FigFlux hands multi-panel figure transitions to AI agents, cutting 16 authors' workload score to 42.86 from 58.44
Synopsis
Multi-panel scientific figures are usually explained orally when presented, and two static panels specify only the start and end, leaving the intermediate keyframes unclear. FigFlux represents a figure in a Figure Space that records panel layout, data transformations and visual encodings, then uses a Transition Space to build and order keyframes that show corresponding data together in a joint coordinate system. In a study with 16 scientific authors, overall workload scored 42.86 versus 58.44 for their familiar tools, and four visualization experts rated its transitions higher on all five criteria than those made with familiar tools or by AI alone.
Interpretation
Scientific authors can now have AI agents generate animated transitions across multi-panel scientific figures, revising intermediate keyframes in natural language before generation. Previously, authors moved from one panel to the next by explaining the data relationships themselves, because two static panels specify only the transition's start and end. In a within-participant study with 16 scientific authors from the eight Nature Portfolio subject areas, FigFlux scored 42.86 on overall NASA-TLX workload versus 58.44 for participants' familiar tools, a significant difference.
The system first represents the figure as a traceable data structure, then decides which data appear together in each keyframe and in which coordinate system. Figure Space uses a Layout to describe panel juxtaposition, repetition and overlay, and a Figure Table to align data fields across panels while retaining each panel's transformation steps and visual encodings. The representation came from iterative analysis of 48 published multi-panel figures across eight scientific fields, with 16 seed figures and 32 more used to check the representation rules and keyframe constraints.
Intermediate frames place the two panels' corresponding data in one joint coordinate system, using a three-dimensional layout when needed to preserve each panel's position encodings. Transition Space first identifies shared dimension-level values between the two panels, constructs Combined Data, and tries a single-panel two-dimensional joint coordinate system first. In the case connecting weekly crime patterns to a city's spatial partition map, the curves' position encodings and the map's geographic layout could not both be preserved in one 2D coordinate system, so the system built a 3D joint coordinate system placing regions by category on separate planes aligned with their curves.
Perspective
The demonstrated scope is published scientific figures with at least two panels, more than half of them charts, and at least one pair of chart panels sharing some entities, categories, locations or time periods; figures dominated by domain-specific scientific imagery are excluded. The system takes the figure with its associated data, source paper and rendering code as input, and currently prefers a two-dimensional joint coordinate system when encodings allow.
Some charts were recreated from approximate data derived from positions in the published panels when materials were unavailable or full reproduction required large datasets or specialized pipelines, so they may not exactly match the originals; generation still takes time, and users can work on other tasks while it runs; agreement among experts on individual clip ratings was limited, with low ordinal Krippendorff's alpha.
