acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

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 , 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 , 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.

Primary area:
applications to robotics, autonomy, planning
Keywords:
multi-agent systems, story generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_6miPOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents One Sentence, One Drama, a hierarchical multi-agent framework that transforms a single-sentence idea into a complete short drama. The system combines retrieval-augmented story planning, multi-agent script review, an Idea Bank that selectively restores discarded creative elements, 3D-grounded first-frame generation, multi-stage quality review, and scene-level transition and background-music planning.

The central technical idea is to maintain a shared 3D reference for scene geometry, character placement, and camera selection across clips. The paper also introduces Short-Drama-Bench, comprising 50 prompts across seven categories, and evaluates narrative engagement, visual continuity, and audiovisual presentation using model-based metrics and a human study.

Critical strengths
  1. A coherent formulation of an important practical problem. The paper recognizes that complete short-drama generation requires more than high-quality individual clips. Its joint treatment of narrative pacing, spatial continuity, and quality control is well motivated.

  2. A technically meaningful approach to spatial continuity. The shared 3D scene enables viewpoint changes while preserving character placement and environmental layout. Geometry-adaptive registration addresses a concrete ambiguity in placing single-image human reconstructions into a scene. Its isolated ablation reports substantial improvements in penetration, floor-contact error, and visible-region overlap.

  3. A thoughtful mechanism for preserving creativity during revision. The Idea Bank addresses the plausible failure mode that repeated correction makes narratives safer but less memorable. Although the reported gains are modest, this is an interesting and potentially transferable design.

  4. Evaluation beyond conventional video quality. The benchmark includes opening hooks, conflict escalation, scene-ending suspense, spatial continuity, music alignment, and transitions. These criteria better reflect the intended application than visual fidelity alone.

  5. Useful supporting evidence and practical reporting. The human study, evaluator sensitivity analysis, confidence intervals, and runtime/cost estimates make the evaluation more informative. The main ablations also show relatively interpretable relationships between removed components and affected metrics.

Critical weaknesses
  1. Risk of "teaching to the test" in the core metrics. The eight narrative/continuity metrics of Short-Drama-Bench are scored by three LLM judges, yet the system's internal reviewer loops (Tables 3–5) optimize against essentially the same rubric. The pipeline effectively performs up to three rounds of retry-and-selection along the direction of the evaluation rubric, while none of the baselines have such a mechanism.

  2. No compute/cost-matched baseline comparison. Neither the main table nor the ablations include a control of "baseline + equivalent retries / best-candidate selection." Given that about half of the method's budget is spent on reviewer-triggered retries (Appendix H), one cannot disentangle how much of the gain comes from the architectural design versus plain best-of-N sampling.

  3. Potential distributional overlap between the benchmark and the retrieval corpus. The Pattern Bank is built from 1,102 "high-performing" short dramas, while all 50 benchmark prompts are highly formulaic tropes. Is the benchmark measuring generalization, or the ability to retrieve nearest-neighbor tropes from the Pattern Bank? The paper reports no retrieval-overlap analysis or novelty metric.

  4. Limited methodological novelty and the contribution is essentially systems engineering. All components are off-the-shelf models, and the contribution lies in orchestration and reviewer-loop design.

Questions for authors
  1. Since the internal reviewers optimize criteria closely aligned with Short-Drama-Bench, how do the authors distinguish general improvements in drama quality from optimization toward the evaluation rubric? Do the gains persist under independent human preference judgments or an alternative rubric not used during generation?

  2. Could the authors compare against baselines augmented with equivalent retries and best-of-N selection under a matched generation budget? How much improvement remains when controlling for candidate count, reviewer calls, and total inference cost?

  3. Could the authors quantify the similarity between benchmark prompts and the retrieval corpus, including overlap in plot structures and dramatic patterns? Does the method retain its advantage on unfamiliar trope combinations or held-out narrative categories, and how does performance change when retrieval is disabled?

  4. Which components constitute methodological advances beyond integrating existing models?

6 comments
AuthorsOpenReview ↗

Response to Reviewer 6miP (1/4): W1 & Q1, W2 & Q2

W1 & Q1: Risk of Optimizing toward the Evaluation Rubric

We thank the reviewer for the comment. Explicitly using the core quality criteria of short dramas as quality-control targets during generation is an intentional design choice. Opening hooks, conflict escalation, ending suspense, and cross-shot continuity are widely recognized as core quality criteria of short dramas: an interview study with short-drama practitioners reports that the typical structure of short dramas consists of a strong opening, rapid plot reversals, and suspense left at the end (Cao et al., CHI 2026); practitioners commonly follow the "golden three seconds" rule, and research on mobile video shows that 47% of a video's value is delivered in the first three seconds (Facebook, 2016). Meanwhile, our evaluation pipeline is independent of the generation pipeline in the following aspects.

1. Independence in the existing setup.

  • (a) The VBench and ViStoryBench metrics are never used during generation, and our method achieves the best results on these metrics.
  • (b) The human evaluation is conducted by 20 annotators in a blind setting, with anonymized samples presented in randomized order.

