Forty captures, forty separate answers
The reason this page exists is the folder. Somebody has forty captures, each carrying a header, a watermark strip or a name they would rather not publish, and clearing them one at a time is forty repetitions of an identical drag.
Drop the whole set. No rectangle is asked for: every capture is scanned for lettering on its own pixels, and the row for it says how many regions were found, what fraction of the frame they covered, whether they were laid over the capture or part of what was photographed, what filled the holes, and how much of the ink inside those regions is actually gone once the fill has run. What comes back is a ZIP, written here in the tab.
What the queue will not do
It will not pretend to have found something it did not. A capture the detector reads as having no lettering in it is reported that way and handed back exactly as it arrived, byte for byte, rather than quietly re-encoded — and the sensitivity control says which way it is biased, because a text remover that silently misses the text is worse than one that says it found nothing.
It will not survive a closed tab. There is no resume in this build, so a run of two hundred frames wants a window you are not about to shut. That is a limitation of what is shipped today rather than a design position.
Flat ground is why this page is the easy one
An interface panel has no texture and, usually, no structure crossing it. Both of those are what make a fill hard. Over a flat or softly graded panel the diffusion fill — the one that downloads nothing — is effectively invisible and finishes before you have let go of the button, and there is very little the trained model can add.
Where a screenshot does get hard is a second line of smaller type immediately under the one you boxed, or a table rule running through the box. Structure crossing a border is the case where a trained fill invents a confident, wrong surface, and the box list marks those regions before you run them.
Questions about screenshots
Why are screenshots the easy case?
Because the ground under the lettering is usually flat. An interface panel is one colour or a gentle gradient, and both fills rebuild that perfectly — the diffusion one does it in a fifth of a second with nothing downloaded. The hard cases on this site are photographs with texture or structure behind the words, and a screenshot rarely has either.
Should I save as PNG or JPG?
PNG, unless something downstream insists otherwise. A screenshot is flat colour and hard edges, which is exactly what JPG handles worst: re-encode a cleared caption bar as JPG and it picks up ringing along every edge, right where you have been looking. PNG is the format this page offers first for that reason.
How does the queue decide what goes together?
It does not group them. Each capture is scanned on its own pixels, so two frames at different sizes, or the same frame with its header in a different place, are each answered separately — which is the opposite of the older behaviour here, where anything that did not match the first frame's dimensions was set aside untouched. What every row reports is that capture's own count of regions, the share of the frame they covered, and the ink left inside them after the fill.
Can I stop a long run part-way?
Yes. Pause takes effect after the frame in progress, and the finished ones are still yours — the ZIP holds whatever completed. Closing the tab ends the run for good, though, because nothing here is written to your disk without you pressing save.
Elsewhere
- Subtitle strips Opens with a wide box along the lower third, which is where a burned-in caption almost always sits.
- Flat ground, busy ground The single best predictor of whether you will be pleased with the result, worked through on three grounds.
- ImageTextRemover Draw a box over the lettering, keep the boxes you meant, and get the picture back with the ground painted across them.
- Lettering somebody else put there What clearing a mark does and does not change about who owns the picture underneath it.
