A watermark has one job: when your picture turns up somewhere you did not put it, people can still see whose it is. Most watermarks fail at that job in the first thirty seconds, because the person taking the picture simply crops the corner away. This article works through the geometry with real numbers, shows what opacity does to the pixels underneath, and ends with the mistakes that make a watermark either useless or ugly.
The numbers come from the layout code of our image watermark tool, which places text or a logo on one or many pictures in the browser, without uploading them.
Size is a share of the picture, not pixels
A fixed pixel size looks huge on a phone screenshot and tiny on a 12-megapixel photo. The tool therefore sets the text height as a percentage of the picture's shorter side:
| Picture | 3% text height | 6% text height |
|---|---|---|
| 4000 × 3000 camera photo | 90 px | 180 px |
| 1080 × 1080 square post | 32 px | 65 px |
| 1080 × 1920 phone story | 32 px | 65 px |
Six percent reads as one clear line of text on a thumbnail, and it stays six percent across a mixed batch, so one setting suits every picture you drop in. For the examples below we take the 4000 × 3000 photo at 6% (180 px tall text) and assume the line © Your name is about six times as wide as it is tall, roughly 1080 px.
How much crop removes a corner mark
Put that mark in the bottom right corner, level, 3% (90 px) from the edges. Its top edge sits at y = 2730 and its left edge at x = 2830. To remove it, a crop only has to take the bottom 270 px off the picture. That is 9% of the height, and the result is a perfectly normal 4000 × 2730 photo with nothing missing that anyone would notice.
Tilting the same mark by 30 degrees makes its bounding box 1025 × 696 px, so its top edge moves up to y ≈ 2214. Now a crop has to remove about 786 px, roughly 26% of the height, or about 1115 px (28% of the width) from the right. That is better, because a quarter of the picture is a real loss, but a determined copier will still pay it for a good photo.
A small corner mark also invites inpainting: today's photo editors can fill a small area from its surroundings in a few seconds, and a corner usually holds sky, wall or floor, which are easy to fill.
Why tiling works
A tiled mark repeats across the whole picture on a rotated grid, with every other row shifted by half a step. With the same 180 px text, 30 degrees of rotation and a gap of 120% of the text height between marks, the tool draws 33 copies on the 4000 × 3000 photo. Change the gap and the count follows:
| Gap between marks | Marks on a 4000 × 3000 photo |
|---|---|
| 60% of the text height | 53 |
| 120% (the default) | 33 |
| 250% | 17 |
There is no crop that removes 33 marks spread over the frame, and inpainting 33 areas that cross faces and products is slow and visible. This is why stock photo previews use a tiled mark. The price is that the picture is covered, so it suits proofs sent to a client before payment, not a portfolio people should enjoy.
What opacity actually does
A semi-transparent mark blends with the pixel underneath. With white text at opacity a, the result is a × 255 + (1 − a) × original. Take an original pixel value of 100 out of 255:
| Opacity | Stored value | Original worked back from it |
|---|---|---|
| 50% | 178 | (178 − 127.5) ÷ 0.5 = 101 |
| 90% | 240 | (240 − 229.5) ÷ 0.1 = 105 |
At 50% the original survives almost exactly: anybody who knows (or estimates) the mark's shape, colour and opacity can reverse the blend and recover the photo underneath to within a level or two. A 2017 Google Research paper (Dekel and colleagues, CVPR) showed that when the same mark is applied identically across many images, its shape and opacity can be estimated automatically from the collection and removed. At 90% the rounding to whole values destroys more of the original, but the mark now hides the picture as well.
So opacity is a compromise, not a lock. Around 25 to 50% keeps the picture pleasant while making the owner obvious. What makes removal hard is covering detailed areas, varying the mark, and tiling, not pushing opacity up.
Common mistakes
- A tiny mark in an empty corner. It is removed with a 5 to 10% crop. Either move it over the subject or tile it.
- Text that disappears on the background. White text vanishes on a bright sky. Turn on the shadow or pick a colour that contrasts with the busiest part of the picture.
- Watermarking the only copy. Always keep the clean original; the tool writes a new file and leaves the source alone.
- Relying on metadata instead. A copyright line in the file's metadata is invisible and many sites strip metadata on upload. A visible mark survives a screenshot; metadata does not.
- One huge, opaque mark on portfolio work. It protects the picture by ruining it. For portfolios, a modest single mark plus a lower resolution upload is usually the better trade.
- Uploading the full-size file. A watermark does not stop someone printing your picture if you publish it at 6000 pixels. Resize for the web first with the image toolkit.
A practical setup
- Client proofs: tiled, 30 degrees, opacity 30 to 40%, gap around 120%.
- Portfolio and social posts: one mark, 3 to 4% size, over a part of the picture that would hurt to crop.
- Product photos for marketplaces: check the marketplace's rules first; some do not allow text on the main image.
The image watermark tool takes up to 50 pictures at once, keeps each file's format, and hands several back as one ZIP. If a picture also contains something you need to hide rather than claim, read why pixelation can be undone first.