2. Per-judge analysis and human–LLM agreement (Appendix K).

  • (a) Human ratings are highly consistent with the LLM-judge scores: across the eight metrics, the median prompt-level Pearson correlations range from 0.878 to 0.960, and the median Spearman correlations range from 0.848 to 0.949 (Appendix K.1).
  • (b) All three judges rank our method first (Appendix K.2). Notably, Seed 2.0 Pro does not belong to the model family of any internal reviewer; under its scores alone, our method obtains 3.97, compared with 3.60 for Xiao Yun Que, the strongest baseline.
  • (c) In the independent human evaluation, our method improves over Xiao Yun Que by 0.349 (95% CI: [0.191, 0.515]), consistent with the improvement of 0.375 in the LLM-based evaluation (Appendix K.3). This shows that our gains also hold under independent human evaluation.

W2 & Q2: Comparison against Best-of-N Baselines

We thank the reviewer for the suggestion. We augment each baseline with best-of-N sampling: for each case, each baseline generates N = 3 candidates (matching the maximum number of retries per stage in our method), and the candidate with the highest average score over the eight final evaluation metrics is selected for comparison. This effectively uses the evaluation itself to select results for the baselines, which favors the baselines, while our results are reported without such selection. Due to the limited rebuttal period, we conduct this test on three cases and report the average scores.

Under the best-of-N setting, our method still outperforms all baselines on all metrics (see the table below), indicating that additional repeated sampling does not close the gap on these cases.

Two further points are worth noting.

  • (1) Our retries differ from best-of-N in mechanism: the reviewers first identify the specific error type and then modify only the problematic part (Appendix T), rather than resampling the whole output and selecting among candidates.
  • (2) In terms of cost, a single run of Xiao Yun Que, the strongest baseline, costs about USD 24.36 per minute (Appendix H), which is already comparable to ours (about USD 25–27 per minute); under the best-of-3 setting, the total budget of the baselines exceeds that of our method.

Comparison with best-of-3 baselines on the Short-Drama-Bench metrics, averaged over three cases. For each baseline, the candidate with the highest average score over the eight metrics is selected.

MethodOpening HookEnd HookEscalationCoherenceChar. SpatialEnv. LayoutBGMTransition
MovieAgent (best-of-3)2.332.071.851.331.721.891.301.44
ScriptAgent (best-of-3)3.003.002.333.333.403.332.672.33
StoryMem (best-of-3)1.562.561.522.222.702.892.152.44
LongLive (best-of-3)2.001.441.521.562.833.211.442.17
HoloCine (best-of-3)3.221.962.071.892.913.042.002.56
ToonFlow (best-of-3)3.893.633.523.562.793.254.193.33
Xiao Yun Que (best-of-3)4.003.933.704.223.493.383.413.67
Ours4.114.333.744.563.744.174.373.72
AuthorsOpenReview ↗

Response to Reviewer 6miP (2/4): W3 & Q3

W3 & Q3: Overlap between Short-Drama-Bench and the Retrieval Corpus

We thank the reviewer for the comment.

1. The Pattern Bank stores abstract beat-level patterns rather than complete plots. The units in the Pattern Bank are beat-level structural cards (opening action, conflict function, and closing hook) rather than complete scripts. Retrieved results are compressed into Pattern Atoms, which serve as pacing priors for story generation; the original text is never reused directly.

2. Similarity between the benchmark and the retrieval corpus. Using the Sentence-Transformers model msmarco-distilbert-cos-v5, we encode the 50 benchmark prompts and the titles of the 1,102 short dramas, and compute their cosine similarities. None of the prompts is identical to any drama in the corpus, and the average cosine similarity between each prompt and its most similar drama is 0.434. As a reference, the average cosine similarity between each drama in the corpus and its most similar other drama is 0.579. In other words, the benchmark prompts are less similar to the corpus than different dramas within the corpus are to each other. This level of similarity mainly stems from high-level elements shared across the short-drama genre (e.g., reversals and time travel).

3. Ablation with retrieval disabled. Due to the limited rebuttal period, we compare scripts generated with and without retrieval on three cases, scored by the same script reviewers in both settings. With retrieval disabled, all scores decrease, indicating that the pacing priors provided by retrieval are effective. Meanwhile, the scores without retrieval remain high (8.22–8.89), indicating that our advantage does not rely solely on retrieval.

Script scores with and without retrieval (0–10 scale), averaged over three cases.

SettingLogicOpeningHook Cont.ClarityReversalPayoff
w/o retrieval8.228.898.788.338.568.67
w/ retrieval8.569.119.338.788.678.89

4. Evaluation on an uncommon theme. We further construct a prompt with a theme that is uncommon in the corpus: "Bonnie-and-Clyde Thieves Have One Day to Steal the Mona Lisa from the Louvre". Its nearest-neighbor similarity to the corpus is 0.283, lower than the benchmark average of 0.434. We compare our method with Xiao Yun Que, the strongest baseline, on this prompt.

Comparison with Xiao Yun Que on an uncommon theme.

