Why faces drift
You have seen the AI clip where someone's jaw widens halfway through, or where the man in shot two is clearly a cousin of the man in shot one. That happens when a model is asked to produce a face from a description. Words can say dark hair and a linen shirt. They cannot hold the exact distance between someone's eyes.
A serial makes it worse. You need to recognise the same two people across rooms, lighting and weeks of episodes, and a small change compounds if each episode copies the last. So the fix is not a better prompt. It is to stop generating faces from text.
| What goes wrong | Why | What Main Character does |
|---|---|---|
| Face morphs mid-clip | The face comes from text, so nothing anchors it. | Every shot with a person starts from an approved still. |
| Different person in shot two | New prompt, no shared reference. | One locked set of reference pictures per person, for the whole serial. |
| "Same outfit" isn't | "Same" means nothing to a model. | The full outfit is written out in every shot. |
| Slow drift across episodes | Each episode copies the last one's frames. | Every shot goes back to the original approved portrait. |
| Nobody checked | No identity check before animation. | A likeness score of 0.60 or more before any video is made. |
1. A locked set of reference pictures
Yours starts with three selfies. The app draws one canonical portrait of you in the show's look, then five more: a three-quarter view, a profile, full body at the build you tapped, and two outfits from your world. You approve the portrait, with one free redo, before anything else is made.
His starts with taps: ethnicity, age band (25 to 34 or 35 to 45), hair, build, style and vibe. Those become a short written description and an invented first name, and the same six-picture set is drawn from it. There is no photo input of any kind.
Supporting cast, like a roommate or a boss, get two reference pictures each, made the first time the story needs them. All of these are locked for the serial.
2. Places that stay put
Each serial has a handful of fixed locations, such as home, work, the commute, a bar or café, a street at night. Each is lit one of two ways, warm window light or night lamps, and a scene never switches lighting halfway. The bar you return to in episode eight is the bar from episode two.
3. A still first, then motion
Each shot with a person begins as a still picture, drawn from the reference pictures of everyone in it, the location, the shot description and a fixed style: a 35 or 50 mm lens look, shallow focus, warm skin, light grain. Only that still is animated, for three to six seconds, with slow camera moves: a push-in, a drift, a rack focus. No whip pans, no running, no crowds. Slow movement is where video models keep a face best.
Shots with no one in them, like an empty street or a cup on a desk, are the only ones made straight from a description.
4. The likeness check
Before a still is animated, a vision model compares the still with your approved portrait and scores the match from 0 to 1. The still needs 0.60 or more. The score is run on every shot you are in; shots of him alone rely on his locked reference pictures. That number measures how strongly the two pictures read as the same person; it is not "60% identical".
| Check | Pass | First fail | Second fail |
|---|---|---|---|
| Still, before animation | Animated | Redrawn | Replaced with a shot that shows no face: hands, over the shoulder, a detail |
| Last frame, after animation | Kept | Clip re-rendered | Kept, and the drift is logged for review |
The gate catches most drift, not all of it. A clip can still lose a little of a face in motion, which is why the shot grammar leans on slow moves and keeps two-person shots to at most two an episode.
5. Thirty seconds, about six shots
Shots are a vocabulary, not a fixed order: an establishing shot, what you see, you, him, the two of you, a detail, and a held final frame. You are on screen in two to four of them. The writing model picks the order that fits the scene.
6. Sound that doesn't restart at every cut
Every clip is rendered silent on purpose, so a drifted picture can be redone without redoing a voice. Then one soundtrack runs under the whole episode: your inner voice, the sound of the place you're in, his voice, and music for the mood. Each character keeps one voice for the whole serial, for the same reason they keep one face.
Questions
How does Main Character keep my face the same in every shot?
Every shot with a person starts as a still picture drawn from your approved portrait, never from a text description. The still is checked against your portrait before it is animated, and only approved stills become video.
Why do AI videos change faces between shots?
Most generate the face from words. A description cannot carry the exact shape of a face, so the model reinvents it each time. Main Character resolves identity in a still image first, where it can be checked and redrawn cheaply.
What is the likeness check?
A vision model compares each new picture with the approved portrait and scores the match from 0 to 1. A picture needs at least 0.60 to be animated. The score measures resemblance; it is not a percentage of you.
What happens when a picture fails?
It is redrawn once. If it fails again, that shot becomes a shot without a face, such as hands, a view over the shoulder or a detail. After animation the last frame is checked again, and a clip that drifted is re-rendered once.
Does the love interest stay consistent too?
He gets the same locked set of six reference pictures, made once for the serial, and every shot of him is drawn from them. The automatic likeness score is applied to shots you are in; his consistency comes from the locked references.