Why a caption band is one box, never eleven
A subtitle is a line of type, but the thing you want gone is almost never just the type. Most players and most burn-in tools lay a translucent band behind the words so they stay readable over a bright shot, and that band has tinted every pixel underneath it. Clear the glyphs alone and you are left with a darkened rectangle with clean letters missing from it, which draws the eye harder than the subtitle ever did.
So box the band. The edges of your box will look for a strong horizontal line within six pixels when you let go, and on a caption band those two lines are usually the clearest edges in the lower third of the frame.
If you cannot see where the band ends, hold Space. The frame as it arrived comes back for as long as you hold (a quick tap switches to it and stays), which is the quickest way to find an edge that a fill has already softened.
One pass or four, and what each costs
One pass, softer
The whole bar is squeezed into the 512 square in one go. It is quick — a single pass at about two and a half seconds — and on a 1080p frame it costs roughly three and a half times the detail along the length of the bar. On a soft or blurred background that is invisible. Against sharp rows of detail immediately above and below, it is not.
Sectioned, sharper
The bar is split into overlapping 512 windows with 64 pixels of overlap (128 over smooth ground) and blended across the joins. Each window sees the frame at close to its own scale, so the strip comes back as sharp as its surroundings. Four sections is about ten seconds, and the panel prints the count before you commit to it.
The cost is the joins. Where two sections disagree about a smooth gradient — a clear evening sky is the worst case — the blend can read as a faint vertical ripple. Each join is corrected for the disagreement measured across its overlap and measured again afterwards; one that still shows is flagged on the region. Over texture it is invisible. Neither of these routes is the right answer for every frame, which is exactly why you are shown both.
Questions about caption bars
Should I box the letters or the whole bar?
The whole bar, every time. A translucent caption band is two problems stacked — the tint of the band and the glyphs on top of it — and clearing only the glyphs leaves a rectangle of altered brightness exactly where the band was. That reads worse than the subtitle did, because a viewer sees a deliberate patch rather than a caption. The exception is a band you have un-blended in the Caption bar panel: once its tint is lifted off the picture, only the letters are left to rebuild.
Why does the strip come back softer than the rest of the frame?
Because a 1,800 by 90 bar does not fit a 512-pixel square without being squeezed about three and a half times along its length. That is what the sectioned route is for: overlapping windows along the bar, each close to the frame's own scale, blended across the joins. It costs four passes instead of one and gives the sharpness back.
The bar sits over a face in one shot. What happens?
You get plausible nonsense, and you will know it when you see it. A face carries structure no fill can guess from its surroundings, so over a mouth or an eye the trained fill invents something confident and wrong. Where a bar crosses a face, cropping the frame or picking a different still is usually the better answer, and this site will not pretend otherwise.
What happens to a translucent backing box?
It can be lifted rather than painted over. Select the band and press "Measure the selected box as the bar" in the Caption bar panel: the band's opacity and tint are fitted from the pixels either side of its edges and printed, and when the fit is good the band is un-blended off the picture first and only the glyphs are rebuilt. When the fit is poor the panel says so and the band is filled as an opaque box.
Can I clear the same bar off a whole set of stills?
Yes. Drop them together and each still is read on its own: the subtitle line is found where it actually sits in that frame, which is what lets stills from different sources, at different sizes, go through the same run. Every row prints the number of regions found, the share of the frame they covered and how much of the lettering is gone afterwards, so a still where the detector found nothing is visible rather than silently passed over.
Read on, or go wider
- Caption bars, and the two ways to clear one Why clearing the glyphs alone leaves a rectangle, and what one soft pass buys against four sharp sections.
- Screenshots, by the folder Flat ground and sharp edges, PNG in and PNG out, and the queue for a set of frames is on by default.
- 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.
