PaperlayerView exact revision on arXiv

Source-linked AI summary

DataMagic: Authoring Data Videos through Declarative Multi-Agent Orchestration

Yupeng Xie, Zhenyang Wang, Liangwei Wang, Jiayi Zhu, Zhouan Shen, Yuyu Luo

arXiv:2609.33403v1cs.HCcs.AIcs.DBcs.MA

TL;DR

Data videos are costly to produce, and existing approaches do not jointly provide data fidelity, narrative coherence, and end-to-end automation. DataMagic combines DVSpec with multi-agent scene generation and orchestration; evaluations report improved video quality and efficiency compared with the described baselines.

  • Problem

    Producing data videos requires expertise in data analysis, narrative design, and video editing, while existing approaches struggle to combine data fidelity, narrative coherence, and end-to-end automation.

    Source: §1 · p. 1

  • Method

    DataMagic uses DVSpec to bind charts, narration, and animations in a shared declarative representation and a Generate-then-Orchestrate strategy to generate and globally organize scenes.

    Source: §1 · pp. 1–2

  • Results

    DataMagic improves video quality and execution stability over direct-generation baselines; on 109 samples, quality is 3.89 (+83%) versus GPT-5’s 2.13/5.

    Source: §1 · p. 1, §9 · p. 9

  • Takeaways & Limitations

    The results suggest declarative representations can coordinate automated generation, data traceability, and fine-grained human control in structured authoring workflows.

    Source: §9 · p. 9

    DataMagic currently targets single-table inputs; support for multitable joins and richer relational data would broaden its applicability.

    Source: §8 · p. 9

Abstract

from arXiv · show

Data videos communicate data insights through dynamic charts, voice narration, and synchronized animations, and have become a widely adopted form of data storytelling. However, producing them requires expertise in data analysis, narrative design, and video editing. Static visualization tools lack narrative and animation capabilities; authoring tools rely on pre-prepared charts rather than raw data; and pixel-level models generate videos end-to-end but cannot guarantee data accuracy or provenance. End-to-end automatic generation faces two core challenges: how to uniformly represent charts, narration, and animations together with their temporal relationships, and how to efficiently search a vast design space for narrative-coherent compositions. We present DataMagic, which authors data videos from raw tabular data through declarative multi-agent orchestration. First, the declarative specification DVSpec unifies charts, narration, and animations with data-bound references and declarative synchronization, ensuring data provenance and automatic audio-visual alignment. Second, a "Generate-then-Orchestrate" multi-agent strategy generates candidate scenes in parallel and then optimizes narrative coherence through global orchestration. DVSpec provides a shared state for three complementary interaction modes, bridging full automation with fine-grained human control. Evaluations on 109 real-world samples show that even the most advanced LLM (e.g., GPT-5) achieves only 2.13/5 with execution success rates between 48.62% and 86.24%; DataMagic improves quality to 3.89 (+83%) with success rates above 95%, with the most significant gains in animation and narrative dimensions. A user study shows that, compared to a conversational LLM workflow, DataMagic improves creation efficiency (79.7% reduction in task time) and reduces perceived cognitive load. Project page: https://github.com/HKUSTDial/DataMagic.

1 INTRODUCTION

Data videos are costly to produce, and existing approaches do not jointly provide data fidelity, narrative coherence, and end-to-end automation. DataMagic combines a declarative video specification with multi-agent orchestration; evaluations report higher quality and execution success than LLM baselines, alongside faster authoring and lower perceived cognitive load.

