The Ba Shu rock carvings, including the UNESCO-listed Dazu sites, depict zhang zang architectural forms that indirectly reflect the xiaomuzuo technique of the Tang and Song periods. However, systematic comparisons between these carvings and textual descriptions in the Yingzao Fashi remain limited, constrained by the deterioration of carvings, which complicate the discernment of fine details and by interpretive gaps regarding whether sculpted images can faithfully represent historical zhang zang forms. This study proposes an AI-enabled virtual restoration method as a cross-media approach to address these challenges. Using the Qwen2.5–32B large model on the Shanghai Jiao Tong University cloud platform, representative carvings from Dazu and other Ba Shu sites were processed for deterioration detection, feature enhancement, and structural restoration. The following research based on restored features, is compared with material evidence and relevant sections of the Yingzao Fashi . Findings reveal (1) discrepancies between Ba Shu carvings and Yingzao Fashi descriptions; (2) AI-based virtual restoration improves the readability of architectural features; (3) the virtual restored materials enable cross-validation across textual, material, and digital interpretations. By restoring the original appearance of stone relics and triangulating reconstructed imagery with regional physical remains and textual sources, this study proposes a localized understanding of Ba Shu zhang zang forms.
Liu et al. (Fri,) studied this question.