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

Generating short dramas from a single sentence raises challenges in narrative pacing, cross-clip spatial consistency, and production-level quality control. One Sentence, One Drama uses a hierarchical multi-agent framework, and its results show improved narrative engagement, visual continuity, and viewing experience over existing pipelines.

  • Problem

    Existing short-drama pipelines often require ready-made stories or one-shot expansion, loosely connect clips, and need substantial manual review across production stages.

    Source: §1 · p. 2

  • Method

    One Sentence, One Drama turns a single-sentence idea into a complete short drama using structured multi-agent story generation, 3D-grounded first-frame synthesis, reviewer loops, and scene-level transition and BGM planning.

    Source: §1 · p. 2, §2 · p. 3, §5 · p. 9

  • Results

    The framework improves narrative engagement, visual continuity, and overall viewing experience over existing pipelines.

    Source: §1 · p. 3, §5 · p. 10

  • Takeaways & Limitations

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

    Source: §5 · p. 10

    Generation costs about $25–$27 per minute, compared with Xiao Yun Que's $24.36 under the evaluation setting.

    Source: I Limitations · p. 23

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

One Sentence, One Drama turns a single-sentence idea into a short drama through structured, self-refining generation that addresses narrative pacing, spatial consistency, and production quality. The paper also introduces Short-Drama-Bench for evaluating short-drama-specific qualities.

1 Introduction

  • Toonflow and Xiaoyunque can depend on ready-made stories or one-shot expansion, loosely connected clip generation, and substantial manual correction across production stages.Source: §1 · p. 2
  • Multi-agent debate and revision model opening hooks, conflict escalation, suspenseful endings, and storyline consistency, while 3D-grounded first frames align clips in a shared spatial coordinate system.The scene-level 3D world model is designed to preserve character positions and layouts even under viewpoint changes or scene re-establishment.Source: §1 · p. 2
  • Reviewer loops across scripts, prompts, keyframes, and videos enforce pacing, spatial, prop-state, physical-plausibility, and action-continuity constraints; scene-level BGM matching and transition planning support immersion.Source: §1 · p. 2
  • Short-Drama-Bench contains 50 prompts across 7 categories and 17 subcategories, with criteria for narrative engagement, spatial continuity, and full-production viewing experience.Source: §1 · pp. 2–3
  • The benchmark outputs total approximately 239 minutes, including 2 long dramas, 5 medium dramas, and 43 short dramas, testing consistency across hundreds of sequential clips.The durations are approximately 30, 10, and 3 minutes per drama, respectively.Source: §1 · p. 2
  • The framework consistently outperforms existing generation pipelines in narrative quality, cross-clip consistency, and overall viewing experience.Source: §1 · p. 3

2 Personalized Short-Form Drama Generation

The framework turns a one-sentence idea into a produced short drama through structured planning, visual and video generation, and post-production, with reviewer loops supporting quality control and cross-stage consistency. It combines short-drama-specific story planning with 3D-grounded shot continuity and scene-aware transitions and music.

2 Personalized Short-Form Drama Generation

  • The pipeline transforms a single-sentence input into story plans and scene- and clip-level scripts, visual assets, 3D-anchored video, and post-production, with reviewer loops across stages.Post-production includes scene transitions and BGM.Source: §2 · p. 3

2.1 Hierarchical Episode Planning

  • A corpus of about 300 scripts yields about 2,923 beat-level Pattern Bank units and a Logic Bank of overlapping chunks, supporting pacing and causal-context retrieval for story planning.The pipeline also retrieves external facts when needed, then combines fact, logic, and pattern atoms into a structured story core and scene plan.Source: §2.1 · pp. 3–4
  • Three LLM judges review the story draft and retrieval atoms; conflicting suggestions go to GPT-5.4 Pro, and selected issues receive local rewrites before clip scripts get targeted pacing review.The pipeline stores potentially useful removed hooks and reversals, then reviews opening hooks, ending suspense, and twist density through revisions limited to the first, last, or middle clips, respectively.Source: §2.1 · p. 4

2.2 Visual Assets and Prompt Generation

  • Scene panoramas, spatial anchors, and multi-view character references provide visual assets, while paired prompts specify each clip’s first-frame composition and subsequent action for review before rendering.The prompt reviewer checks spatial consistency, plausibility, cross-clip continuity, and prop continuity, then suggests targeted prompt revisions when it finds violations.Source: §2.2 · p. 4