MethodOpening HookEnd HookEscalationCoherenceChar. SpatialEnv. LayoutBGMTransition
Xiao Yun Que4.003.224.003.333.042.524.002.67
Ours4.334.894.114.674.214.314.334.00
AuthorsOpenReview ↗

Response to Reviewer 6miP (3/4): W4 & Q4 (Part 1)

W4 & Q4: Methodological Advances beyond System Integration

We thank the reviewer for the comment. We would like to clarify the positioning and academic contributions of this work.

(1) Innovations at the mechanism and algorithm level. The challenge lies not in invoking multiple models, but in coordinating heterogeneous agents along a long pipeline, where errors at any stage propagate downstream and are amplified. We address this challenge with three coupled mechanisms.

  1. Retrieval-augmented multi-agent debate for script generation. Agent roles, structured intermediate states, and debate-based revision, supported by the complementary Pattern Bank and Logic Bank, enable pacing control tailored to short dramas. A concrete algorithmic design is the Idea Bank with final-round revival (Appendix C.4). Iterative revision has a systematic bias: each round removes risks but also deletes striking elements, making the script safer yet blander. We record every deleted or weakened element with its reason, and a final revival step restores those that can return without harming logic or visual executability.
  2. 3D-grounded cross-clip state propagation. Camera trajectories and the spatial states of characters are registered in a shared 3D scene, so the next viewpoint is geometrically constrained rather than relying on the implicit memory of the model. The core algorithm is geometry-adaptive character registration (Section 2.3 and Appendix E). A person mask determines the position and orientation of the body in the image but not its depth: a person twice as large and twice as far away yields the same mask. The algorithm therefore selects a registration strategy according to the geometric cues in the shot: occlusion matching (when an occluder is in front, the visible silhouette changes with depth, and the body is localized at the depth that best matches the mask), ground-depth reading (the scene depth beneath the visible feet determines the body depth, without requiring a standing pose), or floor contact (for standing characters, the lowest point of the body is placed on the scene floor).
  3. Hierarchical multi-agent review and targeted repair. In multi-stage generation, a problem is often detected only in the video, while its cause may lie in an earlier first frame or in a specific sentence of the prompt. Our hierarchical multi-agent review forms a Detection–Localization–Repair loop across the text, image, and video levels (Appendix T).
    • (a) Top-down detection. From prompts to first frames to videos, the generated content becomes increasingly concrete and costly to produce. We place a review gate at each level and verify each level against the content already fixed at the level above: the prompt is checked for whether the states of characters and props connect with the preceding and following clips, the first frame is checked against the character references and the video prompt, and the video is checked against the script. Each error is thus intercepted at the earliest and cheapest level where it becomes observable.
    • (b) Bottom-up localization. The video reviewer maps each problem, with timestamped evidence, to the sentence in the video prompt that induced it; if the problem already exists in the first frame, the localization moves to the first frame, where the system further determines whether the image itself or its underlying text is at fault.
    • (c) Targeted local repair. Only the localized part is repaired: the faulty sentence is replaced, or the first frame is regenerated with the defect appended to its prompt, while the characters, dialogue, speaking order, and plot events remain unchanged.
    • (d) Empirical support. Removing the multi-stage review degrades all eight metrics (Table 2). On 25 clips that failed their first video review, targeted repair fixes more clips than rewriting the entire prompt while better preserving the dialogue and plot (Appendix T).
Repair methodFixed in one roundDialogue preservedPlot events preserved
Full rewriting15 / 25 (60%)40 / 54 (74%)85.3%
Targeted repair (Ours)20 / 25 (80%)54 / 54 (100%)91.7%
Reviewer_KgoZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents One Sentence, One Drama, a hierarchical multi-agent framework that turns a single-sentence idea into a complete short drama. The framework combines retrieval-based story planning with multi-agent debate to improve narrative pacing and coherence, using an Idea Bank to preserve and selectively restore creative elements removed during revision. It also introduces 3D-grounded keyframe generation with geometry-adaptive character registration to maintain spatial consistency across clips. Reviewer loops check and refine scripts, prompts, keyframes, and generated videos, followed by scene transition planning and background music matching. The authors introduce Short-Drama-Bench, containing 50 prompts across seven categories, and report improvements over existing pipelines in narrative quality, cross-clip consistency, and overall viewing experience.

Critical strengths
  1. The paper is well-written, with a clearly organized pipeline and detailed supplementary explanations.
  2. The framework integrates story planning, video generation, and post-production to create short dramas from a single-sentence idea, which is comprehensive in multiple stages.
  3. The proposed 3D-grounded first-frame generation mechanism establishes a shared spatial reference for consistent character positioning and scene layout across video clips, which is reasonable
Critical weaknesses
  1. The novelty over existing multi-agent filmmaking systems needs to be clearer. The paper combines agent collaboration, review, and iterative revision, but related mechanisms have already been explored. For example, EditDuet[1] uses an Editor–Critic loop, while Hollywood Town[2] supports group discussion and feedback-driven refinement. FilMaster[3] and Co-Director[4] are also relevant to cinematic generation with multi-agent pipeline, but are not discussed in the provided manuscript. Although the proposed 3D scene constraints and Idea Bank may offer distinct contributions, the paper should explain more clearly what is new beyond combining existing agent workflows, and which improvements come from these specific designs.

