About · before and after, cleared by this site's own engine
One operation, in depth
This site does one thing: it takes burned-in lettering out of a picture and paints the ground back across it. It does not crop, it does not add type, it does not read what the letters say. Everything it does happens in the tab you are reading this in.
The half that is finished
Working out what should stand where the letters were is a solved problem, and this site solves it two ways. The plain one relaxes the pixels inward from the border of your box with edge-aware weights, downloads nothing, and finishes before you have taken your finger off the button. The other is MI-GAN at 512 — 26,639,447 bytes, MIT-licensed, fetched once from our model file server when you press a control that names the size, and kept on your device afterwards with a delete button next to it.
The half that is not
The name of this site promises that it finds the lettering for you, and since 30 September 2026 it does — with arithmetic on the pixels rather than a trained model. It looks for strokes that stand off their ground, closes them into lines and gates them on stroke evenness and baseline, and shows every region it found as a numbered box you tick or untick before anything changes. It misses faint, curved and hand-written lettering and now and then boxes a patch of foliage, which is why nothing is removed until you have looked.
The class of model that would do it is a scene-text detector — the kind that returns polygons around lettering without reading a word of it. Three things have to be true before one ships here: the licence has to be genuinely open, so a research-only weight file is disqualified however well it performs; the file has to be small and quick enough to sit on top of the fill’s cost rather than double it; and it has to work reliably in single-threaded WebAssembly across the browsers this site supports. No candidate has cleared all three, so the finder here uses no weights at all, and nothing on this site claims a trained one.
Why the rectangles are not a stand-in
Because a caption bar, a date stamp and a screenshot header are rectangles already. Tracing a rectangle with anything rounder is busywork, and for these shapes a drag beats it outright — which is why the manual path stays the product rather than a fallback. The finder’s proposals arrive in the same list as the boxes you draw, each with its confidence written as a number, and a drawn box and a found one are filled exactly alike.
Other things not in this build
The queue cannot pick up where it left off after a closed tab. The translucent-bar estimator lives on the story-text page and fits one opacity and tint per bar, so a bar that fades along its length is beyond it. The joins between sections are corrected for the disagreement measured across each overlap and then measured again; a join that still stands out is flagged on its region rather than shipped silently, but it is flagged, not fixed. Rotated lettering is found with an upright rectangle; you can tilt the box yourself, but the finder does not tilt it for you. There is no learned text detector yet, so faint, curved and hand-written lettering is often missed. The page itself needs the network to open; once it is open, and the model is on your device if you use it, every fill runs without it. All of that is written down here rather than discovered later.
Who made it
A small independent team that builds browser tools which keep your files on your own machine. Every word here, the palette, the mark, the text map and the arithmetic printed before and after a run were made for this site and this job alone. If something here is wrong, the contact page is the way to say so.