Skip to content
VVideoTextRemove

Complex background removal

Removing subtitles over a moving background

Most subtitle removers are demonstrated on a locked-off shot against a plain wall. Real footage is not like that: the camera pans, people walk through the lower third, leaves move in the wind. Whether a tool survives that is the only question worth asking.

This page is the honest version. Two of the three clips we publish are ones we would tell you not to use this tool for, and the measurements say so.

OriginalCleaned

The hard case, unedited: pedestrians cross the caption band and leave soft vertical ghosts where they passed. The text is gone; the people are not restored. Drag to compare.

Doing it for this case

  1. 1

    Run ten seconds of the hardest part of the clip — the moment with the most movement behind the text.

  2. 2

    If people or vehicles cross the caption, expect ghosts and consider a different approach for that shot.

  3. 3

    Switch to Pro erase for textured or moving backgrounds: the standard tier smooths where Pro reconstructs.

Three kinds of movement, three different outcomes

Camera motion over a static scene is the easy case. The background is consistent, merely displaced, so there are neighbouring frames of the same texture to sample. Our foliage test is exactly this, and it scored 28.3 dB.

Subject motion through the caption band is the hard case. A person who crosses the text exists in some frames and not others, so there is nothing stable behind them. Our crowd test scored 19.8 dB and the result is visibly wrong.

Regular texture in a tight crop is the quiet failure. Wood grain, railings and tiled floors give the model a pattern it can average into flat grey — 20.5 dB in our vertical test, only 1.3 dB above a blur box.

What the numbers mean in practice

PSNR is a blunt instrument, but the ordering matches what you see: 28.3 dB holds up at normal size, 19.8 dB shows ghosts of the people who walked through, and 20.5 dB reads as a grey smear.

Untouched areas of the same frames sit at 37.4 dB, so none of these results is perfect — each is some distance from the original pixels. We publish the distance instead of calling it clean.

Getting the best result out of hard footage

Use Pro erase on anything with texture or motion. It costs eight times the standard rate per minute, and on hard footage it is the difference between a soft strip and a plausible one.

Keep the box tight. Covering extra background gives the model more area to invent and makes the repair worse, not better.

Test ten seconds before committing an hour. A failed job refunds its minutes, but your own time is not refunded.

Tested, not assumed

This is the page where we publish the failure — 19.8 dB, pedestrians smeared into vertical ghosts.

The caption is gone and the pavement is rebuilt cleanly, but everyone who crossed the text left a soft vertical trace. At normal size it reads as motion blur; at 100% it is obvious. The same pipeline scored 28.3 dB on camera movement over foliage, so the variable is subject motion, not movement in general.

Where it breaks

  • People or vehicles crossing the caption band: expect ghosts. This is the measured worst case.
  • Regular repeating texture in a tight vertical crop: the repair goes flat, 1.3 dB above a blur box.
  • Text over faces, or covering most of the frame, is beyond what background reconstruction can do.
See the footage and the measurements

Frequently asked questions

Does camera movement break it?+

No. Camera motion over a static scene is the case it handles best — our foliage test has continuous movement and scored 28.3 dB. The background is consistent, only displaced, so neighbouring frames provide what the repair needs.

What happens when people walk behind the caption?+

They get smeared. Measured 19.8 dB, with visible vertical ghosts where each person crossed. If that describes your shot, this is not the right tool for it.

What about water, waves or smoke?+

Untested. Those are non-rigid, constantly changing surfaces with no stable structure to sample, which puts them closer to the crowd case than to foliage. Expect artefacts rather than a clean rebuild.

Can I improve the result on hard footage?+

Two things help: a tighter box, and Pro erase instead of standard. Neither fixes subject motion — the information is not present in the other frames to recover it.

Should I use this on a crowded street scene?+

No. Use it where the caption sits over scenery, sky, walls or foliage. On crowds, cropping or reframing will look better than a reconstruction, and it costs nothing.

Related tools

Try it on your own clip

Sign in and get 1 free credit — enough for a one-minute video. No card required.

Open the editor