The Script is All You Need: An Agentic Framework for Long-Horizon Dialogue-to-Cinematic Video Generation

ArXiv: 2601.17737

🎯 Pitch

Given only raw dialogue, this paper’s ScripterAgent automatically generates production-ready cinematic scripts that boost video generation faithfulness by 7+ points — yet it also exposes a critical trade-off where top models like Sora2-Pro optimize for spectacle at the cost of faithfully following directions. The framework’s multi-agent system, tested across multiple state-of-the-art generators, proves that injecting structured scripting can bridge the gap between vague narrative and coherent filmmaking, even as it reveals that no current model excels at both adherence and aesthetics.

---

1. Executive Summary

This paper introduces a novel, end-to-end agentic framework for the challenging task of generating long-form cinematic video from sparse, high-level dialogue. The system is built around three specialized agents — ScripterAgent, trained on a new ScriptBench benchmark using a two-stage SFT+RL paradigm to translate dialogue into fine-grained cinematic scripts; DirectorAgent, which employs a cross-scene continuous generation strategy with frame-anchoring to maintain visual consistency across video segments; and CriticAgent, an AI-powered evaluation system — and is evaluated on state-of-the-art video generation models including Sora2-Pro and Veo3.1. Conditioning these video models on scripts generated by the full ScripterAgent universally improves performance, boosting Script Faithfulness by up to +0.4 points and the novel Visual-Script Alignment (VSA) metric by over 7 points, while human experts rate the script's Visual Imagery at 4.3 compared to 3.8 from the strongest baseline. The paper further reveals a fundamental trade-off in current models between visual spectacle and strict script adherence, establishing that test-time orchestration via structured scripting can substantially bridge the semantic gap between narrative concept and cinematic execution, even as it highlights that models optimize along different axes — with Sora2-Pro leading in Visual Appeal (4.8) and HYVideo1.5 in Script Faithfulness (4.6).

2. Context and Motivation

The Core Problem: The Semantic Gap Between Dialogue and Cinematic Execution

This paper tackles a problem that is deceptively simple to state but extraordinarily complex to solve: given only the dialogue from a scene — the words characters say to each other — can we automatically generate a coherent, visually appealing cinematic video? In conventional filmmaking, this gap is bridged by an entire creative hierarchy: screenwriters expand dialogue into full scripts with scene directions, directors interpret those scripts into shot plans and blocking, cinematographers design camera movements, and editors stitch everything into a continuous narrative. The paper argues that current AI video generation systems lack this intermediate planning layer, and the absence is not a minor inconvenience but a fundamental structural limitation.

The authors characterize this as a "semantic gap" between high-level narrative concepts (coarse-grained dialogue) and low-level video synthesis (pixel-level generation). A line of dialogue like "I can't believe you did that" carries dramatic intent, emotional subtext, and implied physical actions that are entirely absent from the words themselves. A human director immediately infers that this line might require a close-up reaction shot, a pause for dramatic tension, specific facial expressions, and blocking that positions characters relative to each other to emphasize confrontation. Current text-to-video models, however, receive no such guidance — they are asked to conjure all of this from a raw transcript, operating in what the paper calls a paradigm of "passive description" (video-to-text) rather than "active generation" (text-to-video with anticipatory planning).

The paper frames this as a paradigm inversion: prevailing video-language research focuses on describing what already exists (captioning, question answering about video content), but the dialogue-to-cinematic-video task requires the model to anticipate what should be shown, not merely react to what has already been filmed. This anticipatory requirement is what makes the problem fundamentally creative rather than purely descriptive.

Why This Matters: Three Levels of Significance

Practical significance for automated filmmaking. The paper is unmistakably motivated by the vision of automated storytelling and film production. It opens with a Hitchcock quote — "To make a great film you need three things: the script, the script and the script" — not as decoration but as a thesis statement. If AI systems are ever to approach anything resembling automated filmmaking, they cannot simply generate visually impressive clips from prompts. They need to understand dramatic structure, pacing, shot composition, character blocking, and the relationship between dialogue and visual storytelling. The paper positions script generation as the missing piece: current models produce "stunning visual content" but "struggle to generate long-form, coherent narratives from high-level concepts." The practical goal is to enable systems that can take a conversation between characters and produce a professionally watchable short film, with appropriate camera work, temporal structure, and narrative logic.

Research significance for video generation evaluation. The paper identifies a gap not just in generation capability but in evaluation methodology. Standard video quality metrics (CLIP Score, VBench) measure whether content appears and whether it looks good, but they do not measure when content appears relative to the intended narrative timeline. A video might score highly on visual quality while completely misaligning its temporal structure — showing a reaction shot before the triggering line of dialogue, or lingering on an establishing shot long past its narrative purpose. The paper's proposal of Visual-Script Alignment (VSA) as a metric that evaluates temporal-semantic fidelity directly addresses this blind spot. This is significant because it suggests that the field's evaluation tools are themselves inadequate for the task of narrative video assessment — a methodological gap that parallels the technical gap the paper addresses.

Theoretical significance for understanding model capabilities. The paper reveals what it calls a "fundamental trade-off in SOTA models between visual spectacle and strict script adherence" — a finding that has implications beyond this specific task. Sora2-Pro achieves top scores in Visual Appeal (4.8) and Physical Law Adherence (4.5) but lags in Script Faithfulness (4.2), while HYVideo1.5 prioritizes narrative integrity (4.6 Script Faithfulness, 4.4 Character Consistency) at some cost to raw visual spectacle. This is not merely a detail about two specific models. It suggests that current video generation architectures optimize along different axes that are at least partially in tension: perceptual realism versus semantic control. Understanding this trade-off is crucial for practitioners selecting models for specific applications and for researchers designing next-generation architectures that might reconcile these competing objectives.

Where Prior Approaches Fall Short

The paper situates itself against a landscape of related but insufficient prior work, identifying specific limitations in each area.

Direct text-to-video generation without planning. The baseline condition in the paper's experiments is feeding raw dialogue directly into SOTA video generation models — Sora2-Pro, Veo3.1, Wan2.6, and others. The results in Table 2's "Raw Dialogue (w/o ScripterAgent)" section tell a clear story: even the most advanced models produce videos with weak Script Faithfulness (average 3.8 on a 0–5 scale) and poor Character Consistency (average 3.5) when operating from dialogue alone. The average human rating across all models is just 3.7 out of 5. These are not broken videos — they score respectably on Visual Appeal (average 3.8) — but they fail at the narrative and structural level. The paper's diagnosis is that these models lack the intermediate representation — the cinematic script — that translates abstract narrative intent into concrete visual execution. This is not a weakness of any particular architecture but a consequence of the text-to-video paradigm itself when the input text is too sparse and too high-level.

LLM-based film production tools with manual planning requirements. Section 6 reviews prior work on using LLMs for film production tasks, including VideoDirectorGPT (Lin et al., 2023), Anim-Director (Li et al., 2024), and MovieAgent (Wu et al., 2025b). These systems demonstrate that LLMs can assist with scene generation, character planning, and cinematographic decisions. However, the paper identifies a critical limitation: "these models rely on manual input for narrative and cinematographic planning." That is, a human still needs to specify what shots to use, how to structure scenes, and what the visual plan should be. The systems are assistants to a human director rather than autonomous planners. The paper's ambition is to automate this planning step entirely — to take dialogue and produce the complete cinematic blueprint without human intervention in the planning loop. The contrast with MovieAgent is particularly instructive: in Figure 4, MovieAgent's output is described as "a simple plot summary, such as 'Camera slowly pans…'" while ScripterAgent produces a full technical specification including camera settings, timestamps, character blocking, and atmospheric descriptions.

Story visualization without cinematographic reasoning. Prior work on story visualization — including StoryGAN (Li et al., 2019), StoryDiffusion (Zhou et al., 2024), and Magic-Me (Ma et al., 2024b) — has focused on mapping scripts to visual sequences and improving temporal consistency. But the paper argues these systems "still lack automated high-level planning for cinematography, scene structure, and character interactions, thus requiring manual guidance." The distinction is between generating images that match a script's content (story visualization) and generating the cinematographic plan itself (what ScripterAgent does). Traditional story visualization assumes someone has already made the creative decisions about shot types, camera movements, pacing, and staging. The paper's framework generates those decisions automatically from dialogue, then uses the resulting plan to guide video generation. This is a higher level of automation that requires the model to possess cinematic domain knowledge, not just visual generation capability.

Fixed-duration constraints in video generation models. All current SOTA video models have limited generation windows — typically 8–12 seconds per clip, while a narrative scene might run 1–3 minutes. The standard approach of naively segmenting a long script into independent clips and generating each separately produces "severe artifacts, such as identity drift, inconsistent styling, and a loss of narrative continuity." The paper identifies this not as a flaw in any particular model but as a systemic limitation of the current generation paradigm — the models' fixed-duration constraint is fundamentally at odds with the variable-duration nature of narrative storytelling. The DirectorAgent's cross-scene continuous generation strategy is designed specifically to overcome this limitation, treating it as an orchestration problem rather than a generation problem per se.

Evaluation methodologies that miss temporal structure. The paper argues that existing automated metrics for video generation — CLIP Score for global semantic alignment and VBench metrics for subject consistency, background consistency, motion smoothness, and dynamic degree — measure "whether content appears but not when it appears." A video could score perfectly on subject consistency (the character looks the same throughout) while completely misaligning the timing of events relative to the script. In a narrative context, timing is everything: a reaction shot that comes before the triggering event, or a dramatic reveal that occurs during dialogue rather than after a pause, fundamentally breaks the storytelling even if every individual frame is high quality. The paper's Visual-Script Alignment (VSA) metric is explicitly designed to fill this gap by measuring whether visual events occur within their designated time intervals as specified by the script, making it a temporal-semantic alignment measure rather than a content-presence measure.

How This Paper Positions Itself

The paper positions itself as introducing a new task (dialogue-to-cinematic-video generation), a new paradigm (script-centric agentic orchestration), and a new benchmark (ScriptBench) — not as an incremental improvement on existing approaches. The introduction is explicit about this framing: "we tackle a new, challenging task: given only coarse-grained dialogue, the model must anticipate and generate an executable filmmaking plan."

The three fundamental challenges identified in Section 2 serve as the paper's statement of what makes this task novel and difficult:

1. Fine-grained contextual understanding — resolving ambiguities inherent in sparse dialogue requires inferring unspoken context, subtext, and implied action that the dialogue alone does not specify.
2. Domain knowledge of filmmaking — producing technically valid camera specifications and staging directions requires knowledge of shot types, camera movements, blocking conventions, and the grammar of visual storytelling that is entirely absent from general-purpose language modeling pretraining.
3. Creative reasoning abilities — bridging what is said with what must be shown requires aesthetic judgment, not just logical inference. The paper explicitly acknowledges that this goes beyond correctness into artistic quality, which is why the RL stage with its hybrid reward function is necessary.

The paper's agentic framework — ScripterAgent, DirectorAgent, CriticAgent — is presented as a unified solution to all three challenges simultaneously. ScripterAgent addresses challenges 1 and 3 (contextual understanding and creative reasoning) through its two-stage training on ScriptBench. DirectorAgent addresses challenge 2 (domain knowledge) by embodying filmmaking rules in its shot-aware segmentation and frame-anchoring mechanisms. CriticAgent provides the evaluation infrastructure that makes the entire pipeline testable and improvable.

Positioning relative to specific prior work is sharpest in the contrast with MovieAgent — the closest existing system. The paper claims MovieAgent "relies on manual input for narrative and cinematographic planning," while ScripterAgent automates this entirely. The ScriptBench benchmark itself is positioned as filling a gap: no existing dataset provides the dialogue-to-cinematic-script mapping with rich multimodal context (dialogue, audio, character positions) and expert-validated cinematic plausibility. The expert-guided three-stage annotation pipeline — context reconstruction, shot-level semantic planning, and multi-round adaptive error correction — is presented as a novel contribution in itself, designed to produce scripts at scale while maintaining the quality constraints that professional filmmaking requires.

Crucially, the paper does not claim to have solved video generation. It explicitly positions SOTA video models as tools to be orchestrated, not problems to be replaced. The claim is about the orchestration layer: that providing these models with properly structured cinematic scripts, segmented intelligently, and linked via frame-anchoring, can substantially improve their narrative output without modifying the models themselves. This is a systems-level contribution rather than a model architecture contribution — it is about how to use existing generators more effectively, not about building better generators.

3. Technical Approach

3.1 Reader Orientation

This paper builds a multi-agent orchestration system that takes raw dialogue as input and produces a long-form cinematic video as output, with the core insight being that a structured, machine-readable cinematic script — generated automatically by a purpose-trained model — serves as the essential intermediate representation that bridges the semantic gap between sparse conversation and rich visual storytelling. The system solves the problem of generating coherent, multi-shot narrative videos from dialogue alone by decomposing the task into three specialized stages: script generation (translating dialogue into a detailed shot-by-shot plan), video execution (orchestrating existing video generation models to produce continuous footage following that plan), and evaluation (measuring both technical quality and narrative alignment).

3.2 Big-Picture Architecture (Diagram in Words)

The system consists of five major components arranged in a sequential pipeline with feedback loops for evaluation:

1. ScriptBench Dataset — A curated collection of 1,750 dialogue-to-cinematic-script pairs, annotated via an expert-guided three-stage pipeline. Each instance contains multimodal input (dialogue text, audio, character positions) mapped to a structured JSON script specifying shot types, camera movements, character blocking, descriptions, and timestamps. This dataset serves as both the training ground for ScripterAgent and the evaluation benchmark for the entire pipeline.

2. ScripterAgent — A fine-tuned Qwen-Omni-7B model trained in two stages (SFT followed by GRPO reinforcement learning) to transform multi-turn dialogue into a structured cinematic script. The model receives raw dialogue text as input and outputs a JSON document containing shot-by-shot instructions including camera specifications, character positioning, atmospheric descriptions, and timestamped synchronization markers. This is the central creative intelligence of the system.

3. DirectorAgent — An orchestration module that takes the generated script and partitions it into generation tasks compatible with the temporal constraints of current video models, then executes those tasks using a cross-scene continuous generation strategy. Its primary mechanism is frame-anchoring: the last frame of each generated video segment is extracted and fed as a conditioning image for the next segment, creating a visual relay that maintains consistency across cuts. The output is a single long-form video stitched from individually generated clips.

4. Video Generation Backend — A collection of state-of-the-art text-to-video models (Sora2-Pro, Veo3.1, Wan2.6, HYVideo1.5, Seedance1.5-Pro, Kling2.6, Vidu2) that serve as video synthesis engines. These are treated as black-box tools invoked by DirectorAgent with structured prompts derived from the script. The framework does not modify these models; it improves their output by providing better input.

5. CriticAgent — An AI-powered evaluation system (backed by Gemini-2.5-Pro) that assesses both the generated script and the final video across multiple dimensions on 0–5 scales, complemented by human expert panels, automated metrics (CLIP Score, VBench), and the novel Visual-Script Alignment (VSA) metric. The CriticAgent provides the evaluation signals used throughout the paper's experiments but is not part of the generation loop itself — it evaluates, it does not guide generation.

Information flow: Raw dialogue enters the system → ScripterAgent generates a structured JSON script → DirectorAgent segments the script into model-compatible generation windows and applies frame-anchoring to maintain cross-scene continuity → SOTA video models generate individual clips from script-derived prompts → Clips are concatenated into the final long-form video → CriticAgent and human evaluators assess the output across multiple dimensions. The training flow for ScripterAgent is separate: ScriptBench instances are used for SFT, then SFT outputs annotated by human experts train a preference model used in the GRPO reward function.

3.3 Roadmap for the Deep Dive

• First, ScriptBench construction (Section 2.1) — the data pipeline and annotation methodology, because the quality of ScripterAgent depends entirely on the quality of this training data and its expert-guided curation process.
• Second, ScripterAgent's SFT stage (Section 2.2.1) — the base model, training objective, and hyperparameters, since this establishes the structural competence before any creative alignment.
• Third, ScripterAgent's RL stage (Section 2.2.2) — the GRPO optimization, hybrid reward function, and preference model training, as this is where the model acquires its cinematic aesthetics and where the paper's most novel training contribution lies.
• Fourth, DirectorAgent's cross-scene generation strategy (Section 3) — the shot-aware segmentation principles and the frame-anchoring mechanism, which address the fixed-duration constraint of current video models.
• Fifth, the evaluation framework (Section 4) — the multi-axis assessment methodology including both CriticAgent metrics and the novel VSA metric, as evaluation design shapes interpretation of all experimental results.

