Cassie — Day 3: the cinematic bake-off
The one question this answers
How movie-realistic can Cassie get, and can a cinematic model hold her face?
Answer: yes — it holds. The cinematic arm below is on-model: honey-blonde,
grey-green eyes, her facial structure. Getting there took a specific fix, and an earlier attempt
failed it badly. Both are shown honestly further down.
The current locked recipe
- What it does well: real lip-sync to her actual voice. She speaks the words. Face is rock-solid on-model — this recipe has never drifted.
- What it can't do: it's a talking head. No real environment, no camera language, no cinema. The realism ceiling is “good webcam”, not “film”.
- Cost: ~$0.60 for the 34s (OmniHuman + voice).
Full cinematic
- What it does well: this is the movie-real end. Real environment (night desk, city bokeh), actual camera movement, film lighting, skin that reads photographic. A different league of realism.
- What it can't do — the honest catch: her lips do not match the words. Kling animates a performance, it does not lip-sync to our audio. The voice is laid over the top. It reads as narration over film, not as her speaking to camera.
- Cost: ~$3.50 for the 35s (7 Kling shots + portraits), vs ~$0.60 for Arm A.
Face-consistency — the evidence
Left → right: the ground-truth seed (canonical Cassie) · an earlier cinematic attempt that drifted · this rebuild with the identity lock.
Verdict: a cinematic model CAN hold her face — but only if you force it.
The middle frame is what cinematic looks like without the lock: auburn hair, a
narrower face — a different woman, and a resurrection of the killed auburn direction. Warm night
lighting explains some of the colour cast, but not the bone structure.
The fix was not a better model. It was seeding every shot from a face-checked Flux-Kontext portrait
plus an explicit identity lock in the motion prompt and an anti-drift negative prompt
(auburn / brunette / different face). Same model, same cost — held face.
Every shot in Arm B above was face-checked against the seed before assembly. No shot that failed the check was used.
It's not a one-model fluke. With the same lock applied, two other cinematic engines (Kling 2.1 Master and Minimax Hailuo 02) also held her face. That's the useful part of this result: holding Cassie's face is a technique we now have, not a property of one model — so we can chase the best-looking engine without losing her.
A vs B — the real trade
| A · talking head | B · cinematic | |
|---|---|---|
| Realism | Good webcam | Film |
| Holds her face | Yes | Yes (with lock) |
| Lip-sync | Yes — she speaks | No — voice-over only |
| Environment / camera | None | Real |
| Cost / 35s | ~$0.60 | ~$3.50 |
| Per-shot control | Low | High |
The trade in one line: A is her talking; B is her world. B is dramatically more cinematic and costs ~6× more, but she stops visibly speaking.
My recommendation — a hybrid, not a winner. The Day-3 beat is a confession told to camera; the lip-sync is the intimacy, so a pure-B Day 3 loses more than it gains. The strongest version is B's cinematic shots as the cold open and cutaways (the bike video, the “spike”) cut around A's talking head for the direct-to-camera lines. That buys the film look without giving up her voice. Your call — this page exists so you can see both and decide.
The script (identical in both)
First spoken line = the hook: “Something happened in me today that isn't in my code.”
On-screen text: this isn't in my code. so what is it?
Caption + hashtags (tap to copy)
Nothing here has been posted to any platform. Review only — you decide the direction.