We tested burning captions into video with FFmpeg
FFmpeg can draw an SRT caption file straight into the picture, so the captions show on every platform. We burned the captions from our faster-whisper test into three 60-second clips of Tears of Steel, in FFmpeg's default style and a bigger Shorts style, and checked the result frame by frame.
Tested by us on Nov 5, 2026, on the clips and machine listed below. Every number on this page was measured in that run.
The result
It works, and it is predictable. Every one of the 35 captions in our three clips was in the picture at its time, in both styles, and the frames between captions were untouched. The output kept the exact length of the input, with the audio copied as it was, and came out as H.264 High profile in yuv420p, a widely supported format for phones, browsers and social uploads.
Encoding took a median of 29 to 37 seconds per 60-second 720p clip on a busy laptop CPU. The bigger Shorts style took 0.2 to 2 seconds longer per clip, which is within the swing of a busy laptop.
FFmpeg's default style draws small, thin captions near the bottom edge. For Shorts and Reels set the font size, outline and bottom margin yourself, as in our second command.
What failed or surprised us
- Nothing failed in FFmpeg. The thing that failed was our own first check: averaging the whole lower half of a small preview frame hid the thin caption text and reported most captions as missing. Looking at the frames showed they were there; we now count strongly changed pixels in the caption band at full size, and every caption registers.
- FFmpeg's default caption style is small and sits close to the bottom edge, where Shorts and Reels draw their own buttons. Set FontSize and MarginV for vertical video.
What we tested and how
- Tool: ffmpeg-burnin version ffmpeg version 8.1.2-full_build-www.gyan.dev Copyright (c) 2000-2026 the FFmpeg developers, LGPL/GPL (FFmpeg build) licence.
- Settings compared: FFmpeg default caption style; Shorts style: Arial 22 bold, outline 2, raised 60.
- Input clips: Tears of Steel, (CC) Blender Foundation, mango.blender.org, CC-BY 3.0. 60 s excerpts re-encoded to H.264 720p.
tos_a: 60 s starting at 20 s of the film, SHA-256945262dcbe106f1e...tos_b: 60 s starting at 180 s of the film, SHA-2564097773b615cddfa...tos_c: 60 s starting at 360 s of the film, SHA-2563d3a8cd015310349...
- Machine: AMD Ryzen 7 7735U (8 cores, 16 threads), integrated Radeon 680M, not used (CPU-only runs), 31.2 GB RAM, Windows 11 10.0.26300. The laptop was shared with other work, so timings are what a busy machine gives, not a clean benchmark.
- Runs: one cold run, then three timed runs per clip and setting. Run on 2026-11-05.
We took the captions from our faster-whisper test (small.en model), wrote one SRT cue per transcribed segment, and burned them into the same three clips with FFmpeg's subtitles filter: once in FFmpeg's default style and once in a Shorts style. For every caption we compared the frame at its midpoint with the same frame of the source to confirm the caption is really in the picture.
Measurements by clip
FFmpeg default caption style
| Clip | Captions shown / checked | Length change (s) | Output | Median encode (s) | Peak memory (MB) |
|---|---|---|---|---|---|
| tos_a | 11 / 11 | 0 | h264 yuv420p 1280x534 | 29.1 | 359 |
| tos_b | 8 / 8 | 0 | h264 yuv420p 1280x534 | 35.2 | 360 |
| tos_c | 16 / 16 | 0 | h264 yuv420p 1280x534 | 36.3 | 350 |
Shorts style: Arial 22 bold, outline 2, raised 60
| Clip | Captions shown / checked | Length change (s) | Output | Median encode (s) | Peak memory (MB) |
|---|---|---|---|---|---|
| tos_a | 11 / 11 | 0 | h264 yuv420p 1280x534 | 29.8 | 360 |
| tos_b | 8 / 8 | 0 | h264 yuv420p 1280x534 | 37.1 | 361 |
| tos_c | 16 / 16 | 0 | h264 yuv420p 1280x534 | 36.5 | 350 |
Reproduce it
ffmpeg -i clip.mp4 -vf "subtitles=clip.srt" -c:v libx264 -crf 20 -preset medium -c:a copy out.mp4ffmpeg -i clip.mp4 -vf "subtitles=clip.srt:force_style='FontName=Arial,FontSize=22,Bold=1,Outline=2,Shadow=0,MarginV=60'" -c:v libx264 -crf 20 -preset medium -c:a copy out.mp4We ran FFmpeg from the folder that holds the SRT file and passed just the file name, which keeps the filter's path quoting simple. The first command uses the default style, the second sets a bigger bold style for vertical video.
Method and limits
- Captions shown is measured, not judged by eye: a caption counts when more than 0.2% of the bottom quarter of the frame changed by more than 40 grey levels against the source. Frames between captions are the control and changed by 0%.
- We did not measure caption timing against speech here; the timing comes from the transcript, which our faster-whisper test covers.
- 720p clips, H.264 at CRF 20, preset medium, CPU only on a busy laptop. Larger frames or slower presets take longer.
- Footage: Tears of Steel, (CC) Blender Foundation, mango.blender.org, used under CC-BY 3.0.
Page checked on Nov 4, 2026.