SEEB4 Erase
Remove a watermark
100% free· nothing uploaded· no sign-up

Remove Gemini watermarks instantly

Drop an image or video and download it clean. Your file is never uploaded — the whole conversion runs on your own device, inside this tab, and it stays free forever.

  • Your file never leaves. No server, no API, no account — it works with the network unplugged.

  • Maths, not AI guesswork. The blend that added the mark is inverted, so you get your own pixels back — not invented ones.

  • Video keeps its audio. MP4 in, MP4 out, soundtrack copied packet-for-packet.

Real output · drag to compare

After — cleaned Before — watermarked the mark sits in this corner
Before — still sample, zoomed on the mark
After — same crop, same zoom

Untouched Gemini output on the right, the same file after a drop and a download on the left. Stills come out about 99% clean; video lands near 93%, because it has to be re-encoded.

Clean a file now

Position, size and opacity are measured for you. Most files need nothing but the drop.

Watermark removal

Reverse alpha blending — runs entirely in this browser

Drop an image or video

PNG · JPG · WEBP · MP4 · MOV — or click to browse

Nothing is uploaded — the conversion happens on your device. Pixels are read with Canvas, video frames with WebCodecs, and the file is rebuilt in a tab that never makes a network call. 100% free, forever.

Removes the visible mark only. Invisible provenance signals such as SynthID are not alpha overlays and survive — AI-generated media stays AI-generated, so keep labelling it.

How it works

There is no AI in here. There is algebra.

The generator drew your frame first, then blended a semi-transparent sparkle on top. That blend is one equation per pixel — and an equation that mixed two things can be run backwards to get them apart again.

frame = α·overlay + (1−α)·original

solve for the original

original = (frame − α·overlay) ⁄ (1−α)

Nothing is inpainted, hallucinated or redrawn. The original pixels were still in your file the whole time — scaled and offset, waiting to be divided back out.

  1. 1

    Drop your file

    PNG, JPG, WEBP, MP4 or MOV. Pixels are read with Canvas and video frames with WebCodecs, straight from disk.

  2. 2

    The mark is found

    A matched filter sweeps the whole frame — not just the corner — and locks onto the sparkle's exact position and pixel size, at any resolution.

  3. 3

    α is solved

    Too little correction leaves a bright ghost, too much leaves a dark one. The tool bisects between them until the leftover is zero.

  4. 4

    Download clean

    Images come back as lossless PNG. Video is remuxed to MP4 with the original audio copied across untouched.

Features

Built by measuring, not by guessing

Every default in here had to beat a number on a real file before it shipped.

Finds the mark anywhere

Marks don't always sit in the corner. The search covers the entire frame and lands pixel-exact at 0.95–0.99 confidence — dead centre, top-left, wherever it actually is.

Opacity is solved, not picked

No two-preset menu to gamble with. Given a true α of 1.00, 0.62 or 0.77 it recovers 0.999, 0.618 and 0.769 — including the values presets can't reach.

A score that means something

It reports the percentage of the mark's amplitude actually removed, and warns you when a file came out badly instead of quietly calling it a success.

Video at any resolution

The mark is 48 px at 720p and 72 px at 1080p, so the detector tries each plausible scale rather than assuming one. Audio is copied, never re-encoded.

Calibrate for other marks

A different logo, or next year's sparkle? Feed it one solid-black sample and it measures the overlay's real transparency from your own output.

One file, zero install

No build step, no server, no npm, no extension. Download the folder, open index.html, and read every line of it if you like.

Compare

Why do it locally at all?

SEEB4 Erase Upload-to-a-website removers AI inpainting
Where your file goes Nowhere — stays on your device Uploaded to someone's server Usually uploaded to a GPU
Cost Free, MIT licensed Free tier, then credits Subscription or per-image
What comes back Your original pixels, divided back out Varies — often a blur or a patch Invented pixels that look plausible
Video with its audio Yes — audio copied untouched Often paywalled or absent Rare, slow, expensive
Works offline Yes, network unplugged No No
Can you read the code Yes, all of it, on GitHub No No

The trade-off is honest: this only undoes marks that were alpha-blended on. For anything burned in some other way, an inpainting tool is the right tool.

FAQ

Questions worth asking

Is it really free, or free-until-it-isn't?

Free, permanently. There is no account, no credit counter and no paid tier to upsell you to — there is no server to charge you for. It is MIT licensed, so you can also fork it, host it, or ship it inside something of your own.

Does my photo or video get uploaded anywhere?

No. Your file is decoded, edited and rebuilt by this tab on your own machine, and there is no code path in the project that sends one anywhere. If you want to prove it rather than take our word: open your browser's network tab, or disconnect from the internet entirely and use it as normal.

Which watermarks can it actually remove?

Marks that were alpha-blended over a finished frame — Gemini's corner sparkle being the case it ships tuned for. A mark burned in with a multiply, screen or displacement effect is a different equation and won't come out cleanly. Google's invisible SynthID provenance signal is not an overlay at all and is left completely untouched.

Does video work as well as images?

Almost. Stills reach around 99% of the mark removed; a real 720p clip lands near 93%, because compression and chroma subsampling mean the stored pixels aren't exactly the values that were blended. Video also has to be re-encoded — every tool has that problem, since changed pixels are changed data. Raise VIDEO BITRATE × for busy clips.

A faint mark is still there. What do I change?

Fix the placement, not the strength. Click the preview to move the box, or type X, Y and PX and watch the % removed. If a dark sparkle outline is left behind, switch MARK OUTLINE on. Leave BASE STRENGTH at 1.00 — the solver already assumes it, so anything else scales its answer away from the optimum.

Which browsers does it need?

Images work in any current browser. Video needs WebCodecs, which means Chrome, Edge or another Chromium build. No extension and no install either way.

Is it OK to use this?

Removing a mark doesn't hand you any rights to the work underneath. Use it on output you're entitled to use — your own generations, your own footage, content you have permission for — and check your generator's terms, which may require the mark to stay. AI-generated media is still AI-generated afterwards, so keep labelling it.

Drop a file. Get it back clean.

No upload, no sign-up, no watermark on the way out either.

Measure the mask from your own output

The α mask isn't published anywhere, so measure it instead of guessing.

  1. Ask the generator for a completely solid pure black image; save the watermarked result as PNG.
  2. Drop it here. One black image is enough — over black the blend reduces to α = luma / 255, so α comes out exactly.
  3. Any other background needs a second, clearly different tone (e.g. solid mid-grey) so α and the mark's colour can be separated. PNG only — JPEG noise blurs α.
Drop one solid-black sample (or two tones) — or click to browse