  2. The supplementary videos raise concerns about identity consistency and the effectiveness of the review process. In the Chinese version, the woman at 3:07–3:42 appears noticeably different from her appearance elsewhere. In the English version, the elderly man puts on his glasses and removes them almost immediately at 0:02–0:06, which appears unmotivated in context. The paper includes review and retry mechanisms, but it is unclear why these issues remain in the final outputs. In particular, the character-presence check in Appendix J does not evaluate identity similarity or action quality.

  3. Comparison with representative commercial filmmaking platforms would strengthen the evaluation. The inclusion of XiaoYunQue is useful. However, a comparison with platforms such as Katalist[5] would help assess the practical value of the proposed pipeline. Google Flow[6] may also provide a useful comparison for reference-based consistency and multi-shot generation.

[1] M Sandoval-Castaneda, et al. EditDuet: A Multi-Agent System for Video Non-Linear Editing. SIGGRAPH Conference Papers ’25, Article 2, pp. 1–11, 2025.

[2] Z Wei, et al. Hollywood Town: Long-Video Generation via Cross-Modal Multi-Agent Orchestration. arXiv preprint arXiv:2510.22431, 2025.

[3] K Huang, et al. FilMaster: Bridging Cinematic Principles and Generative AI for Automated Film Generation. arXiv preprint arXiv:2506.18899, 2025.

[4] Y Song, et al. Co-Director: Agentic Generative Video Storytelling. arXiv preprint arXiv:2604.24842, 2026.

[5] Katalist AI. Katalist: AI Storytelling Platform. Commercial AI storyboarding and video production platform. https://www.katalist.ai/storytelling

[6] Google. Google Flow. AI filmmaking and creative production platform. https://flow.google.com/

Questions for authors
  1. How is the ID consistency checked across the generated shots? Were these examples flagged by the reviewers, or did they remain after the retry budget was exhausted? These examples would help clarify whether the main limitation lies in error detection or correction.

  2. For the commercial model indicated in weakness 3, could the authors include a suitable comparison to see the advantages of the proposed method, or explain why these platforms cannot be evaluated under a comparable setting?

8 comments
AuthorsOpenReview ↗

Response to Reviewer KgoZ (1/3): W1 (Part 1)

W1: Compare With Existing Multi-Agent Filmmaking Systems

We thank the reviewer for pointing out these related works. We have cited and discussed EditDuet, Hollywood Town, FilMaster, and Co-Director in the related work section of the revised version (Appendix B, highlighted in blue). Below, we clarify our differences from these works and the contribution of each design.

1. Hierarchical multi-agent review and targeted local repair. Existing methods either review only a single stage or cannot trace an error in the video back to specific upstream content.

  • EditDuet edits existing footage: the videos are converted into text descriptions in advance, and its Critic is a text-only model that reads only the timeline (clips and their text descriptions) and gives text feedback; it does not look at the video frames, nor does it generate or correct images or videos.
  • FilMaster's feedback occurs only in post-production: simulated audiences critique the structure and pacing of the rough cut, after which generated clips are reordered, removed, or adjusted in duration; it has no stage that checks the generated images or videos for errors, and it does not regenerate content.
  • Co-Director refines only the storyline and keyframes locally; for the generated video, it runs the entire pipeline four times with different creative directions and keeps the best result, thereby sidestepping the problem of tracing a final failure back to a specific prompt.
  • Hollywood Town has no separate text or image review stage: when a generated shot is found to violate script continuity after video generation, a revision request is sent back to the scriptwriter to revise the script.

In contrast, our hierarchical multi-agent review forms a Detection–Localization–Repair loop across the text, image, and video levels. In multi-stage generation, a problem is often detected only in the video, while its cause may lie in an earlier first frame or in a specific sentence of the prompt (Appendix T, highlighted in blue).

  • (a) Top-down detection. Each level is checked against the content already fixed at the level above: the prompt is checked for whether the states of characters and props connect with the preceding and following clips, the first frame is checked against the character references and the video prompt, and the video is checked against the script. Each error is thus intercepted at the earliest level where it becomes observable and cheapest to fix, rather than being carried into more expensive generation stages.
  • (b) Bottom-up localization. For each problem, the video reviewer gives timestamped evidence and its cause, and maps it to the sentence in the video prompt that induced it; if the problem already exists in the first frame, the localization moves to the first frame and further determines whether the problem lies in the image itself (e.g., incorrect body structure or identity) or in a specific sentence of the text it is based on.
  • (c) Targeted local repair. For a problem in the image, the defect is appended to the first-frame prompt and the first frame is regenerated; for a problem in the text, only the sentence that induced it is replaced. During repair, the characters present, dialogue, speaking order, and plot events remain unchanged; only blocking, character movement, camera movement, and shot size are adjusted.
  • (d) Empirical support. Removing the multi-stage review degrades all eight metrics (Table 2). On 25 clips that failed their first video review, targeted local repair fixes more clips than rewriting the entire prompt while better preserving the dialogue and plot (Appendix T).
Repair methodFixed in one roundDialogue preservedPlot events preserved
Full rewriting15 / 25 (60%)40 / 54 (74%)85.3%
Targeted local repair (Ours)20 / 25 (80%)54 / 54 (100%)91.7%
AuthorsOpenReview ↗