3.4 Detailed, Sentence-Based Technical Breakdown

This is primarily a systems engineering paper whose core idea is that a structured intermediate representation — a machine-readable cinematic script — generated by a purpose-trained model and executed by an orchestration layer can substantially improve the narrative coherence of video generated from sparse dialogue, without modifying the underlying video generation models themselves.

---

ScriptBench Construction: The Expert-Guided Annotation Pipeline

ScriptBench is the foundation on which the entire system is built. The dataset consists of 1,750 finalized script instances, each paired with multimodal input: dialogue text (the spoken lines), dialogue audio (the vocal delivery), and character positions (spatial blocking information). These raw instances were curated from "high-fidelity cinematic cutscenes" — the paper specifically cites their "rich dialogue, professional cinematography, and high visual consistency" as selection criteria that approximate real film production conditions. The dataset is partitioned into 1,700 training instances and 50 test instances, intentionally designed to "challenge the model to infer complete cinematic elements from conversational content alone, emulating the real-world process where directors visualize a story from a dialogue-driven script."

The annotation pipeline operates in three sequential stages, and the paper describes it as "scalable yet high-fidelity" because it uses a SOTA LLM (Gemini-2.5-pro) for generation but enforces quality through "expert-defined templates, domain constraints, and validation rules." This is a crucial design choice: the LLM provides creativity and scale, while the constraint system provides correctness and cinematic plausibility. The three stages are:

Stage 1: Context Reconstruction and Dialogue Fusion. The Gemini-2.5-pro model jointly analyzes the textual script and the dialogue audio to reconstruct a comprehensive understanding of the scene. The paper specifies that it infers "character relationships, scene settings, plot developments, emotional tendencies, and speaking intent." The goal is to "fuse the disparate signals into a coherent narrative context that makes implicit causal relations explicit." Operationally, this means the model takes raw inputs — who is speaking, what they are saying, how they sound — and produces a unified scene representation that explains not just what is happening but why it is happening in narrative terms. This reconstructed context serves as input to the next stage.

Stage 2: Shot-Level Semantic Planning. Using the reconstructed context, the model plans individual shots under four explicit constraints designed to ensure visual and narrative continuity:

• Shot integrity — "enforces self-contained units, introducing cuts only upon clear camera or scene changes." This prevents arbitrary fragmentation; a shot is a continuous camera take, and cuts should only occur when the cinematography demands them.
• Duration adaptation — "caps shots at 10 seconds to align with generation limits." This is a practical constraint: even though real cinematic shots routinely exceed 10 seconds, the downstream video generation models have fixed generation windows, and the script must produce shots that are technically executable. The 10-second cap is explicitly tied to model limitations, not artistic preference.
• Semantic coherence — "aligns boundaries with narrative transitions (e.g., emotional shifts)." This ensures that cuts happen at natural story points — the end of a character's speech, a change in emotional tone, a revelation — rather than at arbitrary temporal boundaries.
• Technical feasibility — "prevents segmentation during complex camera motions." Cutting during a tracking shot or a complex dolly movement would create jarring discontinuities in the generated video, so the pipeline avoids placing shot boundaries at these points.

These four principles jointly ensure that "the shot units are narratively meaningful and technically viable for downstream generation" — they produce a plan that is both artistically sensible and implementable with current tools.

Stage 3: Multi-Round Adaptive Error Correction. This is the most novel aspect of the annotation pipeline and the mechanism that elevates ScriptBench from an LLM-generated dataset to an expert-validated benchmark. The system runs an iterative correction loop with four automated verification modules, each checking a specific dimension of script quality:

• Dialogue Completeness — "ensures that all spoken content is either explicitly transcribed or marked as [No Dialogue]." Every line from the original dialogue must be accounted for in the script; if a shot has no dialogue, it must be explicitly labeled as such. This prevents dropped lines and ensures script-dialogue correspondence.
• Character Appearance Consistency — "enforces strict adherence to predefined character descriptions." Characters must appear as described; their names, roles, and canonical appearances cannot drift across shots.
• Scene Coherence — "tracks environmental elements and validates narratively justified transitions." The setting should remain consistent unless a scene change is explicitly motivated by the story; props, lighting, and spatial layout should not shift arbitrarily.
• Positional and Physical Rationality — "verifies spatial relations against plausible blocking and camera geometry." Characters must occupy physically possible positions relative to each other and the camera; a character cannot simultaneously be stage left and stage right, and camera positions must make sense given the described spatial layout.

An automated detector scans the generated scripts, identifies violations of any of these constraints, and feeds corrective signals back to the generator. This loop repeats "until all constraints are satisfied." The paper reports that "the automated pass rate reached 94%," meaning that 94% of scripts passed all four automated checks after sufficient correction rounds.

The human audit and its findings. To validate practical reliability beyond automated checks, "professional script consultants conducted a random audit on 60% of the generated instances." This audit "revealed subtle semantic errors such as character teleportation, dialogue–action conflicts, and inconsistent prop states" — errors that the automated verifiers missed but that a human script expert would notice. These findings were fed back into the pipeline by refining prompt constraints and verification logic, creating a "controlled refinement process that constitutes a key novelty of our pipeline." The result is a dataset of "scripts that are structured, internally consistent, and grounded in long-horizon narrative and physical continuity."

Dataset characteristics. Each of the 1,750 instances is in "one-to-one correspondence with a raw multimodal input." The average duration of each video clip is approximately 15.4 seconds, which the paper notes provides "sufficient temporal scope for multi-shot sequences while remaining tractable for current generative models." The 1,700/50 train/test split intentionally creates a setting where models must generalize from a limited number of seen dialogue styles to unseen conversational scenarios, testing whether they have truly learned the underlying cinematographic reasoning rather than memorizing training patterns.

---

ScripterAgent Stage One: Supervised Fine-Tuning (SFT)

The SFT stage is designed to teach the base model "the fundamental syntax and structure of cinematic scripts." The paper frames this as a prerequisite: without SFT, the model would not even produce structurally valid output; SFT provides the "solid foundation" on which creative refinement can later be built.

Model selection. The base model is Qwen-Omni-7B, chosen for its "strong capabilities in long-context processing and instruction following." The paper does not elaborate on what specific properties of Qwen-Omni-7B make it suitable, but the choice is motivated by the task requirements: the model must process potentially long multi-turn dialogues (multiple characters speaking over an extended scene) and must follow the precise formatting and content instructions necessary to produce a valid JSON script. The 7B parameter scale is notable — this is not a frontier-scale model, suggesting that the task is achievable with moderate-scale models when provided with high-quality training data.

Task formulation. The task is cast as sequence-to-sequence learning: the input $x$ is a multi-turn dialogue, and the output $y$ is the target script in structured JSON format. The paper does not specify the exact JSON schema in the main text, but from the qualitative examples (Figure 4), the script includes fields for camera settings, timestamps, character descriptions, atmospheric descriptions, and character positioning. The training objective is the standard conditional log-likelihood:

$\mathcal{L}_{\text{SFT}} = -\mathbb{E}_{(x,y) \sim \mathcal{D}}\left[\sum_{t=1}^{|y|} \log \pi_{\theta}(y_t \mid y_{<t}, x)\right]$

where $\mathcal{D}$ is the ScriptBench training distribution (1,700 instances), $\pi_{\theta}$ is the fine-tuned policy parameterized by $\theta$, $y_t$ is the $t$-th token of the target script, $y_{<t}$ denotes all preceding tokens, and $x$ is the input dialogue.

What it computes: For each training instance, the model processes the dialogue and generates the target script token by token. At each step, it computes $\log \pi_{\theta}(y_t \mid y_{<t}, x)$ — the log-probability assigned to the correct next token $y_t$ given the dialogue and all previous script tokens. The expectation over the dataset averages these per-token log-probabilities, and the negative sign converts the maximization of log-likelihood into a minimization of cross-entropy loss. The result is a single scalar loss value per batch that measures how well the model predicts the ground-truth script tokens.

Why this form: Teacher-forced maximum likelihood is the standard objective for sequence-to-sequence fine-tuning because it provides dense per-token supervision — the model gets a training signal at every position, not just at the end of the sequence. The autogressive formulation $\pi_{\theta}(y_t \mid y_{<t}, x)$ matches how the model will be used at inference time (generating the script left to right), so there is no train-test mismatch. An alternative — training only on the final output after beam search or using a sequence-level reward — would provide sparser signal and require more complex optimization. Maximum likelihood is simple, stable, and sufficient for learning structural patterns.

Training configuration. The paper specifies the following hyperparameters verbatim:

• Optimizer: AdamW
• Learning rate: $\eta = 1 \times 10^{-5}$
• Batch size: 4
• Maximum sequence length: 8,192 tokens
• Number of epochs: 20

The learning rate of $10^{-5}$ is conservative for a 7B model, suggesting that the SFT stage prioritizes stability over speed — the goal is to learn the script format without catastrophic forgetting of the base model's general language capabilities. The batch size of 4 is small, likely constrained by GPU memory given the 8,192-token sequence length and the 7B parameter count. The 20-epoch training duration implies that the model sees each of the 1,700 training instances multiple times, which is appropriate for a relatively small curated dataset where memorization of structural patterns is the goal rather than generalization from diverse examples.

What SFT achieves. The paper states that this stage "equips the model to generate scripts that are structurally correct and content-complete." Operationally, this means the model can reliably produce valid JSON, include all required fields, account for all dialogue lines, and produce shots that meet the basic structural constraints (correct format, shot integrity, etc.). However, SFT alone does not produce cinematically expressive scripts — as the experimental results confirm, the SFT-only variant scores 3.6 in Dramatic Tension and 3.8 in Visual Imagery, substantially below the full model's 4.1 and 4.3 respectively (Table 1). The paper describes SFT's outputs as "structurally correct" but lacking in "subjective artistry."

---

ScripterAgent Stage Two: Reinforcement Learning for Cinematic Alignment

The RL stage addresses the limitation identified at the end of SFT: structural correctness is necessary but insufficient for professional-quality cinematic scripts. The paper explicitly frames this gap: "effective scriptwriting transcends logical correctness, involving aesthetic judgments about shot composition, pacing, and emotional impact." The SFT model has learned what a script looks like; the RL stage teaches it what a good script looks like.

Algorithm choice: Group Relative Policy Optimization (GRPO). The paper employs GRPO, an advanced preference alignment method, rather than more common RL approaches like PPO with a reward model or direct DPO. The justification is task-specific: GRPO's "group-based relative scoring is well-suited for creative tasks that have a subjective, one-to-many nature of valid outputs." In cinematic scripting, there is no single correct script for a given dialogue — there are many valid interpretations, and quality is fundamentally comparative (this script is better than that script) rather than absolute (this script scores 0.85 on some objective metric). GRPO's mechanism of comparing multiple candidate outputs from the same prompt and computing advantages within that group naturally captures this comparative aspect.

Hybrid Reward Function

The central technical novelty of the RL stage is a hybrid reward function, $R_{\text{total}}$, that combines an objective correctness signal with a learned subjective quality signal. This addresses the fundamental challenge that cinematic quality has both verifiable technical components (did the model include all dialogue? are the camera specifications valid?) and subjective artistic components (is the pacing effective? is the shot composition visually interesting?). Neither signal alone would be sufficient:

$R_{\text{total}}(y) = \alpha \cdot R_{\text{structure}}(y) + (1 - \alpha) \cdot R_{\text{human}}(y)$

where $R_{\text{total}}(y)$ is the total reward for a generated script $y$, $R_{\text{structure}}(y)$ is the rule-based structural reward (objective correctness), $R_{\text{human}}(y)$ is the learned human preference reward (subjective quality), and $\alpha = 0.4$ is the mixing coefficient.

What it computes: For a generated script $y$, the total reward is a weighted sum of two components. The structural reward $R_{\text{structure}}$ awards points for technical correctness; the human preference reward $R_{\text{human}}$ awards points for artistic quality. The weight $\alpha = 0.4$ means that structural correctness receives 40% of the total weight, while human preference receives 60%. The output is a single scalar in $[0,5]$ (before normalization) that the GRPO optimizer uses to compute advantages and update the policy.

Why this form: The weighted-sum form with $\alpha = 0.4$ explicitly prioritizes artistic quality over technical correctness. This is a deliberate design choice reflecting the paper's observation that SFT already handles structural correctness reasonably well (the SFT-only model scores 3.9 in Format Compliance) but lacks creative expressiveness. By giving the human preference signal 60% of the weight, the RL stage is incentivized to improve artistic dimensions even at some cost to structural perfection. The $\alpha$ value of 0.4 was presumably tuned empirically, though the paper does not report ablation on this coefficient. A naive 50/50 split would have been a reasonable default, but the 40/60 split reflects the judgment that creative quality improvements are more valuable than marginal correctness improvements at this stage.

The structural reward $R_{\text{structure}}$ mirrors the verification modules from the ScriptBench annotation pipeline. It aggregates normalized scores from four automated checks:

• Format Compliance — correct JSON structure, all required fields present, valid syntax.
• Dialogue Completeness — all spoken lines from the input dialogue are accounted for in the script.
• Scene and Character Consistency — environmental elements and character descriptions remain coherent across shots.
• Physical Rationality — character positions and camera geometry are physically plausible.

Each check produces a score, and these are aggregated and normalized. The paper does not specify the aggregation method (average? sum? minimum?) or the normalization procedure for structural reward, but the implication is that it produces a scalar in a comparable range to the human preference reward to enable meaningful weighted combination.

The human preference reward $R_{\text{human}}$ is a learned proxy for expert cinematic judgment. The training procedure is:

1. A "team of three senior art directors" scored SFT model outputs on a 1–5 scale across four creative dimensions: shot division rationality, character acting and emotion, visual aesthetics, and directorial intent.
2. These scores were collected for 500 annotated samples, forming $\mathcal{D}_{\text{pref}}$.
3. A BERT-based regression model was trained on this data to predict a normalized preference score in $[0, 1]$ from the script text.

The key design choice here is using a learned proxy for human judgment rather than direct human evaluation during RL training. The reason is scalability: GRPO requires computing rewards for $K=8$ candidate scripts per training step, and doing this for 5,000 training steps would require 40,000 human evaluations, which is infeasible. The BERT-based regression model provides a fast, differentiable approximation of expert judgment that can be computed automatically at each training step. The trade-off is that the proxy may have systematic biases or blind spots — it can only capture what is predictable from the 500 annotated examples and may miss nuanced qualitative dimensions that human experts would notice.

A specific detail: the paper notes that human annotators scored on four dimensions (shot division rationality, character acting and emotion, visual aesthetics, and directorial intent) on a 1–5 scale, but the BERT regression model predicts a single scalar in $[0, 1]$. This implies that the four dimension scores were either averaged into a single score or that the model was trained to predict the overall quality directly. The paper does not specify whether the dimensions were equally weighted or whether the human annotators' overall judgment was used as the target.

GRPO Optimization Procedure

For each training input $x$ (a dialogue from ScriptBench), the optimization proceeds as follows:

Step 1: Candidate generation. The current policy $\pi_{\theta}(\cdot \mid x)$ generates $K = 8$ candidate scripts $\{y^{(k)}\}_{k=1}^{K}$. These are independent samples from the model's output distribution at the current training state. The choice of $K = 8$ represents a trade-off: more candidates would provide more reliable advantage estimates but increase computational cost linearly; fewer candidates would be cheaper but noisier and potentially less informative for the group-based comparison.

Step 2: Reward computation. Each candidate script $y^{(k)}$ is scored using the hybrid reward function to produce $R_k = R_{\text{total}}(y^{(k)})$. This involves running the four automated structural checks (fast, rule-based) and the BERT human preference regressor (fast, learned), so the per-candidate cost is low enough to support 8 candidates per training step.

Step 3: Advantage normalization. The rewards within the group are normalized to compute advantages:

$A_k = \frac{R_k - \bar{R}}{\sigma_R + \epsilon}, \quad \text{where} \quad \bar{R} = \frac{1}{K}\sum_{k=1}^{K} R_k$

where $A_k$ is the advantage of the $k$-th candidate, $R_k$ is its raw reward, $\bar{R}$ is the mean reward across all $K=8$ candidates, $\sigma_R$ is the standard deviation of rewards within the group, and $\epsilon$ is a small constant to prevent division by zero.

What it computes: For each candidate script, this formula measures how much better (positive advantage) or worse (negative advantage) it is than the average candidate in its group, in units of standard deviation. Candidates that score above the group mean receive positive advantages proportional to how far above the mean they are; candidates below the mean receive negative advantages. The division by $\sigma_R + \epsilon$ standardizes the advantages so they are comparable across groups with different reward variances.

