PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Yufei Shi, Weilong Yan, Naixuan Huang, Yucheng Chen, Chenyu Zhang, Tao He, Si Yong Yeo, Ming Li

arXiv:2605.22144v1cs.CV

TL;DR

Existing short-drama pipelines struggle with pacing, cross-clip spatial consistency, and production-level quality control. One Sentence, One Drama addresses these gaps with hierarchical multi-agent generation, 3D-grounded frame synthesis, reviewer loops, and structured post-production, outperforming existing pipelines in narrative quality, visual continuity, and viewing experience.

  • Problem

    Existing short-drama pipelines struggle with dramatic pacing, cross-clip spatial consistency, and production-level quality control when expanding brief ideas into videos.

  • Method

    One Sentence, One Drama uses hierarchical multi-agent story generation, 3D-grounded first-frame synthesis, reviewer loops, and planned scene transitions and BGM.

  • Results

    The framework consistently outperforms existing generation pipelines in narrative quality, cross-clip consistency, and overall viewing experience.

  • Takeaways & Limitations

    The results support structured agentic generation as a promising direction for controllable long-horizon video creation.

  • Takeaways & Limitations

    Higher controllability and production quality come with an estimated API cost of about $25-$27 per minute, exceeding Xiao Yun Que's $24.36 under the evaluation setting.

Abstract

from arXiv · show

Existing approaches for digital short-drama production typically rely on one-shot LLM generated scripts and loosely coupled pipelines, which fail to satisfy three key requirements of short-drama generation: (1) narrative pacing, resulting in weak hooks, insufficient escalation, and unattractive endings; (2) spatial consistency, leading to drifting scene layouts and inconsistent character positions across clips; and (3) production-level quality control, requiring extensive manual review and correction across script and visual stages. We present One Sentence, One Drama, a hierarchical multi-agent framework that transforms a user's single-sentence idea into a fully produced short drama through structured intermediate modules and iterative refinement. Our approach is built upon three key components: (1) a multi-agent debate-based story generation module that enforces short-drama pacing and narrative coherence; (2) a 3D-grounded first-frame generation mechanism that establishes a shared spatial reference for consistent character positioning and scene layout across clips; and (3) multi-stage reviewer loops that perform comprehensive error detection and targeted revision across script, visual, and video generation stages. We also introduce scene-level BGM matching and scene transition planning to improve the audience's immersive experience. To systematically evaluate this task, we introduce Short-Drama-Bench, a benchmark that extends standard video quality metrics with short-drama-specific criteria. Experimental results demonstrate that our method significantly outperforms existing pipelines in narrative quality, cross-clip consistency, and overall viewing experience.

1 Introduction

The paper presents One Sentence, One Drama, a hierarchical multi-agent framework that transforms a single-sentence idea into a complete short drama while addressing narrative pacing, spatial consistency, and production-level quality control. It also introduces Short-Drama-Bench to evaluate narrative engagement, spatial continuity, and full-production viewing experience.

  • Framework: Multi-agent debate-based story generation addresses weak dramatic hooks and unsatisfactory storylines by improving short-drama pacing and narrative coherence.Existing workflows often rely on ready-made stories or one-shot LLM expansion when users provide only brief ideas.
  • Framework: One Sentence, One Drama transforms a single-sentence idea into an entire short drama through structured, reversible intermediate modules and hierarchical multi-agent generation.The framework includes multi-agent debate-based story generation, visual and video production stages, and iterative review and revision loops.
  • Benchmark: Short-Drama-Bench augments standard video-quality metrics with short-drama-specific criteria for narrative engagement, spatial continuity, and full-production viewing experience.The benchmark contains 50 diverse story prompts spanning 7 popular categories.
  • Results: The agentic framework consistently outperforms existing generation pipelines in narrative quality, cross-clip consistency, and overall viewing experience.The paper also generates approximately 239 minutes of video across 50 benchmark prompts, including long-, medium-, and short-duration dramas.
  • Technical challenges: The framework jointly models narrative pacing, spatial consistency, and production-level coherence, converting one-shot generation into a controllable and self-refining process.Its pipeline uses structured generation, 3D-anchored visual references, and reviewer loops across script, visual, and video stages.