Response to Reviewer KgoZ (2/3): W1 (Part 2)

W1 (continued)

2. Further algorithmic innovations.

  • (a) Multi-judge debate with an Idea Bank: multiple LLM judges review independently, an arbiter resolves their disagreements, and a Reviser performs only patch-based local edits. Iterative critique-and-revision has a common pitfall: to fix logical issues, critics tend to remove risky but creative content, making the result increasingly bland. The Idea Bank stores such removed high-value ideas and restores them in the final round as long as they do not break the logic. This mechanism applies to other creative generation tasks based on LLM self-refinement.
  • (b) Geometry-adaptive 3D character registration: a human mesh reconstructed from a single frame has an inherent scale–depth ambiguity. We adaptively select among occlusion matching, ground-depth reading, and floor contact according to the geometric cues available in the shot, registering the human, camera, and scene in a common coordinate system. The same procedure applies to placing a single-image human reconstruction into any virtual scene, e.g., for world models or VR/AR content.

3. Designs tailored to short dramas. Unlike works targeting video editing (EditDuet), general long-video generation (Hollywood Town), films (FilMaster), or advertising (Co-Director), our method is designed specifically for short dramas. We distill beat-level pacing patterns (opening action, conflict function, and closing hook) from short-drama scripts as pacing priors for story generation. The judges review scripts with short-drama criteria such as opening strength, hook continuity, reversal pacing, and payoff resolution. The clip-level review separately optimizes the opening hook of the first clip, the ending suspense of the last clip, and the twist density of the middle clips.

4. Contribution of each design. In Table 2, removing the story generation module or the 3D-grounded first-frame generation clearly degrades the corresponding metrics. Appendix L further isolates the two designs in point 2: the Idea Bank improves the memorability-related dimensions without harming logic, and the geometry-adaptive character registration reduces the scene penetration rate and raises the visible-region IoU from 0.27 to 0.67 compared with a projection-only baseline.

5. Comparison with Co-Director. EditDuet, Hollywood Town, and FilMaster have not released their code. Co-Director is concurrent work; although its framework is domain-agnostic, it is instantiated for video advertising in the paper. Due to limited time, we compare with Co-Director on three cases and report the average scores below. Our method outperforms Co-Director on all eight metrics.

Comparison with Co-Director on the Short-Drama-Bench metrics, averaged over three cases.

MethodOpening HookEnd HookEscalationCoherenceChar. SpatialEnv. LayoutBGMTransition
Co-Director3.112.191.832.892.282.003.072.17
Ours4.114.333.744.563.744.174.373.72
AuthorsOpenReview ↗

Response to Reviewer KgoZ (3/3): W2 & Q1, W3 & Q2

W2 & Q1: Identity Consistency and Effectiveness of the Reviewer Loops

We thank the reviewer for carefully watching the supplementary videos.

1. How identity consistency is checked. The character-presence check in Appendix J is a dedicated check that only verifies whether character presence, entrance, exit, and on-screen status follow the clip script; it is not responsible for identity consistency. Identity consistency is checked by a dedicated first-frame identity review (Table 4 in Appendix J): a VLM compares each character in the first frame of every clip with the front view of the character's three-view reference and assigns a score. If the score is below 8 (out of 10), the first frame is regenerated and re-checked, up to the maximum number of retries.

We have added a complete description of the first-frame identity review to Table 4 (highlighted in blue). Action quality is checked by the generated-video spatial audit in Table 5 of Appendix J, where reaction_plausibility checks whether expressions and actions fit the context (e.g., a character should not keep smiling while being publicly humiliated).

2. The two examples in the supplementary videos.

  • (a) The woman at 3:07–3:42 in the Chinese version. Our consistency constraint targets the facial identity of each character, and the character's styling is also kept consistent within the same scene. Across scenes, the system intentionally adapts the character's clothing and styling to the context (e.g., home wear differs from office attire), which avoids a rigid appearance and better fits the plot. This segment takes place in a different scene from the others, so the character's styling changes accordingly.
  • (b) The elderly man at 0:02–0:06 in the English version. In the context of the plot, the elderly professor notices that someone is arriving, puts on his glasses to see who it is, and takes them off once the visitor comes closer.

3. Quantitative results. In the final evaluation, our method achieves the highest scores among all methods on both identity-related metrics, VBench Subject Consistency and ViStoryBench Cross Char. Consist.