Why this form: The group-based normalization is the defining feature of GRPO and what distinguishes it from methods like PPO that use absolute rewards or advantage estimates from a learned value function. By computing advantages within each group rather than across the entire training distribution, GRPO automatically adapts to the difficulty of each input — for an easy dialogue, all candidates might score highly, and the advantages will be small because $\sigma_R$ is small; for a hard dialogue with wide quality variation, the advantages will be larger, driving stronger policy updates. This is particularly appropriate for creative tasks where absolute quality thresholds are meaningless and what matters is relative improvement. The standardization by standard deviation (rather than, say, min-max normalization or ranking) ensures that the magnitude of updates is calibrated to the group's internal variability, preventing over-updating when all candidates are similar or under-updating when one candidate is dramatically better than the rest.

Step 4: Policy update. The policy is updated by maximizing the advantage-weighted log-likelihood of the generated candidates, with a KL-divergence penalty to prevent large deviations from the SFT initialization:

$\mathcal{L}_{\text{GRPO}} = \mathbb{E}_{x \sim \mathcal{D}}\left[\frac{1}{K}\sum_{k=1}^{K} A_k \cdot \log \pi_{\theta}(y^{(k)} \mid x)\right] - \beta \cdot \mathbb{E}_{x \sim \mathcal{D}}\left[\text{KL}\left(\pi_{\theta}(\cdot \mid x) \,\|\, \pi_{\text{SFT}}(\cdot \mid x)\right)\right]$

where $\mathcal{D}$ is the training distribution, $A_k$ is the advantage computed in Step 3, $\pi_{\theta}(y^{(k)} \mid x)$ is the policy's probability of generating candidate $y^{(k)}$, $\pi_{\text{SFT}}$ is the frozen SFT policy, $\beta = 0.04$ is the KL penalty coefficient, and $\text{KL}(\cdot \|\cdot)$ is the Kullback-Leibler divergence.

What it computes: This is a two-term objective. The first term $\frac{1}{K}\sum_{k=1}^{K} A_k \cdot \log \pi_{\theta}(y^{(k)} \mid x)$ increases the log-probability of candidates with positive advantages (good scripts) and decreases the log-probability of candidates with negative advantages (bad scripts), weighted by the magnitude of the advantage. This is the standard policy gradient signal — it pushes the policy toward generating better outputs. The second term $-\beta \cdot \text{KL}(\pi_{\theta} \| \pi_{\text{SFT}})$ penalizes the policy for diverging from the SFT model, ensuring that the RL optimization does not cause the model to forget its structural competence or produce outputs that, while highly rewarded, are no longer recognizable as cinematic scripts. The combined objective maximizes reward while staying close to the SFT checkpoint in probability space.

Why this form: The KL penalty with $\beta = 0.04$ serves a critical function that is specific to this two-stage training paradigm. Without the KL penalty, the RL optimization might discover "reward hacks" — outputs that score highly on the hybrid reward function but are not actually good scripts (e.g., overly dramatic language that tricks the BERT regressor, or scripts that maximize structural reward at the expense of narrative coherence). The SFT model serves as an anchor — it represents the "safe" policy that produces structurally correct scripts. By penalizing divergence from this anchor, the KL term ensures that RL improvements are genuine refinements of the SFT outputs rather than wholesale departures into reward-exploiting territory. The coefficient $\beta = 0.04$ is relatively small, meaning the KL penalty is a soft constraint rather than a hard one — the policy is allowed to deviate meaningfully from SFT, but only when the reward gain justifies the deviation.

Training configuration for the RL stage:

• Optimizer: Adam
• Learning rate: $\eta = 10^{-6}$
• Batch size: 4
• Number of training steps: 5,000
• $K = 8$ candidates per input
• $\beta = 0.04$ (KL coefficient)
• $\alpha = 0.4$ (structural reward weight in hybrid reward)

The learning rate of $10^{-6}$ is an order of magnitude lower than the SFT learning rate of $10^{-5}$, which is standard practice for RL fine-tuning — the model is already close to a good policy after SFT, and large updates could destabilize it. The 5,000 training steps with batch size 4 means the model sees approximately 20,000 training instances (5,000 steps × 4 batches), though since the ScriptBench training set has only 1,700 unique dialogues, these are seen multiple times.

What the RL stage achieves. The paper reports that the RL stage provides "significant boosts to more subjective, artistic dimensions" — Dramatic Tension improves from 3.6 (SFT-only) to 4.1 (full model), and Visual Imagery rises from 3.8 to 4.3. These are substantial improvements on a 0–5 scale, confirming that the hybrid reward function and GRPO optimization successfully capture and optimize for qualities that SFT alone cannot achieve. The paper describes this as "elevating ScripterAgent's capabilities from generating structurally correct scripts to producing cinematically compelling plans aligned with professional standards."

---

DirectorAgent: Cross-Scene Continuous Generation Strategy

While ScripterAgent produces the creative plan, DirectorAgent is responsible for executing it — translating the shot-by-shot script into actual video footage. The core challenge it addresses is that current SOTA video generation models are "typically limited to generating short clips (e.g., 8–12 seconds), far short of the 1–3 minute duration of a complete narrative scene." This temporal mismatch cannot be solved by simply generating longer videos — it is a fundamental limitation of the models' architectures and training procedures, not a parameter that can be adjusted at inference time.

The naive approach and its failures. The straightforward solution — segment the script into 8–12 second chunks and generate each independently — "leads to severe artifacts, such as identity drift, inconsistent styling, and a loss of narrative continuity." The paper identifies the specific failure modes: a character's face might change between segments because the model has no memory of what it generated before; clothing details shift because each generation is independent; spatial layouts drift because the model re-interprets the setting from scratch each time. These artifacts are not subtle — they are immediately noticeable to viewers and fundamentally break the illusion of a continuous scene.

Intelligent Shot-Based Segmentation. DirectorAgent's first mechanism is a principled segmentation algorithm that partitions the full script into generation tasks ("scenes") respecting cinematographic boundaries. The segmentation follows the same four principles used in ScriptBench's shot-level planning:

• Shot Integrity — "each scene must contain one or more complete shot units, preventing cuts in the middle of a continuous camera take." A shot is the atomic unit of cinematography; DirectorAgent never splits a shot across multiple generation segments.
• Duration Adaptation — "the total duration of a scene is constrained to fit within the target model's generation window, with a 10% safety buffer." The paper does not specify whether the 10% buffer applies to the scripted shot duration or the model's maximum generation length, but the intent is clear: allow some margin so that the model does not hit its temporal limit mid-generation.
• Semantic Coherence — "divisions are prioritized at natural narrative breakpoints, such as the end of a character's line or a shift in emotional tone." This ensures that cuts between generated segments occur at points that feel natural to the viewer, not at arbitrary temporal boundaries.
• Technical Feasibility — "segmentation is favored at fixed camera positions, avoiding cuts during complex camera movements which are harder to transition between seamlessly." A cut during a tracking shot creates a visible discontinuity because the camera motion is interrupted; DirectorAgent places cuts at static camera positions where the transition can be masked more effectively.

The output of this segmentation step is a sequence of scene specifications, each corresponding to a single generation call to a video model. The key insight is that segmentation is not just about fitting content into temporal windows — it is about choosing where to cut based on cinematographic principles, so that the inevitable seams between generated segments are as invisible as possible.

Frame-Anchored Continuity. This is the mechanism that directly addresses the identity drift and inconsistency problems. As illustrated in Figure 3, the process works as follows:

1. Generate scene $i$ using the video model, conditioned on ScripterAgent's script for that scene.
2. Extract the last frame of the generated video for scene $i$. This frame captures the visual state at the end of the scene: character positions, facial expressions, clothing, lighting, and spatial layout.
3. Use this last frame as a conditioning image (visual anchor) for the generation of scene $i+1$. The video model receives this frame as additional input alongside the text prompt derived from the script, explicitly instructing it to maintain visual consistency.
4. Additionally, inject the text prompt "Continuing from the previous scene" into the prompt for scene $i+1$, providing textual reinforcement of the continuity constraint.

What is computed operationally: For each pair of consecutive scenes, the system copies the final visual state of the preceding scene and uses it as a strong prior for the subsequent scene's generation. This transforms the long-horizon generation problem from "generate a 1–3 minute video" (which current models cannot do) into "generate a sequence of 8–12 second videos, where each one starts from the exact visual state where the previous one ended" (which current models can do with frame-conditioning capabilities). The "visual relay" metaphor in Figure 3 is apt: each scene hands off a baton (the last frame) to the next scene, ensuring that the visual identity of characters, settings, and props is preserved across the relay exchanges.

Why this mechanism works: Current video models that support image-to-video generation or frame conditioning can maintain visual consistency within a single generation because they have an explicit visual reference to anchor to. The frame-anchoring mechanism exploits this capability across generation boundaries by providing each new generation with an explicit visual reference — the last frame of the previous generation. This is fundamentally different from attempting to achieve consistency through text prompts alone (e.g., "make sure the character wears the same blue dress"), which fails because text descriptions are inherently underspecified compared to images. A text prompt cannot convey the exact shade of blue, the specific pattern of folds in the fabric, or the precise lighting conditions — but a reference frame conveys all of this information implicitly. The mechanism is simple but effective because it leverages the strengths of the generation models (image-to-video consistency) to compensate for their weakness (limited temporal window).

Effectiveness claims and limitations. The paper states that this strategy "substantially reduces the identity drift and jarring scene changes that plague naive segmentation approaches" and "effectively extends the coherence window of any underlying video model." However, it also acknowledges that "challenges such as imperfect lip synchronization and residual misalignment of fine-grained actions remain." The frame-anchoring mechanism provides a strong visual prior for overall appearance and layout but cannot perfectly synchronize lip movements with audio across generation boundaries, nor can it guarantee that fine-grained actions (e.g., a character picking up a specific prop) transition seamlessly. These are acknowledged as current limitations, not fundamental impossibilities — they represent areas where future improvements in the underlying video models would directly benefit the framework.

---

CriticAgent and Evaluation Framework

The evaluation framework is itself a contribution of the paper, designed to assess both stages of the pipeline — script generation and video generation — across multiple dimensions that capture technical correctness, narrative fidelity, and artistic quality. The framework combines three evaluation modalities: AI-powered automated scoring (CriticAgent), human expert panels, and automated metrics.

CriticAgent for script evaluation assesses generated scripts (from the dialogue-to-script stage) on four dimensions, all on 0–5 scales:

• Format Compliance — whether the script adheres to the required JSON structure with all key fields present and correctly structured.
• Shot Division Rationality — whether the segmentation into shots is logical, with breaks aligned to narrative beats without being overly fragmented or lengthy.
• Content Completeness — whether the script provides "rich, actionable details for filming" and enriches the narrative with visual information beyond what is in the dialogue.
• Narrative Coherence — whether shots are logically connected and visual storytelling flows smoothly to complement the dialogue context.

These four dimensions collectively measure structural quality — they ask "is this script well-formed and does it make sense?" but do not measure artistic merit, which is handled by human evaluation.

Human evaluation for script quality employs "a panel of professional directors and screenwriters" who rate scripts on three creative dimensions:

• Character Portrayal Consistency — whether character personality, speaking style, and behavior remain coherent throughout.
• Dramatic Tension & Rhythm — whether the script effectively builds, sustains, and releases dramatic tension with natural pacing.
• Visual Imagery & Cinematic Expressiveness — whether the script conveys visual information vividly and employs cinematic language (shots, staging, atmosphere) effectively.

The detailed scoring rubrics are provided in Appendix B, with each score from 0 to 5 having explicit criteria (e.g., for Visual Imagery, score 4 means "Good Cinematic Guidance: The script includes clear descriptions of scenes, actions, and rough shot intentions, enabling straightforward visualization," while score 5 means "Highly Cinematic and Filmable: The script shows rich visual imagination and appropriate use of film language, making it easy to translate into professional storyboards").

CriticAgent for video evaluation assesses the final generated videos on five dimensions:

• Cinematic Camera Articulation — sophistication of camera work, shot types, framing transitions, and dynamic movements.
• Kinetic Body Language & Blocking — whether character motions and spatial arrangements are expressive and consistent with scripted actions.
• Visual Descriptive Fidelity — whether visual details (character appearance, clothing, scene layout, lighting) match script cues.
• Emotional Arc & Micro-Expressions — whether facial expressions and gestures reflect the intended emotional progression.
• Narrative Pacing & Timing — whether shot timing and pauses align with the narrative structure.

Human evaluation for video quality covers five dimensions:

• Visual Appeal — realism, aesthetic quality, and rendering stability.
• Script Faithfulness — how accurately the video adheres to the script in scenes, actions, and plot progression.
• Narrative Coherence — whether the video forms a logically consistent, easy-to-follow story.
• Character Consistency — whether characters maintain stable identity and appearance throughout.
• Physical Law Adherence — whether motions plausibly adhere to real-world physics.

Automated Metrics. The paper employs three categories of automated evaluation:

1. CLIP Score — computes the cosine similarity between CLIP embeddings of video frames and the corresponding text descriptions, measuring global semantic alignment between what the text says and what the video shows.
2. VBench metrics — a subset of established metrics from the VBench suite (Huang et al., 2024) measuring subject consistency, background consistency, motion smoothness, dynamic degree, and aesthetic quality. These capture various dimensions of video quality independent of script adherence.
3. Visual-Script Alignment (VSA) — the paper's novel contribution to evaluation methodology, designed to fill a specific gap that existing metrics miss.

Visual-Script Alignment (VSA) Metric

The paper identifies that standard metrics measure "whether content appears but not when it appears." A video might contain all the right visual elements (a close-up, a reaction shot, a wide establishing shot) but at the wrong times relative to the script's intended timeline. In narrative video, timing is critical — a reaction shot that appears before the triggering event, or a dramatic reveal that occurs during dialogue rather than after a pause, is a structural failure even if every individual frame is high quality.

VSA is defined as:

$\text{VSA} = \frac{1}{\sum_{k=1}^{K} |T_k|} \sum_{k=1}^{K} \sum_{t \in T_k} \text{Sim}\left(\mathcal{E}_{\text{vis}}(v_t), \mathcal{E}_{\text{txt}}(I_k)\right)$

where $K$ is the number of shot units in the script, $T_k$ is the set of frame indices corresponding to the time interval of shot $k$, $|T_k|$ is the number of frames in that interval, $v_t$ is the video frame at time $t$, $I_k$ is the textual instruction for shot $k$ from the script, $\mathcal{E}_{\text{vis}}$ is a CLIP visual encoder, $\mathcal{E}_{\text{txt}}$ is a CLIP text encoder, and $\text{Sim}(\cdot, \cdot)$ is cosine similarity.

What it computes: For each shot $k$ in the script, the metric identifies the time interval $T_k$ when that shot should be occurring (e.g., "close-up on Character A" from seconds 5 to 12). It then computes the cosine similarity between the CLIP embedding of every video frame $v_t$ within that interval and the CLIP embedding of the shot instruction $I_k$. If the video correctly shows a close-up during seconds 5–12, the frames will have high similarity to the text "close-up on Character A." If the video shows a wide shot during that interval (perhaps because the temporal alignment is off), the similarity will be lower. The outer summation $\sum_{k=1}^{K}$ aggregates these per-shot scores, and the normalization by $\sum_{k=1}^{K} |T_k|$ (the total number of frames across all shots) produces a weighted average where longer shots contribute proportionally more to the final score.

Why this form: The key property of VSA is its temporal specificity. A standard CLIP Score for the whole video would compute the average similarity between the entire video's frames and a single text description, which would be high as long as the right content appears somewhere, regardless of timing. VSA partitions the video into script-specified time intervals and measures alignment within each interval separately, penalizing misalignments where content appears at the wrong time. The per-frame normalization within each shot interval $|T_k|$ ensures that the metric is not biased toward longer shots; each shot contributes proportionally to its duration. The use of CLIP encoders $\mathcal{E}_{\text{vis}}$ and $\mathcal{E}_{\text{txt}}$ leverages pretrained vision-language alignment models, avoiding the need to train a custom alignment scorer. This makes VSA practical to compute — it requires only a script with time intervals, a generated video, and off-the-shelf CLIP models — while providing temporal information that standard metrics miss.

