AI Video Editing Flows

Animated captions with HyperFrames

This workflow turns a transcript with word timing into animated captions using HyperFrames. It covers getting accurate timing, fixing errors, choosing a caption style and rendering the video. The page includes the prompts, the project files and a list of what failed, so you can repeat it on your own footage.

This page is built but kept out of search until it carries the evidence it needs: The lab test needs recorded footage with speech; until a reviewed test exists the page is in testing.

The result

We ran this workflow end to end on a 60-second clip. It took 3.7 minutes on a CPU-only laptop (median of three timed runs), and every caption phrase appeared in the picture. The measurements, the exact script and the frames we checked are on We made animated captions with HyperFrames on a 60-second clip and timed every step.

What we did not test: other caption styles through the same pipeline (see One clip, eight caption styles: which one reads best? for seven styles on five seconds), other clips, or a machine with a GPU.

What failed or surprised us

  • Rendering is slow here: 76% of the 3.7 minutes is the HyperFrames render.
  • The transparent render is 156 MB for one minute.
  • A composition built for a short demo needs its durations changed before it fits a long clip.

The steps

  1. Get word times: faster-whisper with word timestamps on, small.en model.
  2. Fix mishearings in the word list before you build the captions. We did not fix any in this run, so the captions carry the transcript's errors (15.2% word error against the film's subtitles).
  3. Write the words into the caption composition and set its duration to the clip length.
  4. Render with a transparent background: hyperframes render . --format mov --variables '{"plate":false}' --fps 24 --workers 2
  5. Lay the render over your clip with FFmpeg's overlay filter and encode.

The full script is on the test page under Reproduce it.

What it took on this laptop

An AMD Ryzen 7 7735U, 8 cores, no GPU used, 31 GB of memory, shared with other work. Median 225 s for the whole run, peak memory 1.7 GB. Stage medians: transcribe 12 s, render 171 s, overlay and encode 42 s.

Method and limits

The workflow is run on real footage with the tool versions stated on the page. We record timing accuracy, the fixes needed and the steps that did not work first time, and we share the files and prompts used.

Page checked on Nov 4, 2026.