Skip to content
noveltodrama

Character lock: keeping faces stable across 24 episodes

Why AI faces drift between episodes, how character sheets are extracted from a manuscript instead of invented, what lock actually guarantees (about 90%), and the three edits that break it.

noveltodrama editorial 7 min read

Face drift is the most visible failure in AI drama, and the most common cause is not the image model — it is that the character was described differently in episode 3 than in episode 1. “Tall, dark coat, tired eyes” and “lean, long coat, sharp gaze” are the same person to a reader and two different people to a renderer.

Character lock fixes this by making one description canonical and refusing to re-derive it.

Where the description comes from

Not from a prompt you write. The manuscript already contains it: every web novel describes its leads repeatedly, in fragments, across chapters. The extraction step reads the chapter summaries and produces, per named character:

  • a physical description assembled from the strongest textual fragments
  • wardrobe (what the text actually puts them in, not what looks good)
  • age band and build
  • a voice profile (pitch, pace, accent) for the dubbing step
  • a locked reference image, generated once

Twenty credits per character, charged once per project. A five-character cast is 100 credits for a 24-episode season — about 4 credits per episode.

What lock guarantees

Roughly 90% stability. The same face, the same wardrobe, the same grade, across episodes. Not identity-grade consistency: at 1080p vertical, in a two-shot with a 15-degree angle change, you will see drift in the jawline if you look for it. Audiences do not, across 75 seconds, which is the relevant test.

It does not guarantee consistency across a re-cast. Changing a character’s look mid-season re-bills every subsequent shot with that character in it, because the reference changed.

The three edits that break lock

  1. Rewriting a character description after rendering. The lock is keyed to the description. Edit it and the next episode renders a different person. If you must edit, edit before episode 1 renders.
  2. Merging two characters. Common when a novel has three sisters and you keep two. Merge in the cast sheet, not in the script — a script that names a non-existent character produces a new, unlocked face.
  3. Importing a different manuscript into the same project. Re-import replaces chapters; the cast is not automatically rebuilt, so old locks survive against new text. Rebuild the cast after a re-import.

Practical sequence

Import → read summaries → extract cast (20 credits each) → review each sheet against the text → lock → then board. Reviewing before lock costs nothing; reviewing after the first episode renders costs 330 credits per episode to fix.

The one judgement call that matters: which characters get a sheet. A 120-chapter novel may name sixty people. Five to eight get sheets; the rest are “background”, rendered from a generic pool and never speaking in close-up. Choosing those five is the single highest-leverage decision in the whole adaptation, and no pipeline can make it for you.

Your first episode starts with 300 free credits

Enough to import a short manuscript, summarise ten chapters and board one full episode. No card required.

300 credits · No card required · Commercial rights on paid plans