Traditional short-drama production takes months and tens of thousands in cost. In the Kunlun Cave project we industrialized the whole chain: from story generation to deliverable episodes, fully AI-orchestrated, cutting per-episode cost to under 1/10 of conventional shooting. Here is the pipeline breakdown.
| Stage | Tool | Output |
|---|---|---|
| Story evolution | LLM narrative reasoning | Episode scripts + storyboards |
| Character assets | Text-to-image + style constraints | Character sheets / keyframes |
| Scene generation | Text-to-image batch | Scene image sets |
| Motion | Image-to-video | Shot clips |
| Consistency check | Character asset bank | Cross-shot same face |
| Library & release | Works gallery + truth ledger | Deliverable episodes |
Character assets are the root of stability. All frames reference fixed character sheets; style is governed by a unified aesthetic constraint set — model stability plus asset determinism is what makes industrialization possible. Keyframe consistency: video frames are backfilled into the asset bank so later shots reuse the same source, eliminating cross-shot character drift.
We let the LLM continuously evolve the plot: given the world-view and character settings, the model derives arcs, conflicts, and emotional pacing across multiple iterations; humans only do aesthetic sign-off. A 10-minute episode outline can be produced in ~30 minutes with several candidates to compare.
Per episode (~30 shots): roughly 150K text tokens for orchestration plus image/video generation quota — about 1/10 to 1/20 of live-action cost. And assets are reusable: characters, scenes, and style systems persist for the next season.
1) Fix character assets + aesthetic constraints before batch production. 2) Log every production step into a ledger for review and reuse. 3) Visualize the chain with a workbench (see short-drama workbench). Full system: Huodou Cloud AI AIOS whitepaper.