2 Personalized Short-Form Drama Generation

The framework converts a single sentence into structured story plans, scene and clip scripts, visual assets, 3D-anchored video, and post-production, with reviewer loops enforcing quality and consistency throughout. Its modules address pacing, spatial continuity, prompt executability, and scene transitions through hierarchical planning and shared scene references.

  • 2 Personalized Short-Form Drama Generation: A hierarchical sentence-to-video pipeline transforms one sentence into story plans, scene and clip scripts, visual assets, 3D-anchored video, and post-production with transitions and BGM.Reviewer loops are inserted across stages for quality control and cross-stage consistency.
  • 2.1 Hierarchical Episode Planning: An atomized corpus built from about 300 high-performing scripts decomposes each script into about 2,923 beat-level units for reusable pacing and dramatic packaging priors.The beats form the Pattern Bank, while overlapping local chunks form the Logic Bank for causal context.
  • 2.1 Hierarchical Episode Planning: Multi-agent story generation expands the user sentence into a seed skeleton and uses fact, logic, and pattern retrieval routes to support externally constrained content, causality, and dramatic structure.Scene-level review uses three independent LLM judges, with GPT-5.4 Pro deciding conflicting revision suggestions before patch-based rewriting.
  • 2.1 Hierarchical Episode Planning: Scene plans are decomposed into temporally ordered clips specifying narrative descriptions, shots, characters, props, dialogue or audio, actions, interactions, and initial and ending states.Clip-level review separately revises the opening hook, ending suspense, and middle-clip twist density.
  • 2.2 Visual Assets and Prompt Generation: Scene-level 360◦ panoramas and character assets provide environment and layout references, while paired prompts define each clip’s static first frame and temporal development.A prompt reviewer checks spatial consistency, physical plausibility, cross-clip continuity, and prop continuity before targeted rewriting.
  • 2.3 Keyframe-to-Video Generation with 3D Priors: Independent clip generation or tail-frame reuse can cause scene drift and fail under moving views, motivating consistent first-frame synthesis through 3D scene grounding.The method reconstructs a shared 3D world from a panorama and uses geometry-semantic-aware camera selection, character-conditioned generation, and frame repair.
  • 2.3 Keyframe-to-Video Generation with 3D Priors: Generated clips are registered to a shared 3D world, their camera trajectories are anchored and refined, and human meshes are aligned for next-clip planning in a common coordinate system.Tail-frame refinement uses background color, edge, and depth consistency, while human registration uses body keypoints and person masks.
  • 2.4 Post-Production and Assembly: Scene transitions are explicitly designed according to temporal shift, spatial shift, and character movement instead of simply concatenating independently generated scenes.When adjacent scenes continue in both time and space, the system uses a direct cut to preserve action immediacy.

3 Experiments

Experiments evaluate the framework with a short-drama benchmark, multiple quality dimensions, baseline comparisons, qualitative analyses, and component ablations. Results show improved narrative pacing, cross-clip spatial consistency, and production-level quality control.

  • Benchmark: Short-Drama-Bench contains 50 prompts across 7 popular categories and 17 fine-grained subcategories, with 2–3 representative samples per subcategory.The benchmark targets diverse short-drama patterns.
  • Metrics: Evaluation combines VBench low-level video metrics, adapted ViStoryBench multi-frame story visualization, and short-drama metrics for hooks, flow, continuity, and audio-transition quality.Narrative Hook measures opening and ending hooks using scene starts and ends.
  • Qualitative Results: Qualitative comparisons show that baselines exhibit cross-clip drift in character positions and background layouts, whereas our method better preserves spatial continuity.The comparison includes tail-frame and first-frame transitions between adjacent clips.
  • Qualitative Results: The multi-agent debate module strengthens opening conflicts and ending hooks, while reviewer loops automatically detect errors and trigger targeted revisions across script, image, and video stages.These mechanisms jointly address narrative pacing and production-level quality control.
  • Ablation Study: Removing Story Gen mainly degrades opening hooks, ending hooks, and escalation, while removing 3D First-Frame causes the largest drop in continuity metrics.The ablation attributes pacing and suspense construction to multi-agent debate and spatial continuity to 3D grounding.