W3 & Q2: Comparison with Commercial Filmmaking Platforms

We thank the reviewer for the suggestion. We have cited and discussed Katalist and Google Flow in the related work (Appendix B, highlighted in blue). Note that neither platform is designed specifically for short dramas: Katalist mainly targets script-to-storyboard visualization and can export storyboards to videos, while Google Flow is an AI filmmaking tool built on Veo, Imagen, and Gemini that generates videos at the level of individual shot clips. Due to the limited rebuttal period, we evaluate both platforms on three cases and report the average scores. Both platforms take the same one-sentence input as ours, which is expanded by Claude Opus 4.6 into the input format required by each platform (a script for Katalist and shot-by-shot prompts for Google Flow); all other settings follow the platforms' defaults.

Our method achieves the highest scores on all eight metrics. Google Flow performs relatively well on Opening Hook, Escalation, and BGM, but is clearly weaker in cross-clip character spatial continuity and environment layout continuity. Katalist achieves reasonable continuity scores but lower narrative-related scores.

Comparison with commercial platforms on the Short-Drama-Bench metrics, averaged over three cases.

MethodOpening HookEnd HookEscalationCoherenceChar. SpatialEnv. LayoutBGMTransition
Katalist2.002.081.421.673.643.582.251.83
Google Flow3.782.503.662.002.142.143.922.00
Ours4.114.333.744.563.744.174.373.72
Reviewer_9vFTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes One Sentence, One Drama, a multi-agent framework designed to automatically transform a single user-provided prompt into a complete short-drama video. The pipeline consists of four main stages:

  • Hierarchical Episode Planning: Constructs storylines using an atomic script library (pattern bank and logic bank) and multi-agent debate with a "creative bank and final-round revival" mechanism.

  • 3D-Grounded First-Frame Generation: Builds scene-level 3D world models via Marble and employs geometry-adaptive 3D character registration to align depth and scale across multiple views.

  • Consistent Video Generation: Reduces scene drift between adjacent video clips using spatio-temporal camera trajectory planning and 3D anchor projection.

  • Multi-Stage Reviewer Loops & Post-Production: Performs automated quality checks across scripts, keyframes, and composed clips, complemented by scene transition planning and adaptive BGM alignment.

The authors further propose the Short-Drama-Bench and evaluate their method on this benchmark, demonstrating some effectiveness.

Critical strengths

S1. Focus on an emerging and practical multimodal application: The work addresses end-to-end short-drama generation from minimal user input, a timely and practical entry point. The distinctive mechanisms of short dramas (e.g., fast pacing, opening hooks, plot escalation) are well differentiated from traditional long-film generation.

S2. Comprehensive pipeline engineering: The system integrates script planning, 3D guidance, character registration, multi-loop validation, and post-production into a fully automated pipeline, demonstrating strong engineering capability.

S3. Construction of a dedicated evaluation benchmark: The introduction of Short-Drama-Bench and its short-drama-specific evaluation criteria (e.g., opening hooks, spatial coherence, plot escalation) provides a valuable evaluation setup for future research in video storytelling.

Critical weaknesses

W1. Severe deficiency in the evaluation metric system: lack of assessment for professional "cinematic language and audiovisual grammar"

The appeal and dramatic tension of short dramas depend heavily on professional cinematography (e.g., shot-size transitions, shot/reverse-shot framing, camera movement logic, emotional amplification through close-ups). However, the proposed Short-Drama-Bench and its evaluation system almost entirely neglect the evaluation of cinematic language, remaining limited to physical spatial layout alignment and basic editing continuity. How the system decides when to use close-ups or push-in shots, and whether the cinematic language conforms to audiovisual narrative conventions, is entirely a blank in the current evaluation.

W2. Over-reliance on VLM-as-a-Judge without consistency validation

The framework extensively uses commercial VLMs for quality scoring in both the multi-stage reviewer loops and automated evaluation. However, the authors provide no alignment study demonstrating the correlation between VLM scores and genuine human taste/aesthetic preferences. Given that the success of short dramas depends critically on human subjective emotional resonance, emotional payoff, and dramatic rhythm, using VLM scores as the core basis for both internal optimization loops and final evaluation is of questionable reliability and objectivity without evidence that VLM scores correlate highly with human viewing preferences.

W3. Heavy engineering integration with a lack of conceptual and algorithmic innovation

Although the pipeline is highly complex, its core algorithmic components largely rely on off-the-shelf existing components and standard paradigms. Its main purported technical contribution leans more toward laborious engineering orchestration rather than the fundamental novelty.

W4. Confounded sources of performance gains (framework advantages vs. underlying foundation model capabilities)

Baseline comparisons mainly contrast the full integrated solution (equipped with GPT-5.4, Claude Opus 4.6, Kling v3 Pro, and other frontier models) against older or simpler baselines. It remains unclear whether the quantitative performance improvements stem from multi-agent orchestration and 3D grounding logic, or merely from the raw generative power of the underlying state-of-the-art commercial APIs.

