GemClean

GemClean does one thing with mathematical precision: it removes Google Gemini's sparkle watermark from generated images by inverting the exact alpha blending operation Gemini uses to composite it. No AI, no guessing, no quality loss — just pixel arithmetic run entirely inside your browser.

Overview

When Gemini generates an image, it composites a sparkle watermark at a known opacity using standard alpha blending. The formula is deterministic and reversible. GemClean reads each affected pixel, applies the inverse calculation, and writes the original color back — all via the Canvas API, all offline, all in milliseconds.

The tool auto-detects whether the watermark is 48px or 96px (Gemini uses both sizes depending on the output resolution), handles batch drops of multiple files at once, and packages results as a ZIP for convenient download.

How It Works: The Algorithm

Alpha blending composites a foreground color F onto a background color B at opacity α like this:

Output = α × F + (1 − α) × B

Gemini's watermark is composited at a fixed, known alpha. If you know Output, F (the watermark sparkle color), and α, you can solve for B — the original pixel — algebraically:

B = (Output − α × F) / (1 − α)

GemClean applies this formula per channel (R, G, B) for every pixel within the watermark bounding box. Because the math is exact, the restoration is lossless — no estimation, no blur, no smearing.

/**
 * Restores the original pixel color by inverting alpha blending.
 * @param {number} output - Composited channel value (0-255)
 * @param {number} fg     - Watermark channel value (0-255)
 * @param {number} alpha  - Watermark opacity (0-1)
 */
export function reverseAlphaBlend(output, fg, alpha) {
  return Math.round((output - alpha * fg) / (1 - alpha))
}

The watermark region is located by checking the known pixel offset from the image corner, then the full bounding box is processed in a single getImageData / putImageData pass — no per-pixel Canvas reads in a loop.

Features

Tech Stack

Processing Pipeline