Gemini Watermark Remover

Google has shipped several visually distinct watermarks across the Gemini family, and people search for them by different names — Gemini sparkle, Nano Banana, Gemini Omni. They are variations on one mechanism: a small logo composited into the corner of the output at a fixed opacity. Because the compositing is deterministic, all of them can be inverted rather than painted over.

The standard Gemini sparkle is the mark most people mean. It sits bottom-right, occupies roughly 96×96 pixels on a standard render, and blends at around 50% opacity. It is the default on consumer Gemini image output and on Google AI Studio.

The Nano Banana variant is the compact form that lightweight Gemini tiers apply — closer to 48×48 pixels at roughly 35% opacity. It is visually subtler, particularly over busy or low-contrast areas, which is why people often notice it only after zooming in. Mathematically it is no harder to remove; it simply needs the matching alpha map, which Erasio selects automatically.

Gemini Omni output carries the same sparkle family on both images and video. On images the removal is identical to the standard case. On video the mark is burned into every frame, which is a materially different engineering problem and is handled by the dedicated video page rather than here.

You do not need to know which variant you have. Erasio compares the watermarked region against its signature library on load, picks the matching pattern, and applies the corresponding inverse transform. If detection fails — usually because Google has shipped a new variant — the mask controls let you correct it manually.

Most tools that claim to remove watermarks are inpainting: they mask the region and ask a generative model to invent plausible pixels to fill the hole. That works acceptably on flat backgrounds and fails visibly on texture, edges and text, because the model is guessing at content it never saw.

A composited watermark does not require guessing. When Google stamps the mark it computes, for each affected pixel, a weighted blend of the original pixel and the logo colour at a known opacity. That operation is invertible: given the output pixel, the logo colour and the alpha value, the original pixel is recoverable by arithmetic. Erasio solves that equation per pixel.

The practical difference shows up at the boundary. Inpainting leaves a soft halo where invented pixels meet real ones, and smears any detail that ran under the logo. Inversion has no boundary at all, because nothing was replaced — the pixels that were always there are simply restored. Output is written as lossless PNG so the recovered values are not then thrown away by JPEG quantisation.

The limitation is the flip side of the same property: inversion needs the original composited file. A screenshot, a re-compressed copy, or an image that has passed through another editor no longer carries the exact blend, and the arithmetic no longer resolves. Always work from the file Gemini gave you.

Erasio removes the visible logo. It does not touch SynthID, the invisible provenance signal Google distributes across the pixel grid of generated images. SynthID is a separate mechanism, it survives re-encoding, and a file cleaned with Erasio still carries it — the image remains identifiable as AI-generated to anything that reads for it.

That distinction is deliberate. The visible sparkle is a cosmetic mark that interferes with legitimate use of images you generated and own: client deliverables, ad creatives that platforms reject for visible watermarks, product mockups, portfolio work. The provenance signal serves a different purpose, and removing it is not what this tool is for.

Nothing you process leaves your device. Detection and inversion both run in your browser against alpha maps bundled with the page, so the tool works on files you would never upload anywhere, and we never hold a copy of anything you clean.