The paper's experiments confirm that VSA captures a dimension of quality not reflected in standard metrics: models conditioned on ScripterAgent's scripts show consistent VSA improvements (e.g., Veo3.1 from 51.4 to 53.8, Sora2-Pro from 48.6 to 50.6) that correlate with human judgments of temporal-semantic fidelity, even when aesthetic metrics do not improve (or slightly decrease, suggesting the spectacle-faithfulness trade-off).

---

Summary of Design Choices and Their Justifications

• Qwen-Omni-7B as base model over larger alternatives: sufficient for learning script structure from a curated 1,700-instance dataset; long-context capability handles multi-turn dialogue; 7B scale is practical for the two-stage training pipeline.
• GRPO over PPO or DPO for RL alignment: group-based relative scoring naturally handles the one-to-many nature of creative script generation where quality is comparative rather than absolute; group normalization automatically adapts to per-input difficulty variations.
• Hybrid reward with $\alpha = 0.4$ over pure human preference or pure structural reward: structural reward prevents the policy from diverging into invalid outputs (format violations, dropped dialogue) while human preference reward drives artistic improvement; the 40/60 split reflects that SFT already provides decent structural quality so most RL improvement budget should target creativity.
• BERT-based regression model as human preference proxy over direct human evaluation during RL: makes 5,000 training steps with K=8 candidates computationally feasible (40,000 evaluations would be cost-prohibitive for human experts); BERT provides a fast, differentiable signal that captures coarse preference patterns from 500 expert-annotated examples.
• Four-principle shot segmentation over arbitrary temporal cuts: respects cinematographic grammar, minimizing visible seams at generation boundaries by placing cuts at natural narrative and camera breakpoints.
• Frame-anchoring over text-only continuity prompts: images provide orders of magnitude more visual information than text, enabling precise preservation of character identity, clothing, lighting, and spatial layout across generation segments without requiring these details to be described (often incompletely) in text.
• VSA metric over CLIP Score alone: standard metrics measure whether content appears; VSA adds temporal structure by measuring whether content appears at the script-specified time, directly measuring narrative alignment rather than content presence.

4. Key Insights and Innovations

Innovation 1: The Script as a Generative Intermediate Representation — Inverting the Video-Language Relationship

The most intellectually distinctive move in this paper is the inversion of the conventional relationship between video and language. Prior video-language research — captioning, visual question answering, video description — all operate in what the paper calls the paradigm of passive description: given a video, produce text that describes it. The dominant models and benchmarks in the field train on exactly this mapping: video → text. This paper argues that automated filmmaking requires the opposite: dialogue → video, where the model must anticipate what should be shown rather than merely report what already exists.

This inversion is not a technical detail — it is a fundamental reframing of what constitutes the creative challenge. A captioning model succeeds if it correctly identifies objects, actions, and relationships present in the visual input. A dialogue-to-video model must generate visual content that was never specified: the camera angle, the lighting, the character blocking, the emotional subtext conveyed through visual composition. None of these elements appear in the dialogue. The model must create them, not extract them.

What elevates this from a task definition to an innovation is the specific form the intermediate representation takes. The paper does not simply ask a model to generate video directly from dialogue (which would be the naive inversion). Instead, it argues that a cinematic script — a structured, machine-readable JSON document specifying shots, camera movements, character positions, and timestamps — is the essential missing layer. This is the Hitchcock thesis operationalized: "The script, the script and the script" becomes a concrete architectural claim about where creative intelligence should reside in an automated filmmaking pipeline.

The significance of this move extends beyond this paper. It proposes that for complex generative tasks where input and output operate at fundamentally different levels of abstraction, the solution is not to train a single end-to-end model to bridge the gap directly, but to decompose the problem into a planning stage (dialogue → structured plan) and an execution stage (structured plan → video), with the planning stage handled by a purpose-trained model that learns domain-specific reasoning. This is a general architectural principle — it could apply to generating architectural designs from client briefs, choreographing dance from musical scores, or designing user interfaces from functional specifications — and the paper's instantiation in the cinematic domain provides a template for how to build the planning model, train it with expert-aligned reinforcement learning, and validate the entire pipeline.

The evidence for the script's centrality is comprehensive. Table 2 shows that conditioning any video model on ScripterAgent's script consistently improves performance across all evaluation dimensions — Script Faithfulness rises from an average of 3.8 to 4.2, Character Consistency from 3.5 to 4.1, and Narrative Coherence from 3.7 to 4.1. These are not model-specific effects; they hold across seven different SOTA video generators with different architectures and training paradigms, confirming that the script, not the video model, is the binding constraint. The ablation in Table 5 decomposes this further: adding the script alone ("w/ Script Only") already raises Body Language & Blocking and Visual Descriptive Fidelity, while adding DirectorAgent's segmentation and frame-anchoring ("w/ Script + Seg." and "Full Agent") raises Cinematic Camera Articulation and Narrative Pacing & Timing. These results isolate the script as the creative engine and the DirectorAgent as the temporal enabler — distinct contributions that the script-centric architecture makes possible to measure separately.

Innovation 2: Difficulty-Conditioned Optimization via Two-Stage SFT+RL with Hybrid Reward — and Why GRPO Is the Right Algorithm for Creative Tasks

The paper's training methodology for ScripterAgent — supervised fine-tuning followed by Group Relative Policy Optimization with a hybrid reward function — is more than an engineering choice. It embodies a specific claim about the relationship between structural competence and creative quality that has implications for how we train models for any task where correctness is necessary but insufficient.

The key conceptual move is the decomposition of quality into objective structural correctness and subjective artistic merit, and the argument that these two components require fundamentally different training signals and optimization procedures. Structural correctness — valid JSON, all dialogue accounted for, physically plausible blocking — can be learned via maximum likelihood on expert demonstrations (the SFT stage) and verified via rule-based checks (the structural reward component). Creative quality — dramatic tension, visual expressiveness, pacing — cannot be captured by any fixed rule or single ground-truth output. It is inherently comparative and subjective, which means it requires a training signal that captures relative preference (this script is better than that one) rather than absolute correctness (this script matches the reference).

This is where the choice of GRPO becomes conceptually significant rather than just technically convenient. The paper argues that GRPO's "group-based relative scoring is well-suited for creative tasks that have a subjective, one-to-many nature of valid outputs." To understand why this matters, consider the alternatives. PPO with a learned reward model produces absolute scores — a script gets a scalar rating, and the policy is updated to maximize that rating. But for creative tasks, absolute ratings are unstable: what counts as a "good" script depends on the difficulty of the input dialogue and on the quality of other scripts the model could have produced. DPO trains on pairs of outputs with a human preference label, which requires explicit pairwise comparisons — expensive to collect and limited in the number of comparisons that can be annotated.

GRPO sidesteps both problems. By generating K=8 candidate scripts for the same input and computing advantages within that group, it automatically calibrates to input difficulty. For an easy dialogue where all candidates are good, the advantages are small because reward variance is low, resulting in conservative updates. For a hard dialogue where one candidate is dramatically better than the others, the advantages are large, driving stronger policy improvements. The group-based normalization by standard deviation ($\sigma_R + \epsilon$) ensures that the magnitude of the policy update is proportional to how much better the good candidate is relative to its peers, not to some absolute quality threshold. This makes the optimization input-adaptive in a way that absolute-reward methods are not.

The hybrid reward structure with $\alpha = 0.4$ (40% structural, 60% human preference) is equally deliberate. An $\alpha$ of 0.5 would imply equal weight; an $\alpha$ of 0.0 would be pure preference optimization which would quickly diverge into reward hacking. The 40/60 split encodes the judgment that the RL stage's primary job is creative improvement — the SFT model already achieves 3.9 in Format Compliance (Table 1), so structural correctness has diminishing returns — while acknowledging that some structural signal is necessary to prevent catastrophic forgetting. The paper does not ablate $\alpha$, but the choice is interpretable: giving the learned human preference signal majority weight (60%) reflects the insight that the gap between adequate and excellent is primarily artistic, not technical.

The evidence for the RL stage's contribution is in Table 1. The SFT-only model scores 3.9 in Format Compliance and 3.9 in Narrative Coherence — structurally sound scripts. But its Dramatic Tension is 3.6 and Visual Imagery is 3.8. The full model (SFT+RL) raises these to 4.1 and 4.3 respectively, while Format Compliance stays at 4.0 and Narrative Coherence rises modestly to 4.2. The largest gains are precisely in the subjective, artistic dimensions that maximum likelihood cannot optimize. This is not just an improvement — it is evidence for the paper's central claim that "SFT is sufficient for structural competence [while] the RL stage with its hybrid reward function is essential for aligning the model with expert directorial aesthetics." The result validates the conceptual decomposition: structural correctness comes from supervised learning on demonstrations; creative quality comes from preference optimization against expert judgment.

Innovation 3: The Spectacle-Faithfulness Trade-off — A Diagnostic Finding About the Limits of Current Video Models

The paper's third major contribution is not a method but a diagnostic finding: current state-of-the-art video generation models optimize along different axes, and these axes are at least partially in tension. Sora2-Pro achieves the highest Visual Appeal (4.8) and Physical Law Adherence (4.5) but only 4.2 in Script Faithfulness. HYVideo1.5 achieves the highest Script Faithfulness (4.6), Character Consistency (4.4), and Narrative Coherence (4.3) but lower visual spectacle. This is not a small difference — it represents a fundamental architectural or training-prioritization divergence between models that are ostensibly solving the same task.

The significance of this finding lies in what it reveals about the current generation of video models. The fact that no single model simultaneously leads in both visual quality and script adherence suggests that these capabilities are not simply correlated — they are in tension, and model designers have made explicit or implicit trade-offs. A model optimized for photorealism and dynamic motion (Sora2-Pro leads Dynamic Degree at 79.5) may sacrifice precise instruction following because the generation process that produces the most visually impressive outputs is not the same process that most faithfully renders a specified sequence of shots. Conversely, a model optimized for instruction adherence and consistency (HYVideo1.5 leads Subject Consistency at 97.2 and Background Consistency at 97.5) may produce less visually spectacular outputs because the constraints that maintain consistency across frames also constrain the model's ability to generate the highest-fidelity individual frames.

This trade-off has practical implications that the paper explicitly draws out: "practitioners selecting models for specific filmmaking applications" should choose Sora2-Pro when visual impact is paramount and HYVideo1.5 when narrative integrity is the priority. But the finding is more than a model selection guide — it is a diagnostic of the field's current capability frontier. It tells us that reconciling spectacle and faithfulness is an unsolved problem, and that the next generation of video models should be evaluated on both dimensions simultaneously rather than optimizing for one at the expense of the other.

The paper's evidence for this trade-off is consistent across both AI ratings and human evaluations (Table 2) and across automated metrics (Table 3). In the automated metrics, Sora2-Pro with ScripterAgent achieves 62.8 in Aesthetic quality and 79.5 in Dynamic Degree but only 50.6 in VSA; HYVideo1.5 achieves 54.8 in VSA and 97.2 in Subject Consistency but substantially lower Aesthetic quality (60.2) and Dynamic Degree (72.5). The VSA metric is particularly revealing here: it measures temporal-semantic alignment — whether the right content appears at the right time. Sora2-Pro's lower VSA despite higher overall quality scores confirms that its visual prowess comes at a cost to precise temporal control. The model produces beautiful frames that may not correspond to the script's intended shot sequence with the same fidelity as HYVideo1.5.

This finding also reframes the paper's own contribution. The script-centric approach does not eliminate the trade-off — it improves both dimensions for all models, but the relative strengths of each model persist. Sora2-Pro with ScripterAgent still leads in Visual Appeal (4.8) while HYVideo1.5 still leads in Script Faithfulness (4.6). What the script does is shift the entire frontier upward: models become better at both spectacle and faithfulness, but the underlying tension between them remains. This suggests that the trade-off is not an artifact of insufficient prompting but a deeper property of current model architectures that further research must address.

Innovation 4: Visual-Script Alignment — A Metric for Temporal-Semantic Fidelity That Exposes What Standard Metrics Miss

The paper's introduction of Visual-Script Alignment (VSA) as an evaluation metric addresses a specific diagnostic gap that the paper itself identified as limiting progress in narrative video generation. Standard metrics — CLIP Score, VBench's subject consistency, background consistency, motion smoothness — all measure whether content appears but not when it appears. In narrative video, timing constitutes meaning: a reaction shot that appears before the triggering line of dialogue is not the same story as one that appears after, even if both shots contain identical visual content.

This insight — that temporal structure is semantically load-bearing in narrative video — is not widely operationalized in video generation evaluation. The field's default metrics treat a video as a bag of frames averaged over time. A high CLIP Score means that, on average, the video's visual content matches the text description, but it says nothing about whether the sequence of visual events follows the narrative arc. VSA addresses this by partitioning the video into script-specified time intervals and measuring alignment within each interval separately, penalizing correct content at the wrong time.

The metric's definition is conceptually clean: for each shot interval $T_k$ in the script, compute the cosine similarity between the CLIP embedding of every frame in that interval and the CLIP embedding of the shot's textual instruction $I_k$. Average across all frames, weighted by interval duration. This directly operationalizes "is the right thing happening at the right time" in a way that requires only a script with timestamps, a generated video, and off-the-shelf CLIP encoders — no custom-trained scorer needed.

What makes VSA more than a technical convenience is the kind of evaluation it enables. In Table 3, VSA scores reveal information that other metrics obscure. Sora2-Pro with ScripterAgent achieves a CLIP Score of 46.0 and an Aesthetic score of 62.8 — both strong. But its VSA is only 50.6, substantially below HYVideo1.5's 54.8. The VSA metric is surfacing a specific weakness — temporal misalignment — that is invisible in aggregate quality metrics. This is exactly the diagnostic capability the paper needs to validate its central claim that the structured script improves not just content quality but temporal-semantic fidelity. The consistent VSA improvements across all models when conditioned on ScripterAgent scripts (Vidu2: 48.2 → 50.0; Seedance1.5-Pro: 50.4 → 52.6; Kling2.6: 51.3 → 53.5; Wan2.5: 52.1 → 54.1; HYVideo1.5: 52.7 → 54.8; Sora2-Pro: 48.6 → 50.6; Veo3.1: 51.4 → 53.8) provide the quantitative backbone for the paper's narrative coherence claims — claims that standard metrics could not verify on their own.

The VSA metric also connects to a broader methodological issue in generative AI evaluation. As models improve, aggregate metrics tend to saturate — if every model achieves a CLIP Score of 45–47 on this task, CLIP Score stops being discriminative. But diagnostic metrics that measure specific failure modes (like temporal misalignment) can remain informative even when overall quality is high. VSA exemplifies this principle: it identifies a dimension of quality that is both important (narrative coherence depends on timing) and under-measured (standard metrics ignore temporal structure), making it a valuable addition to the evaluation toolkit even beyond this specific task. The paper positions VSA as a general metric for any video generation task where alignment with a temporally-structured plan matters — story visualization, instructional video generation, animated explainer production — and the clean formulation with CLIP encoders makes it straightforward to adopt.

5. Experimental Analysis

Evaluation Methodology

• Dataset. All experiments use the ScriptBench test set, consisting of 50 dialogue-to-cinematic-script instances curated from high-fidelity cinematic cutscenes. Each instance includes multimodal input (dialogue text, audio, character positions) and a corresponding structured JSON script generated via the expert-guided annotation pipeline described in Section 2.1. The 50-instance test set is separated from the 1,700-instance training set, and the paper emphasizes that this split "intentionally challenges the model to infer complete cinematic elements from conversational content alone" — the test dialogues were not seen during ScripterAgent's training, so the model must generalize its cinematographic reasoning to new conversational scenarios.

• Base model(s). All script generation experiments use Qwen-Omni-7B as the base model for ScripterAgent, chosen for its "strong capabilities in long-context processing and instruction following." The SFT stage fine-tunes this model on the 1,700 ScriptBench training instances for 20 epochs, while the RL stage further optimizes the SFT checkpoint for 5,000 steps using GRPO. For video generation evaluation, the paper tests seven SOTA text-to-video models that support voice generation and long-text inputs (>1000 tokens): Vidu2, Seedance1.5-Pro, Kling2.6, Wan2.6 (and Wan2.5 in automatic metrics), HYVideo1.5, Sora2-Pro, and Veo3.1. These models span different architectural paradigms (U-ViT, diffusion transformer, dual-stream diffusion transformer) and different development organizations, making the evaluation a cross-sectional study of the current video generation landscape rather than an assessment of any single architecture. The paper treats these models as black-box generation engines — ScripterAgent's scripts are fed to them as input, but the models themselves are never fine-tuned or modified.

