Paper #16: A Typography Benchmark for Co-Creative Graphic-Design Agents

Typography serves as a central medium for expressing visual intent in human–AI design collaboration, where font, size, weight, color, alignment, spacing, curvature, and inline style jointly encode hierarchy, tone, and brand identity. Yet whether current frontier vision-language models (VLMs) can actually read or render this medium reliably remains largely untested. We introduce TYPOBENCH, a typography-focused benchmark of 12 tasks grounded in a layered-composition corpus of 989 real-world design templates and 2,568 text elements. Ten understanding tasks probe fine-grained perception of typographic properties, and two generation tasks evaluate whether models can insert styled text into masked layouts and remove text while preserving the design context. Evaluating four state-of-the-art models, we find an uneven capability profile: VLMs can localize styled spans, estimate some visual attributes, and detect many curved or rotated text instances, but typography remains far from design-grade reliability. Font identification tops out at 29.9%, exact recovery of styled spans remains at 0%, generated text routinely spills outside prescribed masks, and text removal remains only partially solved. We release the benchmark as a diagnostic tool for the human-AI co-creativity community to quantify how well future design partners can perceive, reproduce, and manipulate the typographic layer designers care about most. The benchmark and code are available at https://github.com/lica-world/GDB.