2.3 Keyframe-to-Video Generation with 3D Priors

  • Next-shot first frames use geometrically feasible cameras and rendered scene and human references, with geometric and semantic filters selecting views that preserve scene anchors, layout, and character continuity.A vision-language model checks frame quality, and image generation repairs background defects while preserving the human, camera view, and scene layout.Source: Figure 3 · p. 5, §2.3 · pp. 5–6
  • For each scene, the system reconstructs a shared 3D world and registers the first frame and generated clip camera trajectory to it, aligning the tail frame and human model for next-clip planning.The first-frame pose is initialized from the selected panorama view and refined against the 3D world; clip trajectories are scale-calibrated and tail-frame poses are refined using background consistency.Source: §2.3 · p. 5

2.4 Post-Production and Assembly

  • Transitions are selected by temporal and spatial shifts and character movement, using direct cuts, temporal overlays, location-establishing shots, or motion bridges to support continuity without unnecessary narrative burden.Source: §2.4 · p. 6
  • Scene-level BGM is chosen from 8,122 tracks organized into 16 functional buckets, then mixed with generated audio using dialogue-aware volume adjustment, loudness calibration, and speech-preserving compression.An LLM selects primary and backup buckets per scene, and GPT-Audio scores candidate segments for emotional, narrative, rhythm, and transition fit.Source: §2.4 · p. 6

3 Experiments

The experiments evaluate the framework on Short-Drama-Bench using video, story-visualization, short-drama-specific, and human-rating measures. Results and ablations examine its overall performance and the contributions of its components to pacing, continuity, and production quality.

3.1 Script Corpus and BGM Library Setup

  • The narrative-planning database distills 300 high-performing scripts into 2,923 beat cards and 6,984 logic chunks, alongside a BGM library of 8,122 tracks.The library covers 8 high-level categories and 40 fine-grained subcategories.Source: §3.1 · p. 6

3.2 Experimental Settings

  • Short-Drama-Bench contains 50 prompts across 7 categories and 17 subcategories, and comparisons include story-visualization pipelines and short-drama products.The evaluation uses VBench [22], an adapted multi-frame ViStoryBench [57] protocol, short-drama-specific measures for hooks, flow, continuity, audio and transitions, model-based judges, and ratings from 20 annotators; baselines are evaluated on the same prompts, with Claude 4.6 Opus expanding prompts for pipelines not designed for single-sentence ideas.Source: §3.2 · pp. 6–7

3.3 Qualitative Analysis

  • Qualitative examples show that 3D-grounded first-frame generation better preserves scene layout and character blocking, while debate and reviewer loops strengthen hooks and trigger targeted revisions.The paper contrasts these results with baseline cross-clip drift and commercial-platform workflows that often require manual involvement.Source: §3.3 · p. 7, Figure 4 · p. 8

3.4 Quantitative Analysis

  • The method reports strong performance across standard video, story-visualization, and short-drama metrics, improving visual quality and story-level consistency without trading off low-level video quality.The analysis attributes baseline weaknesses to insufficient short-drama pacing, cross-clip spatial drift, and limited audio and transition planning; Table 1 also reports human ratings across the same short-drama criteria.Source: §3.4 · pp. 7–8, Table 1 · p. 9

3.5 Ablation Study

  • Ablations show that each component targets a distinct failure mode: story generation supports pacing, 3D first-frame synthesis supports continuity, review improves all metrics, and transition/BGM planning supports audio-transition quality.Removing Transition & BGM leaves clip-level continuity nearly unchanged because continuity is evaluated on adjacent clip frames before post-production assembly.Source: §3.5 · p. 8, Table 2 · p. 9

4 Related Work

Prior video-generation and narrative-planning methods extend clip or scene synthesis but leave short-drama pacing and production coherence unresolved. Existing short-drama systems also rely on sparse references and manual inspection, whereas this framework combines multi-agent story generation, 3D-grounded first frames, reviewer loops, BGM matching, and transition planning.

4 Related Work

  • Video foundation models [32] [49] [23] [37] deliver strong per-clip fidelity but are limited to 5–15 seconds; autoregressive methods [50] [21] [12] [56] [17] [53] [25] [46] [11] reach minute-scale generation without multi-shot pacing or production-level coherence.The autoregressive approaches use teacher- or self-forcing rollouts and support interactive or drama-oriented generation, but remain shot-level backends.Source: §4 · p. 9
  • Narrative-level layout, agent-based, and memory-conditioned methods [26] [27] [55] [54] [4] [41] [29] [20] [51] [2] [18] [14] target movie-style stories, assume curated inputs, and leave dense hooks, reversals, and compressed payoffs unmodeled.They often rely on fine-tuned visual modules and produce loosely coupled five-second shots with absent or post-hoc audio.Source: §4 · p. 9
  • Short-drama systems SkyReels-V1, Toonflow, and Xiaoyunque rely on full novels or one-shot LLM scripting, sparse-reference keyframes, manual inspection, hard cuts, and no scene-level audio or transitions.In contrast, this framework unifies retrieval-augmented multi-agent story generation, 3D-grounded first-frame synthesis, multi-stage reviewer loops, scene-level BGM matching, and space-time-aware transition planning.Source: §4 · p. 9

5 Conclusion … B.2 Story Visualization

One Sentence, One Drama turns a single-sentence idea into a complete personalized short drama using structured agents, 3D-grounded first frames, reviewer loops, and scene-level transition and BGM planning. The paper also introduces Short-Drama-Bench and reports improved narrative engagement, visual continuity, and viewing experience over existing pipelines.

5 Conclusion

  • The framework combines retrieval-augmented multi-agent story generation, 3D-grounded first-frame synthesis, multi-stage review, and scene-level transition and BGM planning, evaluated with Short-Drama-Bench.Experiments across automatic metrics, adapted story-visualization evaluation, and human ratings show improved narrative engagement, visual continuity, and overall viewing experience over existing pipelines.Source: §5 · pp. 9–10

Appendix Overview

  • The appendices detail related work, generation and review methods, benchmark and evaluation protocols, implementation settings, limitations, and responsible-use considerations.They also include prompt templates, library details, time and cost analysis, and supplementary code and dubbed video showcases.Source: Appendix Overview · p. 16

A Broader Impacts

  • Single-sentence-to-production generation may broaden creative access, lower cost and technical barriers, and support rapid prototyping, multilingual production, and personalized storytelling.Potential users include independent creators, educators, small studios, and people without professional filmmaking resources.Source: A Broader Impacts · p. 17
  • Deployment raises copyright and licensing concerns when generated stories, styles, voices, or music resemble protected works or commercial assets.The paper recommends copyright-aware training and retrieval sources, licensed audio-visual assets, and clear ownership and attribution policies.Source: A Broader Impacts · p. 17

B.1 Video Generation

  • Text- and image-to-video models have advanced visual fidelity, motion realism, and prompt adherence, but their typical 5–15-second clips cannot ensure multi-minute planning or cross-clip consistency.The framework therefore uses them as per-shot rendering backends; examples include Sora, Veo, Kling, Seedance [32], CogVideoX [49], HunyunaVideo [23], and Wan [37].Source: §B.1 · p. 17

B.2 Story Visualization

  • Prior story-visualization systems often require curated scripts, character references, detailed shot prompts, or local fine-tuning, while loosely coupled clips and limited pacing models leave short-drama needs unmet.Examples include MovieAgent [41], StoryMem [51], VideoDirectorGPT [26], VideoStudio [27], StoryDiffusion [55], VideoGen-of-Thought [54], StoryAgent [20], and ScriptAgent [29].Source: §B.2 · p. 17

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

The framework turns a single logline into a short drama using retrieval-supported story planning, iterative multi-agent review, and production methods for spatial continuity, scene transitions, and music. Its modules address weak pacing and causal coherence, visual drift, and manual quality-control burdens identified in existing short-drama systems.