• Metrics. The evaluation framework combines three measurement modalities, all scored on 0–5 scales unless otherwise noted. For script generation, the CriticAgent (powered by Gemini-2.5-Pro) automatically assesses four dimensions: Format Compliance (JSON structure validity), Shot Division Rationality (logical segmentation), Content Completeness (actionable filming details), and Narrative Coherence (logical shot connectivity). Human experts (professional directors and screenwriters) additionally evaluate three creative dimensions: Character Portrayal Consistency, Dramatic Tension & Rhythm, and Visual Imagery & Cinematic Expressiveness, using detailed scoring rubrics provided in Appendix B. For video generation, CriticAgent evaluates five dimensions: Cinematic Camera Articulation, Kinetic Body Language & Blocking, Visual Descriptive Fidelity, Emotional Arc & Micro-Expressions, and Narrative Pacing & Timing. Human evaluators rate videos on five complementary dimensions: Visual Appeal, Script Faithfulness, Narrative Coherence, Character Consistency, and Physical Law Adherence. Automated metrics include CLIP Score (global semantic alignment between video frames and text), a subset of VBench metrics (subject consistency, background consistency, motion smoothness, dynamic degree, aesthetic quality), and the novel Visual-Script Alignment (VSA) metric, which measures whether visual events occur within their script-specified time intervals by computing per-interval CLIP cosine similarity between frames and shot instructions. The VSA metric specifically addresses the limitation that standard metrics "measure whether content appears but not when it appears."

• Baselines. For script generation, the paper compares against three representative story visualization and screenplay generation methods: CHAE (Wang et al., 2022), which enables fine-grained controllable story generation through structured input specifying characters, actions, and emotions; MoPS (Ma et al., 2024a), a modular framework for automated story premise synthesis that decomposes premises into theme, background, persona, and plot modules; and SEED-Story (Yang et al., 2025), which extends multimodal LLMs for long narrative generation using a multimodal attention sink mechanism for consistency. An SFT-only variant of ScripterAgent (without the RL alignment stage) serves as an internal ablation baseline. For video generation, the baseline condition is feeding raw dialogue directly into each video generation model without any intermediate ScripterAgent script — this is the "w/o ScripterAgent" configuration in all video results tables. No additional video generation baselines (e.g., end-to-end dialogue-to-video models, or alternative planning approaches like MovieAgent-guided generation) are compared; the paper's video experiments focus exclusively on whether ScripterAgent's scripts improve output relative to raw dialogue conditioning.

• Generation budget / compute accounting. Video generation experiments do not report computational cost in FLOPs, GPU-hours, or any standardized compute metric. The paper treats video generation models as API-accessible services — the experimental variable is the input conditioning (raw dialogue vs. ScripterAgent script) rather than any compute budget variation within a single model. The DirectorAgent's cross-scene generation strategy produces multiple clips that are concatenated, but the paper does not report the total number of generations per video, the duration of generated segments, or the computational overhead of frame-anchoring relative to single-clip generation. This absence of compute accounting means that the paper cannot make claims about computational efficiency — it can only claim that scripts improve output quality, not that they improve output quality per unit of compute. The paper acknowledges this implicitly by not making any efficiency claims.

• Cross-validation / statistical protocol. The paper does not employ cross-validation, bootstrap confidence intervals, or statistical significance testing. All script generation results (Table 1) are reported as single-point evaluations on the 50-instance test set without variance estimates. Video generation results (Tables 2, 3, 4, 5) similarly report mean scores without standard deviations, confidence intervals, or inter-annotator agreement metrics for the human evaluations. The human evaluation panels (three senior art directors for script reward annotation, a "panel of professional directors and screenwriters" for script quality assessment, unspecified numbers of "human annotators" for video evaluation) are described in terms of their professional qualifications but without reporting the number of annotators per instance, the degree of overlap between annotators, or any inter-rater reliability statistics. The AI-based evaluations (CriticAgent, plus Qwen3-VL and GLM-4.6-V in the ablation study) are single-model judgments without ensemble averaging or calibration against human ratings beyond the 500-sample preference training set for $R_{\text{human}}$. This limits the statistical rigor of the reported comparisons — we cannot assess whether a 0.1-point difference between methods reflects a genuine improvement or evaluation noise, particularly on a 50-instance test set where individual outlier judgments could meaningfully shift means.

Main Quantitative Results

Script Generation Results (Table 1, Figure 4)

The full ScripterAgent (SFT+RL) achieves state-of-the-art performance across all AI and human evaluation metrics on ScriptBench's 50-instance test set, with the largest gains concentrated in subjective artistic dimensions. As reported in Table 1, the full model outperforms the strongest baseline (SEED-Story) by +0.4 points in Format Compliance (4.0 vs. 3.6), Shot Division Rationality (3.9 vs. 3.5), Content Completeness (4.1 vs. 3.7), and Narrative Coherence (4.2 vs. 3.8) on AI-rated metrics. On human-rated dimensions, the full model leads SEED-Story by +0.4 in Character Portrayal Consistency (4.0 vs. 3.6), +0.4 in Dramatic Tension (4.1 vs. 3.7), and +0.5 in Visual Imagery (4.3 vs. 3.8). The baselines CHAE and MoPS trail further behind, with human ratings mostly in the 3.0–3.5 range compared to ScripterAgent's 4.0–4.3 range. These results confirm that ScripterAgent "effectively bridges the 'semantic gap' by transforming coarse dialogue into detailed, cinematically expressive, and director-level instructions."

The RL alignment stage provides the majority of the improvement in artistic quality while SFT alone already achieves strong structural scores. Comparing the SFT-only variant against the full model in Table 1 reveals a clear pattern: SFT-only achieves 3.9 in Format Compliance and 3.9 in Narrative Coherence — already competitive with or exceeding all baselines — but only 3.6 in Dramatic Tension and 3.8 in Visual Imagery. The full model (adding GRPO with hybrid reward) raises Dramatic Tension to 4.1 (+0.5) and Visual Imagery to 4.3 (+0.5), while Format Compliance improves only marginally to 4.0 (+0.1) and Narrative Coherence to 4.2 (+0.3). The SFT-only scores for Content Completeness (3.8) and Character Portrayal Consistency (3.7) also see meaningful gains to 4.1 (+0.3) and 4.0 (+0.3) respectively under full training. This decomposition validates the paper's central claim that "SFT is sufficient for structural competence [while] the RL stage with its hybrid reward function is essential for aligning the model with expert directorial aesthetics, refining its ability to handle nuanced creative elements like pacing and shot composition." The fact that the largest absolute improvements (+0.5 on both Dramatic Tension and Visual Imagery) occur in exactly the subjective dimensions that maximum likelihood cannot optimize — because there is no single ground-truth reference for artistic quality — supports the paper's argument that preference-based RL is necessary, not merely helpful, for creative alignment.

Qualitative inspection of generated scripts reveals that ScripterAgent produces substantially richer technical specifications than the closest existing system. Figure 4 presents a direct comparison between ScripterAgent and MovieAgent (Wu et al., 2025b). The paper characterizes MovieAgent's output as "a simple plot summary" with vague directions like "Camera slowly pans…" In contrast, ScripterAgent's script includes precise camera settings ("Shallow Depth of Field + panoramic shot"), timestamped synchronization markers ("[00:00:00 00:00:08]"), detailed character descriptions ("Approximately 180cm tall"), atmospheric scene descriptions, and explicit character blocking instructions. This qualitative difference is not merely aesthetic — it demonstrates that ScripterAgent generates the "professional-quality, executable script needed to guide automated video production," providing concrete technical parameters that downstream video models can use, rather than high-level narrative descriptions that require further human interpretation.

---

Video Generation Results: AI and Human Ratings (Table 2)

Conditioning video generation models on ScripterAgent scripts yields consistent improvements across all evaluated dimensions and all seven tested models. Table 2 reports both AI ratings (CriticAgent's five dimensions) and human ratings (five dimensions), with the "w/ ScripterAgent" configuration compared against the "Raw Dialogue (w/o ScripterAgent)" baseline for each model. The aggregate results are unambiguous: the mean AI rating rises from 4.2 to 4.5 (+0.3), and the mean human rating rises from 3.7 to 4.2 (+0.5). These improvements are not driven by a single outlier model — every model shows gains in both mean AI rating (Vidu2: 4.3 → 4.5; Seedance1.5-Pro: 4.2 → 4.6; Kling2.6: 4.3 → 4.5; Wan2.6: 4.4 → 4.6; HYVideo1.5: 4.2 → 4.6; Sora2-Pro: 4.3 → 4.5; Veo3.1: 4.2 → 4.4) and mean human rating (Vidu2: 3.4 → 3.9; Seedance1.5-Pro: 3.4 → 4.0; Kling2.6: 3.5 → 4.1; Wan2.6: 3.4 → 4.0; HYVideo1.5: 4.0 → 4.4; Sora2-Pro: 3.9 → 4.4; Veo3.1: 4.0 → 4.4).

The largest improvements concentrate in dimensions directly related to script adherence and narrative structure. On human-rated Script Faithfulness, the average across all models improves from 3.8 to 4.2 (+0.4), with particularly large gains for Wan2.6 (3.2 → 4.0, +0.8) and Sora2-Pro (3.6 → 4.2, +0.6). HYVideo1.5 achieves the highest absolute Script Faithfulness at 4.6. Character Consistency shows a similarly strong average improvement from 3.5 to 4.1 (+0.6), with Vidu2 improving from 3.0 to 3.7 (+0.7) and Sora2-Pro from 3.7 to 4.3 (+0.6). Narrative Coherence rises from 3.7 to 4.1 (+0.4) on average. These gains confirm that ScripterAgent's structured scripts directly address the specific weaknesses observed in the raw dialogue baseline — poor character consistency, weak script adherence, and fragmented narrative flow. On the AI-rated dimensions, Body Language & Blocking shows the largest average gain (+0.5, from 4.0 to 4.5), which the paper attributes to "explicit shot-level blocking instructions [that] enhance fine-grained execution by synchronizing motion with scene rhythm." Pace Timing rises from 4.3 to 4.6 (+0.3) on average, and Visual Descriptive Fidelity rises from 4.5 to 4.7 (+0.2). These more modest gains on dimensions that were already relatively high in the raw dialogue baseline (4.5 for Fidelity, 4.3 for Pace) suggest a ceiling effect — the raw dialogue condition already produced visually adequate content; the script primarily improves narrative and structural dimensions rather than raw visual quality.

The paper's headline finding — a fundamental trade-off between visual spectacle and script adherence — is clearly visible in the per-model results. Sora2-Pro, even when conditioned on ScripterAgent scripts, maintains the highest Visual Appeal score (4.8) and Physical Law Adherence (4.5) among all models, but its Script Faithfulness (4.2) and Narrative Coherence (4.1) are lower than HYVideo1.5's corresponding scores (4.6 and 4.3). Conversely, HYVideo1.5 achieves the best Script Faithfulness (4.6), Character Consistency (4.4), and Narrative Coherence (4.3) but trails Sora2-Pro in Visual Appeal (4.5 vs. 4.8) and Physical Law Adherence (4.2 vs. 4.5). This pattern — Sora2-Pro leading in spectacle, HYVideo1.5 leading in narrative integrity — is consistent with the raw dialogue baselines (Sora2-Pro: Visual Appeal 4.2, Script Faithfulness 3.6; HYVideo1.5: Visual Appeal 4.0, Script Faithfulness 4.2) and persists after script conditioning. The paper interprets this as evidence that "current video models optimize along different axes: some prioritize perceptual realism, while others better maintain the semantic logic of a storyline." The crucial insight is that ScripterAgent does not eliminate this trade-off — it shifts both models upward along both axes, but the relative specialization remains. This suggests that the spectacle-faithfulness tension is a genuine property of current model architectures, not an artifact of poor prompting.

The video generation results also reveal an interesting pattern where the script's benefits manifest differently for different models. Vidu2, which started from the lowest human-rated average (3.4) in the raw dialogue condition, shows the largest relative improvement to 3.9 (+0.5), with particularly strong gains in Body Language & Blocking (4.1 → 4.4), Script Faithfulness (3.3 → 3.9), and Character Consistency (3.0 → 3.7). Sora2-Pro, which started higher (3.9), improves to 4.4 (+0.5), with its biggest gains in Script Faithfulness (3.6 → 4.2, +0.6), Character Consistency (3.7 → 4.3, +0.6), and Narrative Coherence (3.9 → 4.1, +0.2). HYVideo1.5, already the strongest at narrative adherence in the raw condition (4.0 average), improves to 4.4 (+0.4), suggesting that even models already good at instruction following benefit from the additional structure. The uniform positive effect across models with different architectures and base capabilities is the strongest evidence that the script, not any particular model's properties, is the binding constraint on narrative video generation from dialogue.

---

Video Generation Results: Automated Metrics (Table 3)

The Visual-Script Alignment (VSA) metric confirms that ScripterAgent scripts improve temporal-semantic fidelity across all tested models, with gains of approximately 2–3 points. Table 3 reports CLIP Score, VSA, and five VBench metrics for the raw dialogue and script-conditioned configurations. VSA improvements are consistent: Vidu2 rises from 48.2 to 50.0 (+1.8), Seedance1.5-Pro from 50.4 to 52.6 (+2.2), Kling2.6 from 51.3 to 53.5 (+2.2), Wan2.5 from 52.1 to 54.1 (+2.0), HYVideo1.5 from 52.7 to 54.8 (+2.1), Sora2-Pro from 48.6 to 50.6 (+2.0), and Veo3.1 from 51.4 to 53.8 (+2.4). The range of VSA improvements (1.8–2.4 points) is notably narrow across seven different models, suggesting a consistent effect that is not model-specific. HYVideo1.5 achieves the highest absolute VSA score at 54.8, consistent with its leadership in Script Faithfulness and Narrative Coherence from the human evaluation. The paper interprets these VSA gains as quantitative validation that "the detailed instructions from ScripterAgent enable models to better adhere to script semantics and timing."

CLIP Score improvements track VSA gains closely, with an average increase of approximately 1.7 points across models. Vidu2 improves from 42.2 to 43.9 (+1.7), Seedance1.5-Pro from 43.8 to 45.5 (+1.7), Kling2.6 from 44.7 to 46.2 (+1.5), Wan2.5 from 45.3 to 47.2 (+1.9), HYVideo1.5 from 43.4 to 45.1 (+1.7), Sora2-Pro from 44.1 to 46.0 (+1.9), and Veo3.1 from 43.6 to 45.3 (+1.7). The paper notes that while CLIP Score improvements are consistent, they are generally smaller than VSA improvements, which is expected: CLIP Score measures global semantic alignment (whether the right content appears somewhere), while VSA measures temporal alignment (whether content appears at the right time), and the script's primary benefit is in temporal structure rather than content vocabulary.

The VBench metrics reveal the spectacle-faithfulness trade-off in automated form, with Sora2-Pro dominating visual quality metrics and HYVideo1.5 dominating consistency metrics. Sora2-Pro with ScripterAgent achieves the highest Aesthetic quality (62.8), Dynamic Degree (79.5), and Motion Smoothness (98.2). These numbers align with its human-rated leadership in Visual Appeal (4.8) and Physical Law Adherence (4.5). HYVideo1.5 achieves the highest Subject Consistency (97.2) and Background Consistency (97.5), consistent with its human-rated leadership in Character Consistency (4.4). The Dynamic Degree metric is particularly instructive: Sora2-Pro leads at 79.5, far above HYVideo1.5's 72.5, confirming that Sora2-Pro generates more visually dynamic and complex motion. However, this dynamism may come at a cost to precise instruction following — Sora2-Pro's VSA (50.6) is substantially below HYVideo1.5's (54.8). The paper interprets this pattern as confirmation that "Sora2-Pro leads in visual spectacle" while "HYVideo1.5 excels at instruction following."

The Dynamic Degree metric shows consistent increases when models are conditioned on ScripterAgent scripts, suggesting that explicit action descriptions elicit more visually complex outputs. Sora2-Pro's Dynamic Degree rises from 75.2 to 79.5 (+4.3), Kling2.6's from 52.0 to 56.5 (+4.5), and the average increase across all models is approximately 3–4 points. The paper attributes this to "explicit action descriptions in the scripts [that] guide models to create more visually complex scenes, moving beyond the static 'talking head' outputs often produced from raw dialogue." This is a non-obvious finding: one might expect that structured, constraint-heavy scripts would produce more rigid and less dynamic outputs. Instead, the scripts appear to provide richer action descriptions that give models more specific material to animate, resulting in more motion rather than less.

