AI ビデオの 24 エピソード全体で顔の一貫性を保つ
AI ビデオがショット間で変動する理由、キャラクターを実際にロックする 3 つのレバー (ルック プロンプト、参照画像、シード)、まだ壊れているもの、エピソードの代わりにシングル ショットを再撮影する方法。
このガイドは英語で発行されています。タイトルと概要はローカライズされています。全文は元の言語のままです。
The first question every studio asks about AI drama is not “does it look real”. It is “will it be the same person in episode 19”. A viewer forgives a slightly plastic hand. They do not forgive a lead whose jaw, hairline and eye colour change between scenes — that reads as a mistake, and mistakes break the spell.
Here is what actually controls consistency, in order of impact.
Lever one: the look prompt, written once and never edited
A look prompt is a short, boring, specific paragraph describing only what a camera can see:
Woman, late twenties, East Asian, oval face, straight black hair to the shoulder blade, thin brows, small mole under the left eye, grey tailored blazer over white shirt, no jewellery, neutral expression.
Three rules make look prompts work:
- Only visible attributes. No personality, no backstory, no “confident”. The model cannot render confident; it can render a chin angle.
- Permanent features only. Anything that changes per scene (wet hair, a bandage, a coat) belongs in the shot prompt, not the look prompt.
- One identifying detail. A mole, a scar, a specific earring. This is what the viewer uses to confirm identity across cuts, and it is what the model latches onto.
Every shot prompt then references the character by name and inherits the look prompt verbatim. If you edit the look prompt in episode 12, episodes 1–11 are now a different person.
Lever two: a locked reference image
Text alone drifts. A reference image anchors the face. We store one canonical frame per character — neutral light, neutral expression, front-facing — and pass it with every shot.
Practical notes:
- Generate the reference at the highest resolution you will ever need. A reference that is softer than your output becomes the ceiling.
- Keep the reference expression neutral. A smiling reference pushes every subsequent shot toward a smile, including the funeral scene.
- One reference per character, not a folder. Multiple references give the model a choice, and it will choose differently on different days.
Lever three: a fixed seed per character
Seeds are not magic, but they remove one source of variance for free. Fix the seed per character and change it only when you deliberately want a variation (a dream sequence, a flashback with different lighting).
| Setting | Fixed | Varies |
|---|---|---|
| Look prompt | ✅ per character | — |
| Reference image | ✅ per character | — |
| Seed | ✅ per character | shot-level only when intentional |
| Shot prompt | — | ✅ per shot |
| Lighting / lens | — | ✅ per scene |
| Wardrobe | — | ✅ per scene, described explicitly |
What still breaks
Honesty matters more than marketing here. These are the failures we still see:
- Extreme angles. Profile and low-angle shots drift more than front-on. Budget a re-render for hero profile shots.
- Hands in close-up. At 1/8 frame, hands still deform. Write around it: cut to a face, or hold the object out of frame.
- Crowds. Any shot with more than three people loses identity assignment — the model will happily swap two leads. Keep group shots wide and short.
- Long takes. Beyond about 12 seconds in a single shot, features start to slide. Our shot unit is 5 seconds for exactly this reason; long takes are assembled in the edit, not generated in one pass.
- Voice across languages. A cloned voice does not automatically speak a second language with the same timbre. If you localise, expect to re-cast the voice per language, and say so to your client up front.
Re-shoot one shot, not one episode
Because shots render individually, a failure is cheap to fix: identify the shot, adjust one variable (angle, lighting, or a single adjective in the prompt), re-render. The cost is one shot — 35 credits for 5 seconds at 1080p — not the 350–550 credits an episode costs.
The workflow rule that follows: never accept a shot you would not put in front of a client, just because the rest of the episode is finished. A drifting face in shot 9 will be the only thing anyone mentions.