A new AI model can take a blurry screenshot of a flowchart and hand back a fully editable vector file.
Researchers built VFIG, a family of vision-language models trained specifically to convert rasterized diagrams (flowcharts, model architectures, technical schematics) into structured SVG files. To train it, they released VFIG-Data, a 66,000-pair dataset that mixes real paper figures converted through a describe-and-generate pipeline with programmatically generated diagrams for clean supervision on arrows, fonts, and geometry. They also built VFIG-Bench, an evaluation suite that checks whether arrows connect to the right boxes and lines stay unbroken, instead of just comparing pixels. VFIG itself was trained with a curriculum that starts on simple shapes and escalates to complex diagrams, then fine-tuned with reinforcement learning tied to how well the rendered SVG actually looks.
This targets a real, unglamorous problem: countless technical diagrams exist only as flattened images because the original vector file got lost somewhere along the way, and redrawing them by hand eats hours nobody has. Existing raster-to-vector tools mostly handle logos and icons; they choke on diagrams where arrows, connectors, and precise geometry actually need to stay correct. The more interesting shift is the benchmark itself, which grades structural accuracy rather than whether an image merely looks similar at a glance.
On the numbers, VFIG essentially matches Claude Sonnet 4.6, and even edges past it on two of the paper's own scoring rubrics, though it still trails GPT-5.2 and Gemini-3. That gap is a reminder that in AI research right now, being state of the art mostly means best of the open-source options, not best overall.