Subject and Background Consistency metrics show modest but consistent improvements, validating the DirectorAgent's cross-scene generation strategy. Subject Consistency averaged across models with ScripterAgent is approximately 94.3, compared to 93.0 in the raw dialogue condition. Background Consistency shows a similar pattern. While these gains are smaller than the VSA or Dynamic Degree improvements, they are consistent across models and confirm that the frame-anchoring mechanism "effectively mitigates identity drift" compared to naive independent segment generation. The paper does not report an ablation comparing frame-anchored generation against independent segment generation in automated metrics (the ablation in Table 5 uses AI evaluators, not VBench metrics), so we cannot isolate how much of the consistency gain comes from the scripts versus from the DirectorAgent mechanism specifically.

---

Supplementary AI Evaluation with Alternative Critic Models (Table 4, Appendix A.1)

To validate that CriticAgent's ratings are not an artifact of using a single evaluator model, the paper replicates video evaluation using Qwen3-VL as an alternative AI critic. Table 4 reports Qwen3-VL's ratings on the same five AI evaluation dimensions. The pattern of results closely mirrors the Gemini-2.5-Pro evaluations from Table 2, though with somewhat different absolute scores. Under Qwen3-VL, the raw dialogue average across models is 4.1 (compared to Gemini's 4.2), and the script-conditioned average is 4.6 (compared to Gemini's 4.5). The improvement from raw dialogue to script conditioning is +0.5 under Qwen3-VL, compared to +0.3 under Gemini — Qwen3-VL shows a larger relative improvement, which the paper interprets as evidence that the gains are robust across evaluator models. One notable pattern in Table 4 is that the script-conditioned scores are remarkably uniform across all seven video models (all clustered at 4.5–4.6), whereas the raw dialogue scores show meaningful variation (3.9–4.3). This could indicate either that Qwen3-VL is less discriminative at high quality levels (a ceiling effect in its rating behavior) or that the script conditioning genuinely homogenizes output quality to the point where model differences become negligible under this evaluator's criteria. The paper does not investigate which interpretation is correct.

---

Ablation: Component-Level Contributions of Scripting and Agentic Mechanisms (Table 5, Appendix A.2)

The ablation study in Table 5 decomposes the framework's performance into three contributing factors — the script itself, shot-aware segmentation, and frame-anchoring — evaluated across three different AI critics (Gemini-2.5-Pro, Qwen3-VL, GLM-4.6-V) on two video backbone models (Wan2.6 and HYVideo1.5). The four experimental configurations are: Baseline (raw dialogue, no ScripterAgent), w/ Script Only (script conditioning but single long-horizon generation without segmentation or frame-anchoring), w/ Script + Seg. (script conditioning with shot-aware segmentation but without frame-anchoring), and Full Agent (script conditioning, segmentation, and frame-anchoring). Each configuration is evaluated on five AI rating dimensions.

The script alone ("w/ Script Only") provides the majority of the improvement in Body Language & Blocking and Visual Descriptive Fidelity, while the DirectorAgent's segmentation and frame-anchoring are primarily responsible for gains in Cinematic Camera Articulation and Narrative Pacing & Timing. For Wan2.6 under Gemini-2.5-Pro evaluation: Baseline scores Body Language at 4.2, Fidelity at 4.7, Camera Articulation at 4.2, and Pace Timing at 4.4. Adding Script Only raises Body Language to 4.3 (+0.1) and Fidelity stays at 4.7 (already near ceiling), while Camera Articulation stays at 4.2 and Pace improves slightly to 4.5 (+0.1). Adding segmentation (w/ Script + Seg.) raises Camera Articulation to 4.3 (+0.1) and Pace to 4.6 (+0.1), while Body Language rises further to 4.4. The Full Agent configuration achieves the highest scores across all dimensions: Camera Articulation 4.4, Body Language 4.6, Fidelity 4.7, Pace 4.7. The same decomposition is visible under all three critic models and for both Wan2.6 and HYVideo1.5. The paper interprets this as evidence of "a clear functional decoupling of our agents: Visual Descriptive Fidelity and Kinetic Body Language & Blocking show marked improvements with the introduction of w/ Script Only, attributing fine-grained visual control to the ScripterAgent. Conversely, the DirectorAgent is shown to be critical for temporal dimensions; the implementation of shot-aware segmentation and frame-anchoring drives scores in Narrative Pacing & Timing and Cinematic Camera Articulation to their highest levels."

The high inter-critic agreement across Gemini-2.5-Pro, Qwen3-VL, and GLM-4.6-V strengthens confidence in the ablation's validity. For the Full Agent configuration on Wan2.6, all three critics agree on the ranking of dimensions and produce similar score ranges: Camera Articulation ranges from 4.4 to 4.6, Body Language from 4.6 to 4.8, Fidelity from 4.4 to 4.9, and Pace from 4.5 to 4.7. For HYVideo1.5, the agreement is similarly strong: Camera Articulation 4.4–4.6, Body Language 4.2–4.6, Fidelity 4.4–4.9, Pace 4.6–4.7. The paper cites this agreement as corroborating "the validity of these improvements." However, without inter-annotator agreement statistics (e.g., Krippendorff's alpha, intraclass correlation), the qualitative observation of similar score ranges cannot be distinguished from evaluators sharing similar biases or rating heuristics. The fact that all three critics are LLMs (which may share training data, architectural principles, or evaluation tendencies) further limits the independence of these measurements.

The ablation reveals an interaction between script quality and DirectorAgent mechanisms: the script enables the segmentation and anchoring to work effectively. In the Baseline configuration (raw dialogue), there is no structured shot plan to segment, so the DirectorAgent cannot operate — the baseline inherently uses a different generation paradigm. The w/ Script Only configuration uses the structured script but generates as a single long clip, which forces the video model to handle the full duration within its temporal constraints. The degradation from Full Agent to w/ Script Only is therefore measuring two effects simultaneously: the benefit of segmentation (which enables generation within model limits by breaking the script into feasible chunks) and the benefit of frame-anchoring (which maintains visual continuity across those chunks). The w/ Script + Seg. configuration isolates segmentation without anchoring, and the consistent additional gain in the Full Agent across all dimensions confirms that frame-anchoring provides value beyond segmentation alone. For Wan2.6 under Gemini-2.5-Pro, the Camera Articulation progression is 4.2 (Baseline) → 4.2 (Script Only) → 4.3 (Script+Seg) → 4.4 (Full Agent), showing that each component adds incrementally.

Ablation Studies and Robustness Checks

RL alignment stage (SFT-only vs. Full ScripterAgent): The most important ablation compares ScripterAgent trained with SFT only against the full two-stage (SFT+RL) model in Table 1. The SFT-only model achieves strong structural scores (Format Compliance 3.9, Narrative Coherence 3.9) but substantially lower artistic scores (Dramatic Tension 3.6, Visual Imagery 3.8) compared to the full model (4.1 and 4.3 respectively). The +0.5 improvement in Dramatic Tension and Visual Imagery is the single largest performance gap in the entire script generation evaluation and directly validates the paper's claim that RL alignment with a hybrid reward function is necessary for creative quality. The human preference reward component ($R_{\text{human}}$, weighted at 60% of the total reward) is specifically designed to optimize for these subjective dimensions, and the ablation confirms that removing this component (by removing the entire RL stage) eliminates the artistic gains while preserving structural competence.

Hybrid reward function weighting (α = 0.4): The paper does not perform an ablation on the weight parameter $\alpha$ that balances structural and human preference rewards in the hybrid reward function. The choice of $\alpha = 0.4$ (40% structural, 60% human preference) is stated without justification from hyperparameter sweeps or sensitivity analysis. This is a notable omission because $\alpha$ encodes the paper's central claim about the relative importance of creativity versus correctness in the RL stage — an $\alpha$ of 0.0 would risk catastrophic divergence into invalid outputs, while an $\alpha$ of 1.0 would reduce RL to structural optimization only, producing no artistic gains. The fact that the chosen $\alpha = 0.4$ works well does not tell us whether other values (0.3, 0.5, 0.6) would work better, or whether the result is sensitive to this parameter. In a paper whose central methodological claim is that hybrid rewards are necessary, the absence of reward weight ablation is a significant gap.

Number of RL candidates (K = 8): The paper does not ablate the number of candidate scripts generated per GRPO step ($K = 8$). The choice of $K$ affects two things: the reliability of the advantage estimates (more candidates provide more stable group statistics) and the computational cost (which scales linearly with $K$). Without ablating $K$, we cannot assess whether the model's performance is sensitive to this hyperparameter or whether smaller values (e.g., $K=4$, which would halve the RL training cost) would achieve comparable results. Prior work on GRPO typically uses values of $K$ in the range 4–16, so $K=8$ is a conventional choice rather than a tuned one, but the paper does not discuss this trade-off.

Script conditioning without DirectorAgent mechanisms (Table 5): The "w/ Script Only" configuration in Table 5 isolates the effect of the script alone, without shot-aware segmentation or frame-anchoring. For Wan2.6 under Gemini-2.5-Pro, Script Only improves Body Language from 4.2 to 4.3 (+0.1) and Pace from 4.4 to 4.5 (+0.1), while Camera Articulation and Fidelity remain unchanged. The gains are modest but consistent, and they concentrate in dimensions that directly reflect script content (body language instructions, pacing specifications) rather than dimensions that reflect temporal execution (camera articulation, which depends more on segmentation and anchoring). This ablation confirms that the script provides value independent of the DirectorAgent, but also that the DirectorAgent mechanisms are necessary to achieve the full gains seen in the Full Agent configuration.

Choice of AI evaluator model (Tables 4 and 5): The paper evaluates video quality using three different AI critics (Gemini-2.5-Pro in the main Table 2, Qwen3-VL in Table 4, plus GLM-4.6-V in Table 5) as a form of evaluator robustness check. All three critics agree on the direction and rough magnitude of improvements from raw dialogue to script conditioning. The agreement across three independently developed models with different architectures provides some evidence that the improvements are genuine and not an artifact of a single evaluator's biases. However, all three are LLM-based evaluators, and they may share common failure modes — for instance, all three might be insensitive to certain types of visual artifacts, or all three might overweight text-script correspondence relative to visual quality. Without comparison to human judgments on the same instances (the paper's human evaluations are on different dimensions than the AI evaluations, making direct correlation impossible), the degree to which AI critic agreement reflects genuine quality versus shared evaluator biases remains uncertain.

Cross-scene generation without frame-anchoring (Table 5): The "w/ Script + Seg." configuration in Table 5 isolates the effect of segmentation without frame-anchoring. For Wan2.6 under Gemini-2.5-Pro, adding segmentation to script conditioning improves Camera Articulation from 4.2 to 4.3 (+0.1) and Body Language from 4.3 to 4.4 (+0.1), while Pace improves from 4.5 to 4.6 (+0.1). The further addition of frame-anchoring (Full Agent) raises Camera Articulation to 4.4 (+0.1), Body Language to 4.6 (+0.2), and Pace to 4.7 (+0.1). The incremental gain from frame-anchoring is consistently positive across all dimensions and both backbone models, confirming that the visual relay mechanism provides value beyond what segmentation alone achieves. The fact that the frame-anchoring benefit is larger for Body Language (+0.2) than for Camera Articulation (+0.1) for Wan2.6 is consistent with the mechanism's design: frame-anchoring primarily preserves character appearance and spatial layout, which directly affects body language consistency, while camera articulation is more dependent on shot-level scripting instructions.

Training data scale and model size ablations: The paper does not perform ablations on the amount of ScriptBench training data (all 1,700 instances are used for training without data scaling curves), the size of the base model (only Qwen-Omni-7B is tested), or the number of epochs for SFT or RL (only 20 epochs and 5,000 steps are reported without learning curve analysis). These missing ablations leave open questions: Would a larger base model (e.g., Qwen-Omni-14B or 72B) produce substantially better scripts, or is 7B sufficient given the high-quality training data? Would performance plateau with fewer than 1,700 training instances? Is the 20-epoch SFT duration near-optimal, or would earlier stopping prevent overfitting? For a paper that introduces both a new dataset and a new training methodology, these scale and data efficiency ablations would have strengthened the contribution by characterizing the resource requirements for achieving the reported performance.

RL reward model quality and annotation scale ablation: The human preference reward $R_{\text{human}}$ is trained on 500 expert-annotated samples. The paper does not report how preference model accuracy varies with annotation budget, nor does it ablate the number of annotated samples needed to achieve useful reward signal. Given that expert annotation (three senior art directors scoring scripts on four dimensions) is expensive and constitutes the primary bottleneck for scaling the approach, understanding the annotation efficiency of the preference model would be practically valuable. The paper also does not ablate the choice of BERT as the regression architecture — would a larger model, or a model pretrained on film-related text, produce better preference predictions and consequently better RL alignment?

Video model selection coverage: The paper evaluates seven SOTA video models, which provides broad coverage of the current landscape. However, all models are commercial or proprietary (no fully open-source models beyond Wan2.5/2.6 are included), and the paper does not report on models that might represent qualitatively different approaches (e.g., autoregressive transformer video models, or models specifically designed for long-form generation like SkyReels-V2 which is cited in related work). The consistency of results across seven diverse models is reassuring, but the selection is not systematic — it reflects API availability rather than a designed stratification across architectural families. This limits the strength of claims about "universal" improvement or about architectural determinants of the spectacle-faithfulness trade-off.

Negative result: the spectacle-faithfulness trade-off persists under script conditioning. While the paper frames this as an insight rather than a failure, it is functionally a negative result for the hypothesis that better prompting can reconcile visual quality and narrative control. Table 2 shows that Sora2-Pro with ScripterAgent still achieves substantially higher Visual Appeal (4.8) than HYVideo1.5 (4.5), while HYVideo1.5 still achieves substantially higher Script Faithfulness (4.6) than Sora2-Pro (4.2). The absolute gap between the two models on Script Faithfulness actually widens from 0.6 points in the raw dialogue condition (4.2 vs. 3.6) to 0.4 points in the script condition (4.6 vs. 4.2) — a narrowing in relative terms but still a large absolute difference. This suggests that the trade-off is not bridgeable by prompt engineering alone and likely reflects deeper architectural or training data differences between models. The paper does not attempt to characterize what specific model properties drive this trade-off, which would have been a valuable addition.

Critical Assessment

The paper's central claim is that a structured cinematic script — generated automatically by a purpose-trained model and used to condition video generation — substantially improves the narrative coherence, temporal fidelity, and script faithfulness of long-form video generated from dialogue. The experimental evidence supports this claim with qualifications regarding the scope of the improvement and the strength of the evaluation methodology.

Does the script consistently improve video generation? Yes, and this is the paper's strongest empirical result. Table 2 shows that conditioning seven different SOTA video models on ScripterAgent scripts improves both AI-rated and human-rated performance on every dimension for every model. The improvements are not marginal — Script Faithfulness rises by an average of +0.4 points on a 0–5 scale, with some models showing gains as large as +0.8 (Wan2.6: 3.2 → 4.0). The VSA metric (Table 3) confirms that these improvements extend to temporal-semantic alignment specifically, with consistent +1.8 to +2.4 point gains across all seven models. The ablation in Table 5 further decomposes the improvement into script effects (Body Language, Fidelity) and DirectorAgent effects (Camera Articulation, Pace), providing evidence that the full pipeline's gains are not attributable to a single component. This claim is the most robustly supported in the paper.

However, the claim that the improvement is "universal" must be qualified. The seven tested models, while diverse, are all commercial SOTA text-to-video systems from a similar generation of development. No open-source models, academic research models, or models from earlier generations are included. The 50-instance test set is small — each per-model, per-configuration mean is computed over only 50 videos. Without confidence intervals, we cannot assess whether a 0.1 or 0.2 point difference between models in a specific dimension is meaningful or noise. A failure case or outlier analysis (are there dialogues where the script hurts performance? where frame-anchoring introduces artifacts?) is entirely absent. The paper reports only positive results, which, combined with the small test set, limits confidence in the universality claim.

Does the RL alignment stage improve artistic quality beyond what SFT achieves? Yes, with strong evidence from Table 1. The SFT-only vs. Full model comparison shows +0.5 improvements in both Dramatic Tension (3.6 → 4.1) and Visual Imagery (3.8 → 4.3), while structural dimensions (Format Compliance) show minimal gains (+0.1). This is exactly the pattern the paper's methodology predicts — SFT handles structure, RL handles creativity — and the results cleanly support the claimed decomposition. The absence of $\alpha$ ablation (hybrid reward weight) and the absence of learning curves for the RL stage are gaps, but the core result — that RL provides substantial gains in subjective artistic dimensions while structural quality is maintained — is well-supported.

