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

2026-05-30

AI Summary

Existing models struggle to perform fine-grained, interpretable evaluation of multidimensional visual design principles—such as readability, contrast, and alignment—and often rely on coarse heuristic scores. To address this limitation, this work introduces PRISM, a benchmark comprising 100,000 training and 10,000 validation samples systematically generated from the Crello professional layout system, where each sample perturbs only a single design principle, enabling the first decoupled and controllable assessment of individual principles. We develop a multi-scale evaluation framework that integrates lightweight scorers, instruction-tuned vision-language models (e.g., Qwen-2.5-VL, GPT-4o), and prompt engineering to jointly support holistic quality judgment and localized defect diagnosis. Experiments reveal that prevailing multimodal models are largely insensitive to design degradations, whereas our framework delivers interpretable feedback and effectively guides targeted refinements that significantly enhance layout quality.

Abstract

Effective visual communication stems from the harmony of multiple design principles, such as readability, contrast, alignment, overlap, and coherence, which collectively govern clarity and intent of the communicator. While human designers reason holistically over these principles, machine agents typically condense them into a single heuristic score, offering limited interpretability and diagnostic precision. To address this gap, we introduce PRISM (PRinciple-aware, Interpretable, and Structure-guided Design Modifications), a benchmark that systematically perturbs professional layouts from the Crello dataset along measurable design principles. The benchmark comprises 100K perturbed training samples and 10K perturbed validation designs, each isolating a specific principle violation for controlled analysis of multimodal reasoning about design quality. We show that models like Qwen-2.5-VL and GPT-4o-mini are largely insensitive to targeted principle degradations, whereas GPT-4o exhibits global awareness without fine-grained disentanglement. Building on these insights, we propose a multi-scale evaluation framework that integrates lightweight scorers for quantitative assessment, instruction-tuned vision-language models for localised feedback, and prompt-based methods for global reasoning. Our framework provides interpretable explanations of design failures. Using these localised insights, we show targeted refinements that improve layout quality. Together, PRISM and our framework lay the foundation for interpretable design-literate multimodal reasoning systems.

Problem

• visual design evaluation
• design principles
• interpretability
• multimodal reasoning
• principle violation

Innovation

• PRISM
• design principles
• interpretable evaluation
• multimodal reasoning
• multi-scale assessment