B.3 Short-Drama Generation

  • The framework addresses existing systems’ weak hooks, brittle narrative logic, spatial drift, manual review, and hard-cut transitions with retrieval-augmented generation, 3D-grounded synthesis, reviewer loops, and scene-level audio and transitions.Toonflow requires a complete novel, while Xiaoyunque appears to use one-shot LLM expansion; both generate keyframes independently from a few references and rely on manual inspection.Source: §B.3 · p. 17

C.1 Atom Script Corpus Construction

  • Two retrieval banks provide complementary pacing and causal support, built from approximately 300 short-drama scripts distilled into script cards, roughly 3,000 beat units, and overlapping local text chunks.The Pattern Bank encodes reusable short-drama structures such as opening actions and closing-hook visuals; the Logic Bank preserves local narrative context, including motivations and consequence transitions.Source: §C.1 · p. 18

C.2 Problem-Driven Retrieval

  • A structured retrieval plan combines web evidence, top-k Logic Bank chunks, and top-k Pattern Bank beats, which are summarized as Fact, Logic, and Pattern Atoms for downstream planning.The routes target factual grounding, local narrative validity, and dramatic packaging, respectively; abstraction into atoms avoids directly copying source scripts.Source: §C.2 · p. 18

C.3 Story Drafting

  • Story drafting creates a scene-by-scene core with hooks and escalation, guided by five cross-scene progression lines and stable character, location, and prop assets.The knowledge-state line tracks what audiences and characters know, what remains hidden, and newly introduced evidence or state changes.Source: §C.3 · p. 18

C.4 Multi-Agent Debate Polishing

  • Three independent LLM judges score six story dimensions and flag must-fix issues, while aggregated disputes go to a Final Decider before targeted patch-based revision.The judges also assess visual executability; revisions preserve unaffected story elements, loop through review until quality thresholds are met or at most N rounds, and use an Idea Bank to selectively revive strong ideas that can return without harming logic or executability.Source: §C.4 · p. 19

D.1 Diverse Transition through Scenes

  • Transitions are selected according to temporal shift, spatial shift, and character movement, using direct cuts, temporal transitions, location-establishing shots, or motion bridges to clarify scene continuity.The approach aims to improve interpretability and viewing smoothness without adding narrative burden.Source: §D.1 · p. 19

D.2 BGM Planning & Mixing

  • Scene-level BGM selection uses 16 functional buckets and GPT-Audio scores candidate tracks for emotional, narrative, rhythm, and transition fit; adaptive mixing lowers music during dialogue to preserve clarity.BGM levels are calibrated using the LUFS gap between scene audio and the selected segment, with speech-preserving dynamic compression.Source: §D.2 · p. 20

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

Multi-character first-frame generation registers required characters in a shared 3D world, then adapts camera sampling and review to preserve their layout across clips. Human ratings use anonymized samples, a 5-point scale, and scores aggregated across 20 volunteers.

E Multi-Character 3D-Consistent First-Frame Generation.

  • The multi-character extension retains the scene-level 3D world, trajectory anchoring, geometry-aware camera sampling, conditioned synthesis, and frame review, but registers all required characters before selecting the next camera.This shared-world registration is the main change from the single-character pipeline.Source: E Multi-Character 3D-Consistent… · p. 20
  • For adjacent clips, the system registers characters needed in both clips, using the tail frame or scanning backward for sufficiently visible characters.For selected frames, SAM 3D Body [47] reconstructs meshes and keypoints, while SAM3 [10] supplies person masks.Source: E Multi-Character 3D-Consistent… · pp. 20–21
  • CUT3R [39] recovers each selected frame’s camera pose; mesh depth is refined along the camera ray to match the person mask while preserving keypoint alignment.All characters are thereby represented as meshes in the same world coordinate system.Source: E Multi-Character 3D-Consistent… · p. 21
  • Camera candidates are sampled around the center of all involved characters, helping keep them in view while preserving their relative spatial arrangement.Candidates vary in radius, azimuth, and elevation.Source: E Multi-Character 3D-Consistent… · p. 21
  • Geometric and semantic filters reject unsuitable views based on character visibility, scene geometry, and fit to the next-clip prompt before ranking views for synthesis.The semantic filter checks scene anchors, interaction direction, and relative blocking.Source: E Multi-Character 3D-Consistent… · p. 21
  • Synthesis conditions on rendered multi-character layouts, identity references, and prior-clip context; a reviewer checks identities, relative positions, and background, repairing or resampling failures.This supports spatial continuity across adjacent clips.Source: E Multi-Character 3D-Consistent… · p. 21

F Human Rating Details

  • The human study recruited 20 unpaid volunteers, used anonymized and randomly shuffled samples with a 5-point Likert scale, and averaged participant scores by method and metric.Evaluation units matched the model-based protocol, and final human ratings are reported in Table 1.Source: F Human Rating Details · p. 22

G Detailed Experiment Settings

Experiments combine API-based generation modules with local 3D and vision inference, using a common evaluation set and consistent review procedures. The local components run on a 48GB NVIDIA RTX A6000 or, more slowly, on CPU.

