Exact math, not generative guessing

1. Detect

We scan likely corner placements at multiple scales and score each candidate with normalized cross-correlation across luminance, inverse-luminance (for bright backgrounds) and gradient edges. The best match sets the region of interest and a confidence score.

2. Unblend

Visible AI overlays are alpha-blended: watermarked = original × (1 − α) + 255 × α. Since the mask (α) is known, we solve for the original: original = (watermarked − α × 255) / (1 − α). This only touches the detected region.

3. Tune

Detection seeds the tuner (Strength, Size, X, Y). When a source has been cropped or resized, small adjustments align the mask precisely. Reset to Auto restores the detected values.

4. Export

Images export as lossless PNG; video re-encodes to MP4 with audio preserved when possible. Everything happens on your device — there is no upload step, by design.

Privacy by architecture

There is no server endpoint that accepts your media. You can verify this in your browser’s network inspector: selecting a file triggers no upload request.