Tap the thing you want gone. It selects itself, and fills with the colour that was already there.
Click on desktop or tap on a phone, and everything touching that point in a similar colour is selected. Fill it with the surrounding background and it disappears. Before you commit, the page measures whether the fill will actually be invisible rather than telling you it used AI and hoping.
Drag, drop or paste · JPG, PNG, WebP · nothing is uploaded
Best on flat backgrounds: watermarks on a plain plate, logos on white, labels on a solid colour, a sign against a clear sky. On texture it will read as a patch, and the page says so before you commit rather than after.
Clicking or tapping a pixel starts a flood fill: every pixel touching that point within a colour tolerance joins the selection, and the fill spreads until it meets an edge. It is the same mechanism as the magic wand in a desktop editor, running locally on a canvas.
The tolerance slider is the control that matters. Low values select only what is nearly identical to the point you tapped, which is right for a logo on a clean plate. High values reach across gentle shading, which is right for a sky. If the selection leaks into something you wanted to keep, lower it; if it leaves a fringe behind, raise it or add a pixel of edge grow.
The default fill is not a colour you pick, it is the colour that was already there. The page reads the ring of pixels immediately outside the selection, takes the median, and fills with that. On a flat background the erased area becomes indistinguishable from what surrounds it, which is what makes the removal read as removal rather than as a sticker.
You can override it with any colour, or erase to transparent and download a PNG with a real alpha channel, which is the right choice when the image is going on top of something else.
Before you commit, the surrounding ring is measured for uniformity and the result is stated in plain terms: the fill will be invisible, it will mostly blend, or it will read as a patch. That number is the honest predictor of whether this technique suits your particular image.
Above roughly 90% uniform, a flat fill cannot be detected. Below about 65%, the background is textured and a flat fill will show no matter how carefully you select. That case genuinely needs inpainting, which requires a model, and no browser-only tool can do it honestly. Saying so is more useful than calling a flat fill "AI" and letting you find out afterwards.
Good: watermarks over a plain area, logos on white, a date stamp in a corner, a label on a solid product, a sign against clear sky, a colour swapped out of a flat graphic. Anything where the surroundings are one colour.
Not good: a person removed from grass, an object over a detailed pattern, anything where the background continues behind the thing you are deleting. There is no information about what was hidden, so nothing can fill it truthfully. Use the background remover if the goal is to isolate a subject rather than patch a hole.
Keep going without re-uploading — your image carries over to the next tool automatically.