One phone shot.
A multicam edit.
Left: the real selfie take. Right: the same take, cut between the real camera and AI cameras that say the same words in the same voice, re-shot from your own take with ComfyUI (LTX-2.3 + CrossView IC-LoRA): same head moves, same mouth shapes, frame for frame, over the real audio, which is never altered.
How it works
1 · Transcribefaster-whisper word timings (snapped to your script).
2 · PlanInsert shots on phrase boundaries, B/C alternation, 5 s generation chunks.
3 · Re-shootComfy Cloud: each 5 s chunk of your real take is the reference video for LTX-2.3 + CrossView IC-LoRA; the prompt is only a camera move. A fresh angle per insert.
4 · CutYour real audio is frozen into the generation, so inserts are time-locked to the source (0–1 frame). ffmpeg cuts A/B/C on word boundaries over the real audio.