Gemini Watermark Remover
Clean the visible sparkle mark from Gemini images locally in your browser. Focuses on visible mark cleanup and does not claim to remove invisible SynthID watermarks.
Gemini Watermark Remover — Features
Clean visible sparkle mark
Standard & non-standard sizes
100% browser-local processing
The Gemini Watermark Remover is designed to clean the visible ✦ sparkle mark that may appear in the bottom-right corner of Gemini images. Separately, Google may embed invisible SynthID watermarks for AI provenance. This tool focuses on visible mark cleanup and does not claim to remove SynthID. Alpha reversal runs on every image size, not just a list of standard ones: the sparkle is anchored at a fixed inset from the bottom-right corner rather than scaled to the image, so the dimensions do not matter. The Gemini watermark remover picks the 48px template at 32px margins, or the 96px template at 64px margins when the shorter side exceeds 1024px. Inpainting is a fallback for when the alpha pass fails outright, not a path chosen by size. Your image is not uploaded.
How to Use
- Upload your Gemini-generated image (supports PNG, JPG, WebP formats)
- The Gemini watermark remover reverses the alpha blend at the fixed bottom-right anchor, solving original = (watermarked − α×255) / (1−α) per pixel
- Download the cleaned image — the alpha path restores the original pixels exactly, it is not an AI repaint
Why Choose This Tool
- Visible mark cleanup — Targets the bottom-right sparkle mark only
- Honest fallback — if the alpha template cannot load or the canvas is tainted, the Gemini watermark remover detects the sparkle within a 150px scan and inpaints an 80px patch instead
- Two templates, picked by the shorter side — 48px with 32px margins, or 96px with 64px margins above 1024px
The Gemini watermark remover at a glance
Removal here is arithmetic, not repainting: original = (watermarked − alpha x 255) / (1 − alpha), solved per pixel against a known overlay template. On an untouched original the recovered pixels are exact, and nothing outside the mark is modified.
| Input formats | PNG, JPG, WebP. The original PNG straight out of Gemini gives the only exact result — see the note on re-saved files below. |
|---|---|
| Method | Reverse alpha blending. The overlay is a known constant, so the pixel underneath is solved for: original = (watermarked − alpha x 255) / (1 − alpha). |
| Quality | Mathematically lossless on an untouched original — the recovered pixels are computed, not repainted. No model, no inference, no hallucinated detail. |
| Overlay templates | Two: a 48 px mark inset 32 px from the right and bottom edges, and a 96 px mark inset 64 px. The 96 px template is used when the shorter side of the image is over 1024 px. |
| Pixels modified | Only those where the template alpha is at least 0.002. Everything outside that footprint is written back unchanged. |
| Sizes covered | The standard Gemini output grid, including 1024 x 1024, 1536 x 1024 and 2816 x 1536, plus any other size that keeps the same corner inset. |
| Fallback | If the template pass cannot be applied, the Gemini watermark remover locates the sparkle and inpaints it from surrounding pixels. That path is a reconstruction, not a recovery. |
| SynthID | Not removed. The invisible provenance watermark is untouched and still detectable. |
| Price / account / upload | Free, no sign-up, no upload. Processing runs in your own browser tab. |
How the Gemini watermark remover works
Gemini blends a white sparkle into your image with a fixed per-pixel alpha map, which leaves exactly one unknown in the compositing equation. The Gemini watermark remover ships that alpha map as a template — 48 px or 96 px — and solves for the original pixels.
The algebra: the original pixels were never thrown away
Gemini does not paste an opaque logo over your image. It blends a white sparkle into it with a per-pixel opacity map, which in compositing terms means watermarked = alpha x 255 + (1 − alpha) x original. That equation has exactly one unknown. Rearranged, original = (watermarked − alpha x 255) / (1 − alpha), and since the sparkle and its alpha map are fixed for a given Gemini output size, the Gemini watermark remover ships the alpha map as a template and simply solves for every pixel it covers. This is the whole difference from a generic AI watermark remover: inpainting invents a plausible background and gets it slightly wrong in a way you notice at 100% zoom, while this recovers the values that were actually there.
The two guards: an alpha floor of 0.002 and a ceiling of 0.99
Pixels whose template alpha is below 0.002 are skipped entirely rather than being run through the formula, so the faint outer fringe of the mark and everything around it leave the Gemini watermark remover byte-identical to the way they arrived. At the other end, alpha is clamped to 0.99 before the division, because the divisor is 1 − alpha and it approaches zero as the overlay approaches opaque — at alpha = 1 the original pixel was replaced rather than blended and no amount of arithmetic brings it back. The clamp keeps the divisor at or above 0.01 so a near-solid core of the sparkle degrades into a slightly flat patch instead of exploding into noise.
Where the mark is anchored, and what happens when it is not there
The template is applied at a fixed inset from the bottom-right corner: a 48 x 48 px box 32 px in from both edges, or a 96 x 96 px box 64 px in when the shorter side of the image exceeds 1024 px. So a 2816 x 1536 image uses the large template, and a 1024 x 1024 one uses the small template. If the template cannot be applied, the Gemini watermark remover falls back to detection: it scans the bottom-right 150 x 150 px for a bright connected blob between 50 and 800 px in area whose bounding box is no more than twice as long as it is wide, then replaces those bright pixels with the average of the darker pixels within 8 px of each. That fallback is a guess, it works on smooth dark backgrounds, and it smears on busy ones. It exists so you get something rather than nothing — not because it matches the quality of the template pass. The Gemini watermark remover does not branch on which model produced the image, so Nano Banana and Nano Banana Pro output goes through the same pass; the only thing that picks a template is whether the shorter side clears 1024 px.
When people reach for a Gemini watermark remover
Print sellers, bloggers and game developers, mostly — people for whom the sparkle survives everything downstream. It prints at roughly 4 mm on a 20 cm poster, collides with theme furniture on 3:2 heroes, and ghosts through a packed texture atlas.
A print-on-demand seller preparing a mockup
A 1024 x 1024 Gemini image printed on a 20 cm poster at 300 dpi renders the corner sparkle at about 4 mm across — small on screen, unmistakable on paper, and the kind of thing that gets a listing flagged as using someone else's asset. Because the removal is exact rather than repainted, the file that goes to the printer is the same image everywhere except inside a 48 or 96 px square, which matters when the print process will already add its own dot gain.
A blogger generating 1536 x 1024 hero images
The 3:2 hero size is the one where the sparkle collides most often with a caption bar, a play button or a rounded card corner in the theme. Cropping the bottom 60 px away is the usual workaround and it wrecks the composition on an image the model framed deliberately. Recovering the corner instead keeps the full frame, and since the shorter side is 1024 px the small 48 px template applies — the touched area is under 0.15% of the image.
An app or game developer using Gemini output as placeholder art
Icons, tile textures and sprite sheets get scaled, atlased and tinted downstream, and a semi-transparent white sparkle survives all of it as a ghost in the alpha channel or a bright smudge on a tiled edge. Clearing it at the source, before the asset enters the build pipeline, is far cheaper than finding it later in a packed atlas. Batch several files at once; each is processed independently in the tab.
Anyone who will zoom to 100% and check
If the image is going into a retouching workflow, a print proof or a client review where someone will inspect the corner, the distinction between recovered and repainted stops being academic. An inpainting tool leaves a region that is plausible but not correct, and it changes every time you run it. This produces the same output on every run, and outside the mark's footprint the pixels are the ones you uploaded. The full walkthrough of the maths is in this article on reverse alpha blending.
What the Gemini watermark remover will not do
The maths only holds on an untouched original still anchored at its corner. A re-saved JPG amplifies quantisation error into a faint square; a crop or a resize moves the mark off the template entirely. SynthID stays in the pixels either way.
SynthID is not removed, and nothing here pretends otherwise
Every image Gemini generates carries SynthID, a watermark embedded during generation into the pixel signal itself rather than added as a layer or stored in metadata. It is designed to survive cropping, resizing, filters, lossy compression and screenshots, which is exactly why deleting EXIF or clearing a corner does nothing to it. Removing the visible sparkle makes an image look clean; it does not make it undetectable as AI-generated, and any tool advertising one-click SynthID removal is either confusing the two watermarks or overstating what it does.
A re-saved JPG or a screenshot no longer inverts cleanly
The recovery divides by 1 − alpha, which means it also multiplies any error in the input by 1 / (1 − alpha). Where the sparkle is faint that factor is close to 1 and nothing bad happens. Where alpha reaches 0.8, an error of only 2 grey levels — well within what JPEG quantisation introduces — comes out as 10 levels; at 0.9 it comes out as 20. That is the mechanism behind the faint square people report on re-saved files. Always feed the original PNG download; do cropping, resizing and format conversion afterwards.
Cropped or resized images move the mark off its anchor
The template is applied at a fixed inset from the corner, not searched for. Crop 100 px off the bottom, upscale to 2x, or export at a size where Google positioned the mark differently and that anchor no longer lands on the sparkle. The visible result is the worst kind: the sparkle survives and a faintly lightened square appears where the Gemini watermark remover expected it. Process the untouched file first, then edit.
If you can regenerate the image, that beats cleaning it
Google removes the visible sparkle at the source for AI Ultra subscribers and in Google AI Studio, while free and AI Pro tiers keep it. A regenerated output has no mark to recover from and no corner to inspect, so if you still have the prompt and the surface, use it. This Gemini watermark remover is for the case you actually have: an image that already exists, in a deck or a draft you cannot rerun, where the corner is the only thing in the way. Note that regenerating changes nothing about SynthID — that stays either way.
Troubleshooting the Gemini watermark remover
Nearly every problem here traces back to one cause: the file was edited after it left Gemini. The template is applied at a fixed 32 px or 64 px inset from the bottom-right corner, so any crop, resize or re-save breaks it.
A faint square outline is left where the sparkle was
Almost always a re-saved file. If the image passed through a JPEG save, a screenshot, a messaging app or a social platform, the pixel values shifted by a few levels and the division by 1 − alpha amplifies that error by up to 100x at the mark's densest points. Re-download the original PNG from Gemini and run that. If the original is genuinely gone, the outline is as good as the arithmetic can get and the remaining fix is a manual patch.
Nothing changed — the sparkle is still there
The image was almost certainly altered after it left Gemini. The template is placed at a fixed 32 px or 64 px inset from the bottom-right corner, so any crop, resize, rotation or canvas extension moves the real mark away from where the maths is applied. Check the original download; if the sparkle in your file is not roughly 48 px wide at 32 px from each edge (or 96 px at 64 px on images whose shorter side is over 1024 px), the file has been edited. If what you are looking at is a wordmark reading "Created with Gemini Notebook" rather than a four-point sparkle, that is a different watermark entirely and needs the infographic remover, which fills rather than inverts.
There is a lightened square in the corner but the sparkle is somewhere else
Same root cause as the previous item, seen from the other side: the template ran at the anchor and brightened an area that had no watermark in it. It is cosmetic and undoing it is easy — discard the result, go back to the unedited download, and do the removal before any cropping or resizing step.
The corner looks washed out or flat after processing
The pixels under the mark were already very bright. Solving for the original can produce a value above 255 on a near-white background, and the result is clamped into the 0-255 range, so a highlight that was blown out in the corner comes back slightly flatter than its surroundings. This is a property of 8-bit output, not a bug in the detection. It is most visible on white product shots and skies, and least visible on mid-tone or dark backgrounds.
My WebP came back as a PNG and the file got bigger
Expected behaviour. Only JPG input is written back as JPG (at quality 0.95); PNG and WebP both come out as PNG, which is lossless and has no quality setting. Re-encode to WebP afterwards if size matters — but always as the last step, since a lossy re-encode before removal is exactly what causes the faint-square problem above.
An AI detector still flags my cleaned image
It should. SynthID is still in the pixels, and detectors also key on generation artefacts, C2PA metadata and file-level signals that have nothing to do with the corner mark. Removing the sparkle is a presentation change. If you need the image not to be identifiable as AI-generated, this Gemini watermark remover does not achieve that, and neither does anything else that claims to work in one click.
Privacy, copyright and acceptable use
The image is decoded into a canvas in your tab and the two alpha templates are small PNGs served from this site, so nothing is uploaded. SynthID stays embedded, which means a cleaned image is still verifiable as Gemini output.
Your own generations, or images you have the right to edit
The clear case is an image you prompted yourself in Gemini or Nano Banana, which you are cleaning up for your own deck, listing or draft. The unclear cases are the ones worth avoiding: an image someone else generated and shared, an asset a client or platform requires to stay marked, or any use where a clean corner would misrepresent who made the image or how. The technical question of whether a mark can be removed is separate from whether it should be, and this Gemini watermark remover only answers the first one.
No upload, and you can check that in DevTools
The image is decoded into a canvas in your tab and the alpha templates are two small PNGs served from this site. There is no upload request, no processing queue and no stored copy — so there is no retention policy you have to believe. Verify it directly: open DevTools, go to the Network panel, drop in a file, and confirm nothing carrying the image leaves the page. Once loaded, the Gemini watermark remover keeps working with the network disconnected.
Disclosure is a separate obligation from the badge
Removing a visible marker does not remove a duty to say the image is AI-generated. Major platforms require creators to label synthetic or materially altered media at upload, and the EU AI Act places transparency obligations on whoever puts such content in front of the public rather than on the model provider. Because SynthID remains embedded, a cleaned Gemini image also stays verifiable as Gemini output through Google's own detection tooling. Clean the corner for presentation; disclose the origin where it is expected of you.