1 INTRODUCTION

  • Existing tools automate charts, add narration and animation to pre-made visualizations, or generate videos end-to-end, but struggle to combine fidelity, coherence, and automation.Static tools lack narrative and animation; authoring tools start from prepared visualizations; pixel-level models cannot trace visuals to data records.Source: §1 · p. 1
  • The authors frame generation as hierarchical orchestration across macro narrative structure, scene content, and animation details, with challenges in unified representation and global search.Scenes center on distinct analytical insights and may follow logic such as macro-to-micro or phenomenon-before-cause.Source: §1 · p. 1
  • DataMagic uses DVSpec to bind visual and animation elements to data fields and trigger animations from narration, while agents generate candidate scenes in parallel for global selection and ordering.Context-aware narration supports coherence; the shared editable state also supports canvas, script, and natural-language interaction.Source: §1 · pp. 1–2
  • 3.89 (+83%) video quality, with execution success rates above 95%, compared with GPT-5’s 2.13/5 and LLM success rates of 48.62%–86.24%.The evaluation covers 109 real-world samples.Source: §1 · p. 1
  • 79.7% reduction in task time versus a conversational LLM workflow, alongside reduced perceived cognitive load.The study also reports improved creation efficiency.Source: §1 · pp. 1–2

2 RELATED WORK

Related work includes human-in-the-loop authoring and automated generation for data videos, automated data storytelling, and declarative visualization and animation specifications. DataMagic’s stated distinction is to start from raw tabular data and coordinate multi-scene video content in a shared declarative representation.

2.1 Data Video Authoring

  • Human-in-the-loop tools support clip composition, narration-linked authoring, and chart-animation design, while automated systems generate videos from infographics, datasets, visualizations, or prepared narratives.Examples include DataClips, WonderFlow, AutoClips, Data Player, and Narrative Player; the approaches vary in their inputs and design focus.Source: §2.1 · p. 2

2.2 Automated Data Storytelling

  • Automated storytelling has used fact extraction and search to compose narratives, while LLM-based systems generate narrative and visualizations; the cited systems produce static outputs rather than motion-and-voice videos.Examples include DataShot, Calliope, DataNarrative, and InReAcTable.Source: §2.2 · p. 2

2.3 Declarative Visualization and Animation Specifications

  • Existing declarative specifications describe individual charts and animations but do not cover multi-scene data videos integrating charts, narration, and synchronized animation.DVSPEC adds data-driven semantic references, narration-indexed triggers, and scene-level organization.Source: §2.3 · p. 2

3 DESIGN REQUIREMENTS

The design requirements address recurring workflow difficulties in cross-modal references, synchronization, scene structure, and global narrative orchestration. They specify modular scenes, data-bound declarative synchronization, global scene planning, and multimodal human refinement.

3 DESIGN REQUIREMENTS

  • Scenes should be independently generated, rendered, and edited as semantic units, supporting modular production and flexible narrative composition.The requirement responds to the opacity and costly whole-video regeneration associated with pixel-level generation methods.Source: §3 · p. 2
  • A unified representation should bind visual elements to data attributes rather than fragile rendering identifiers and defer temporal alignment to declarative triggers instead of manual timestamps.This is intended to preserve traceability and references across data, chart, or narration changes while reducing rework.Source: §3 · pp. 2–3
  • Global orchestration should select scenes by insight value and query coverage, plan their playback order, and generate context-aware narration to support coherent transitions.The requirement addresses redundancy and discontinuities that can arise from greedy scene-by-scene generation.Source: §3 · p. 3
  • A shared representation should support direct visual manipulation, structured-script editing, and natural-language commands so users can switch interaction modes without losing context.The requirement is framed as a way to bridge full automation and fine-grained human control.Source: §3 · p. 3

4 DVSPEC: A DECLARATIVE SPECIFICATION FOR DATA VIDEOS

DVSPEC provides a declarative, scene-based representation for data videos, with data-bound visual references and narration-indexed animation triggers. Its organization supports automated generation and interactive editing while keeping cross-modal elements traceable and synchronized.

