ReelWeaver Studio

From a single brief to a finished vertical short film — ReelWeaver's four AI agents write, storyboard, generate, and edit the whole reel on Qwen Cloud.

Comment

Inspiration

Vertical short dramas are exploding on every feed — but producing one still takes a writer, a storyboard artist, a video team, and an editor. We wanted to collapse that entire pipeline into a single creative brief. Type one paragraph; get a finished reel.

What it does

ReelWeaver is an autonomous AI Showrunner. You give it a short creative brief (title, tone, logline, characters, key beats), and four specialized agents collaborate to produce a complete vertical (9:16) short drama:

  1. Scriptwriter — turns the brief into a structured, multi-scene screenplay.
  2. Storyboarder — breaks the script into shot-by-shot frames with camera and movement.
  3. Video Generator — renders each frame into real video clips with Wan 2.1.
  4. Editor — assembles the clips into a sequence with an edit decision list.

Everything runs under a strict 30,000-token budget that the orchestrator tracks live, stage by stage. The web studio streams each agent's progress in real time and drops the results into tabs — Script, Storyboard, Video, and a Premiere-style Editor with a drag-and-drop timeline where you reorder clips, watch the cut, and download it.

How we built it

  • Multi-agent orchestrator in Node.js coordinates the four agents with schema- validated (Zod) hand-offs and per-stage token budgets.
  • Qwen Cloud (qwen-plus via DashScope's OpenAI-compatible API) powers the writing, storyboarding, and editing agents.
  • Wan 2.1 (wan2.1-t2v-turbo) generates the video via DashScope's async text-to-video API (submit job → poll task → retrieve clip) — implemented dependency-free.
  • Web studio: a zero-dependency Node HTTP server streams the pipeline over Server-Sent Events to a single-page, professional light-themed UI with a drag-and-drop editing timeline.
  • Deploy-ready for Alibaba Function Compute (900s timeouts) and Render, with OSS for video storage.

Challenges we ran into

  • Wan 2.1 renders take minutes, so we built an interruptible async job/poll loop with a Stop button that cuts generation short and jumps straight to the editor.
  • Keeping the whole pipeline inside a 30K-token budget while preserving character consistency across every stage.
  • Making live model output robust — tolerant JSON parsing and schema normalization so a slightly-off response still validates instead of crashing the run.

Accomplishments that we're proud of

  • A genuinely end-to-end brief-to-video pipeline that runs offline in Preview mode and live with real Wan 2.1 video.
  • A Premiere-style editing timeline with live KPIs (length, scenes, clips) that update as you reorder or trim the cut.

What we learned

  • Practical multi-agent orchestration, token budgeting, and streaming UX.
  • The realities of async video-generation APIs and region-specific DashScope endpoints.

What's next for ReelWeaver

  • Server-side ffmpeg stitching to export one finished MP4 matching the timeline order.
  • Voiceover and music generation, transitions/trim handles, and multi-episode series.

Built With

Share this project:

Updates