The Gemini ✦ mark is already built in (above), so for Gemini images you can skip straight to the Remove tab. Use this only to remove a different fixed watermark: drop ~20–60 images that all carry the same mark in the same corner, and the tool measures it statistically — the watermark is the one thing common to every image, so the varied picture content cancels out. No clean originals or labelling needed.
Learning replaces the active profile above. To get the built-in Gemini one back, just reopen the page in a fresh tab or clear site data.
Click a result to open it in the single-image editor below for fine-tuning.
Nothing here is taken from any watermark-removal product. The watermark profile is measured from your own images using a published, generic statistical idea:
- Learning. The watermark is a fixed, mostly-white overlay added the same way to every image:
observed = original·(1−α) + 255·α. Across many varied images, at each watermark pixel the darkest observed values occur when the underlying picture is dark there, so the low percentile of observed luminance ≈255·α. Taking a low percentile per pixel across the batch (and subtracting the surrounding background level) recovers α directly; picture content, being different in every image, cancels out. - Removal. With α known, the composite is reversed exactly:
original = (observed − 255·α)/(1−α). The watermark's position in each new image is found by correlating the learned α-shape against the local brightening.
Give it enough varied images (20+) and the recovered watermark is essentially identical to the true one.