Cre8Motion

Cre8motion is a continuity-aware AI showrunner that builds animated series, not one-off clips, preserving characters and worlds while planning, reviewing, repairing, and editing every episode.

Comment

Inspiration

AI video tools can generate an impressive clip. But ask them for episode two and everything falls apart, your character has a new face, the kitchen became a bedroom, the necklace from the last scene simply doesn't exist anymore. Every AI-generated "show" today is really a folder of disconnected clips held together by a human manually copying prompts between tools.

We wanted to build the missing layer: not a video generator, but a showrunner, an agent that understands an ongoing series, remembers what happened, protects what things look like, and turns one sentence into a finished episode that belongs to a larger story.

What it does

Cre8Motion is a continuity-aware AI showrunner for silent, 45-second vertical microdramas.

You give it an idea, or press "Let the showrunner decide" and it continues the story from where the last episode ended. From there the agent autonomously runs a 17-stage production pipeline:

  1. qwen-max writes the show bible: premise, characters with locked visual descriptions, and a season-ready world
  2. qwen-image-plus paints the show's poster key art; wan2.2 draws each character's reference sheet
  3. For every episode, qwen-max writes a silent-story script under hard narrative constraints (one visible objective established in 8 seconds, flaw-driven complications, reaction shots, setup-before-payoff, the episode must work with the sound off)
  4. The plan is compiled into an 8–12 shot storyboard, each shot with framing, lighting, continuity locks, and generation-risk fallbacks
  5. wan2.5 generates keyframes using the locked character references; HappyHorse animates each shot from its keyframe
  6. qwen3-vl-plus reviews every artifact against the shot spec, failures are diagnosed by qwen-max, which rewrites the prompt and regenerates only the broken shot
  7. ffmpeg assembles the episode; continuity state (who knows what, who owns which object) carries into the next episode

Every stage is visible in the UI: live pipeline status, real budget tracking, elapsed time, and human-in-the-loop pause/resume at any point.

How we built it

  • Backend: FastAPI + SQLAlchemy + Alembic on a 17-stage validated state machine, deployed with Docker. All AI runs on Qwen Cloud (DashScope), five model families orchestrated by one production engine (backend/app/providers/qwen.py)
  • Storytelling engine: we encoded professional silent-film craft (object → eyeline → reaction → action grammar, one dominant emotion per beat, rule-of-three objective setup) directly into the agent's prompts, so quality is enforced, not hoped for
  • Frontend: React + Vite + TypeScript studio interface, shows, episodes, characters, productions, and assets all live from the API, with SSE events streaming pipeline progress
  • Persistence: PostgreSQL for narrative state; generated artwork is stored in-database so it survives redeploys

Challenges we ran into

  • Silent storytelling is a constraint engine, not a vibe. Early generations were pretty but unreadable. The fix was treating our story guide as rules the agent must satisfy, visible physical objectives, immediate cause-and-effect, no dialogue crutches.
  • Continuity is a database problem wearing a creative costume. Character identity, object ownership, and open questions have to survive across episodes, redeploys, and retries, that pushed us from file storage to full relational state.
  • Long-running AI pipelines fail in creative ways. A single stalled model call froze productions forever until we built timeouts, budget-aware retries, and selective regeneration, the agent repairs the one broken shot instead of re-rolling the episode.

Accomplishments we're proud of

  • A one-sentence idea becomes a planned, storyboarded, generated, QC'd episode with zero human steps, and you can watch the agent work stage by stage
  • The "Let the showrunner decide" button: episode N+1 written purely from accumulated story state
  • Character faces, props, and lighting that persist across shots because references are locked assets, not prompt suggestions

What we learned

The hard part of AI video isn't generation, it's production management: state, budgets, retries, and taste. Encoding editorial judgment (what makes a story readable without sound) into an agent's constraints taught us more about prompting than any amount of model-picking.

What's next for Cre8Motion

  • Wire the remaining review surfaces (references, assembly, final review) to the live pipeline
  • A Silent Legibility Score: automated muted-viewer comprehension testing per episode
  • Multi-agent writers' room: separate ideator, story editor, and continuity critic negotiating each episode
  • Creator monetization: publish a show once, let the showrunner keep it running

Built With

Share this project:

Updates