Rescript Studio
Type a new line and the actor says it — voice, lips, and all.
The problem
Video models will generate anything except the one thing you asked for. Putting one exact sentence into one exact shot — same face, same voice, same cut — is the hard version of the problem, and the only version that matters for real editing.
The build
The editor extracts the selected region frame-accurately with FFmpeg and stages it into a ComfyUI workflow built on MiniMax H3 reference-to-video models.
Edits are boundary-conditioned: up to 22 frames of real footage and audio are pinned before and after the new line, so identity, motion, and sound stay continuous across the splice.
Selections under 1.5 seconds switch automatically to a word-replacement strategy; delivery trims the handles, overlaps four frames, and rebuilds audio with de-click crossfades.
Every take records its seed and settings, so any result can be reproduced or interrogated later.
The result
- A dialogue rewrite takes minutes on local hardware. Footage never leaves the machine.
- Quality and 4-step turbo modes trade fidelity for iteration speed inside one workflow.
- Interface, API, and generation pipeline built end to end in a day, then tuned.
Film scenes are brief excerpts used to demonstrate a personal research tool.