W5. High computational overhead and expensive trial-and-error cost

As shown in Appendix H, generating a 10-minute short drama requires approximately 74 to 90 minutes, at a cost of about $25 to $27 per minute, primarily due to frequent reviewer retries and video regeneration loops. This extreme resource consumption severely limits real-world deployability, interactive use, and reproducibility in academia.

W6. Fragility of 3D registration and cascading failure risk

The geometry-adaptive character registration mechanism relies heavily on monocular depth estimation and occlusion heuristic matching. In real dynamic short-drama scenes involving complex multi-character occlusions, fast motion, or non-standard camera angles, monocular depth recovery frequently fails. Failures at the 3D first-frame stage can cascade and amplify errors in subsequent video clips, and automated reviewer loops can only address this through costly brute-force regeneration. I would expect to see how the proposed method works on generating videos with more complex interactions, e.g., 2 people fighting.

W7. "Personalized" is a title-level claim with no supporting experiment

The title and abstract claim personalized short-form drama generation, but personalization (user-uploaded portrait cast as protagonist) is described only in Appendix Q, with no experiment, no identity-fidelity metric, and no user study.

W8. Overly long appendix obscures rather than clarifies

Finally, I understand the authors' intention to demonstrate the workload of their work and preempt possible criticisms through a sufficiently extensive appendix. However, in my personal view, an excessively long appendix tends to hide the effective information rather than highlight it, and may thus produce a counterproductive effect. This is just my personal opinion and I won't reject paper because of this. Just some personal advice.

Questions for authors

I hope authors can clarify my proposed weaknesses above.

8 comments
AuthorsOpenReview ↗

Response to Reviewer 9vFT (1/6): W1

W1: Evaluation of Cinematic Language and Audiovisual Grammar

We thank the reviewer for highlighting the importance of cinematic language to the appeal of short dramas. Accordingly, we have added a corresponding evaluation.

1. Existing metrics already partially capture cinematic language and audiovisual grammar. Short-Drama-Bench is designed around audience retention, the central goal of short dramas. An interview study with 28 short-drama practitioners and screenwriters (Cao et al., CHI 2026) reports that the typical structure of short dramas consists of a strong opening, rapid plot reversals, and suspense left at the end; accordingly, the benchmark includes three narrative metrics: Opening Hook, End Hook, and Escalation Effect. Beyond these, several metrics directly or indirectly involve cinematic language and audiovisual grammar:

  • (a) Shot continuity. Film editing requires characters to keep their screen positions and facing directions across cuts (the 180-degree rule; Bordwell & Thompson, Film Art). Character Spatial Continuity checks exactly this between adjacent shots, and Environment Layout Continuity checks whether the background layout and camera direction remain consistent.
  • (b) Audiovisual alignment. BGM Emotion Alignment evaluates whether the background music matches the mood, tension, pacing, and dramatic intent of each scene, i.e., whether sound and image work together to create an immersive experience.
  • (c) Visual presentation of dramatic beats. Opening Hook and End Hook are scored on the final video rather than on the script, so how the opening conflict and the ending suspense are presented on screen (e.g., the choice of shot size or close-ups of character reactions) directly affects their scores.

2. Shot selection is driven by narrative content. During script generation, the LLM selects a shot size (six levels) and a camera movement (eight types) for each clip based on the narrative structure of each scene in the story core, including the opening attractor, scene goal, escalation beats, ending hook, and ending last shot, as well as the plot and dialogue of the clip. For example, one escalation beat in an office scene is that an administrator asks to inspect the protagonist's hard drive, and he flips it over to reveal the engraving "Private Property" on its bottom. For this reveal moment, the generated script assigns a medium close-up with a slow push-in onto the engraving, focusing the viewer's attention on the key evidence that turns the situation around.

3. A new cinematic-language metric. Following the reviewer's suggestion, we add a new metric, Camera Narrative, to Short-Drama-Bench. Each segment is scored on a 1–5 scale by the same three judge models used in Short-Drama-Bench. The metric jointly evaluates three aspects:

  • (a) whether it is always clear who is where and who is doing what;
  • (b) whether each camera move, reframing, or cut follows the action and the dialogue; and
  • (c) whether cuts and camera moves keep the viewer oriented, without disorienting jumps in position, direction, or framing.

The full scoring prompt is provided in Appendix R. Results are reported below.

MovieAgentScriptAgentStoryMemHoloCineLongLiveToonFlowXiao Yun QueOurs
Camera Narrative ↑1.442.122.442.331.673.674.424.44
AuthorsOpenReview ↗

Response to Reviewer 9vFT (2/6): W2

W2: Alignment between VLM-as-a-Judge and Human Preference

We thank the reviewer for this comment. We have systematically validated this in Appendix K of the original submission, and summarize the main results below.

1. VLM scores are highly consistent with human ratings (Appendix K.1). For each of the 50 prompts and each of the eight metrics, we separately average the ratings from the 20 human raters and the scores from the three LLM judges for each method, and then compute the correlation across the eight methods within each prompt. Across the eight metrics, the median prompt-level Pearson correlations range from 0.878 to 0.960, and the median Spearman correlations range from 0.848 to 0.949 (Table 8). In particular, the median Pearson correlations for Opening Hook, Escalation Effect, Narrative Coherence, and End Hook are 0.960, 0.950, 0.955, and 0.909, respectively.

