VFIG: Vectorizing Complex Figures in SVG with Vision-Language Models
The episode discusses the paper "VFig: Vectorizing Complex Figures in SVG with Vision-Language Models," which trains a vision-language model to convert raster images of complex scientific figures into editable SVG code. The hosts detail the massive dataset creation, two-stage training process using reinforcement learning, and a multi-level evaluation benchmark called VFig-Bench. They conclude that this technology could democratize editing scientific visuals.