A team from Nanyang Technological University has developed a hierarchical agent framework that transforms a single creative idea into a fully produced short drama, addressing major gaps in automated video generation that have plagued the industry. The system, called "One Sentence, One Drama," uses artificial intelligence agents to handle narrative structure, visual design, scene consistency, and post-production in a coordinated workflow, moving short drama production toward standardized, controllable quality.

What Problems Does This Solve in AI Video Generation?

Video foundation models like Sora, Kling, and Veo have made impressive strides in generating videos from text prompts, but the current production of short dramas still relies on loosely connected workflows that create three persistent problems.

• Weak Narrative Rhythm: Openings often fail to capture attention, and plots lack the conflict and tension needed to keep viewers engaged throughout the story.
• Insufficient Spatial Consistency: When cameras move between scenes, characters and objects shift positions unpredictably, breaking the viewer's sense of the physical space.
• Immature Quality Control: The generation process still requires extensive manual review and correction, making automation impractical at scale.

The Nanyang Technological University research team built their agent framework to solve these issues by breaking video production into four coordinated stages, each with built-in quality checks.

How Does the AI Agent Framework Actually Work?

The system operates through a structured four-stage pipeline that coordinates multiple AI agents working together. In the first stage, story generation, the framework uses multiple agents to debate and refine narrative structure. The agents then consult libraries built from about 300 high-quality short dramas, extracting patterns in three dimensions: factual accuracy, logical coherence, and narrative rhythm. This creates a controllable framework for building stories with proper pacing and emotional arcs.

The second stage handles visual material and prompt generation. Agents create panoramic scene views and character reference images, then generate specific prompts for each video segment. A review module checks spatial relationships and props for consistency before generation begins, rewriting problematic descriptions automatically.

The third stage uses 3D scene anchoring to maintain spatial consistency across shots. The framework reconstructs the scene space based on panoramic views, then unifies character movements, camera positions, and scene relationships. When multiple characters appear, the agent fine-tunes camera angles to keep all characters visible and preserve their standing relationships.

In post-production, the final stage, agents handle transitions, background music, and voice connections according to plot progression, integrating all segments into a cohesive drama with consistent rhythm and emotional flow.

What Do the Benchmark Results Show?

The research team created a specialized evaluation benchmark called Short-Drama-Bench to measure performance on 7 major genres and 17 sub-genres, including revenge stories, ancient palace intrigue, suspense, time-travel, romance, and workplace dramas. The benchmark covers approximately 239 minutes of generated video across long, medium, and short formats.

The evaluation system uses three measurement approaches. VBench measures general video quality, ViStoryBench evaluates how well the story translates visually, and 8 short drama-specific indicators examine opening hooks, narrative coherence, spatial consistency between characters and environments, and the naturalness of background music and transitions.

In quantitative comparisons against existing methods like MovieAgent, ScriptAgent, StoryMem, and commercial products like Toonflow, the One Sentence, One Drama framework showed leading performance across short drama-specific indicators, standard video quality metrics, and story visualization benchmarks. Ablation studies revealed that each production stage serves a distinct function: story generation affects opening appeal and plot progression, 3D first-frame generation improves cross-shot spatial coherence, multi-stage review enhances overall quality, and transitions plus background music make emotional connections more natural.

What Are the Practical Limitations?

Despite strong performance, the framework faces real-world deployment challenges. The average API cost to generate one minute of video is approximately $25 to $27, compared to about $21.53 per minute for Toonflow. Time costs are also significant: generating a complete 10-minute short drama takes approximately 74 to 90 minutes.

The research team acknowledged that achieving large-scale deployment requires addressing cost reduction. Additionally, the current framework focuses primarily on automatic generation, though the researchers noted that future versions could incorporate interactive interfaces to provide users with review scores and diagnostic feedback, enabling more human-computer collaboration in the creative process.

Steps to Understand AI Video Generation Quality

• Evaluate Narrative Structure: Look beyond visual quality to assess whether stories have compelling openings, clear conflicts, and satisfying resolutions that maintain viewer engagement throughout.
• Check Spatial Consistency: Watch how characters and objects maintain their positions and relationships across camera transitions and scene changes, as this is a key indicator of production sophistication.
• Consider Production Workflow: Understand whether video generation relies on loosely connected steps or coordinated agent systems that can catch and correct errors automatically before final output.
• Review Quality Control Mechanisms: Assess whether the system includes built-in review stages throughout production or requires extensive manual correction after generation completes.

The One Sentence, One Drama framework represents a significant shift in how AI approaches video production. Rather than treating video generation as a single end-to-end task, the system breaks production into coordinated stages where agents can specialize in narrative, visuals, consistency, and finishing. This structured approach has moved short drama production from a largely manual, unpredictable process toward something more standardized and controllable, even if cost and time constraints still limit widespread adoption.