TextureBench: A Privacy-Preserving Multi-Genre Benchmark for Mapping Unstructured Text to Editable SVG Textures

We introduce TextureBench, a privacy-preserving benchmark for evaluating systems that map unstructured text to editable SVG visual textures for downstream creative refinement. Instead of treating vector output as a finished chart, TextureBench treats it as provisional material whose micro-variation, including jitter, opacity and structural layering, remains open to later editing. The benchmark combines fully synthetic multi-genre corpora, a unified segment schema and controllable metamorphic operators to test whether textual changes propagate consistently into extracted signals and rendering parameters without human participants or private datasets. In a prototype study on 180 synthetic documents spanning monologue, hierarchical and threaded genres, signal-level tests showed perfect directional consistency across five operators, while renderer-level tests exposed recurrent saturation caused by parameter clamping. TextureBench provides a reproducible, low-burden way to evaluate text-to-SVG pipelines and diagnose design boundaries.