A caption bar is not part of the photo
Most captions people want gone were never in the scene. A meme caption sits on a white bar stacked above the picture; a subtitle from a video frame sits on a dark strip at the bottom; a news screenshot carries a coloured ribbon with a headline. In every one of those cases the pixels under the words are not missing detail — they are simply more of the bar. That is why this page does not send the area to an AI that would invent texture. It measures the bar instead and writes the bar back.
How the band finder decides what a band is
For every row of the photo the finder fits a straight colour line from one side to the other, so a horizontal gradient counts as flat, and checks whether most of the row sits close to that line and whether both ends of the row agree with it. A caption bar runs edge to edge; a sky does too, but a sky does not hold lettering. Rows that pass are grouped into bands as long as their colour follows on smoothly — the rows that carry the lettering are kept in, because their two ends still sit on the bar's colour. A band then has to look like a caption bar: touching the top or bottom of the picture, or, in the middle of the picture, a solid strip with sharp borders. Finally it must hold lettering, meaning many small off-colour marks rather than one big shape — a sun over the sea is a shape, a line of words is many marks. Tall gradient areas are rejected, because a tall gradient is usually sky.
Erase the words, keep the bar
When you choose Erase the words, each row of the band is modelled from its two ends — the left and right edges, where lettering almost never reaches — as a colour that may change steadily across the row. Those row models are smoothed over five rows so the result has no stripes, the band's own fine grain is measured and added back, and only the lettering pixels (grown by two pixels to take their soft edges and outlines) are replaced. Every other pixel of the photo is copied through untouched. On a solid or evenly graded bar the result is exact: there is nothing to guess.
Or crop the band off
Sometimes the bar itself is the problem — you want the picture without the meme frame. For a band at the top or bottom, Crop off removes it. The button shows the size you will end up with, and the confirmation repeats the before-and-after dimensions, so a 1200 × 800 meme whose caption bar is 151 pixels tall becomes 1200 × 649 and you know it before you press anything. Cropping is one step in History and Undo restores the bar.
When the words sit straight on the photo
If the caption is printed over the scene itself — white letters across a sky, a date stamp on a beach — there is no bar to refill, and the finder says so. The Remove text from image tool is the right one then: its finder locates the lettering, flat surroundings are fitted exactly, and busier ones are rebuilt from their surroundings. To change the words rather than remove them, Edit text in image erases the old text and gives you a text box in its place.
Your own pictures
Use this on memes you made, screenshots of your own videos, and photos you have the right to edit. Removing a caption that credits a photographer or carries a copyright notice is not something this tool is for.