If you have ever generated assets with Google Gemini (Imagen 3, Nano Banana) or Veo 3 for a UI mockup, presentation, or side project, you have probably run into the corner sparkle watermark.
Most people immediately reach for an "AI Magic Eraser" or a diffusion-based inpainting tool to clean it up. As a developer, that always bugged me. Why run a heavy, server-side neural network to hallucinate replacement pixels—usually leaving a blurry, smudged patch—when the original pixels are still mathematically sitting right there underneath the badge?
I built UnWatermarkGemini (and open-sourced the core engine under FalconX) to solve this purely with client-side JavaScript and the HTML5 Canvas API.
Here is how alpha-blended watermarks actually work under the hood, and how you can invert them with zero quality loss.
1. The Math: Inpainting vs. Reverse Alpha Blending
Unlike stock photo sites that stamp a 100% opaque logo over an image (destroying the underlying RGB values), Google Gemini applies a semi-transparent alpha-blended overlay in the bottom-right corner (48x48 or 96x96 depending on image resolution).
When any compositor renders a semi-transparent overlay onto a base image, it computes each output pixel channel ($C_{\text{out}}$) using standard linear alpha blending:
$$C_{\text{out}} = (1 - \alpha) \cdot C_{\text{orig}} + \alpha \cdot C_{\text{wm}}$$
Where:
- $C_{\text{orig}}$ = The true, untouched pixel generated by the model
- $C_{\text{wm}}$ = The watermark's luminance color (e.g.,
255for pure white) - $\alpha$ = The opacity value ($0.0 \le \alpha \le 1.0$) at that exact $(x, y)$ coordinate
Because $\alpha < 1.0$, zero background information is actually lost. Instead of guessing $C_{\text{orig}}$ with AI inpainting, we can simply solve the equation for $C_{\text{orig}}$:
$$C_{\text{orig}} = \frac{C_{\text{out}} - \alpha \cdot C_{\text{wm}}}{1 - \alpha}$$
2. Implementing the Restoration Loop in HTML5 Canvas
Once you extract the exact alpha mask (alphaMap) for the badge dimensions, restoring the corner in browser-native JavaScript takes less than 2 milliseconds on a standard 2D Canvas context:
javascript
/**
* Restores original RGB pixel values by inverting alpha blending
* @param {ImageData} imageData - Canvas ImageData of the corner region
* @param {Float32Array} alphaMap - Pre-calibrated alpha values (0.0 to 1.0)
* @param {number} wmColor - Watermark channel value (255 for white overlay)
*/
function reverseAlphaBlend(imageData, alphaMap, wmColor = 255) {
const data = imageData.data;
for (let i = 0, p = 0; i < data.length; i += 4, p++) {
const alpha = alphaMap[p];
// Skip completely transparent pixels outside the badge shape
if (alpha <= 0.001) continue;
// Clamp divisor to prevent division-by-zero spikes
const invAlpha = Math.max(1 - alpha, 0.05);
// Solve C_orig = (C_out - alpha * C_wm) / (1 - alpha) for R, G, B
data[i] = Math.min(255, Math.max(0, Math.round((data[i] - alpha * wmColor) / invAlpha)));
data[i + 1] = Math.min(255, Math.max(0, Math.round((data[i + 1] - alpha * wmColor) / invAlpha)));
data[i + 2] = Math.min(255, Math.max(0, Math.round((data[i + 2] - alpha * wmColor) / invAlpha)));
}
return imageData;
}
Top comments (0)