4 DVSPEC: A DECLARATIVE SPECIFICATION FOR DATA VIDEOS

  • DVSPEC unifies cross-modal content and temporal coordination in a declarative intermediate representation for data videos.It organizes the video as scenes, with semantic references connecting visuals and animations to data and narration indices defining trigger timing.Source: §4 · p. 3

4.1 Task Definition

  • Data video generation maps a dataset and user query into a coherent audiovisual narrative composed of ordered, self-contained scenes.Each scene contains an identifier, type, visualization content, narration, and animation.Source: §4.1 · p. 3

4.2 Scene-Driven Organization

  • DVSPEC structures a video as global metadata plus an ordered scene sequence, enabling varied scene roles and independent rendering.Scene-level style and layout allow role-specific organization while supporting a coherent visual identity across the video.Source: §4.2 · p. 3

4.3 Data-Driven Semantic Referencing

  • Data-keyed semantic references bind animations to the intended chart elements, avoiding fragile rendering-layer identifiers.At render time, matching data records resolve to chart nodes, preserving the reference across changes in data ordering or chart type as long as the data structure is preserved.Source: §4.3 · p. 4

4.4 Narration-Indexed Declarative Triggering

  • Narration-indexed triggers let the renderer compute animation timing from the generated audio, so narration revisions need no manual timestamp adjustments.The trigger expresses which narration segment fires an animation, separating logical intent from the renderer’s precise timing calculation.Source: §4.4 · p. 4

5 SYSTEM FRAMEWORK

DataMagic separates accurate scene-content generation from global narrative orchestration to address the coherence and search challenges of data-video generation. Its framework generates candidate scenes in parallel, then selects, orders, narrates, and synchronizes them, with interactive refinement through a shared DVSPEC state.

5 SYSTEM FRAMEWORK

  • Generate-then-Orchestrate separates per-scene content generation from global sequence planning to balance scene accuracy with narrative coherence.The framework first generates diverse candidates in parallel, then globally selects and orders scenes while aligning narration and animation.Source: §5 · p. 4

5.1 Candidate Scene Generation

  • Candidate generation decomposes the query into independent analytical subtasks, extracts a targeted micro-dataset for each, and designs a matching visualization.Data extraction is separated from visualization design for independent verification; narration and animation are deferred to orchestration.Source: §5.1 · p. 4, Input · p. 5

5.2 Global Narrative Orchestration

  • Global orchestration selects scenes for insight value and query coverage, orders them into a narrative pattern, and generates context-aware narration and synchronized animations.A sliding window supplies adjacent-scene context for narration; animation targets are bound to mentioned data entities and triggered by narration indices.Source: §5.2 · p. 5

5.3 Interactive System

  • The interactive system applies scoped changes to shared DVSPEC state, while supporting canvas edits, data-driven Q&A, and scene generation from surfaced insights.Within-scene changes can be rerendered locally without rerunning generation; Q&A executes structured queries against underlying tabular data and returns a reasoning trace.Source: §5.3 · p. 6

6 EXPERIMENTS

The experiments compare DataMagic with direct-generation LLMs on real-world data-video samples, then examine module ablations and direct-generation error patterns. DataMagic improves quality and execution stability, while the user study is reported separately.

6.1 Experimental Setup

  • The evaluation used 109 samples from 60 single-table datasets, comparing four direct-generation LLMs with model-matched DataMagic variants using execution rate and five 1–5 quality dimensions.Automated scores correlated with expert ratings at Pearson r = 0.91 (p < 0.001; MAE = 0.38).Source: §6.1 · pp. 6–7

6.2 End-to-End Video Generation Performance

  • DataMagic raised average quality from 1.91–2.22 to 3.38–3.89, with animation increasing from 1.84 to 4.03 (+120%) and narrative from 2.00 to 3.43 (+72%).Table 1 compares end-to-end video-generation performance.Source: Table 1 · p. 7, §6.2 · p. 7

6.3 Ablation Study

  • Removing the Story Planner lowered average score from 3.89 to 3.44 (11.6% decrease); removing Orchestration lowered it to 3.54 (9.0% decrease).Without Orchestration, Intent fell 12.4%; Animation remained 3.95, only 10.0% lower, with narration-indexed triggering intact.Source: Table 2 · p. 7, §6.3 · p. 7

