Subtitles are predictable: white text, bottom strip, standard sizes. Watermarks are the opposite — a logo can be any shape, any colour, any opacity, parked in any corner or ghosted across the middle. Detectors tuned to find text miss them entirely or erase only the letters and leave the emblem, which is exactly the half-cleaned look you’re trying to avoid.
So this tool doesn’t guess. You tell it precisely where the watermark lives by drawing a box, and it spends its entire effort inside that box: tracking the mark across every frame, erasing it, and reconstructing the background that belongs underneath — while every pixel outside your box passes through untouched. That precision is also why the box needs a margin: the rebuild blends inward from the box’s edges, and it needs clean pixels there to work from.
The same logic applies to time. If the mark only exists for the first five seconds, upload just those five seconds — trim the clip in any editor before it comes here. Five seconds of careful reconstruction beats a minute of unnecessary work: faster, fewer credits, and the footage you didn’t upload can’t be altered by anything.