2. The conclusions do not depend on a particular judge (Appendix K.2). Although the three judge models assign different absolute scores, all of them rank our method first, followed by Xiao Yun Que and ToonFlow (Table 9). The median pairwise item-level Spearman correlation among the three judges is 0.823, and Krippendorff's α is 0.804.

3. The main conclusions are supported by independent human evaluation (Appendices K.3 and K.4). The human study involves 20 raters on 50 prompts, yielding 1,000 ratings per method. The ICC(2,20) values range from 0.79 to 0.86, indicating that the aggregated human ratings are reliable. Our method achieves the highest human-evaluation mean among all methods (3.52). Compared with Xiao Yun Que, the strongest baseline, our method improves the average human score by 0.349 (95% CI: [0.191, 0.515]), consistent with the improvement of 0.375 in the LLM-based evaluation (95% CI: [0.282, 0.470]). Both confidence intervals lie entirely above zero. Therefore, our main conclusions do not rest solely on VLM scores.

AuthorsOpenReview ↗

Response to Reviewer 9vFT (3/6): W3

W3: Heavy Engineering Integration

We would like to clarify the positioning and academic contribution of this work.

(1) Mechanism- and algorithm-level novelty. The challenge is not invoking multiple models, but coordinating heterogeneous agents across a long pipeline where errors propagate between stages. We address it through three coupled mechanisms:

  1. Retrieval-augmented multi-agent debate for script generation. Agent roles, structured intermediate states, and debate-based revision enforce short-drama pacing with support from the complementary Pattern and Logic Banks. One concrete design is the Idea Bank with final-round revival (Appendix C.4). Iterative revision has a systematic bias: each round removes risk and deletes memorable elements along with it, so scripts become safer but flatter. We record every deleted or weakened element with its reason, and a final revival step restores those that can return without harming logic or visual executability.
  2. 3D-grounded cross-clip state propagation. Camera trajectories and character spatial states are registered in a shared 3D scene, so the next viewpoint is geometrically constrained rather than left to the model's implicit memory. Its core algorithm is geometry-adaptive character registration (Section 2.3 and Appendix E). The person mask fixes the body's position and orientation but not its depth: a person twice as large and twice as far away produces the same mask. The algorithm therefore picks a strategy from the shot's geometric cues: occlusion matching (with an occluder in front, the visible silhouette changes with depth, and the body is localized where it matches the mask), ground-depth reading (the scene depth beneath the visible feet fixes the body's depth; standing is not required), or floor contact (for standing characters, the body's lowest point is placed on the scene floor).
  3. Hierarchical multi-agent review and targeted repair. In multi-stage generation, a problem is often detected only in the video, while its cause may lie in an earlier first frame or in a specific sentence of the prompt. Our hierarchical multi-agent review forms a Detection–Localization–Repair loop across the text, image, and video levels (Appendix T, highlighted in blue).
    • (a) Top-down detection. From prompts to first frames to videos, the generated content becomes increasingly concrete and costly to produce. We place a review gate at each level and verify each level against the content already fixed at the level above: the prompt is checked for whether the states of characters and props connect with the preceding and following clips, the first frame is checked against the character references and the video prompt, and the video is checked against the script. Each error is thus intercepted at the earliest and cheapest level where it becomes observable.
    • (b) Bottom-up localization. The video reviewer maps each problem, with timestamped evidence, to the sentence in the video prompt that induced it; if the problem already exists in the first frame, the localization moves to the first frame, where the system further determines whether the image itself or its underlying text is at fault.
    • (c) Targeted local repair. Only the localized part is repaired: the faulty sentence is replaced, or the first frame is regenerated with the defect appended to its prompt, while the characters, dialogue, speaking order, and plot events remain unchanged.
    • (d) Empirical support. Removing the multi-stage review degrades all eight metrics (Table 2). On 25 clips that failed their first video review, targeted repair fixes more clips than rewriting the entire prompt while better preserving the dialogue and plot (Appendix T).
Repair methodFixed in one roundDialogue preservedPlot events preserved
Full rewriting15 / 25 (60%)40 / 54 (74%)85.3%
Targeted repair (Ours)20 / 25 (80%)54 / 54 (100%)91.7%

(2) Task and benchmark contribution. We formulate one-sentence-to-drama generation as a structured task that jointly models narrative pacing, spatial consistency, and production-level coherence, with Short-Drama-Bench as its first systematic benchmark. Agentic pipelines with companion benchmarks are recognized contributions, e.g., EvoPresent (ICLR 2026), Paper2Poster (NeurIPS 2025), AutoPresent (CVPR 2025), and PPTAgent (EMNLP 2025); our work additionally makes the algorithmic contributions above.

(3) Transferability to related directions. The spatial state propagation also applies to consistency problems in long video generation and world models (e.g., WorldPlay, ICML 2026), and the review loop applies to other multi-stage generation pipelines.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.