Multi-Metric Evaluation of Physical–AI Correspondence in Reference-Based Velvet Upholstery Visualization Using a Gemini Consumer Web-Interface Workflow Under Different Lighting Conditions for Interior Design

This exploratory technical case study evaluates correspondence between physical and artificial intelligence (AI) images in reference-based visualization of velvet upholstery surfaces under three illumination conditions. Cotton, polyester, and acrylic velvet specimens were photographed under daylight, cool white (6500 K), and warm white (2700 K) using standardized acquisition geometry. Twenty-seven separately captured physical reference images were each used once to obtain one corresponding AI output through the Gemini consumer web-interface workflow labeled “Gemini 3 Flash,” yielding 27 matched acquisition and generation pairs rather than repeated generations from identical inputs. Correspondence was evaluated using CIEDE2000 color difference, multiscale structural similarity, learned perceptual image patch similarity, and gray-level co-occurrence matrix features. For inference, the three acquisition and generation pairs within each specimen and lighting condition were averaged, yielding nine condition-level observations. Color difference and multiscale structural similarity differed significantly by specimen but not by lighting, whereas learned perceptual image patch similarity differed significantly by both. After Holm adjustment, sign-flip tests showed significant overall differences between AI and physical images in contrast, energy, homogeneity, and correlation, whereas blocked permutation tests showed no robust main effects of specimen or lighting. Overall, correspondence was metric- and scale-dependent within the evaluated workflow.