Questions about removing the Gemini watermark

It removes the visible watermark that Gemini, Nano Banana and Veo composite onto their output: the small mark in the corner of the frame. It does not touch anything else in the picture, and it does not remove invisible provenance marks such as SynthID. If the visible mark is not found and confirmed, your file comes back exactly as you supplied it.

It is free, with no account, no credits and no per-file limit. There is no paid tier because there is no cost to cover: the work runs on your computer rather than on a server someone has to pay for. The removal engine is a fork of an MIT-licensed open source project, credited on the terms page.

No. There is no sign-in, no email field and no cookie banner, because there is nothing to sign in to and nothing being collected. Open the page, drop a file, take the result.

Yes, after the first load. Everything needed to detect and remove the mark from images ships with the page, so once it is cached you can disconnect entirely and keep working. The video path is the exception: it downloads its denoising model the first time you use it, and needs that one fetch to have completed.

The image tool works in current versions of Chrome, Edge, Firefox and Safari, on desktop and on mobile. The video tool needs WebCodecs with H.264 encoding, which in practice means desktop Chrome or Edge. nomark checks for that before it lets you start, rather than failing halfway through an export.

No. nomark has no backend at all. It is a static page, and the removal runs in your browser. Dropping a file produces no network request, which you can verify yourself: open the developer tools network panel, clear it, drop an image, and watch nothing happen.

Your files are held in memory for as long as the tab is open and are released when you close or reload it. Nothing is written to cookies, local storage or IndexedDB. The only persistent thing is your browser’s ordinary cache of the page assets, which clearing your cache removes.

No. There is no analytics script, no tag manager, no pixel, no session recorder and no error reporting service. There is no third-party request of any kind, which is also why the fonts are self-hosted. The privacy page and the terms say the same thing, and both claims are checkable in the network panel.

Load the page, disconnect from the network, then use the tool. If files were being uploaded it would stop working. For a stricter check, open the network panel with request logging preserved and confirm that processing a file adds no entries at all.

For an unmodified original from Gemini, yes. The reconstruction is arithmetic rather than a repaint, and the result is encoded as PNG without lossy compression. The qualifier matters: lossless describes the pixels nomark writes and the file it saves, not a source file that had already lost information before you dropped it in.

Because saving as JPEG would immediately re-compress a picture that was just reconstructed pixel by pixel, undoing part of the work. PNG stores the result exactly. The trade-off is file size: a JPEG in usually becomes a noticeably larger PNG out. Convert it yourself afterwards if you would rather have the smaller file.

The removal is the inverse of the blend that put the mark there, and it needs the composited pixels intact. When an image is re-compressed, whether by saving it as JPEG again, sending it through a messaging app or screenshotting it, those pixel values shift slightly. The inverse then reconstructs from shifted inputs: the result is close but no longer exact, and in bad cases the safety check refuses to write anything at all.

No. Only the pixels inside the detected mark region are rewritten; every other pixel is copied through unchanged. For images that region is typically a square between 36 and 96 pixels on a side, in the bottom right corner. The result card shows you the exact region it used.

It is not carried over. The image is decoded to raw pixels and re-encoded as PNG, so camera data, GPS tags and embedded colour profiles from the source do not survive. Keep the original alongside the cleaned file if you need them. One consequence worth knowing: a JPEG that relies on an EXIF rotation flag is processed in its rotated orientation, which can move the mark out of the corner nomark looks at, and the file will come back untouched.

No, and it does not try to. SynthID is Google’s invisible provenance watermark, embedded in the structure of the image itself rather than composited on top of it. nomark removes the visible mark and leaves everything else alone, so a cleaned image remains identifiable as generated by tools that look for those signals. Any browser tool claiming otherwise is overstating what it does.

The visible marks that current Gemini, Nano Banana and Veo output carry, at the sizes and positions those products use. Detection is built from that known geometry, which is what makes it precise on the cases it covers. If Google changes the mark, the tool needs updating before it works on the new one.

No. This is not a general watermark remover. Stock photo marks, broadcaster logos, social media overlays and hand-added text are all different problems, and the arithmetic that makes this tool exact depends on knowing the precise alpha mask that was used. Anything else would need the kind of generative guessing nomark deliberately avoids.

Usually not. Cropping moves the mark away from the position detection expects and changes the dimensions away from Gemini’s official sizes; screenshotting re-samples every pixel and often re-compresses them. nomark may still find the mark through its local search, but it is much more likely to decide it cannot confirm one and hand the file back untouched.

That is workable, just harder. The size catalogue is a prior rather than a requirement: when a file does not match a known resolution, detection falls back to searching near the corner and validating the reconstruction before writing anything. Whether it succeeds depends on how much of the mark survived whatever resize happened.

The file was returned unchanged, which is the deliberate behaviour whenever a watermark cannot be confirmed. The usual causes are: the image has no Gemini mark, it was cropped or resized after generation, it has been re-compressed enough to blur the mark, or removing the mark would have visibly damaged the picture and the safety check stopped it. The result card names which of those applied to your file.

As many as your machine can hold. Files are processed in parallel across background threads, and the practical ceiling is memory rather than a quota: a batch of 4K images runs out sooner than a batch of small ones. There is no server-side limit, because there is no server.

Yes. Cleaned results can be packed into a ZIP that is built inside the tab and saved directly to your downloads folder. Files that were skipped or that failed are left out of the archive, so what you get is only what actually changed.

Yes, on a separate page, and with a clear warning attached. MP4, WebM and MOV go in and MP4 comes out. It needs desktop Chrome or Edge for WebCodecs, it takes minutes rather than seconds, and 1920 x 1080 is the resolution it handles best.

Three reasons. The mark has to be located across a moving picture rather than in a single still. The clip has to be fully re-encoded, so unlike images the output is not lossless outside the mark region. And residual edges are cleaned by a small denoising model running on your device, which is the one part of the pipeline that is not pure arithmetic. It works, it is simply a heavier and less exact process.

The image tool does. The video tool does not, and nomark blocks it rather than letting it fail: the export holds the entire finished file in memory while a model runs on every frame, and mobile browsers run out of memory well before that completes.

Working with output you generated yourself. The visible mark sits in the corner of every Gemini image, which is a problem when you are using your own generation as a texture, a background plate, a mock-up or a thumbnail. Removing the mark from someone else’s work, or using a cleaned image to pass generated content off as a photograph, is not what this is for, and the terms say so.

That depends on the content, on your jurisdiction and on what you do next, and nothing here is legal advice. Editing an image you generated and own is ordinarily unremarkable. Removing a mark in order to misrepresent authorship, evade a licence or strip provenance from work that is not yours may not be. nomark deliberately leaves invisible provenance signals in place, so a cleaned file remains identifiable as generated content.

The removal engine is a fork of an MIT-licensed project, shipped under that licence with its attribution intact; the notice file in the repository names the original authors. The site around it is written and maintained by Vladimir Mihalevich.