Research on the Digital Heritage of Xiangyun Sha Products Based on LoRA Model and Controllable Workflow

Abstract

As the fourth industrial revolution continues to deepen, the digital inheritance of intangible cultural heritage (ICH) has gradually become a prominent focus in academic research. The digitization of Xiangyun Sha fabric holds significant commercial value and serves as a crucial practice for the dynamic preservation of ICH. However, existing approaches face persistent challenges, including inadequate texture restoration, unstable model outputs, and a lack of systematic, replicable workflows applicable to ICH fabric digitization. To address these issues, this paper proposes and constructs a controllable digital design workflow that integrates LoRA (Low-Rank Adaptation) model training with the ComfyUI node-based pipeline. The workflow is structured in three progressive stages: first, a high-quality GuiLie crackle texture dataset is established through targeted data collection and preprocessing, with optimized labeling strategies and customized trigger words to capture the core material characteristics of Xiangyun Sha accurately; second, an advanced LoRA model is constructed through multi-element joint training that integrates fabric texture, product form, and decorative pattern; third, the complete workflow is validated and optimized through systematic ComfyUI-based experiments, through which a model weight of 0.8 and the LoRA5 version are identified as the optimal configuration. The proposed workflow provides a technically replicable framework for building digital gene libraries of ICH fabrics. It offers a transferable reference paradigm for the intelligent design and digital innovation of other traditional ICH textiles.