Dramaforge

langing page

Inspiration

Traditional short-form vertical drama production is slow and expensive, requiring crews, actors, and hours of editing. We wanted to build a zero-crew AI studio that handles the entire pipeline autonomously, producing a finished mobile episode for under $0.20.

What it does

DramaForge is an autonomous, vertical short-drama generation platform. From a single-line premise, a 7-stage agent pipeline writes the script, storyboards, casts characters with locked visual references, generates video clips, scores audio, and compiles a final vertical MP4 with subtitles. It features an Economy Mode that dynamically downgrades LLM models if costs approach the user's budget cap.

How we built it

DramaForge is a Next.js 15 app deployed natively on Alibaba Cloud ECS, using Nginx and PM2. Data is stored in ApsaraDB RDS MySQL, and media on Alibaba Cloud OSS.

The 7-stage pipeline uses:

Hook Strategist (qwen-max): Hooks & title.
Story Architect (qwen-max): Scene outlines.
Scriptwriter (qwen-plus): Dialogue & actions.
Storyboard Director (qwen-max): Camera framing.
Casting & Video (Wan Models): wan2.7-image creates reference portraits saved to RDS, ensuring face consistency when generating clips with wan2.1-i2v-720p.
Audio & Music: Vocals and background scores.
Post-Production (FFmpeg): Scales to 9:16 vertical, burns subtitles, and stitches the video.

🛡️ Budget Safeguard

Cumulative cost $C_t$ is tracked. If $C_t \ge 0.8 \times B$ (where $B$ is the budget), Economy Mode triggers, switching subsequent qwen-max calls to qwen-plus to save tokens.

Challenges we ran into

Character Consistency: We solved this by caching wan2.7-image portraits in RDS and passing them as image references to the wan2.1-i2v model.
Aspect Ratio & Subtitles: We used complex FFmpeg filters to crop, pad, and burn text into a 9:16 format.
Stateless Progress: WebSockets dropped during long runs, so we switched to stateless HTTP polling against a JSON cache.

Accomplishments that we're proud of

Zero character drift across scenes.
Robust Economy Mode for strict cost containment.
Production-grade deployment on Alibaba Cloud ECS.

What we learned

Strict JSON contracts between agent stages are critical. We also found qwen-plus is incredibly efficient for scriptwriting, proving a hybrid multi-model approach beats relying on a single expensive LLM.