Original title: SVG-Score: Human-Aligned Evaluation of Text-to-SVG Generation
Marco Cipriano, Leonardo Zini, Alexandra Schild, Valentin Teutschbein +4 · September 2026
Computers can now turn written text into Scalable Vector Graphics, or SVGs, which are digital drawings made of math instead of pixels. However, we lack good ways to check if these drawings match their descriptions. People usually use tools built for regular photos, like CLIPScoreAn AI metric that measures how well an image matches text, trained primarily on natural images., but these miss common SVGA type of image that uses mathematical descriptions of shapes, allowing it to scale without quality loss. mistakes. To solve this, researchers introduced SVG-Score. First, they tweaked text and images to test current tools. They found CLIP barely notices errors in colors, counts, or layout, while standard AI vision models react unevenly across drawing styles. Next, the researchers collected human ratings on how well SVGs match text. Using these ratings, they built fast, tuned CLIP tools and a detailed AI judge, then tested major SVG generators on independent captions.
Imagine you ask a friend to draw a city map from a list of landmarks. If you only check whether the drawing is neat, you might miss that a park is labeled as a lake. Current AI evaluation tools for graphics work similarly, ignoring content errors. In this paper, the new evaluation framework is like having an expert who verifies each landmark against the list to ensure accuracy.
This matters because designers and developers rely on SVG generation, but poor evaluation means they cannot trust the tools. Improved metrics help advance AI that creates accurate graphics.
The abstract does not state specific limitations of SVG-Score, but it highlights that previous metrics were inadequate and the framework focuses on semantic alignment.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
SVG-Score: Human-Aligned Evaluation of Text-to-SVG Generation — Marco Cipriano, Leonardo Zini, Alexandra Schild et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.