6.2 End-to-End Video Generation Performance

  • Execution rates exceeded 95% versus 48.62%–86.24%; DataMagic took approximately 176 seconds versus 57 seconds for direct generation at parallelism 3.The authors describe this longer runtime as a trade-off for higher quality and robustness.Source: §6.2 · p. 7

6.4 Error Pattern Analysis of Direct Generation

  • Direct-generation failures span rendering breakdowns, visualization defects, and audio-visual misalignment, reflecting cross-modal coordination demands beyond a single LLM inference pass.Examples include code errors that prevent rendering, cropped or overlapping elements, static-image-like animation, and narration that does not match the displayed scene.Source: §6.4 · pp. 7–8

7 USER STUDY

In a within-subjects study, participants compared DataMagic with a conversational LLM workflow for data-video creation. DataMagic reduced task time and perceived workload, while interviews described the value of synchronization and scene-level editing.

7.1 Study Design

  • Twelve participants completed counterbalanced authoring tasks in DataMagic and a same-model conversational baseline, covering attribution, evolution, and comparison tasks.Each task included the full creation workflow and at least one editing operation.Source: §7.1 · p. 8

7.2 Quantitative Results

  • Mean per-item usability was 5.97 (SD = 0.35) with DataMagic versus 3.22 (SD = 0.91) with the baseline (p < .001).Ratings used a 1–7 scale, with higher values indicating better perceived usability.Source: §7.2 · p. 8
  • Five NASA-TLX dimensions fell (all p < .001), with no significant Performance difference; feedback favored narrative coherence (6.6 vs. 2.8) and animation effectiveness (6.3 vs. 3.0).The other workload dimensions were Mental Demand, Physical Demand, Temporal Demand, Effort, and Frustration.Source: §7.2 · pp. 8–9, Figure 5 · p. 9
  • Task time fell 79.7%, from 39.2 minutes with the baseline to 8.0 minutes with DataMagic.Seven participants exceeded the 40-minute target with the baseline; all 12 completed DataMagic within it.Source: §7.2 · p. 8, Figure 5 · p. 9

7.3 Qualitative Feedback

  • Participants described manual handoffs and timeline alignment as baseline pain points; all 12 completed requested edits within the target scene without disturbing the rest of the video.Novices preferred natural-language commands, while experts favored direct script editing.Source: §7.3 · p. 9
  • Participants suggested a scene-template library and support for brand visual guidelines in exports.These were proposed directions for future iterations.Source: §7.3 · p. 9

8 DISCUSSION

DataMagic exposes its declarative state for user refinement, while its current expressive range depends on the renderer’s visual vocabulary and the models’ analytical capabilities. Its structured animation guidance improves reliability but constrains creative diversity, and several scope limitations remain.

Expressiveness and practical scope

  • DataMagic’s practical expressiveness depends on its rendering vocabulary and model capabilities, though these layers can expand while DVSpec remains a stable core.Extending rendering components can add visual forms, and stronger models can support more complex analyses; the design preserves chart reliability and traceability to source data.Source: §8 · p. 9

Designing for user intervention

  • Three interaction modes let users refine an inspectable shared DVSpec state through natural-language commands, script editing, or canvas actions.Scoped edits update only the relevant fields and re-render the target scene, preserving cross-modal bindings and other user-authored decisions without rerunning the full pipeline.Source: §8 · p. 9

Balancing animation reliability and creative diversity

  • Structured animation patterns keep motion aligned with chart semantics and narration, but their conservative guidance may suppress more expressive animations.Guideline-free prompting could yield more creative videos, but makes legibility, pacing, and narration alignment harder to control consistently.Source: §8 · p. 9

Limitations and future work

  • DataMagic currently targets single-table inputs and a single renderer, while image-generation assets and cross-backend synchronization remain future directions.The paper also identifies the lack of a commonly accepted data-video benchmark as a field-wide evaluation limitation.Source: §8 · p. 9

9 CONCLUSION

DataMagic combines a shared DVSpec state with Generate-then-Orchestrate to separate scene generation from global narrative organization. Case studies report stronger quality and synchronization than direct generation, alongside examples spanning agriculture, environmental science, and compensation analysis.

