Scalability of existing figure-to-SVG methods to complex scientific diagrams

Determine how well existing figure-to-SVG generation approaches—including classical raster-to-vector tracing, learning-based techniques, and LLM/VLM-based SVG code generation—scale to structurally complex scientific figures featuring multi-panel layouts, dense annotations, hierarchical grouping, and precise connectivity.

Background

Prior vectorization and SVG-generation methods have largely been developed and evaluated on simple graphics such as icons or small diagrams. Whether these methods maintain fidelity and editability on real-world, complex scientific figures—those with nested layouts, dense labels, and precise connectivity—has not been systematically established. This uncertainty motivates the need for targeted datasets and benchmarks to assess scalability to more challenging diagram structures.

The paper positions this gap as a central uncertainty in the field and builds VFig-Data and VFig-Bench to address it, but the quoted text explicitly frames the baseline state of knowledge as unclear regarding scalability to complex figures.