AI Video Editing Flows

One clip, eight caption styles: which one reads best?

Most caption pages show each style on its own demo text. Here every style gets the same five seconds of the same film, with the same 20 words, so the differences are the style and nothing else.

Made and measured by us on Nov 14, 2026, from our own renders over one clip of Tears of Steel (CC-BY). Every number on this page comes from those files.

The result

We put the same five seconds of the film Tears of Steel (film seconds 25 to 30, 20 spoken words) through seven caption styles we built ourselves, plus FFmpeg's plain burn-in as a baseline. Watch the eight videos below, then use the numbers to pick.

By contrast against the footage around the captions, camera follow came out highest (median 4.14:1) and word highlight lowest (2.76:1). The spread is narrow, 2.76 to 4.14, because every style here puts light text over the same bright daylight scene, and none reached 4.5:1. The karaoke pill is the odd one out by design: it carries its own white background, so its dark text measured 4.16 to 4.47:1 against the pill whatever the footage does.

What decided it for us was less the ratio than what each style does to the picture. The karaoke pill covers the most of the frame (up to 8.3% for a caption) and kinetic slam 7.3%, while matrix decode covers the least (1.4%) but shows scrambled characters before each phrase resolves. Camera follow stacked all five lines of this fast speech and its block reached up to 19% of the frame height from the top, over the speakers.

The styles side by side

The same five seconds, the same words, the same crop. Each tile is muted until you press play, and carries the clip's own sound. The last two tiles are the baselines: FFmpeg's plain burn-in and the clip with no captions.

What we measured

StyleContrast against the footage around it (median, lowest)Most of the frame covered (%)Caption block, top to bottom (% of frame height)Transparent render (s for 5 s)Transparent file (MB)
Word highlight2.76:1, lowest 2.72:13.976 to 8957.226
Karaoke pill3.07:1, lowest 3.01:18.373 to 9253.418
Neon glow3.9:1, lowest 3.75:13.269 to 8965.768
Kinetic slam3.65:1, lowest 3.44:17.371 to 915931
Particle burst3.57:1, lowest 3.51:13.568 to 7959.522
Camera follow4.14:1, lowest 3.71:15.219 to 8962.676
Matrix decode3.48:1, lowest 3.32:11.475 to 8467.926

For every caption group we took the frame where the caption covers the most pixels, and read the caption's own pixels from the transparent render. The contrast is the WCAG ratio between the mean luminance of those pixels and the footage in a ring 12 to 40 pixels around the caption. The web accessibility guidelines ask for 4.5:1 for normal text and 3:1 for large text; video captions are not web text, so treat that as a yardstick, not a pass mark.

What failed or surprised us

  • None of our caption compositions takes a transcript. Each has its demo words written into the file, so we copied all seven and rewrote their word lists with a script. The seven files hold the words in seven different shapes (word lists with start and end times, start times only, whole phrases, key-word flags), so there is no one input you can feed them from a transcript.
  • Twenty words in five seconds, about four a second, is fast speech. Styles built for three short phrases got five groups, and camera follow ended up with all five lines on screen at once, covering the speakers from 19% of the frame height down.
  • The matrix decode style shows scrambled characters before each phrase resolves, so for part of every phrase the words cannot be read. Its low screen coverage (1.4%) comes with that cost.
  • Contrast ratios cannot see text shadows, glows or outlines, which all help legibility, so they can understate a style such as neon glow. We checked the frames behind the numbers by eye: the pixels measured are the caption text and highlight boxes. The FFmpeg baseline has a black outline, so we did not give it a ratio.
  • With the settings from our burn-in test the FFmpeg baseline came out smaller on this 1920 by 1080 frame than on the 1280 by 534 clips of that test. We did not tune it for this frame.
  • The transparent renders are large: 18 to 76 MB for five seconds each, against about 1.4 MB for the 1280 by 720 web tile of the same style.

How we made it

Words and timing come from our faster-whisper small.en run on the clip (see the faster-whisper test). We took the 20 words inside the window, grouped them into phrases of up to four words, broke them at punctuation, and wrote them into a copy of each HyperFrames composition. Each copy was rendered with a transparent background (HyperFrames 0.8.10, MOV with alpha, once each, 23 to 35 seconds per render on a busy laptop CPU) and laid over the footage with FFmpeg's overlay filter.

The film clip is 1280 by 534, so it was scaled to 1080 pixels high and cropped at the sides to fill the 1920 by 1080 composition frame. The web tiles are 1280 by 720, H.264 at CRF 21 (rebuilt on 14 Nov 2026 from the same renders; every measured number above came out the same). The eighth tile is FFmpeg's subtitles filter with the Shorts style from our burn-in test, one cue per phrase, over the same crop.

Method and limits

  • One clip, one five-second window, one bright daylight scene. In a dark scene the ratios would change, and the ranking could too.
  • Each style was rendered once; the render times are single runs, not a benchmark.
  • The words are the model's transcription, not the film's official subtitles, so a word or two may differ from the film's script.
  • Contrast is a yardstick for legibility, not a verdict: motion, size, position and how long a caption stays up matter as much.
  • Footage: Tears of Steel, (CC) Blender Foundation, mango.blender.org, used under CC-BY 3.0. All caption styles are our own renders.

Page checked on Nov 4, 2026.