Material-dependent structure of roughness representation in generative AI-based texture analysis

Abstract Generative image models can produce visually different expressions of surface roughness depending on material context, but the structure of these variations remains unclear. Here, we systematically examined roughness-related image features in textures generated for six material categories at ten prompt-defined roughness levels. Three prompting conditions were compared: a Quantitative Model, a Generic Linguistic Model using common roughness descriptors across materials, and a Material-Specific Linguistic Model using category-specific descriptions. Nine interpretable image features comprising eight luminance-distribution statistics and the FFT spectral centroid were analyzed. Feature–roughness correlations varied substantially across materials, indicating that no single feature within the examined feature set consistently captured roughness-related variation across all categories. Hierarchical clustering revealed significant associations between material category and cluster assignment for both linguistic models but not for the Quantitative Model, with the strongest associations observed for the Material-Specific Linguistic Model. Silhouette scores in the full nine-dimensional feature space were −0.065, −0.001, and 0.271 for the Quantitative, Generic Linguistic, and Material-Specific Linguistic Models, respectively, indicating clearer category-dependent organization with material-specific linguistic conditioning. These findings demonstrate that roughness-related image-feature organization in the examined generative model depends on both material context and linguistic prompting. This framework provides testable hypotheses for future comparison with physically characterized surfaces and human perceptual judgments.