9 CONCLUSION

  • DataMagic improves video quality and execution stability over direct-generation baselines, while a user study reports greater authoring efficiency and reduced cognitive load.The conclusion frames declarative representations as an intermediate layer for coordinating automated generation, data traceability, and fine-grained human control.Source: §9 · p. 9

A.1 Cross-Method Comparison

  • 4.2 versus 1.5 average score: DataMagic outperforms direct generation in the Q4 2025 revenue case, with consistent gains across evaluation dimensions.Direct-generation methods frequently show missing charts, visual overlap, and narration–visual content or timing mismatches.Source: §A.1 · p. 12

A.2 High-Quality Generations

  • The agriculture example scores 4.6 overall and synchronizes a narration-triggered highlight of the 74.1% animal-fodder share without manual timeline adjustment.Its four-scene narrative moves from an overview through irrigation yield comparisons, residue-disposal distributions, and industry price disparities; narrative quality and animation effectiveness each score 5/5.Source: §A.2 · p. 12

A CASE STUDIES

  • Additional environmental-science and gender-compensation cases are described as preserving numerical fidelity, fine-grained comparisons, and coherent narration–visual coordination across domains.The appendix frames its case studies as covering cross-method comparisons, high-quality generations, and failure modes in several data-storytelling scenarios.Source: §A.2 · p. 12, A CASE STUDIES · p. 12

A.2 High-Quality Generations

  • The environmental-science case identifies a 12.8°C temperature peak on May 28, 2016, and depicts solar-radiation changes alongside humidity and temperature conditions.Radiation recovers from 106.2 W/m² to 326.0 W/m² within two days; it drops near zero above 90% humidity and reaches 9.4 W/m² as temperatures cool to 2.7°C overnight.Source: §A.2 · p. 12
  • The gender-compensation case represents fine-grained disparities: female employees average $70,582, exceeding males by $7,232, and EX1-level females reach $236,000.Departmental comparisons further show the distribution of gender pay gaps across administration and operational divisions.Source: §A.2 · p. 12

A.3 Analysis of Low-Quality Generations

The low-quality case analyses identify failures under extreme distributions and complex multimodal instructions. The accompanying evaluation-set passages describe its dataset dimensions and broad query coverage.

A.3 Analysis of Low-Quality Generations

  • Extreme input disparities can clip preset layouts and produce a narration–visual value mismatch, while a complex query can yield content that misses the requested analytical focus.The agriculture example reports a chart value of 40.9% against narration stating 40.6%; the energy baseline also exhibits missing visuals, overlap, and temporal asynchrony.Source: §A.3 · p. 12

B.1 Dataset Scale Distribution and Dimensions

  • Table 3 is scoped to dataset sources and scale distribution, while Table 4 summarizes dataset dimensions.The accompanying text describes the evaluation set’s range of dataset sizes and physical dimensions.Source: Table 3 · p. 12, Table 4 · p. 12, §B.1 · p. 12

B EVALUATION DATASET STATISTICS

  • The evaluation set contains 60 datasets and 109 samples, with sizes ranging from fewer than 100 rows to more than 1K rows.The largest dataset has 150,000 rows and some tables reach 555 columns, including up to 552 numeric and 57 categorical columns.Source: B EVALUATION DATASET STATISTICS · p. 12, §B.1 · p. 12

B.2 Query Type Analysis

  • The 109 test samples cover multiple analysis perspectives, with comprehensive analysis accounting for more than one quarter and correlation, trend, comparison, and distribution tasks also represented.The authors say this task mix is intended to assess logical reasoning and narrative quality across data-video composition scenarios.Source: §B.2 · p. 12

B.3 Domain and subdomain of evaluation data

The supplied passages identify a comparative video-generation-quality figure and a representative high-quality DataMagic example emphasizing synchronized visual highlights.