4 Related Work

Related work includes video foundation models that provide strong short-clip fidelity and autoregressive methods that extend generation to minute-level videos. However, these approaches remain shot-level backends without multi-shot pacing or production-level coherence.

  • Video Generation Models: Modern video foundation models achieve strong per-clip fidelity but are limited to 5–15 seconds.The cited works include references [32] [49] [23] [37].
  • Long-Video Generation: Autoregressive long-video methods use teacher- or self-forcing rollouts to scale generation to minute-level interactive or drama-oriented videos.The cited methods include references [50] [21] [12] [56] [17] [53] [25] [46] [11].
  • Limitations: Despite longer generation, autoregressive methods remain shot-level backends without multi-shot pacing or production-level coherence.This limitation motivates approaches designed for coordinated short-drama production rather than isolated shot generation.

5 Conclusion … B.2 Story Visualization

The paper presents a hierarchical multi-agent framework for transforming a single-sentence idea into a complete personalized short drama, addressing pacing, spatial consistency, and production-level quality control. The appendices discuss broader impacts, related video-generation and story-visualization work, and associated copyright and licensing concerns.

  • 5 Conclusion: One Sentence, One Drama combines retrieval-augmented multi-agent story generation with 3D-grounded first-frame synthesis to address pacing, cross-clip spatial consistency, and production-level quality control.The framework generates complete personalized short dramas from single-sentence ideas.
  • 5 Conclusion: Experiments using automatic metrics, adapted story-visualization evaluation, and human ratings show improved narrative engagement, visual continuity, and overall viewing experience over existing pipelines.The results position structured agentic generation as a promising direction for controllable long-horizon video creation.
  • Appendix Overview: The appendix provides additional details on related work, the generation pipeline, benchmark construction, evaluation protocol, implementation settings, prompts, and responsible use.Its sections cover broader impacts, related work, story generation, transition clips, BGM planning, and other implementation details.
  • A Broader Impacts: The framework may broaden short-drama access by converting a human idea into scripts, visual assets, coherent clips, transitions, and BGM while reducing production costs and technical barriers.Potential beneficiaries include independent creators, educators, small studios, and users without professional filmmaking resources.
  • A Broader Impacts: Automated short-drama generation raises copyright and licensing concerns when stories, visual styles, voices, or music resemble protected works or commercial assets.The appendix recommends copyright-aware training and retrieval, licensed audio-visual assets, and clear generated-content ownership and attribution policies.
  • B.1 Video Generation: Foundation models have advanced text-to-video and image-to-video fidelity, motion realism, and prompt adherence, but typically remain limited to 5–15 seconds per clip.Representative systems include Sora, Veo, Kling, Seedance, CogVideoX, HunyunaVideo, and Wan.
  • B.2 Story Visualization: Prior narrative-video work explores LLM-guided planning, memory conditioning, and multi-agent collaboration to organize scripts, storyboards, layouts, and cross-scene consistency.Examples include VideoDirectorGPT, VideoStudio, StoryDiffusion, VideoGen-of-Thought, StoryAgent, MovieAgent, and ScriptAgent.
  • B. Related Work: Related work on story visualization includes layout planning and shared self-attention for cross-scene consistency, alongside decomposition methods for organizing narrative production.These approaches extend single-clip generation toward longer narrative videos.

B.3 Short-Drama Generation … D.2 BGM Planning & Mixing

The framework converts a single logline into a structured short drama through retrieval-supported drafting, multi-agent debate polishing, and scene-level transition and BGM planning. It targets pacing, causal coherence, visual continuity, emotional consistency, and dialogue clarity across generated clips.

