Cassie — Talking-Head Bake-Off

Pick the engine for her real videos.

Heads up on the voice: the audio in all three clips is the placeholder fal voice (voice D @ 0.85), just so the mouths have something to move to. That is not the final voice — Cassie's real voice will be ElevenLabs. Judge these purely on the face and motion: which one looks most like a real person talking, and least like a photo with a moving mouth.
Current baseline · near-static

SadTalker

The "photo that talks" engine we've been using. The face barely moves — it animates the mouth on a still image. This is the bar to beat.

Candidate · real motion

Kling

fal-ai Kling image-to-video. Generates genuine head movement and expression from the reference frame — more lifelike, higher motion.

Candidate · real motion

OmniHuman

fal-ai OmniHuman avatar. Purpose-built talking-avatar model — drives natural head/lip motion from audio, higher realism than the baseline.

Your call: reply with which engine to lock for Cassie's real videos — SadTalker, Kling, or OmniHuman.