Title: Through the PRISM: Principle-Aware, Interpretable, and Multi-Scale Evaluation of Visual Designs

Abstract:

Effective visual communication relies on the synergistic balance of various design principles—such as contrast, alignment, readability, overlap, and coherence—which collectively dictate the clarity and intent of the message. While human designers evaluate these elements holistically, machine agents typically reduce them to a single heuristic score, resulting in limited interpretability and diagnostic precision. To bridge this gap, we present PRISM (PRinciple-aware, Interpretable, and Structure-guided Design Modifications), a benchmark designed to systematically perturb professional layouts from the Crello dataset according to measurable design principles. This benchmark includes 100,000 perturbed training samples and 10,000 perturbed validation designs, with each instance isolating a specific principle violation to facilitate controlled analysis of multimodal reasoning regarding design quality.

Our experiments reveal that models such as Qwen-2.5-VL and GPT-4o-mini are largely insensitive to targeted principle degradations, while GPT-4o demonstrates global awareness but lacks fine-grained disentanglement. Leveraging these findings, we propose a multi-scale evaluation framework that combines lightweight scorers for quantitative assessment, instruction-tuned vision-language models for localized feedback, and prompt-based methods for global reasoning. This framework generates interpretable explanations for design failures and enables targeted refinements that enhance layout quality based on localized insights. Together, PRISM and our proposed framework establish a foundation for interpretable, design-literate multimodal reasoning systems.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC