1. What the watermark is

The Gemini sparkle is not painted into the render. It is composited on top of the finished image afterwards, with ordinary source-over alpha blending:

watermarked = α × logo + (1 − α) × original

The logo is white, so logo = 255 on every channel. α is a per-pixel coverage map: near 1 in the solid core of the glyph, tapering to 0 at its antialiased edges, and exactly 0 everywhere else in the frame.

That last part matters more than it sounds. Because α is zero outside the glyph, the overwhelming majority of the image is provably untouched by the watermark — and a remover has no business modifying any of it.

2. Finding the box

Gemini anchors its mark to a corner with a margin that scales with output size, so the starting point is a lookup rather than a search:

Shortest edge ≥ 1400 px → a 128 px box, 84 px from each edge
Shortest edge ≥ 768 px → a 96 px box, 64 px from each edge
Shortest edge ≥ 384 px → a 48 px box, 32 px from each edge

The catalogue guess is then verified rather than trusted. If it does not hold up, the other four plausible anchor positions are swept before giving up. Giving up is a real outcome here: the tool says it found nothing rather than mangling a corner on a hunch.

3. Recovering α without the logo asset

Inverting the composite needs α, and α is not shipped with your image. It gets recovered from the picture itself, in two moves.

First, estimate the clean background. The ring of pixels immediately outside the watermark box is known to be untouched. Those pixels are pinned as a boundary condition and Laplace's equation is relaxed across the interior, producing a smooth membrane that approximates what the area looks like without any glyph on it.

Then, read off the excess. Wherever the observed pixel is brighter than that membrane, the difference is the overlay's contribution:

α = (observed − background) ÷ (255 − background)

This is computed per channel and combined with a median, which throws out the channel that disagrees. A white overlay lifts red, green and blue together; if one channel drops instead, the pixel is image content rather than watermark, and α is forced to zero.

4. Why the α map gets smoothed

This is the step that separates the result from an inpaint. Used raw, the recovered α would invert back to exactly the membrane — a smooth blob with none of the texture that was there before. Useless.

Smoothing the α map first breaks that circularity. α is a property of the glyph, which is smooth and low-frequency; the image underneath is not. Once α is regularised, inverting the composite keeps the observed pixel's high-frequency content and simply rescales it, so grain, edges and texture survive underneath the mark.

5. The unstable core

The inversion divides by 1 − α. As α approaches 1 that denominator approaches zero, and any sensor noise in the numerator gets amplified without limit. At α = 0.95 the noise is magnified twentyfold; at α = 1 the original pixel is simply gone — it was fully painted over and no arithmetic brings it back.

So above α ≈ 0.7 the result cross-fades from the inverted value to a second Laplace solve, this one seeded from the ring of pixels that were recovered cleanly. The fade means there is no visible seam between the two methods.

For the Gemini sparkle this affects a few hundred pixels in the centre of the glyph. On any normal photographic background it is invisible. On a hard geometric edge running straight under the core, look closely — that is the one case where the reconstruction is a guess.

6. Video

The mark occupies the same box on every frame, so α is solved once rather than thirty times a second. Nine frames are sampled across the clip and a per-pixel median is taken. Because the glyph is constant and the scene is not, whatever the moving background contributed to any single estimate cancels out — the merged map is cleaner than any individual one, and vastly cheaper.

Playback is then driven in real time and captured with MediaRecorder, which is why a clip takes about its own duration to process. Audio is routed through Web Audio into the recording graph but never connected to your speakers, so the original track is preserved without the clip playing out loud.

7. What quality actually means here

Dimensions are unchanged. Nothing is cropped, scaled or padded.
PNG output is lossless. Pixels outside the watermark box are byte-identical to the input.
JPEG and WEBP are re-encoded at quality 0.96, because the format gives no other option. If that matters, export PNG.
Video is re-encoded once by your browser's recorder. This is a real generational loss and it is unavoidable in a browser.

8. What this does not remove

SynthID is still in your file. Google embeds an invisible watermark that is spread statistically across the entire image, not drawn in one corner. It survives cropping, resizing, compression and colour adjustment, and it is unaffected by anything described on this page.

Any tool advertising “zero traces” or “completely undetectable” is describing the visible logo and hoping you read it as provenance. Removing the sparkle makes an image look unmarked; it does not make it unidentifiable as AI-generated, and it should not be relied on as though it did.

Try it on a file

How the removal actually works