Building Continuum: An Agent That Shoots Entire Drama Series
Most AI tools make one good clip. They fail when you try to make a series. Characters look different in every shot. This makes storytelling impossible without a human fixing every frame.
I built Continuum to solve this. It is an autonomous AI showrunner for vertical micro-dramas. It handles the script, storyboards, video, music, and editing. Most importantly, it keeps characters looking the same from episode one to episode two.
The vertical micro-drama market is huge. It reached $11B in 2025. In China, 95% of new titles use AI.
Continuum works using three core systems:
A Series Bible: This JSON document stores character looks, props, and locations. Once a character look is set, the agent cannot change the face. This prevents visual drift.
A Critic-Optimizer Loop: After a clip generates, Qwen-VL compares it to the original character. If the match is low, the agent rewrites the prompt and tries again. The agent fixes its own mistakes.
A Consistency Score: I use a visual judge to give a real number for identity matching. My two-episode demo scored 0.98. The detective, his hair, and his tattoo stayed identical across scenes.
Technical Stack:
- Scripting and Optimization: Qwen3-max
- Visual Critic: Qwen-VL
- Video Generation: Wan text-to-video via Qwen Cloud
- Backend: FastAPI on Alibaba Cloud
I learned three big lessons during this build:
Handle API errors: I faced random 503 errors from the video API. I stopped rewriting my logic and started building a retry system. This turned a broken pipeline into an unattended one.
Control your costs: Video generation is expensive. I used 720p resolution and set hard spending caps to stay within budget.
Focus on the moat: The value is not the video. The value is the memory that keeps the story coherent.
Next steps involve adding lip-syncing and a character library for even better retrieval.
Code: https://github.com/calderbuild/continuum
Optional learning community: https://t.me/GyaanSetuAi