B.3 Domain and subdomain of evaluation data

  • Figure 6 is presented as a comparative analysis of video-generation quality across methods.The caption identifies the comparison’s subject but does not state an outcome.Source: Figure 6 · p. 13
  • Figure 7 shows a representative agriculture generation with an overall score of 4.6 and narration-triggered visual highlights synchronized without manual timeline adjustment.Four key frames are shown with corresponding narration; underlined narration spans mark the DVSpec index triggers.Source: Figure 7 · p. 13

C AGENT PROMPTS

The appendix documents DataMagic’s agent prompts and code examples for generating video configurations and executable scenes. It specifies how agents plan analyses, prepare data, and create data-grounded visual, narrative, and animated components with coordinated styling, scene ordering, and timing.

C AGENT PROMPTS

  • The pipeline first generates video configuration files and then generates executable files from those configurations.The framework uses prompts to establish clear responsibility boundaries, standardized input–output formats, and collaboration protocols.Source: C AGENT PROMPTS · p. 13

C.1 Scene Planner Agent

  • The Scene Planner selects 2–4 distinct analysis angles, gives each a priority from 0.0–1.0 and context explaining its importance, and specifies what to analyze rather than how to visualize it.Source: Critical Rules · p. 16
  • The Scene Planner chooses among five narrative patterns, including chronological, comparison-driven, drill-down, hook-first, and default single-peak structures.The selected pattern is based on the story’s shape and is used later by the Narrative Director to sequence scenes.Source: Narrative Pattern · pp. 15–16, Critical Rules · p. 16

C.1.1 Scene Planner Agent Example

  • For a flight-delay query, the example plan uses comparison analysis for average departure delays by carrier and distribution analysis for average arrival delays by destination city.The two chart scenes receive priorities of 1.0 and 0.9, respectively.Source: §C.1.1 · p. 16

C.2 Data Preparation Agent

  • The Data Preparation Agent selects transformations for the analysis, including aggregation, time-series aggregation, top-N, correlation, filtering, and representative sampling.It uses existing metadata fields, applies necessary filters with actual values, and returns a single JSON object.Source: §C.2 · p. 16, Transformation Types · p. 16

C.2.1 Data Preparation Agent Example

  • For the Punjab wheat-yield comparison, the example groups by irrigation method, averages yield, filters to Punjab and wheat, excludes nulls, and sorts descending.The output uses a group_by_aggregate transformation.Source: §C.2.1 · p. 17

C.3 Visual Designer Agent

  • The Visual Designer selects chart types to match analysis and data characteristics, must follow the planned scene type, and uses only provided data with exact numbers.Examples map trends to line or area charts, comparisons to bar charts, and correlations to scatter or bubble charts; insufficient data calls for stat cards.Source: §C.3 · p. 17, Input · p. 17, Chart Selection & Data Requirements · p. 17
  • Visual Designer prompts require query dimensions and data values to remain faithful, while chart type, title, insight summary, and theme support clear visual communication.They also prohibit adding narration or timing fields to the visualization configuration.Source: Critical Rules · p. 18

C.3.1 Visual Designer Agent Example

  • Examples contrast specific, evidence-based insight summaries with vague descriptions such as “This chart shows sales data.”The examples identify a standout result and comparison rather than merely describing the chart or noting unspecified variation.Source: §C.3.1 · p. 18

C.4 Narrative Director Agent

  • The director writes opening, scene, and closing narration, may add 2–4 stat cards, and must use chart-grounded values and the exact data for any named year or time.Narration rules also require references to scene insights, continuous transitions, and inclusion of all scene IDs in the ordered script.Source: §C.4 · p. 18, Your Tasks · p. 19, Critical Requirements · p. 19
  • The Narrative Director orders scenes by the selected pattern, defaulting to a Freytag arc from setting through rising evidence and climax to resolution.Alternative patterns specify a hook before evidence, adjacent contrasts, chronological progression, or a drill-down from overview to drivers; stat cards belong in the resolution.Source: Your Tasks · p. 18, Narrative Pattern Strategy · p. 18

