A Study on Material Fidelity Evaluation of Feicui "Qiaose Qiaodiao" (Clever Colour Carving): A V-O-I Protocol for Generative AI Application

Qiaose Qiaodiao (clever colour carving) in Feicui has strict requirements for material compliance. However, generative artificial intelligence may lead to distortion when processing heterogeneous material information, including texture, translucency, colour zoning, and fractures. To establish a comparable evaluation standard, this study proposes a V-O-I(Visual-Object-Intention) protocol centred on visual material evidence, object-level semantics, and design intention and craft constraints. Material fidelity is operationalized through four core evaluation dimensions: colour-zone utilization, outline retention, surface consistency, and intention consistency.Following the protocol, a controlled experiment was conducted using five AIGC tools, three control levels, and three independent generations under each condition, producing 45 outputs. These outputs were evaluated through structured expert scoring and qualitative failure-mode analysis. The results show that, under the material and task conditions examined, the systems exhibited varying limitations in preserving visible material characteristics, maintaining semantic correspondence, and controlling two-dimensional craft risks. The principal failure types included outline deviation, misplacement of colour zones, and fabricated textures. The addition of the rough-stone image, outline constraints, and colour-zone constraints improved some evaluation dimensions, although the overall gains remained limited. The proposed protocol provides a reproducible, comparable, and diagnostic method for evaluating material fidelity and offers an initial reference for AI-assisted design in Feicui carving and related subtractive crafts. Based on on a single material sample and two-dimensional inputs, this study provides an initial validation of the protocol under controlled conditions.The results serve to characterize the system performance under current controlled settings and lay the foundation for subsequent validation using heterogeneous materials and multi-view inputs.