Short-form video from the same engine: planned as beats, moved by b-roll grown from real photos, narrated by a real voice with word-timed captions when the format calls for it, and gated by a judge that watches motion, not keyframes. It ships Taka's reels in production today.
Rendered by the production engine
unretouched · sound on
Generated for Driftwood Roasters, a fictional roaster, by the same production pipeline that cuts Taka's reels. Voice, captions, motion, and cuts exactly as the engine returned them.
How a reel gets cut
the video lane of the v2 engine
01BeatsmodelA reel is planned as a narrative in one of five native formats: a hook, value beats, a turn. At most six beats and eight to thirty seconds of story, not a slideshow.
02B-rollmodelReal brand photos lead here too: image-to-video motion from your own shots. Where the pool is thin, text-to-video candidates fill in, name-free as always.
03VoiceovermodelWhen the format calls for narration, a real voice reads the script and the engine keeps the word-level timestamps, because captions that drift half a second read as cheap. Formats that work silent stay silent.
04Captions + overlayscodeOn-screen text is composited by the same deterministic design layer as everything else, timed to the voiceover word by word.
05Motion QCmodelEvery motion candidate faces a vision gate that watches the clip, not a keyframe. Unnatural physics, jitter, and AI slop get rejected.
06The floorcodeWhen every candidate for a beat fails QC, the beat falls back to a branded, deterministic card. The engine never ships slop to avoid shipping nothing.
07AssemblecodeThe final 9:16 master is cut and mixed by ffmpeg in an isolated container: clips, voiceover, captions, overlays, one file.
A timeline, not a black box
beats are layers too
A finished reel carries its composition: the beat timeline, each beat's background clip, the overlay text layers, and the voiceover with its word timings. That is what powers the reel editor in Taka, where every beat can be adjusted and the voiceover round-trips, and it is the same revisable-spec discipline as images and decks: change a beat, keep the rest.
Where video stands
in production · not yet public
Powers Taka today. Coming to the API.
The video lane runs in production behind Taka: every reel Taka ships for its customers is cut by this pipeline. It is not on the public API yet; it joins when it clears the same bar as everything else on this site, and it will land in the changelog when it does. Until then, what you can hold today: 9:16 vertical, eight to thirty seconds, real-photo-first motion, and a QC gate that would rather ship a clean branded card than a warped clip.
