A new system called DataMagic can automatically convert raw spreadsheet data into polished narrative videos complete with animated charts, voice narration, and synchronized storytelling. Researchers at Hong Kong University of Science and Technology and China Unicom developed the tool to solve a persistent problem: creating high-quality data videos typically requires expertise spanning data analysis, narrative design, and video production, making it inaccessible to most organizations.

Why Is Automating Data Video Creation So Difficult?

Data videos are increasingly popular for communicating insights because they combine dynamic charts, voice narration, and synchronized animations into temporal narratives that help audiences understand complex information faster. However, existing tools fall short in different ways. Static visualization platforms like business intelligence dashboards produce charts without narrative flow or animation. Video authoring tools require users to manually prepare visualizations first rather than working directly from raw data. Meanwhile, general-purpose video generation models like Sora can create videos but often produce numerical errors and cannot trace visual elements back to the underlying data.

The core challenge is that effective data videos are fundamentally structured narratives, not just collections of visual elements. This means the system must handle two critical problems: designing a structured format that precisely describes all components while ensuring data accuracy and traceability, and efficiently searching through countless design possibilities to balance both individual scene quality and overall narrative coherence.

How Does DataMagic Transform Data Into Video?

DataMagic works through a three-part architecture. First, it accepts raw tabular datasets and natural language queries as input. Second, it processes these through a multi-agent engine that generates the final video specification. Third, it compiles that specification into a complete narrative video.

The system introduces a new technical standard called DVSpec, or Data Video Specification. Think of DVSpec as a detailed blueprint that decouples the logical structure of the video from how it actually gets rendered. This separation is crucial because it allows the generation engine to focus on creating the right content while the rendering engine handles the visual output. Each scene in the video is defined by four components: its type, the visualization content, the narration text, and animation effects. By binding visual and animation elements directly to underlying data fields, DVSpec ensures complete data traceability and prevents the numerical hallucinations that plague general-purpose video models.

To handle the enormous design space of possible videos, DataMagic uses what researchers call a "Generate-then-Orchestrate" multi-agent architecture. This approach generates many candidate scenes in parallel, then applies global optimization to select the best scenes, order them logically, and ensure the overall narrative flows coherently. The system currently renders charts using D3.js, a popular visualization library, and synthesizes videos with Remotion, a web-based video creation tool.

What Makes DataMagic Different From Existing Tools?

• End-to-End Automation: Unlike static visualization tools or manual authoring platforms, DataMagic works directly from raw data and natural language questions, requiring no pre-prepared visualizations or manual scene creation.
• Data Fidelity Guarantees: By using DVSpec's declarative specification and data-driven semantic references, the system ensures numerical accuracy and maintains a complete audit trail linking every visual element back to the source data.
• Narrative Coherence: The multi-agent architecture optimizes not just individual scenes but the entire video's narrative flow, ensuring scenes are ordered logically and connect meaningfully rather than appearing as disconnected visualizations.
• Interactive Exploration: Beyond generating static videos, DataMagic supports three interaction modes and provenance-based question-answering, allowing viewers to refine videos and explore underlying data directly rather than passively watching.

How Well Does the System Actually Work?

The research team evaluated DataMagic on 109 real-world data samples to validate its effectiveness. The evaluation confirmed that the system successfully transforms raw tabular data into coherent narrative videos while maintaining data accuracy and enabling interactive exploration. This validation across a substantial number of real-world examples suggests the approach is practical beyond academic proof-of-concept.

The system's ability to decouple logic from rendering also opens possibilities for future improvements. Because DVSpec is rendering-library agnostic, developers could swap in different visualization or video synthesis tools without redesigning the core generation engine. This modularity makes the approach more flexible and adaptable as new video generation technologies emerge.

What Are the Practical Implications?

DataMagic addresses a real bottleneck in how organizations communicate data insights. Currently, creating a polished data video requires hiring specialists or spending significant time learning multiple tools. By automating this process, the system could democratize data storytelling, allowing analysts and business teams to generate professional narrative videos directly from their data without specialized video production skills. This is particularly valuable in fast-moving industries where insights need to be communicated quickly and frequently.

The emphasis on data fidelity and provenance also matters for regulated industries and high-stakes decision-making. Unlike black-box video generation models that might hallucinate numbers, DataMagic maintains a complete chain of custody from raw data through visualization to final video, making it suitable for contexts where accuracy and auditability are non-negotiable.