AI Video Editing Flows

Captions and on-screen text: put them on a card, and know when to skip them

Text that sits loose on busy footage is hard to read. We put captions on a backing card, and we have three cases where we add none. The FFmpeg commands below were run, and one of our first attempts failed.

The rules we follow

  • Text never floats loose on footage. A caption, a stat or a title sits on a backing card, a band or a bubble.
  • Before you add captions, look at the source. If the picture already carries burned-in text, or a face and its captions must stay visible, add nothing on top.
  • Keep captions out of the zones a platform covers with its own buttons. Our free Shorts, Reels and TikTok safe zone overlay draws those zones.

These are our habits for the videos we make. They are not platform rules, and the tests below only show that the commands do what we say.

Put the caption on a card

Busy footage makes loose white text hard to read, so we put it on a dark card. FFmpeg's subtitles filter can draw the card itself, or a drawbox band can sit behind a plain caption. We ran both on one frame of the clip and looked at the result.

A card behind each caption (subtitles filter, box style)
ffmpeg -y -i clip.mp4 -vf "subtitles=cap.srt:force_style='FontName=Arial,FontSize=22,Bold=1,BorderStyle=3,Outline=8,Shadow=0,BackColour=&H99000000,MarginV=40'" -c:a copy card.mp4

BorderStyle=3 asks libass for a box instead of an outline. The box takes its colour from BackColour, written as &H then an alpha byte and the colour in blue-green-red order, so &H99000000 is black at roughly 60% opacity.

A band behind a plain caption (drawbox, then subtitles)
ffmpeg -y -i clip.mp4 -vf "drawbox=x=0:y=ih*0.78:w=iw:h=ih*0.16:color=black@0.65:t=fill,subtitles=cap.srt:force_style='FontName=Arial,FontSize=22,Bold=1,Outline=0,Shadow=0,MarginV=60'" -c:a copy band.mp4

The band covers the lowest part of the frame, which is also where Shorts and Reels put their own buttons, so check it against the safe zone before you ship.

VariantMean brightness of the caption strip (source 0 to 255)Pixels in the strip that changed by more than 40 grey levels
No caption (source)105.80
Plain caption, no card110.23%
Card behind each caption96.513.2%
Band behind the caption62.750.2%

The strip is the rows from 72% to 96% of the frame height, measured on the frame at 2 seconds. The brightness figure shows the card or band really darkens what sits behind the text; it is not a readability score.

When we add no captions

Three cases where we leave captions off. These are our rules of thumb.

  1. The source already has captions. If a creator's footage carries its own burned-in text, a second layer on top is unreadable. We also keep split-screen layouts clear of the speaker's face and caption band.
  2. Nobody on the team can proofread the language. We do not burn captions in a language we cannot check, because a burned-in mistake cannot be fixed without a re-render.
  3. Tour and teardown videos with callout cards. Captions compete with the callouts, so we ship one plain file and add captions only when asked.

What failed or surprised us

  • Our first card command used OutlineColour for the box colour, which is what the ASS format description suggests. It ran without an error and drew no box at all: only 7.6% of the strip changed against 3% for a plain caption. We saw it only when we looked at the frame. FFmpeg's libass takes the box colour from BackColour.
  • A measurement that passes is not the same as a card you can see. Look at the frame.

Method and limits

Every command on this page was run with FFmpeg 8.1.2-full_build-www.gyan.dev on a Windows laptop, from one folder, with the file names shown. The clip is eight seconds of Tears of Steel, (CC) Blender Foundation, mango.blender.org, CC-BY 3.0; other inputs were generated by FFmpeg or by a short script. The measurements are on our test clips, not a survey of footage. Where a rule is our habit and not something we measured, the page says so. Our method is on {link:VID-trust-method}.

  • One clip and one font (Arial). The commands do not depend on the footage, but how readable the card looks does.
  • We measured that the card and band change the picture where the caption sits. We did not run a reading test with viewers.

Page checked on Nov 4, 2026.