Producing data videos — dynamic charts paired with voice narration and synchronized animation — demands a blend of data analysis, narrative design, and video editing skills that few creators possess. Existing tools fall short in different ways: static visualization software lacks narrative and animation support, authoring tools depend on pre-prepared charts rather than raw data, and pixel-level generative models cannot guarantee data accuracy or provenance.
Researchers behind a new paper, DataMagic: Authoring Data Videos through Declarative Multi-Agent Orchestration, tackle two core challenges: uniformly representing charts, narration, and animations along with their temporal relationships, and efficiently searching a vast design space for narrative-coherent compositions.
Their system rests on two designs. DVSpec, a declarative specification, unifies charts, narration, and animations through data-bound references and declarative synchronization, which the authors say ensures data provenance and automatic audio-visual alignment. On top of it, a "Generate-then-Orchestrate" multi-agent strategy produces candidate scenes in parallel before optimizing narrative coherence through global orchestration. DVSpec also acts as a shared state supporting three complementary interaction modes, bridging full automation with fine-grained human control.
Evaluation on 109 real-world samples found that even the most advanced LLM tested (GPT-5) scored only 2.13 out of 5, with execution success rates between 48.62% and 86.24%. DataMagic raised quality to 3.89 — an 83% improvement — with success rates above 95%, with the largest gains in animation and narrative dimensions. A user study reported a 79.7% reduction in task time versus a conversational LLM workflow, along with lower perceived cognitive load.
The project page is available at github.com/HKUSTDial/DataMagic.