Imagine transforming a blurry, uneditable photograph of a complex engineering diagram into a pristine, scalable vector file with a single command. This is no longer science fiction but a tangible reality made possible by VFIG, a new family of vision-language models designed to reverse-engineer the lost original source code behind digital illustrations. For decades, designers and researchers have faced a frustrating bottleneck: once professional vector files are rasterized into JPEGs or PNGs, their precise geometry and editability vanish, leaving only images that cost an arm and a leg to reconstruct manually.

VFIG tackles this crisis by training on VFIG-DATA, a massive dataset comprising 66,000 high-quality pairs of paper figures and procedurally generated diagrams. Unlike previous attempts that struggled with scale, this system employs a sophisticated coarse-to-fine training strategy. It starts by learning basic atomic shapes through supervised fine-tuning before leveraging reinforcement learning to refine the global layout, ensuring perfect topological consistency even in complex diagrams. The results are staggering; the model not only outperforms other open-source tools but also matches the performance of cutting-edge models like GPT-5.2, scoring a remarkable 0.829 on the new VFIG-BENCH evaluation suite. By democratizing access to high-fidelity vector conversion, this technology promises to reshape how we preserve and manipulate technical documentation in an era where original data is increasingly fragile. Researchers and engineers can now recover geometric intent that was previously lost forever, turning static images into editable assets instantly.

Source: VFIG: Vectorizing Complex Figures in SVG with Vision-Language Models by Qijia He et al., https://arxiv.org/abs/2603.24575