Google's current pricing rather than relying on this page. The limitation is that a subscription does nothing for images you have already generated and saved.

Photoshop is far more capable at everything except this one job. For this specific task the difference is what happens in the corner: generative fill invents plausible new content, while this reverses the known blend to recover the original pixels. Photoshop also costs money and processes images on Adobe's servers rather than locally.

Both are more flexible and both are better inside their own workflows. For this task, each replaces the region rather than reconstructing it — Canva by patching from surrounding pixels, ChatGPT by regenerating, which gives a different result on each run. Both require an account and an upload; this runs locally and gives the same output every time.

Cropping is a perfectly reasonable answer, and sometimes the right one — it always works, adds no artefacts and takes seconds. The cost is losing real image area and changing the aspect ratio. For a 96 px mark set 64 px from the edge you are cutting well into the frame, which matters on a tight composition.

For images, yes — the tool runs in mobile browsers and the interface adapts to small screens. Very large images may be slower or run out of memory on a phone. For video, no in practice: the video tool needs desktop-class browser features and enough memory that mobile devices are not a realistic target.

For images, any current version of Chrome, Edge, Firefox, Safari or Opera, on desktop or mobile. For video the requirement is narrower: the WebCodecs API, which currently means desktop Chrome, Edge or Opera. If a browser lacks it, the video page names the missing features and stops rather than failing partway through.

Two reasons. It needs the WebCodecs API for decoding and encoding, which mobile browsers largely do not expose. It is also memory-heavy: frames are decoded, modified and re-encoded continuously, and a phone will typically run out of memory and drop the tab. The tool warns you on small screens and low-memory devices but lets you continue anyway.