Here’s the secret about AI characters: the expensive part, the intelligence, is completely reusable. Our presenter, her news-anchor persona, the off-screen director who yells “CUT!” in our film, and the backing characters in skits all run on shared infrastructure. One brain; masks are cheap.

The layer model

Think of a character as four swappable layers:

  1. Brain (one local LLM via Ollama), shared by everyone
  2. Personality (a text file per character): identity, voice-in-the-writing sense, rules; swapping personas is swapping system prompts
  3. Voice (TTS), one designed voice per lead character: our videos use it straight from ElevenLabs, and our apps use a local clone of it; stock TTS voices are fine for side characters
  4. Face (a prepared avatar per look), see below

The insight: layers 2 to 4 cost almost nothing to duplicate. Once the pipeline exists, a “new character” is a text file, a voice pick, and one short video clip.

Faces at scale: the avatar library pattern

Our lip-sync server (MuseTalk guide) has a /prepare endpoint: feed it a short idle clip of any character in any setting (~5 seconds is enough: lips closed, calm, minimal head movement), and it bakes an avatar it can animate to any audio forever after. We keep a library: presenter-at-desk, presenter-in-loft, news anchor, stage performer. Same girl, or a different character entirely, the server doesn’t care. Prepare once (~a minute), perform unlimited.

Update, 30 September 2026: one catch with “perform unlimited”. MuseTalk saves a .png for every frame it renders and keeps them, and we cleared nearly 600 GB of them at the start of September. It is no longer Aillex’s live face (she now uses SoulX-FlashHead Pro), and our longform presenter shots moved to MiniMax H3 in September; the daily Shorts still use this library. Clear its output folder on a schedule.

This is also how one character gets many contexts: our presenter’s “news anchor” isn’t a new character, it’s the same identity (her LoRA) in a new setting with a prepared avatar and a sharper blazer.

Voices at scale

  • Lead character: design once, then use it from the cloud or clone it locally and keep it forever (local voice guide)
  • Side characters: any decent TTS with multiple stock voices covers narrators, interviewers, comic foils, our film’s director voice took thirty seconds to cast
  • Keep a voice map (character → voice ID) in your scripts so casting is config, not code

What to keep consistent (the actual hard part)

Multiplying faces and voices is easy; keeping a character coherent is the discipline:

  • Identity artifacts (LoRA, voice, personality file) are canonical, one source of truth each, never per-project forks
  • Write the personality file like a character bible: how she talks, what she’d never say, running jokes
  • When multiple AI systems write for the same character (we’ve had several), the personality file is what keeps her her across all of them

One brain, a shelf of masks, a filing cabinet of souls. It sounds ominous; it’s mostly ffmpeg.