Gemini Watermark Remover

Remove the visible Gemini AI watermark (star/sparkle logo) from generated images using mathematically accurate reverse alpha blending.

Fully offline — pure Python, no external binary downloads, no network access.

When to Use

• Remove the Gemini watermark from AI-generated images
• Batch process a directory of Gemini-generated images
• Clean images before publishing or sharing
• Automate watermark removal in pipelines

Quick Start

Install Dependencies (one-time)

pip install Pillow numpy

# Recommended: use uv for faster, isolated installs
uv pip install Pillow numpy

Requires: Python ≥ 3.9. No Rust toolchain, no compiled binaries, no downloads.

Basic Usage

# Single image (auto-detect watermark, save as photo_cleaned.jpg)
python3 scripts/remove_watermark.py photo.jpg

# Specify output path
python3 scripts/remove_watermark.py photo.jpg -o clean_photo.jpg

# Batch process directory
python3 scripts/remove_watermark.py ./input_dir -o ./output_dir

# Force removal without detection
python3 scripts/remove_watermark.py photo.jpg -o clean.jpg --force

How It Works

Gemini adds a semi-transparent white star/sparkle logo to generated images using alpha blending:

watermarked = alpha * 255 + (1 - alpha) * original

This tool reverses the equation to recover the original pixels:

original = (watermarked - alpha * 255) / (1 - alpha)

The alpha map (watermark transparency pattern) is generated mathematically as a 4-pointed star (central Gaussian core + 4 elongated cardinal rays) at two sizes:

• 48×48 with 32 px margin — images where either dimension ≤ 1024 px
• 96×96 with 64 px margin — images where both dimensions > 1024 px

For improved accuracy you can supply your own alpha map derived from a background capture of the Gemini watermark on a white background (--alpha-map).

Detection

Before removal, a three-stage algorithm checks whether a watermark is present:

1. Spatial NCC (50% weight) — normalised cross-correlation with the alpha map
2. Gradient NCC (30% weight) — edge signature matching via Sobel operators
3. Variance Analysis (20% weight) — texture dampening detection

Images without detected watermarks are automatically skipped.

CLI Parameters