C.4.1 Narrative Director Agent Example

  • A Freytag example progresses from an overall revenue setting to regional growth comparisons, an Asia growth climax, and a closing recap.The example assigns the stages to a continuous narration that links scenes with transitions.Source: §C.4.1 · p. 20

C.5 Scene Animation Generator Agent

  • The animation agent adds an entrance and an emphasis animation for each entity mentioned, triggering them through narration indices rather than manually specified times.The prompts describe common effects by chart type and allow custom sequences through declarative synchronization.Source: §C.5 · p. 20, Process · p. 20, Critical Rules · p. 20

C.6.1 Static Charts Scene

  • Static-chart prompts prioritize clean, data-focused scenes with distinct visual identities, readable subtitle space, and consistent typography, spacing, and theme.They also direct designers to highlight key data in the chart itself and retain strong text and axis contrast.Source: §C.6.1 · p. 21
  • Static-chart prompts require JSX-safe characters, avoid redundant information cards or tooltips, and vary color schemes across consecutive scenes.The static rendering instructions also require generated output to be complete TypeScript code without explanations.Source: §C.6.1 · p. 21

C.6.2 Opening, Closing, Stat_cards Scene

  • Opening, closing, and stat-card prompts specify scene content, typography, subtitle space, and distinct layouts, while deferring animation to a later stage.Stat cards use a single-row layout; the prompt reserves bottom space for narration subtitles and requests inline styles for later animation.Source: §C.6.2 · p. 21

C.6.3 TSX Static Component Agent Example

  • Static component examples illustrate TSX structures for opening/closing scenes and stat cards, including card props and a non-wrapping row layout.The stat-card example defines number, label, and color fields and styles the cards as a flex row.Source: §C.6.3 · pp. 21–22

C.7.1 Static Charts Animation

  • Chart-animation prompts add Remotion hooks and animation logic without changing static rendering, using absolute-time timing, precise selectors, and narration subtitles.They also require forcing elements to their final state after entrance animations and handling simultaneous highlights.Source: §C.7.1 · p. 22

C.7.2 Opening, Closing, Stat_cards Scene Animation

  • Opening, closing, and stat-card animation prompts preserve component content, sync subtitles using scene-relative timing, and apply animations to individual elements.They prescribe fade-and-slide entrances for opening and closing scenes and staggered card entrances with pulsing emphasis driven by the animation configuration.Source: §C.7.2 · p. 23

C.7.3 Animation Adding Agent Example

  • Animation examples demonstrate scene-time and narration calculations, opening and closing title motion, and sequential stat-card entrances with emphasis effects.The stat-card example uses staggered progress and illustrates pulse, scale, and border styling when a card is emphasized.Source: §C.7.3 · p. 23
  • The styling fragment applies the card color and an emphasis box shadow.The code references ${card.color} and emphasis.boxShadow.Source: §C.7.3 · p. 24
  • The rendering fragment places card content inside an AbsoluteFill component.The fragment shows a card-content comment followed by JSX closing and return syntax.Source: §C.7.3 · p. 24

D.1 Detailed Evaluation Framework and Scoring Criteria

This section defines a five-dimension, 1–5 rubric for data-video quality and specifies a Gemini-2.5-Pro judging prompt to apply it. The prompt documents observed audiovisual content and requires evidence-based scoring calibrated for issue severity and cross-dimension effects.