  • B.3 Short-Drama Generation; C Details of the Multi-Agent Debating-based Story Generation Framework: Existing short-drama systems such as Toonflow and Xiaoyunque remain limited in their input requirements and production capabilities.Toonflow converts a complete novel through sequential character extraction, script generation, storyboarding, and video synthesis, while Xiaoyunque is a closed-source commercial product built on Seedance 2.0.
  • C.1 Atom Script Corpus Construction: Single-logline expansion commonly produces weak openings, underpowered ending hooks, and unstable local causal coherence.Reported failures include under-motivated actions, unclear evidence timing, and disconnected consequences between scenes.
  • C.2 Problem-Driven Retrieval: The method addresses missing story-planning support by expanding the logline into seed text and generating a structured plan across three complementary retrieval routes.The seed text contains a preliminary narrative skeleton and key conflict cues, while the retrieval plan identifies missing support.
  • C.3 Story Drafting: Story drafting first constructs a story core containing drama-level metadata and a structured scene plan for the entire narrative.The story core defines the title, theme, genre, overall framing, and scene-specific planning information such as boundaries, outlines, attractors, progression, and goals.
  • C.4 Multi-Agent Debate Polishing: Three independent frontier-LLM judges review each draft using six dimensions, must-fix issues with severity levels, and a visual-executability gate.The rubric covers logical integrity, opening strength, hook continuity, narrative clarity, reversal pacing, and payoff resolution.
  • C.4 Multi-Agent Debate Polishing: Deterministic aggregation merges the reviews into retained strengths, average scores, candidate fixes, and disputed items for revision.Disputes reflect score or severity discrepancies, conflicting fix directions, and high-risk signals such as failed visual-executability checks.
  • D Details of Diverse Transition Clips and BGM Planning & Mixing; D.1 Diverse Transition through Scenes: Adjacent scenes are connected through explicitly modeled transitions rather than simple concatenation to reduce abrupt pacing and ambiguous temporal or spatial context.The transition design addresses difficulties viewers face when inferring when and where the next scene occurs.
  • D.2 BGM Planning & Mixing: Scene-level BGM planning uses a 16-bucket drama-oriented library and adaptive mixing to maintain emotional consistency while preserving dialogue clarity.BGM levels are adjusted for dialogue density and LUFS gaps, with speech-preserving compression reducing music during dialogue-heavy regions and strengthening it elsewhere.

E Multi-Character 3D-Consistent First-Frame Generation. · F Human Rating Details

The framework extends 3D-grounded first-frame generation to multi-character clips by jointly registering characters, sampling cameras around their shared center, and reviewing synthesized frames. Human evaluation complements model-based assessment through anonymized 5-point Likert ratings from 20 volunteers.

  • E Multi-Character 3D-Consistent First-Frame Generation.: Multi-character generation first registers all required characters in the shared scene-level 3D world before sampling the next first-frame camera.The pipeline retains video trajectory anchoring, geometry-aware camera sampling, character-conditioned generation, and frame-level review from the single-character case.
  • E Multi-Character 3D-Consistent First-Frame Generation.: Characters absent from the tail frame are recovered from earlier visible frames using SAM 3D Body for meshes and keypoints and SAM3 for person masks.The primary character is registered from the tail frame, while other characters are located by scanning backward through the current clip.
  • E Multi-Character 3D-Consistent First-Frame Generation.: Each character is localized in the shared 3D world by recovering camera pose with CUT3R and refining mesh depth along the camera–character ray.The character transform is initialized from reconstructed body keypoints and refined by rendering translated meshes from the selected-frame camera.
  • E Multi-Character 3D-Consistent First-Frame Generation.: After registration, cameras are sampled around the center of all involved characters across varying radii, azimuths, and elevations.This replaces single-character-centered spherical sampling and encourages the next first frame to retain the required characters.
  • E Multi-Character 3D-Consistent First-Frame Generation.: Geometric filtering rejects candidate views with missing, tiny, occluded, or insufficiently visible characters, poor backgrounds, or excessive proximity to scene surfaces.Remaining candidates are passed to a semantic VLM filter to assess support for the next clip.
  • E Multi-Character 3D-Consistent First-Frame Generation.: The synthesis stage conditions on rendered multi-character geometry, multi-view character references, and previous clip context, followed by review of identities and relative positions.Rendered meshes provide pose, scale, and spatial-layout constraints, while references preserve identity and clothing.
  • F Human Rating Details: 20 volunteers rated anonymized samples using written instructions and a 5-point Likert scale, where 1 denotes very poor quality and 5 denotes excellent quality.The study was voluntary and provided no monetary compensation.

G Detailed Experiment Settings

The experiments combine API-based generation with local 3D and vision modules, and evaluate the method and baselines under specified hardware, prompt, and retry settings.

