DEV Community

Nuco · CleanVideoText
Nuco · CleanVideoText

Posted on Edited on

Why word-by-word captions are harder to remove than subtitles

Someone sends me a clip with big centered captions and asks whether remove subtitles will work. The honest answer is maybe, and the reason has nothing to do with the tool. It's the format.

Word-by-word captions and classic subtitles look similar in a screenshot and behave nothing alike in a repair pipeline.

A subtitle line usually sits in one place, one or two lines at a time, with an outline or a box behind it. Each cue stays for a second or more.

Captions from CapCut, Descript, Opus Clip or TikTok's auto-captions do something else. One or two words per beat, centered or mid-frame, often animated, often with a highlight on the spoken word. The text changes two to four times a second, and it usually sits over the subject's chest or face.

From the repair side, that changes three things at once.

The mask area. A subtitle line might cover five percent of the frame. A centered caption with a highlight box can cover a quarter to a third of it.

The stability. A static subtitle lets the model reuse context across many frames. A caption that pops word by word forces the mask to change constantly, and small inconsistencies read as flicker.

What's behind it. Subtitles sit over clothing, a table, the lower third of the frame. Captions sit over faces and hands, the two things viewers notice first.

Blur and crop fail differently here too. A blurred caption leaves a soft rectangle that moves with the text. Cropping a centered caption means cropping the subject. A sticker recreates the same problem in another color.

Rebuilding the background is the only approach that tries to keep the framing. It works best when the caption sits over something flat or repetitive, and worst when a face is behind the words for most of the clip.

The test takes two minutes. Take three frames, one over a plain background, one over a face or detailed clothing, one where the caption changes mid-motion. Scrub back and forth across the repaired area. If the patch breathes or the texture stops at an edge, the tool is guessing too much. Then confirm the audio and the resolution survived.

Sometimes the honest answer is to go back to the source. If captions cover a face for most of the video, or the shot is a fast pan across a detailed scene, reconstruction is the wrong budget. Turn off the caption track in the edit and re-export. It takes minutes and looks perfect. If the source is gone, fixing a short segment is usually better than trying to repair the whole thing.

I build CleanVideoText. It takes an MP4 or MOV, a 5 to 90 second segment, and one of three modes, subtitles, all text, or a drawn area. It keeps the source frame size and audio, failed cleanups don't use credits, and new accounts get 10 free seconds without a card. That's enough to run the three-frame test on your own footage before deciding.

If your clip has word-by-word captions, remove captions from a video. Judge the result on the three frames above, not on the demo.

Top comments (0)