D.1 Detailed Evaluation Framework and Scoring Criteria

  • D.1 defines five 1–5 criteria: intent fulfillment, data insight, narrative quality, animation effectiveness, and aesthetic quality.The framework considers user intent alongside the video’s visualization, narration, and animation components.Source: §D.1 · pp. 24–25
  • Intent fulfillment measures whether a video answers the user’s problem, while data insight evaluates meaningful findings and whether viewers can verify them visually.Intent fulfillment concerns relevance, completeness, and clarity; insight quality concerns depth, novelty, relevance, and value.Source: §D.1 · p. 24
  • Narrative quality assesses story coherence, structure, transitions, and links to insights, whereas animation effectiveness assesses synchronization and pointing accuracy across the video.Animation scores focus on timing and pointing, with significant downgrades reserved for recurring or severe problems.Source: §D.1 · pp. 24–25
  • Aesthetic quality rates visual polish and professionalism through color, typography, layout, design details, and consistency.The top score requires a polished, cohesive, complete presentation; the rubric evaluates visual presentation rather than data accuracy.Source: §D.1 · p. 25

D.2 MLLM-as-a-Judge Prompt

  • The Gemini-2.5-Pro judging prompt requires an objective description of what is seen and heard before assigning 1–5 scores on all five dimensions.The description covers the data, visualizations, narration, video flow, scene sequence, and observed comprehension issues, without speculation.Source: §D.2 · pp. 25–26
  • The framework scores user intent, data insight, narrative quality, animation effectiveness, and aesthetic quality on a 1–5 scale.Each score requires a brief justification citing observable evidence from the video; visual failures cannot be overridden by narrative claims.Source: §D.2 · p. 26
  • Judges must describe and score only what is actually visible or audible, avoiding fabricated details, speculation, and unsupported assumptions.When evidence is unclear, they should omit the claim or use cautious language, and prefer a lower evidence-based score to an assumed higher one.Source: §D.2 · p. 26
  • Scores should reflect comprehensive viewing quality, with major penalties reserved for severe problems rather than localized or minor errors.Minor numerical differences from rounding or unit conversion are acceptable, while order-of-magnitude errors, opposite trends, incorrect categories, or missing key elements warrant heavier penalties.Source: §D.2 · p. 26
  • Aesthetic scores of 2 or below require serious harm to readability or professionalism, while minor flaws can still merit a 3.For animation, synchronization, timing, and highlight accuracy are distinct from missing data or incomplete charts, which should affect F1/F2/F3 rather than F4.Source: §D.2 · p. 26
  • Judges should account for how major issues in one dimension affect others, such as animation–narration misalignment reducing comprehension and disrupting narrative flow.The dimensions should not be evaluated in isolation.Source: §D.2 · p. 26

E DETAILS OF HUMAN EXPERT VALIDATION AND EVALUATION INTERFACE

The expert-validation procedure used three specialists and a stratified 60-video set, then compared automated scores with expert ratings. Overall agreement was strong, with dimension-level correlations above 0.75.

Expert validation

  • Three experts with data-analysis and video-production backgrounds independently reviewed a balanced validation set of 60 videos, with 15 sampled from each of four score intervals.The intervals were 1–2, 2–3, 3–4, and 4–5 points, sampled across methods and base models.Source: E DETAILS OF HUMAN EXPERT VALIDATION… · p. 27

Consistency Results

  • Overall automated–expert agreement was Pearson r = 0.91 (p < 0.001), with MAE = 0.38; all dimension-level correlations exceeded 0.75.Animation had the strongest correlation (r = 0.90), while Aesthetic had the largest absolute error (MAE = 0.73).Source: Table 7 · p. 27, E DETAILS OF HUMAN EXPERT VALIDATION… · p. 27

Evaluation interface

  • Experts first checked the original query, then inspected playback for audio-visual synchronization and narrative coherence before rating five dimensions on a 1–5 scale.The platform also included qualitative feedback; its interface combined the query, synchronized playback, scoring guide, and feedback module.Source: E DETAILS OF HUMAN EXPERT VALIDATION… · p. 27, Figure 12 · p. 28

Expert validation

  • Experts spent approximately 15–30 minutes reviewing each video and recorded qualitative defects alongside their quantitative scores.Reviews could identify logical deviations, cross-modal numerical conflicts, and temporal misalignments.Source: E DETAILS OF HUMAN EXPERT VALIDATION… · p. 27
Loading 2609.33403v1…