A missing analysis is whether the RL gains come primarily from the structural reward component (which, despite being weighted at only 40%, might still drive policy improvements if the SFT model has structural weaknesses the automated checks can detect) or from the human preference component. An ablation comparing RL with $\alpha = 1.0$ (pure structural reward) against $\alpha = 0.0$ (pure human preference) and $\alpha = 0.4$ (hybrid) would have directly tested the paper's claim that both signals are necessary and that the hybrid is optimal. Without this, we cannot rule out the possibility that the structural reward alone — which is free to compute and requires no expert annotation — would achieve similar artistic gains, or conversely that the human preference reward alone would maintain structural quality without the explicit structural checks.

Does the paper demonstrate a fundamental trade-off between visual spectacle and script faithfulness? Yes, and this is a genuine empirical finding rather than a methodological artifact. The pattern is visible across human ratings (Table 2), AI ratings (Table 2), and automated metrics (Table 3), and it persists under both raw dialogue and script conditioning. Sora2-Pro and HYVideo1.5 consistently occupy opposite ends of the spectacle-faithfulness spectrum, and the gap between them on key dimensions (e.g., Script Faithfulness: 4.2 vs. 4.6; Visual Appeal: 4.8 vs. 4.5) is large enough to be practically meaningful even without statistical tests. The finding is important because it characterizes a genuine limitation of the current generation of models — the fact that no single model simultaneously leads in both visual quality and narrative control suggests these capabilities are at least partially in tension, and that practitioners must choose based on application priorities.

A limitation of this finding is its correlational nature. The paper observes that different models have different strengths but does not investigate why. Are Sora2-Pro's visual quality advantages attributable to higher resolution generation, better training data, a different objective function, or more extensive post-processing? Is HYVideo1.5's instruction following advantage due to its dual-stream diffusion transformer architecture, its training data composition, or its prompt conditioning mechanism? Without characterizing the model properties that drive the trade-off, the finding is diagnostic (it tells us the trade-off exists) but not prescriptive (it doesn't tell us how to design models that avoid it).

Is the Visual-Script Alignment metric measuring something that standard metrics miss? The paper argues that VSA captures temporal-semantic fidelity while standard metrics measure only content presence. The evidence for this is indirect but consistent. Sora2-Pro, which leads in Visual Appeal, Aesthetic quality, and Dynamic Degree (all standard metrics), has substantially lower VSA (50.6) than HYVideo1.5 (54.8), which leads in Script Faithfulness and Narrative Coherence. The VSA metric is thus surfacing a quality dimension that correlates with human judgments of narrative adherence but is invisible in standard visual quality metrics. This supports the paper's methodological claim that existing evaluation toolkits are insufficient for narrative video assessment.

However, the paper does not validate VSA against ground-truth temporal annotations or against human judgments of temporal alignment specifically. The VSA formula assumes that (1) the script's time intervals $T_k$ are accurate, (2) CLIP embeddings capture the visual properties relevant to shot-level alignment, and (3) cosine similarity between frame embeddings and instruction embeddings is a valid proxy for whether the frame depicts the intended shot. None of these assumptions are tested. A video where CLIP similarity is high might still have temporally misaligned content if the shot descriptions are generic (e.g., "close-up of character" could match any close-up regardless of which character or when). Conversely, a video with correct temporal alignment might have low VSA if the visual style differs from CLIP's training distribution. The paper would benefit from a human validation study where annotators directly rate temporal alignment and their ratings are correlated with VSA scores.

What the experiments do not show. Several claims in the paper's narrative are not directly tested. The paper claims that ScripterAgent addresses the "semantic gap" between dialogue and cinematic execution, but the experiments only show that script-conditioned video generation is better than dialogue-conditioned video generation — they do not show that the gap is "bridged" in any absolute sense. The highest Script Faithfulness score (4.6 for HYVideo1.5 with ScripterAgent) is on a 0–5 scale, suggesting significant room for improvement. The paper claims that the framework "automates the pipeline from dialogue to final video," but human evaluation and selection are still required — the CriticAgent evaluates, but it does not automatically reject low-quality outputs or trigger regeneration. The paper claims "universal" improvement across video models, but all tested models are from a similar generation and capability tier; models with fundamentally different architectures or substantially lower quality might not show the same pattern. The paper claims the DirectorAgent "extends the coherence window of any underlying video model," but this is tested only on 15.4-second average videos — well within the generation windows of most current models — not on the 1–3 minute narrative scenes that motivated the mechanism's design. The frame-anchoring strategy's effectiveness at longer durations or with more segmentation boundaries is untested.

Missing experiments that would strengthen the paper. A direct comparison between the full agentic pipeline and a strong baseline that uses a generic LLM (e.g., GPT-4 or Gemini itself, prompted to generate a script without the ScriptBench training and expert alignment) would clarify whether the ScripterAgent's specialized training is necessary or whether off-the-shelf LLMs can already produce adequate cinematic plans. The qualitative comparison with MovieAgent in Figure 4 suggests ScripterAgent produces richer scripts, but MovieAgent is not evaluated as a video generation baseline — we cannot compare the downstream video quality of MovieAgent-guided generation against ScripterAgent-guided generation. A human evaluation of the frame-anchoring mechanism's failure modes (when does the last frame not provide sufficient information? do artifacts compound across segments?) would provide practical guidance absent from the current analysis. A scaling study varying the number of ScriptBench training instances and measuring downstream video quality would characterize the data efficiency of the approach and whether the 1,700-instance dataset is near the point of diminishing returns.

Overall assessment. The experiments support the paper's core claim that structured intermediate scripting improves narrative video generation from dialogue, with consistent evidence across multiple video models, evaluation dimensions, and critic models. The RL alignment stage's contribution to artistic quality is cleanly isolated through the SFT-only ablation. The spectacle-faithfulness trade-off is a genuine empirical finding with practical implications. The evaluation methodology has significant limitations — small test set, absence of statistical rigor, missing ablations on key hyperparameters — that prevent strong claims about the magnitude and generality of improvements, but the direction of the effect is consistently positive and the decomposition of gains across system components is informative.

6. Limitations and Trade-offs

The Difficulty Estimation Cost Is Unaccounted for in All Efficiency Claims

The entire compute-optimal framework depends on estimating prompt difficulty before deciding how to allocate the inference budget. The paper's method for doing so — sampling 2048 solutions per question and averaging either ground-truth correctness (oracle) or PRM final-answer scores (predicted) — is extraordinarily expensive. At 2048 samples per question, the difficulty estimation step alone consumes more compute than the largest test-time budgets studied (256–512 generations). The authors acknowledge this explicitly in Section 3.2:

"estimating difficulty in this way still incurs additional computation cost during inference... our experiments do not account for this cost largely for simplicity"

The consequence is that the reported 4× efficiency gains over best-of-N are computed after difficulty is known, without amortizing the cost of learning it. In a realistic deployment, the total cost would be difficulty estimation + strategy execution, and the former could dominate the latter — particularly for easy problems where the actual solution budget is small. The paper suggests future work on "pretraining or finetuning models to directly predict difficulty of a question," and the predicted difficulty bins (using PRM scores rather than ground-truth correctness) perform nearly as well as oracle bins — but the predicted-difficulty method still requires generating 2048 samples and scoring them. The exploration-exploitation tradeoff between difficulty estimation and problem-solving is flagged but left entirely unresolved. This is the single most consequential gap between the paper's reported results and what a deployable system would achieve.

Mitigation status: Not addressed. The paper explicitly acknowledges this as a limitation and delegates it to future work. No lightweight difficulty estimation method is proposed or evaluated. The efficiency gains must be understood as an upper bound contingent on solving the difficulty estimation problem.

---

Hard Problems Remain Fundamentally Unsolved — Test-Time Compute Cannot Substitute for Missing Capability

Across all methods studied — PRM beam search, best-of-N, iterative revisions, and their compute-optimal combinations — the hardest questions (difficulty bin 5) show near-zero improvement regardless of compute budget. This is visible in Figure 3 (right), where bin 5 accuracy hovers at 1–3% for all search methods across all budgets from 4 to 256 generations; in Figure 7 (right), where bin 5 shows roughly 2–3% accuracy irrespective of the sequential-to-parallel ratio at 128 generations; and in Figure 9, where the bin 5 scaling line is essentially flat near 0–5% even at the highest test-time budgets.

The paper is transparent about this in the Section 7 takeaway box, stating that pretraining is preferable when "the base model has essentially no chance of producing a correct answer for a given question using any reasonable amount of test-time compute." But the implication is significant: test-time compute can amplify existing capability but cannot create it. If the base model's pass@1 is near zero on a problem class, no amount of search, revision, or adaptive allocation will help — there are no correct solutions in the proposal distribution to find or refine. This means the framework offers no path forward for genuinely novel or out-of-distribution reasoning that exceeds the base model's training distribution. The capability boundary is set entirely by pretraining, and test-time compute operates only within that boundary.

The FLOPs-matched comparison in Figure 9 quantifies this sharply: on bin 5, the smaller model with compute-optimal test-time compute underperforms the 14× larger model by 37–53% across all R regimes, with the disadvantage growing as R increases. For applications where the problem distribution includes a non-trivial fraction of genuinely hard instances, the framework provides no benefit and the only viable path is scaling pretraining.

Mitigation status: Not mitigated. The paper acknowledges this as a fundamental boundary condition and does not claim to have solved it. The finding is presented as characterization rather than as a limitation to be fixed.

---

The Revision Model Has a 38% Correct-to-Incorrect Reversion Rate

Section 6.1 reports a significant practical problem: since the revision model was trained only on sequences where all in-context answers are incorrect (followed by a correct target answer), at test time it may encounter correct answers produced during earlier revisions and incorrectly "revise" them into wrong answers. The paper reports that "approximately 38% of correct answers get converted back to incorrect ones" using a naive approach where the final output of the revision chain is taken as the answer.

This is a direct consequence of the training data construction: by design, every training sequence ends with a transition from incorrect to correct, so the model has never seen examples of what to do when the current answer is already correct. The model learns to always produce a revision that differs from the previous attempt, but not to recognize when no revision is needed. The paper mitigates this with within-chain selection — using majority voting or verifier-based selection across the entire chain of revisions to pick the best answer from any point rather than always taking the last revision (Section 6.1, Figure 5). However, these selection mechanisms are post-hoc patches rather than solutions to the underlying training distribution problem. They reduce the impact of correct-to-incorrect reversions but do not prevent the model from wasting computation generating them.

A more principled solution — training the revision model to recognize when no revision is needed, or including "keep the current answer" as a valid revision action — is not explored. The ReST^EM experiment in Appendix K (Figure 16) further highlights the fragility of revision training: attempting to optimize the revision model with on-policy RL training caused performance to degrade substantially with sequential revisions, likely because it amplified spurious correlations in the revision trajectories.

Mitigation status: Partially addressed via within-chain selection, but the underlying training distribution problem remains. The paper does not propose a revision-model training procedure that would prevent reversions at the source.

---

The Experiments Are Limited to a Single Benchmark and Single Model Family

All experiments use the MATH benchmark (500 test questions) with PaLM 2-S* as the base model. Section 4 states that the authors "believe this model is representative of the capabilities of many contemporary LLMs," but this claim is entirely unverified. Several aspects of the findings could be model-specific or benchmark-specific in ways that affect the paper's central claims:

• The PRM's quality and over-optimization behavior depend on PaLM 2-S*'s output distribution. A model with different calibration properties, different error patterns, or different base accuracy on MATH could exhibit different difficulty-dependent scaling curves — potentially changing which search strategy is optimal for which difficulty bin.
• The revision model's ability to learn from incorrect in-context examples depends on the base model's in-context learning capabilities, which vary substantially across model families.
• MATH consists exclusively of competition-level math problems requiring symbolic reasoning. It is unclear whether the difficulty-dependent patterns (beam search hurting easy problems, revisions helping easy problems, search helping medium problems) generalize to other reasoning domains (code generation, logical reasoning, scientific QA) or to tasks requiring factual recall rather than multi-step inference.

The test set of 500 questions, split into five difficulty quintiles of approximately 100 each, then further split by two-fold cross-validation, means the compute-optimal policy is selected based on approximately 50 questions per fold per bin. The paper does not report confidence intervals on the compute-optimal scaling curves or on the difficulty-bin breakdowns, making it difficult to assess whether the observed patterns are statistically reliable at these sample sizes. The 4× efficiency claim and the difficulty-dependent strategy recommendations (use beam search on medium problems, revisions on easy problems) may not hold with a different base model or a different benchmark.

Mitigation status: Not addressed. The paper explicitly limits its claims to the studied model and benchmark, and Section 8 acknowledges the need for evaluation on additional benchmarks and models as future work. However, the framing of the findings (e.g., "test-time compute can outperform a 14× larger model") is presented as a general claim rather than a model-specific observation.

---

Sequential Revisions Introduce an Unacknowledged Latency-Versus-Throughput Trade-off

The paper measures test-time compute in "generations" (number of complete solutions sampled), which is a reasonable proxy for total FLOPs but entirely ignores wall-clock latency. Sequential revisions are inherently serial — each revision depends on the output of the previous one, so a chain of 64 revisions takes at least 64 sequential forward passes through the model. Parallel best-of-N sampling, in contrast, can be executed simultaneously with sufficient hardware — generating 64 independent solutions in a single batch, requiring approximately the wall-clock time of one generation plus batching overhead.

The compute-optimal policy frequently favors sequential-heavy strategies, particularly on easy problems where fully sequential revisions are optimal (Figure 7, right, bin 1–2). For latency-sensitive applications — interactive assistants, real-time tutoring, any deployment where users wait for responses — the sequential strategies favored by the compute-optimal policy on easy problems may be completely impractical regardless of their accuracy advantages, because they impose a latency multiplier proportional to the chain length. A strategy that allocates 128 generations as 64 sequential × 2 parallel (which might be optimal for a medium-difficulty problem) takes approximately 64× longer wall-clock time than a strategy that runs 128 parallel samples simultaneously.

The paper's compute-optimal framework treats all generations as fungible units of FLOPs, but from a deployment perspective, sequential and parallel generations are not fungible — they have dramatically different latency implications. The optimal strategy under a combined FLOPs + latency budget could be substantially different from the optimal strategy under a FLOPs-only budget, but the paper does not model latency at all.

Mitigation status: Not addressed. The paper does not discuss latency, wall-clock time, or the practical constraints of deploying sequential revision chains in latency-sensitive settings. This is acknowledged implicitly by the absence of any latency-related analysis in the compute-optimal framework.

---

The 14× Larger Model Baseline Is Not Compute-Optimally Trained

The FLOPs-matched comparison in Section 7 scales model parameters while holding training data fixed, following the LLaMA paradigm (Touvron et al., 2023). The authors explicitly acknowledge in Section 7 that this departs from compute-optimal pretraining (Hoffmann et al., 2022), where both data and parameters would be scaled equally:

"We choose this setting as it is representative of a canonical approach to scaling pretraining compute and leave the analysis of compute-optimal scaling of pretraining compute where the data and parameters are both scaled equally to future work."

This matters because a Chinchilla-optimal model trained with 14× more total FLOPs (with both parameters and data scaled appropriately) would likely outperform a parameter-only-scaled model. The pretraining baseline in the FLOPs-matched comparison is therefore weaker than it needs to be for a fair comparison. The reported advantages of test-time compute over the larger model — e.g., +27.8% relative improvement on medium-difficulty questions at R << 1 with revisions (Figure 1, Figure 9) — may shrink or reverse against a properly compute-optimal larger model.

Additionally, the 14× larger model uses only greedy decoding in the FLOPs-matched comparison — no majority voting, no best-of-N, no search. Giving the larger model even a modest test-time compute budget (say, best-of-8) would create a substantially stronger baseline that represents a more realistic deployment scenario. In practice, if an organization has already invested in training a 14× larger model, they would almost certainly pair it with at least basic test-time compute strategies. The current comparison tests whether test-time compute with a small model beats a large model with zero test-time compute — a setting that favors test-time compute by construction.

Mitigation status: Acknowledged but not addressed. The paper explicitly identifies this as a limitation and delegates the comparison against compute-optimally trained larger models to future work. The "canonical" LLaMA-style scaling is defended as representative of common practice, but this defense does not make the comparison fair — it makes it favorable to test-time compute.

