One Sentence, One Drama: Personalized Short-Form Drama Generation via Multi-Agent Systems

🕒 Published (v1): 2026-05-21 00:00 UTC · Source: HuggingFace · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

"One Sentence, One Drama" presents a hierarchical multi-agent pipeline that converts a single-sentence prompt into a fully produced short-form drama video through structured intermediate modules and iterative reviewer loops. The system addresses three simultaneous failure modes of prior work—narrative pacing, cross-clip spatial drift, and manual quality-control burden—by decomposing generation into debate-driven script synthesis, 3D-grounded keyframe anchoring, and multi-stage automated review. A new benchmark, Short-Drama-Bench, is introduced with 50 prompts and drama-specific evaluation metrics.

Problem

Existing short-drama pipelines (Toonflow, Xiao Yun Que) rely on one-shot LLM script expansion and loosely coupled generation units, producing three interrelated deficiencies: (1) weak dramatic hooks and insufficient narrative escalation/reversal density; (2) cross-clip spatial drift in character positions and scene layouts; (3) production outputs that require substantial manual inspection and correction before release. General long-form video pipelines (MovieAgent, ScriptAgent, StoryMem) are not designed for the compressed, high-hook pacing demanded by short dramas.

Method

The pipeline has four hierarchical stages (A→D):

A. Hierarchical Episode Planning. A corpus of ~300 high-performing scripts is atomized into a Pattern Bank (2,923 beat-level units encoding pacing priors) and a Logic Bank (6,984 overlapping causal chunks). Given a logline, three retrieval routes (fact via web search, logic from Logic Bank, pattern from Pattern Bank) supply atoms to a structured story-core generator. Three independent LLM judges debate the draft; conflicts are resolved by a designated decider (GPT-5.4 Pro); a reviser performs patch-based local rewriting. Valuable discarded ideas are stored in an Idea Bank and restored if safe. Clip-level scripts are then synthesized, with partitioned reviewers targeting opening hook (first clip only), ending suspense (last clip only), and twist density (middle clips only).

B. Visual Assets and Prompt Generation. Scene-level 360° panoramas and multi-view character references are generated. Paired keyframe-video prompts are produced per clip, then passed through a prompt-level reviewer loop that checks spatial consistency, physical plausibility, and prop continuity before any rendering occurs.

C. 3D-Grounded Keyframe-to-Video Generation. A scene-level 3D world \(W\) is reconstructed from a person-free panorama using Marble. For each new clip, background candidates \(B_k = \Pi(P; v_k)\) are sampled, character-conditioned first frames \(I_k\) are synthesized, and a VLM selects the optimal view. \(I^*\) is registered back into \(W\) via VGGT-estimated relative transform \(\Delta T_{I^* \to B^*}\) with scale resolved by depth alignment. After clip generation, CUT3R recovers the camera trajectory anchored as \(T_t = T_{I^*} \Delta T_t\). A human mesh (SAM 3D Body) is registered from the tail frame into shared coordinates. Geometrically feasible cameras for the next clip are sampled on a local spherical shell, filtered by geometric and VLM-based semantic filters, then used to synthesize the next first frame with human mesh pose constraints.

D. Post-Production Assembly. Scene transitions are selected from four types (direct cut, temporal overlay, location-establishing shot, motion-bridge) based on temporal/spatial shift and character movement. BGM is selected from a library of 8,122 tracks organized into 16 functional buckets; an LLM selects primary/backup buckets per scene, and GPT-Audio scores candidate segments for emotional and narrative fit; adaptive mixing applies dialogue-aware LUFS calibration.

Key Contributions

• One Sentence, One Drama framework: hierarchical multi-agent pipeline converting a single sentence into a fully produced short drama via structured, reversible intermediate modules.
• Multi-agent debate-based story generation: three-judge debate loop with patch-based revision and Idea Bank recycling enforcing short-drama pacing (hooks, escalations, reversals).
• 3D-grounded first-frame generation: shared 3D spatial coordinate system for consistent cross-clip character positioning and scene layout under arbitrary viewpoint changes.
• Multi-stage reviewer loops: automated error detection and targeted revision at script, prompt, keyframe, and video levels, reducing manual QC burden.
• Short-Drama-Bench: 50 prompts across 7 categories and 17 subcategories, ~239 minutes of generated video, with drama-specific metrics (Opening Hook, End Hook, Escalation Effect, Narrative Coherence, Continuity, BGM Emotion Alignment, Transition Naturalness).

Results

• Outperforms all five baselines (ScriptAgent, MovieAgent, StoryMem, Toonflow, Xiao Yun Que) on VBench standard video metrics, ViStoryBench story-visualization metrics, and all proposed short-drama-specific metrics.
• Long-form baselines (MovieAgent, ScriptAgent, StoryMem) score worse on Opening Hook and End Hook, confirming they are not optimized for short-drama pacing.
• Short-drama commercial platforms (Toonflow, Xiao Yun Que) achieve better narrative scores than long-form baselines but show visual inconsistency and lack explicit BGM/transition planning.
• Ablation confirms all four key components (Story Gen, 3D grounding, reviewer loops, BGM/transitions) each contribute independently to overall quality; removing Story Gen primarily degrades narrative-related metrics; exact delta numbers are in Table 2 (truncated in provided text).
• Human study with 20 annotators on 5-point Likert scale corroborates model-based evaluation rankings.

Limitations

• Depends on multiple large external models (GPT-5.4 Pro as debate decider, Gemini 3 Pro / Qwen3.5-Omni / Seed 2.0 Pro for evaluation), introducing cost and API dependency.
• 3D reconstruction via Marble and camera registration via VGGT/CUT3R may degrade on complex indoor/outdoor scenes with ambiguous depth or textureless backgrounds.
• Multi-character scenes require reconstructing additional 3D meshes from previous visible frames, adding complexity and potential error propagation.
• Short-Drama-Bench covers 50 prompts; generalization to non-Chinese short-drama conventions or other cultural narrative forms is unverified.
• Runtime cost of the full pipeline (multi-round debates, 3D reconstruction per scene, multi-stage reviewer loops) is not reported; production-scale latency is unknown.

Relevance to Harnesses / Meta-Harnesses

This paper is a concrete instantiation of a production-grade multi-agent meta-harness: it orchestrates heterogeneous specialized agents (debaters, judges, deciders, reviewers, VLM filters, BGM selectors) across multiple abstraction levels (story → scene → clip → frame → audio) with explicit reviewer loops functioning as automated harness-level quality gates. The pattern of inserting self-correcting reviewer loops between pipeline stages—where each loop outputs an issue list, root-cause analysis, and targeted revision—directly mirrors the iterative-refinement architecture central to meta-harness design. The Idea Bank mechanism (preserving discarded outputs for possible later reuse) is a novel stateful buffer pattern not commonly formalized in harness literature. For researchers tracking harnesses, this work provides empirical evidence that staged, multi-agent reviewer loops reduce manual intervention in a complex, multi-modal generation pipeline, and demonstrates how a shared 3D world model can function as persistent cross-agent state analogous to a harness-level memory store.