  • Implementation: The method’s local components run on a single NVIDIA RTX A6000 GPU with 48GB memory, primarily for CUT3R-based trajectory estimation and frame pose anchoring.The local 3D modules can alternatively run on CPU, but substantially slower runtime is expected.
  • Baseline Evaluation: Baselines use the same 48GB A6000 setup except StoryMem, whose experiments require an H200 GPU because of higher memory and runtime demands.All baselines are evaluated on the same 50 Short-Drama-Bench prompts under the main-paper protocol.
  • Retry and Selection Policy: Failed text, first-frame, and video-review items may be revised or regenerated at most three times, after which the highest-scoring available candidate is retained.For 3D-consistent first-frame generation, the system retains the top-8 camera views after geometric and semantic filtering.

H Time and Cost Analysis

The framework has a practical generation runtime dominated by video synthesis and achieves substantially lower production costs than professional short-drama production, despite added review and 3D-construction expenses. Parallelized asset generation and limited local 3D inference help keep runtime practical, while estimated API costs remain close to automated platforms.

  • Time Analysis: Nearly one hour is required for video generation of a typical complete short drama, making it the dominant wall-clock cost.Story generation and script refinement take about 10–15 minutes, image and visual-asset generation about 2–4 minutes, and each scene-level 3D world about 2–4 minutes.
  • Time Analysis: 1.5–2 hours, 2–3 hours, 4 hours, and 35 hours are the reported runtimes for Xiao Yun Que, Toonflow, ScriptAgent, and MovieAgent or StoryMem, respectively.The comparisons use the same benchmark prompts and evaluation setting.
  • Cost Analysis: $10.08 is the estimated cost of 60 seconds of video generation using Kling v3 Pro image-to-video at $0.168/s.The one-minute, 1080P estimate also includes approximately 30 generated images costing about $4.02 at $0.134 per image, alongside text/model review and 3D world construction costs.
  • Cost Analysis: $24.36/min and $21.53/min are the reported costs for Xiao Yun Que and Toonflow, respectively, while the proposed method is slightly higher but in a similar range.The additional cost mainly comes from multi-stage review, regeneration after failed review, and 3D world construction.
  • Cost Analysis: $293/min, $439/min, $732/min, and $1,464/min are reported professional production costs for Dramaland A-level, S-level, S+ productions, and live-action productions, respectively.The estimated API cost of about $25–$27/min is higher than existing automated short-drama platforms but remains far below these professional production costs.

I Limitations

The framework improves controllability and production quality but has higher generation costs, limited human-in-the-loop interaction, and practical audio-licensing constraints. These limitations motivate more interactive workflows and broader licensed music access.

  • Generation Cost: Estimated API cost is about $25-$27 per minutes for a one-minute generated drama, compared with about $24.36 per minutes for Xiao Yun Que.The higher cost reflects multi-stage generation, 3D world construction, automatic review, and retry mechanisms.
  • Human Interaction: The framework emphasizes automatic generation but provides limited human-in-the-loop interaction for reviewing and revising clips.Future workflows could expose reviewer scores, regenerate low-scoring clips, accept high-scoring clips, and route borderline cases to human creators.
  • Audio Licensing: Audio licensing constraints limit BGM style and emotional diversity because the current library mainly contains royalty-free or commercially usable music.A larger licensed library with explicit purchase or licensing options could improve audio quality while supporting commercial publishing requirements.

J Multi-Stage Review Metrics and Judge Models … J.4 Audio and BGM Review

The framework combines internal reviewer models for generation-time rewriting, regeneration, and candidate selection with external judges for post-generation benchmark evaluation. Review loops assess text, images, video, and audio to support narrative, spatial, temporal, character, and soundtrack quality.

