How Watermark Eraser Works

Media Ingestion & Parsing

When you drop an image (PNG, JPG, WebP) or video (MP4, MOV) into the studio, our engine reads the binary stream directly into memory. Video files are parsed for container headers, video track dimensions, and audio stream codecs without immediate destructive transcoding.

Sub-Pixel Target Detection

The detection engine performs Normalized Cross-Correlation (NCC) across designated corner coordinates (or the entire canvas). It correlates the image against calibrated template tensors to locate target watermarks down to 0.1-pixel sub-pixel offsets, extracting bounding boxes and contour masks.

Algorithmic Deblending & Inpainting

Depending on the overlay characteristics, the appropriate mathematical pipeline is applied:

• Reverse-Alpha Deblending: Inverts the standard alpha compositing equation: I_clean = (I_blended - α × C_watermark) / (1 - α) to restore authentic background colors without blurring.
• Poisson Gradient Reconstruction: For solid or opaque stamps, Laplace equations propagate boundary gradients inward to reconstruct surrounding scene geometry.
• Film Grain Re-Synthesis: Optical noise is synthesized matching the image sensor's ISO profile to ensure seamless organic blending.

Interactive Before / After Verification

The result is rendered on an interactive comparison canvas. You can toggle between Before and After states, inspect pixel-level edge diffusion, or adjust drift gain correction if necessary.

Fidelity-Preserving Export

Still images can be exported losslessly as PNG, or compressed as high-quality JPG or WebP. Videos are multiplexed with original audio tracks copied directly (-c:a copy), delivering master MP4 files with zero audio degradation.