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 — a consistent cloned voice (ours is documented here). 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.