A Trustworthiness Assessment Method for AIGC-Based Animation Translation of Traditional Narrative Images in Architectural Decoration

Artificial intelligence-generated content (AIGC) image-to-video systems have lowered the technical threshold for presenting traditional cultural content in motion, while generated motion may alter characters, props, narrative relationships, craft features, composition, and symbols. This study implements a trustworthiness assessment workflow for traditional narrative images in architectural decoration, linking cultural evidence dossiers, sample-adapted prompts, computer-vision measurements, and blinded human coding. Eight source images were used to generate 120 videos under three prompt conditions, with five outputs per image–condition cell. Four coders assessed the videos, five experts reviewed the coding instrument, and two new coders independently assessed 48 videos. Primary deviation rates were 7.12%, 7.37%, and 9.07% for conditions A, B, and C, respectively, with no significant blocked difference after correction (Friedman χ2(2) = 1.400; Holm-adjusted p = 1.000). Mean item-level Krippendorff’s alpha was 0.250 across 14 estimable items; the independent-review composite ICC (A,1) was 0.402. Additional analyses examined coder sensitivity and the grouped generation structure. The original association involving inter-frame stability SSIM was not supported in within-image–condition comparisons (ρ = −0.052; permutation p = 0.610). An exploratory association was observed between source-image SSIM and human-coded deviations within these cells (ρ = −0.366; Holm-adjusted permutation p = 0.021). The study provides a documented procedure for comparing generated outputs and identifies limits in prompt effectiveness and the reproducibility of human judgments.