G Detailed Experiment Settings

  • Local components run on one 48GB NVIDIA RTX A6000; CPU execution is available but substantially slower.CUT3R, used for video trajectory estimation and frame pose anchoring, accounts for most GPU memory use; VGGT, SAM 3D Body, SAM3, and rendering also run locally.Source: G Detailed Experiment Settings · p. 22
  • Baselines are evaluated on the same 50 Short-Drama-Bench prompts under the main-paper protocol, except StoryMem [51], which uses an H200.Most baseline runs use the same 48GB A6000 setup; StoryMem uses an H200 because of its higher memory and runtime requirements.Source: G Detailed Experiment Settings · p. 22
  • Failed text, first-frame, and video reviews allow up to three revisions or regenerations, after which the best reviewer-scored candidate is retained.For 3D-consistent first-frame generation, the system retains the top eight camera views after geometric and semantic filtering, then selects the final frame by reviewer score.Source: G Detailed Experiment Settings · p. 22

H Time and Cost Analysis

The analysis estimates practical runtime and API costs for producing short dramas, comparing the pipeline with automated platforms and professional production.

Time Analysis

  • Our method produces a complete 10-minute drama in about 74–90 minutes, versus 1.5–2 hours for Xiao Yun Que and 2–3 hours for Toonflow.ScriptAgent [29] takes about 4 hours, while MovieAgent [41] and StoryMem [51] take about 35 hours in the evaluation setting; runtime estimates vary with API latency, queueing, video duration, and reviewer-triggered retries.Source: H Time and Cost Analysis · p. 22
  • Video generation takes near one hour; story generation and script refinement take 10–15 minutes, image generation 2–4 minutes, and 3D world construction 2–4 minutes per world.Image and scene-level asset generation can be parallelized, while local 3D inference is a small fraction of total pipeline time.Source: H Time and Cost Analysis · pp. 22–23

Cost Analysis

  • About $17.3/min is the base API cost, rising to an average $25–$27/min with reviewer-triggered retries or repairs.The one-minute 1080P estimate includes $10.08 for 60 seconds of video generation, about $4.02 for 30 images, $2.0 for text generation and review, and $1.2 for one 3D world; around half of the text, image, and video budget goes to retries or repair.Source: H Time and Cost Analysis · p. 23
  • At about $25–$27/min, our estimated API cost is slightly higher than Xiao Yun Que’s $24.36/min and Toonflow’s $21.53/min.The additional cost mainly comes from multi-stage review, regeneration after failed review, and 3D world construction, which the passage identifies as targets for future optimization.Source: H Time and Cost Analysis · p. 23
  • Our estimated $25–$27/min API cost is far below professional production estimates: $293/min for A-level, $439/min for S-level, $732/min for S+ productions, and $1,464/min for live-action productions.The passage says the method provides stronger controllability, reviewer-based refinement, and cross-clip spatial consistency than professional-cost comparisons alone convey.Source: H Time and Cost Analysis · p. 23

I Limitations

The framework’s limitations include higher generation cost, limited human-in-the-loop interaction, and restricted BGM diversity due to licensing constraints. The paper outlines interactive review and expanded licensed music options as possible future improvements.

I Limitations

  • A one-minute drama costs about $25-$27 per minutes on average, versus Xiao Yun Que’s $24.36 per minutes under the evaluation setting.The authors note that Xiao Yun Que is closed-source, so its internal pipeline and true operating cost are not fully observable.Source: I Limitations · p. 23
  • The framework offers limited human-in-the-loop interaction; a hybrid interface could route borderline clips to creators and reduce unnecessary retries.The proposed interface would expose reviewer scores and diagnostic feedback, enabling low-scoring clips to be regenerated and high-scoring clips accepted directly.Source: I Limitations · p. 23
  • The BGM library’s focus on royalty-free or commercially usable music limits available styles and emotional expressions.The authors propose a larger licensed library with explicit purchase or licensing options for matched tracks.Source: I Limitations · p. 23

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

The section distinguishes reviewers that guide generation from judges used only for benchmark evaluation, then outlines review across text, images, video, and scene-level BGM.

