WildSVG: Towards reliable SVG generation under Real-Word conditions

ICLR 2026 Conference Submission21011 Authors

Keywords: SVG, VLLM, LLM

TL;DR: We present a new dataset and benchmark for a new SVG generation task

Abstract: We introduce SVG extraction, the task of translating specific visual inputs into scalable vector graphics. Existing multimodal models such as StarVector achieve strong results when generating SVGs from clean renderings or textual descriptions, but they fall short in real-world scenarios where natural images introduce noise, clutter, and domain shifts. To address this gap, we extend StarVector’s capabilities toward robust vision-to-SVG translation in the wild. A central challenge in this direction is the lack of suitable benchmarks. To fill this need, we develop two complementary datasets: Natural WildSVG, consisting of real-world images paired with SVG annotations, and Synthetic WildSVG, which integrates complex and elaborate SVG designs into real-life scenarios to simulate challenging conditions. Together, these resources provide the first foundation for systematic benchmarking SVG extraction. Building on them, we benchmark StarVector and related models. Our study establishes SVG extraction as a new problem domain, introduces datasets and evaluation protocols for its study, taking initial steps toward extending multimodal LLMs to handle reliable SVG generation in complex, natural scenes.

Primary Area: datasets and benchmarks

Submission Number: 21011