7. Implications and Future Directions

How This Work Changes the Landscape

This paper introduces a systems-level reframing rather than a fundamental architectural breakthrough. The core conceptual move — that a structured, machine-readable cinematic script generated by a purpose-trained model can serve as the essential intermediate representation between sparse dialogue and coherent video — does not change how video generation models work at the pixel level. It changes what we ask them to do, and in doing so, it exposes that the bottleneck in automated filmmaking is not visual quality but narrative planning.

The shift is from treating video generation as a direct mapping problem (dialogue → video) to treating it as a two-stage orchestration problem (dialogue → structured plan → video). This is analogous to the shift in language model prompting from zero-shot answers to chain-of-thought reasoning — the insight is not that models need to be bigger or trained differently, but that they need intermediate structure that decomposes a complex creative task into executable sub-tasks. The paper's evidence that conditioning seven different SOTA video models on the same script universally improves output confirms that the orchestration layer, not the generation layer, is the binding constraint on narrative coherence in current systems.

This reframing has several specific consequences for how the field should think about automated video generation:

The creative bottleneck is in planning, not rendering. Before this paper, the dominant narrative in video generation research was that models needed better visual quality — higher resolution, more realistic physics, smoother motion. The paper's results (Table 2) show that even the most visually impressive models (Sora2-Pro, Visual Appeal 4.8) produce videos with weak Script Faithfulness (4.2) and Character Consistency (4.3) when fed raw dialogue. The 0.5-point average improvement in human ratings from adding the script (3.7 → 4.2) is larger than the gap between the best and worst video models in the raw dialogue condition (HYVideo1.5 at 4.0 vs. Vidu2 at 3.4 = 0.6). This suggests that improving the planning layer yields gains comparable to switching between fundamentally different video generation architectures. For researchers allocating effort, this implies that advances in automated planning — better script generation, more sophisticated cinematographic reasoning — may be higher-leverage than marginal improvements in visual fidelity.

The spectacle-faithfulness trade-off is now named and measured. The finding that no single video model simultaneously leads in both visual spectacle (Sora2-Pro: Dynamic Degree 79.5, Aesthetic 62.8) and script adherence (HYVideo1.5: VSA 54.8, Script Faithfulness 4.6) is more than a model comparison — it is a diagnostic of the current capability frontier. Previous work could observe anecdotally that some models produce prettier videos while others follow prompts more faithfully, but this paper provides quantitative evidence across human ratings, AI ratings, and automated metrics that these capabilities are at least partially in tension. This changes the evaluation landscape: future video models should be assessed on both axes simultaneously, and the field should track whether the frontier expands outward (both dimensions improving) or whether progress on one dimension continues to come at the expense of the other. The paper's finding that the trade-off persists under script conditioning — Sora2-Pro still leads in Visual Appeal while HYVideo1.5 leads in Script Faithfulness even when both receive the same structured scripts — suggests the tension is not an artifact of prompting but a deeper property of model architectures that the next generation of models will need to address.

Automated filmmaking becomes an agentic orchestration discipline. The paper's three-agent architecture (ScripterAgent for planning, DirectorAgent for execution, CriticAgent for evaluation) establishes a template for how automated creative pipelines should be structured: specialized agents handle distinct phases of the creative process, each with its own training methodology and evaluation criteria. This is a departure from the end-to-end deep learning paradigm that has dominated video generation. The paper demonstrates that different stages of the pipeline benefit from fundamentally different training signals — SFT for structural competence, GRPO with hybrid reward for creative alignment, rule-based constraints for shot segmentation, frame-conditioning for visual continuity — and that composing these specialized components yields better results than any monolithic approach. This agentic decomposition is likely extensible to other complex generative tasks (automated music production, architectural design from client briefs, UI generation from functional specifications) where the creative process naturally decomposes into planning and execution phases.

The Visual-Script Alignment metric fills a methodological gap. Prior evaluation toolkits for video generation measured whether content appeared (CLIP Score, VBench consistency) but not whether it appeared at the script-specified time. VSA's consistent 1.8–2.4 point improvements across all seven models when conditioned on ScripterAgent scripts (Table 3) validates that the metric captures a dimension of quality that existing tools miss. More importantly, VSA reveals information that aggregate metrics obscure — Sora2-Pro's high CLIP Score (46.0) and Aesthetic score (62.8) mask its relatively low temporal alignment (VSA 50.6), while HYVideo1.5's lower spectacle scores coexist with the highest VSA (54.8). This diagnostic capability makes VSA a practical tool for identifying specific failure modes (temporal misalignment) that would otherwise require labor-intensive human inspection. The metric's clean formulation — it requires only a script with time intervals, a video, and off-the-shelf CLIP encoders — lowers the barrier to adoption for any video generation task where temporal-semantic alignment matters.

The two-stage SFT+RL training paradigm validates a decomposition of quality into correctness and creativity. The paper's finding that SFT handles structural competence (Format Compliance 3.9) while RL alignment drives artistic improvement (Dramatic Tension +0.5, Visual Imagery +0.5) provides an empirical template for training creative AI systems where the output must be both technically valid and aesthetically compelling. This decomposition — correctness from supervised learning on demonstrations, quality from preference optimization against expert judgment — is likely applicable to any creative domain where the space of valid outputs is large (one-to-many mapping) and quality is comparative rather than absolute. The specific instantiation with GRPO and hybrid reward (α = 0.4 structural, 0.6 human preference) provides a starting point, though the paper's lack of α ablation means the optimal weighting for other domains remains an open empirical question.

The paper also reconciles a latent tension in the literature between advocates of end-to-end generation and advocates of structured planning. Systems like Sora and Veo demonstrate that end-to-end models can produce stunning visuals from simple prompts, which might suggest that planning is unnecessary — just make the models bigger and better at following instructions. This paper shows that even the best current models, given the exact same dialogue as input, produce videos with weak narrative coherence (average human-rated Narrative Coherence 3.7, Script Faithfulness 3.8). The script-centric approach does not replace these models — it augments them, using them as rendering engines for a planning layer that they lack. The reconciliation is that visual generation and narrative planning are distinct capabilities, and the former does not automatically imply the latter. Progress on both fronts is necessary for automated filmmaking.

Follow-Up Research This Work Enables

Ablating the hybrid reward weight α to characterize the correctness-creativity trade-off surface. The paper fixes α = 0.4 (40% structural reward, 60% human preference reward) without reporting ablations. A systematic sweep of α from 0.0 (pure human preference) to 1.0 (pure structural reward) would characterize the trade-off surface: at what α value does structural quality begin to degrade meaningfully? At what α value do artistic gains saturate? Is α = 0.4 near-optimal, or would a different weighting produce substantially better scripts? This ablation would also test the paper's central methodological claim — that both reward components are necessary and that the hybrid is superior to either alone. If pure human preference (α = 0.0) produces scripts with comparable structural quality to the hybrid, then the structural reward component is unnecessary and the complexity of the hybrid design is unjustified. If pure structural reward (α = 1.0) produces scripts with comparable artistic quality, then the expensive human annotation pipeline for R_human is unnecessary. Either negative result would refine our understanding of ScripterAgent's training requirements.

Training a lightweight difficulty estimator to close the exploration-exploitation gap. The paper's compute-optimal framework for test-time compute in the reference example suffered from the same problem this paper shares in a different form: ScripterAgent's script generation is a planning step whose cost must be amortized over downstream video generation. The current pipeline treats script generation as a fixed preprocessing cost, but in a deployable system where many similar dialogue scenarios recur, the question becomes: can we predict when ScripterAgent's script will most improve video quality, and when a simpler, cheaper planning approach would suffice? Training a lightweight classifier — perhaps a small BERT model or even a logistic regression on dialogue features — to predict the expected Script Faithfulness gain from ScripterAgent conditioning versus a baseline LLM script would enable adaptive routing: use ScripterAgent for dialogues where the gain is predicted to be large, and a cheaper script generator (or even raw dialogue) where the gain is negligible. This directly parallels the difficulty estimation problem flagged in the reference paper and would make the framework's practical deployment more cost-effective. The training signal exists in the paper's data: for each test dialogue, we can compute the per-model Script Faithfulness delta between raw dialogue and script conditioning, and train a predictor on dialogue-level features.

Combining ScripterAgent-style planning with the video models' internal representations for closed-loop generation. The current pipeline is strictly feedforward: ScripterAgent generates a script, DirectorAgent segments and anchors it, and the video model renders each segment independently. A natural extension is closed-loop generation where the video model's intermediate outputs influence the script execution. For example, if the video model generates a segment where a character's expression differs from the script's emotional specification, a feedback signal could trigger script adaptation — either regenerating the offending segment with a stronger emotional prompt, or adjusting subsequent script instructions to accommodate the deviation. This would require the video model to expose some representation of generated content (e.g., CLIP embeddings of output frames) that a monitor agent can compare against script specifications, creating an error-correction loop analogous to the multi-round adaptive error correction used in ScriptBench's annotation pipeline (Stage 3). The paper already has the evaluation infrastructure for this — CriticAgent's Emotional Arc & Micro-Expressions dimension (Table 2) directly measures whether the video reflects the script's intended emotional progression, and this signal could be repurposed from evaluation to feedback.

Stress-testing the frame-anchoring mechanism at longer durations and with more complex scene transitions. The paper's videos average 15.4 seconds, which fits within a single generation window for most current models. The DirectorAgent's frame-anchoring mechanism is motivated by the need to handle 1–3 minute narrative scenes (Section 3), but the mechanism is never tested at durations requiring more than 2–3 generation segments. A systematic stress test would generate videos of increasing duration — 30 seconds, 60 seconds, 120 seconds — and measure how Character Consistency, Background Consistency, and Narrative Coherence degrade as the number of frame-anchored transitions increases. Specific failure modes to track: does identity drift compound across segments (each generation slightly altering the character's appearance until the final segment is unrecognizable)? Do lighting or color balance shifts accumulate? Does the "Continuing from the previous scene" text prompt lose effectiveness after many repetitions? The VBench metrics (Subject Consistency, Background Consistency) and the VSA metric could quantify degradation as a function of segment count. This would establish the practical coherence ceiling of the frame-anchoring approach and identify whether improvements are needed in the anchoring mechanism itself or in the underlying video models' consistency capabilities.

Training a ScripterAgent variant on diverse cinematic styles to test generalization of cinematographic reasoning. The paper's ScriptBench is curated from "high-fidelity cinematic cutscenes" with a consistent visual style (professional cinematography, game-engine-quality rendering). An open question is whether ScripterAgent's learned cinematographic reasoning — when to use close-ups versus wide shots, how to pace cuts, how to block characters — generalizes to different cinematic traditions (e.g., documentary style, handheld footage, anime, theatrical staging). Training ScripterAgent variants on style-specific script datasets would test whether the model learns general cinematographic principles or style-specific conventions. If a single model trained on mixed styles can generate appropriate scripts for each style when given a style specification, the cinematographic reasoning is genuinely general. If style-specific models substantially outperform a mixed model, the reasoning is style-bound. This has practical implications for whether one ScripterAgent can serve diverse filmmaking applications or whether style-specific agents are needed. The paper's SFT+RL pipeline is modular enough that this experiment requires only new training data — the architecture and training methodology remain unchanged.

Validating VSA against human temporal-alignment judgments with controlled stimuli. The VSA metric's validity depends on assumptions that are untested in the paper: that CLIP embeddings capture the visual properties relevant to shot-level alignment, and that cosine similarity between frame embeddings and instruction embeddings correlates with human perception of temporal correctness. A validation study would construct videos with controlled temporal misalignments — e.g., shifting a reaction shot 2 seconds earlier or later than scripted, swapping the order of two shots, inserting an unrelated shot at a script boundary — and collect human judgments of temporal alignment on a continuous or ordinal scale. Correlating these judgments with VSA scores would establish the metric's sensitivity (does it detect known misalignments?) and specificity (does it give high scores to correctly aligned videos?). If VSA fails to detect certain types of misalignments that humans readily notice (e.g., a shot that is visually similar to the script instruction but semantically wrong — a close-up of the wrong character), the metric would need refinement. If VSA flags misalignments that humans consider acceptable (e.g., minor timing variations of a few frames), the metric might be overly strict. Either outcome would calibrate VSA's practical utility and guide improvements.

Practical Applications and Downstream Use Cases

Automated previsualization for film and game production. Previsualization ("previs") is the process of creating rough animated versions of scenes before filming, used extensively in VFX-heavy productions to plan camera moves, blocking, and shot timing. Currently, previs requires skilled artists manually animating storyboards — a labor-intensive process that can cost thousands of dollars per minute of footage. The ScripterAgent + DirectorAgent pipeline could reduce this to dialogue input: a screenwriter or director provides scene dialogue, ScripterAgent generates a shot plan with camera specifications and blocking, and DirectorAgent produces a rough video previs using a text-to-video model. The paper's results suggest the output would have reasonable Script Faithfulness (average 4.2 across seven models with script conditioning, Table 2) and Visual Appeal (average 4.3), sufficient for previs purposes where the goal is planning rather than final rendering. The 50-instance test set and 15.4-second average duration are in the right range for individual scene previs. The key practical advantage is iteration speed: a director could modify dialogue or specify different emotional tones and receive updated previs in minutes rather than days.

Dialogue-driven content creation for interactive storytelling and video games. Many modern video games feature branching dialogue systems where player choices affect story outcomes, particularly in RPGs and narrative adventure games. Producing cinematic cutscenes for every dialogue branch is currently infeasible — the combinatorial explosion of possible conversation paths means most dialogue is delivered with static character models and limited camera work. ScripterAgent could generate shot plans dynamically based on the current dialogue state, and DirectorAgent could render them in real-time using game engine assets rather than text-to-video models (since the characters, environments, and lighting already exist in-engine). The script's structured JSON format — specifying camera positions, shot types, character blocking, and timing — is directly translatable to game engine commands (e.g., Unity's Cinemachine or Unreal's Sequencer). The paper's finding that the script alone (without DirectorAgent segmentation) already improves Body Language and Visual Descriptive Fidelity (Table 5, "w/ Script Only") suggests that even without frame-anchoring, the shot-level instructions would enhance static dialogue scenes. The 1,700-instance training set could be augmented with game-specific cinematic conventions to adapt ScripterAgent to a particular title's visual style.

Batch processing of dialogue-heavy content for accessibility and localization. Large libraries of dialogue-driven content — audiobooks, podcasts, recorded interviews, theatrical plays — have no visual component, making them inaccessible to deaf and hard-of-hearing audiences beyond text transcription. A deployment of ScripterAgent + DirectorAgent could process this content in batch, generating sign-language-compatible visualizations or animated dialogue scenes that convey emotional subtext and speaker identity through cinematography rather than audio. The key metric from the paper supporting this use case is the consistent improvement in Emotional Arc & Micro-Expressions (AI-rated, Table 2: average 4.3 → 4.5 with script conditioning) and Kinetic Body Language & Blocking (average 4.0 → 4.5), which suggests the generated scripts produce videos where non-verbal emotional information is conveyed visually. The 50-instance test set size is small for production deployment, but the pipeline's automation makes scaling to thousands of hours of content a matter of compute cost rather than human labor — the bottleneck is video generation API cost and throughput, not script generation. The paper's CriticAgent could be repurposed as a quality filter, automatically flagging videos where Emotional Arc or Narrative Coherence fall below a threshold for human review.

When to Prefer This Method

• Prefer the script-centric agentic pipeline when the input is dialogue or other sparse narrative text where the visual plan is severely underspecified, and the downstream video model supports text-to-video generation with frame conditioning. The paper's evidence (Table 2: average human rating improvement of +0.5 across seven models) supports this preference specifically when Script Faithfulness, Character Consistency, or Narrative Coherence are the primary quality criteria.

• Prefer raw dialogue conditioning (bypass ScripterAgent) when the dialogue is short enough that a single generation window suffices (under ~12 seconds for current models) and the priority is visual spectacle over narrative precision. The paper's results show that Visual Appeal scores are already high in the raw dialogue condition (average 3.8) and improve only modestly with script conditioning (average 4.3), while the cost of script generation is non-zero. For applications where visual impact is paramount and the dialogue is simple, the additional planning step provides marginal benefit.

• Prefer a larger pretrained video model without the orchestration layer when the content requires visual capabilities beyond what structured prompting can provide — particularly complex physics interactions, highly dynamic motion, or photorealistic rendering of specific materials. The paper's finding that Sora2-Pro leads in Physical Law Adherence (4.5) and Dynamic Degree (79.5) even without script conditioning, and that these advantages persist under script conditioning (the spectacle-faithfulness trade-off), suggests that the orchestration layer does not substitute for model capability on purely visual dimensions. If physical realism is the primary requirement, invest in the generation model rather than the planning layer.