J Multi-Stage Review Metrics and Judge Models

J.1 Text-Level Review

  • Text-level reviewers assess story quality, prompt executability, prop continuity, and whether the next clip needs added scene information for 3D-consistent first-frame generation.Source: §J.1 · p. 24

J.2 First-Frame and Tail-Frame Image Review

  • Image-level reviewers select 3D-consistent first-frame candidates and decide whether a previous tail frame has enough visual context for direct reuse.Source: §J.2 · p. 24

J.3 Video-Level Review

  • Video-level reviewers assess visual physics, temporal continuity, reaction plausibility, and whether character entrances, exits, and presence states match the script.Source: §J.3 · p. 25

J.4 Audio and BGM Review

  • Audio reviewers select suitable BGM buckets from scene text, then score candidate segments to choose the final soundtrack.Source: §J.4 · p. 25

J.5 External Benchmark Evaluation

External benchmark evaluation used separate judge models and metrics on generated videos from the same 50 Short-Drama-Bench prompts, distinct from the internal reviewer loops.

J.5 External Benchmark Evaluation

  • Table 7 summarizes the judge models and metrics used only for final evaluation, separate from internal reviewer loops and applied to videos from the same 50 prompts.The prompts were from Short-Drama-Bench.Source: §J.5 · p. 25

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

The appendices document Short-Drama-Bench’s prompt coverage and generated examples, the script and BGM libraries, and prompt-based evaluation and review stages for scripts, images, and videos.

K Short-Drama-Bench Prompts and Generated Videos

  • Short-Drama-Bench covers seven high-level short-drama genres, each divided into fine-grained subcategories with representative story prompts, designed to evaluate diverse narrative settings, character relationships, conflict structures, and genre-specific storytelling patterns.Source: K Short-Drama-Bench Prompts and… · p. 25
  • Figures 7 and 8 present generated video examples from Short-Drama-Bench across different prompt topics and narrative settings.The passages identify these as representative examples, without reporting further comparative results.Source: K Short-Drama-Bench Prompts and… · p. 28, Figure 7 · p. 31, Figure 8 · p. 32

L Script Library and BGM Library Details

  • The script database distills 300 high-performing original short-drama scripts into 2, 923 beat cards and 6, 984 logic chunks for retrieval-based narrative planning.The references support plot rhythm, conflict escalation, and genre-specific storytelling patterns.Source: L Script Library and BGM Library Details · p. 28
  • The BGM library contains 8, 122 tracks across 8 high-level categories and 40 fine-grained subcategories, paired with descriptions for scene-level audio matching.Matching is guided by scene rhythm and emotional intent.Source: L Script Library and BGM Library Details · p. 28, Figure 9 · p. 33

M.1 Evaluation Prompt

  • The unified evaluation prompt varies metrics, scopes, rubrics, and reference contexts to assess narrative quality, continuity, audio alignment, and transition naturalness.Figure 10 shows the shared template used for model-based Short-Drama-Bench evaluation.Source: §M.1 · p. 28, Figure 10 · p. 34

M.2 Text Review

  • The clip-level script review separately assesses first-clip hooks, last-clip ending hooks, and middle-clip twist density to target rewrites without changing unrelated scene parts.Figure 11 shows the review prompt used before visual generation.Source: §M.2 · p. 28, Figure 11 · p. 35

M.3 Image Review

  • First-frame candidate review scores temporal continuity, coarse layout, background quality, character integrity, color continuity, and person-scene interaction before video generation.Figures 12 and 13 show the prompt and scoring criteria for selecting 3D-consistent candidates.Source: §M.3 · p. 28, Figure 12 · p. 36, Figure 13 · p. 37

M.4 Video Review

  • Generated-video review checks physical realism, temporal continuity, reaction plausibility, and character-presence consistency; failed clips go to prompt revision or regeneration.Figure 14 shows the video review prompt.Source: §M.4 · p. 28, Figure 14 · p. 38

Related papers

Loading 2605.22144v1…