  • J Multi-Stage Review Metrics and Judge Models: Internal reviewers trigger rewriting, regeneration, or candidate selection during generation, while external judges evaluate completed outputs only for benchmarking.The system summarizes reviewer metrics, judge models, and decision rules across both settings.
  • J Multi-Stage Review Metrics and Judge Models: The review architecture spans text, image, video, and audio stages, enabling targeted quality checks before, during, and after visual production.Each stage applies reviewers to modality-specific generation or evaluation decisions.
  • J.1 Text-Level Review: Text-level reviewers assess story quality, prompt executability, prop continuity, and additional scene information needed for 3D-consistent first-frame generation.These checks occur before visual generation.
  • J Multi-Stage Review Metrics and Judge Models: The reviewers collectively enforce content continuity across story elements, scene layouts, temporal clips, character states, and soundtrack choices.These functions are distributed across the text-level, image-level, video-level, and audio review modules.
  • J.2 First-Frame and Tail-Frame Image Review: Image-level reviewers select 3D-consistent first-frame candidates and route clips based on whether tail frames contain sufficient visual context for reuse.The review covers both first-frame selection and tail-frame routing.
  • J.3 Video-Level Review: Video-level reviewers evaluate visual physics, temporal continuity, reaction plausibility, and scripted character entrances, exits, and presence states.These reviews occur after clip generation.
  • J.4 Audio and BGM Review: Audio reviewers support scene-level BGM planning and selection by choosing suitable BGM buckets from textual scene context and scoring candidate audio segments.The highest-scoring candidate becomes the final soundtrack.

J.5 External Benchmark Evaluation

Final evaluation uses benchmark judge models and metrics on generated videos from the same 50 Short-Drama-Bench prompts, separately from the internal reviewer loops.

  • Evaluation setup: Table 7 defines the judge models and metrics used exclusively for final evaluation.These benchmark judges are distinct from the internal reviewer loops.
  • Evaluation setup: The benchmark judges assess generated videos from 50 Short-Drama-Bench prompts.The prompts are the same set used for the generated videos in final evaluation.

K Short-Drama-Bench Prompts and Generated Videos · L Script Library and BGM Library Details · M Prompt Templates for Each Stage

The appendix specifies Short-Drama-Bench’s diverse prompt coverage, the script and BGM retrieval libraries, and stage-specific prompt templates for evaluation and iterative review. These materials operationalize narrative, visual, video, audio, and transition quality assessment.

  • K Short-Drama-Bench Prompts and Generated Videos: Short-Drama-Bench spans seven high-level genres, each subdivided into fine-grained categories with representative prompts for diverse narrative evaluation.The genres include underdog comeback, social realism, ancient court intrigue, suspense and thriller, time travel and rebirth, sweet romance, and corporate/business war.
  • K Short-Drama-Bench Prompts and Generated Videos: Figures 7 and 8 provide representative generated videos across different prompt topics and narrative settings.They illustrate the visual results produced under Short-Drama-Bench conditions.
  • L Script Library and BGM Library Details: The script library derives 2,923 beat cards and 6,984 logic chunks from 300 high-performing original short-drama scripts.These structured references support plot rhythm, conflict escalation, genre-specific storytelling patterns, and stronger short-drama style priors.
  • L Script Library and BGM Library Details: The BGM library contains 8,122 tracks across 8 high-level categories and 40 fine-grained subcategories, paired with textual descriptions for scene-level audio matching.Matching is guided by scene rhythm and emotional intent.
  • M.1 Evaluation Prompt: The evaluation template instantiates different metrics, scopes, rubrics, and reference contexts to assess narrative quality, continuity, audio alignment, and transition naturalness.The benchmark judge uses a 1–5 integer scoring scale, with 5 reserved for outstanding cases with almost no visible weakness.
  • M.2 Text Review: The clip-level script review separately scores the first-clip hook, last-clip ending hook, and middle-clip twist density for targeted rewriting.The prompt excludes the first and last clips when evaluating reversal density in the middle clips.
  • M.3 Image Review: The first-frame reviewer selects among three candidates using temporal continuity, coarse layout consistency, background quality, character integrity, color continuity, and person-scene interaction.Scores range from 0 to 5, with scores of 2 or below requiring rejection.
  • M.4 Video Review: Generated video review checks physical realism, temporal continuity, reaction plausibility, and character presence consistency, routing failed clips to prompt revision or regeneration.The criteria detect artifacts including floating objects, gravity violations, object penetration, flicker, disappearance, teleportation, and abrupt state changes.
Loading 2605.22144v1…