Captions

All five actions live under Composition ▸ … and in the command palette:

A caption is a text layer#

There is no caption object, no caption track and no caption renderer. A caption is an ordinary text layer whose clip bar is its cue window, marked with one hidden prop (__caption) so it can be found again.

That is the whole design, and everything follows from it:

• Captions get per-glyph animators, layer styles, 3D, the graph editor and motion blur, because they are text layers and text layers have those.
• Re-timing a caption is dragging its bar. Export reads the bars, so the file agrees with the picture — nothing stores the original cue times, deliberately, because two sources of truth for "when does this show" is exactly how an exported .srt ends up disagreeing with the render.
• Restyling one caption is restyling a layer. Restyling all of them is selecting them all and restyling a layer.

Import and generate both replace any captions already there. A second import over an unremoved first is forty layers of doubled text, which reads as a rendering bug rather than as the user's own second import.

Styling#

New captions are sized and placed from the composition, not from fixed pixels: cap height is 5% of comp height and the baseline sits a tenth of the frame off the bottom, clear of the UI chrome every social platform draws there. The same style therefore reads correctly on a 1080×1920 vertical cut and a 4K master, which a hardcoded 48px does on neither.

Long captions are wrapped in the model, at ~42 characters over at most two lines (broadcast practice), rather than being left to the renderer to wrap on width. That means the line breaks survive an export and are identical on every machine, whatever fonts are installed.

Formats#

One parser reads both SRT and WebVTT, because every real file blurs at least one of the differences between them. It handles what actual exporters produce:

• CRLF line endings and a leading BOM
• .vtt written with SRT's comma, and .srt written with a period
• MM:SS.mmm timestamps with no hour field
• WebVTT cue settings trailing the end time (align:middle line:90%)
• NOTE and STYLE blocks
• files with no blank line between cues
• two-line captions, kept as two lines

A cue with no words is dropped; a zero-length cue is given a one-frame floor rather than having its words discarded. Overlapping cues are trimmed against each other on the way in — two captions on screen at once is a transcript artefact that renders as text over text.

Generating from audio#

Composition ▸ Generate Captions from Audio mixes the composition down and sends it for transcription.

It transcribes the composition, not a footage file. Layer trims, levels and mutes are all applied first, so the cues line up with the picture; transcribing the source file of a clip that appears at 00:42, trimmed, would produce captions timed to the file instead. If a work area is set, only that range is transcribed and the cues are re-based onto composition time.

The audio is prepared as 16 kHz mono — six times smaller than an export mixdown, which is the difference between "transcribe this" and "the provider refuses anything over 25 MB". At that rate the cap is about 13 minutes; set a work area over a longer piece and run it in passes.

Provider: OpenAI, via the key in Settings ▸ Assistant. The call is made from the Electron main process, so the key never enters the renderer — the same custody as image generation (electron/aiProxy.ts). Anthropic has no audio API, and Gemini returns prose without timings, which cannot become captions; a request naming either says so rather than failing obscurely. The command is disabled, not hidden, where the shell cannot transcribe at all.

There is no hosted transcription route, so this is a local-edition feature today.

Source map#