Drive the face
Engine: AnalyserNode RMS jaw + HeadAudio MFCC visemes (MIT, vendored, in-browser). Photo frames: gpt-image-2 identity edits, one-time generation. No per-minute vendor. No server GPU.
Also on this box
The whole face roster on one screen — the hybrid vector portraits, the Alive cast and the
photo-real set, in three labelled groups — with the state, size and tuning controls beside
them. Click a face and it mounts and drives right there, which is the fastest way to catch
a bad line before a customer does. It is also where face-tuning.json is
corrected and exported.
Worth knowing before you open it: the bench is free to drive, including its ▶ $0 talk control — but its voice copilot has a TALK button that opens a live voice session, and the bench labels that one billed usage. Nothing there spends until you press it.