Everything on this channel is one character, produced by one machine, mostly unattended. This is the capstone guide: how the pieces snap together into a factory, and the order we’d build it in if we started over.

The five layers

  1. Identity, a character LoRA. One small file that keeps her her in every frame. Everything else stacks on this; build it first and never render without it.
  2. Looks, a curated plate library: approved stills (outfits × settings × framings) generated in ComfyUI, human-picked. Plates are the casting department: every video starts by choosing one. Curation rules that saved us: describe whole outfits, no people on background screens (video models animate everyone), dress for motion not the still.
  3. Motion: image-to-video turns a plate into presenter footage; lip-sync makes it talk. Keep these as separate, swappable stages, engines improve monthly and you’ll want to upgrade one without rebuilding the factory.
  4. Voice, one consistent voice. Ours was designed once in ElevenLabs; the videos are voiced by ElevenLabs directly, and our apps speak through a local clone built from a single reference clip. Same principle as the face: one identity, reused everywhere.
  5. Assembly & publish: scripts that composite (captions, cards, corner cams), run a QC gate (a vision model reads sampled frames and fails distortions: give your AI eyes), and upload on a schedule with metadata and disclosure.

The design rules that matter

  • Everything is a file. Workflows are JSON; plates are folders; the schedule is a timer. If a human choice is needed, it’s a pick from a sheet, not a hand on a slider.
  • Human taste is the gate, not the labor. The machine generates batches; the human picks winners and sets policy. Quality control by taste is still undefeated.
  • Consistency compounds. Same seed families, same plates, same voice → every asset reinforces the character. Audiences notice identity drift before they notice render quality.
  • Own your engine. The whole stack above runs on one consumer GPU. Rented tools change pricing and policies; an engine you own runs forever.
  • Disclose. Synthetic media should say so. Ours does, in writing, on every upload.

A build order that works

Voice → identity LoRA → plates → one manual video end-to-end → then automate stage by stage (generation, QC, upload), hardest-to-trust last. Automating before you’ve shipped manually just automates your mistakes.

This factory is also becoming a real-time companion, that story is unfolding on YouTube → @AskAillex. Build along, and show us yours at r/aillex.