A watermark is a deterrent, and the honest version says so
Every tiled watermark makes exactly one promise that can be checked: that the mark is still there after somebody crops the picture. It is worth stating in those terms because the alternative claim, that a watermark protects an image, is not true and never has been. Anything drawn on top of pixels can be painted back out, and the tools for doing it went from specialist to one-click somewhere around 2023. What a mark actually buys is friction and provenance. Removing it is now a deliberate act rather than a right-click, and the copy that turns up somewhere else still says where it came from.
So the tiling here is built around the crop, and the geometry falls out of that. Marks are laid on a lattice in their own rotated frame rather than stamped onto the image and turned afterwards, which is what keeps the spacing even when the angle is not zero. Alternate rows are offset half a step, because an aligned grid leaves clean vertical corridors between columns and a narrow crop can fall straight down one. And the lattice is sized so its cell diagonal never exceeds thirty percent of the shorter side of the image. That number is the promise: any point in the picture is then within half a cell of a mark, so any crop keeping thirty percent of the width and thirty percent of the height contains one.
The consequence is that the mode overrules the size you asked for, and it is the only place in this tool that does. Ask for a mark half the height of the image and tile it, and no arrangement of it survives an aggressive crop, because the marks are too far apart to be dense and too big to be close. The engine solves for the largest mark that still satisfies the guarantee and uses that instead. A signature in the corner has no such constraint and gets exactly the size it was given.
Sizing is relative to the image throughout, never in pixels. A mark is a fraction of the shorter side, and the margin is too. That is what makes one setting work across a batch where a 12 megapixel photograph sits next to a 600 pixel thumbnail, and it is also why the live preview can be trusted: the preview is the same code painting the same relative geometry onto a smaller copy, not an approximation of it. Text is measured by its ink rather than its font size for the same reason, since the em box of a line of lowercase and a line of capitals is identical and the visible height is not.
Colour is chosen per mark from what is under it. A coarse luma map of the image is sampled beneath each placement, using the same weights the document scanner uses, and the mark goes dark on light ground and light on dark. This matters more than it sounds: a scan of an ID card is a bright sheet of paper with a dark photograph in the corner of it, and a single global choice of white or black is invisible over one half or the other. A halo in the opposite colour goes underneath, so the mark holds an edge either way.
The one thing this tool cannot avoid is the re-encode. Painting means the pixels change, so unlike putting an image into a PDF there is no byte-for-byte path. JPEG comes back as JPEG at quality 0.92 and PNG comes back as PNG with its transparency, while HEIC, WebP and AVIF become JPEG because a canvas cannot write them back. A side effect worth knowing about is that the EXIF block does not survive, so the GPS coordinates and the camera serial number in the original are gone from the copy you hand out. That is usually what you wanted anyway.