摘要
克拉克瓷作为16~17世纪海上丝绸之路的重要文化载体,其装饰纹样蕴含着跨文化的交融特征,但传统研究方法难以系统解析其纹样生成规律与文化基因。本文提出了一种基于稳定扩散模型的克拉克瓷纹样智能生成方法,通过构建多维度文化特征数据库,结合低秩自适应微调与ControlNet结构控制技术,突破了现有AI生成模型的文化语义脱节与结构失序瓶颈。实验采用709件克拉克瓷图像数据,通过LoRA实现参数高效微调,保留了纹样风格特征;引入ControlNet线稿约束机制,精准控制主纹样、副纹样与附加纹样的开光布局。结果表明,该方法生成的纹样在保持青花瓷美学特征的同时,能稳定呈现中心对称、八开光等典型结构。本研究构建了文化遗产数字化再生技术范式,通过跨模态语义嵌入实现了纹样文化基因的可控表达,为传统工艺的活态传承与创新设计提供了方法论支持。未来研究将深化跨文化符号解构算法,拓展复杂纹样的生成维度。
Kraak porcelain,as an important cultural carrier of the Maritime Silk Road in the 16th and 17th centuries,its decorative patterns integrate cross-cultural characteristics.However,traditional methods are difficult to systematically analyze its generation rules and cultural genes.This study proposes an intelligent generation method based on the Stable Diffusion model,combining low-rank adaptive fine-tuning(LoRA)and structure control technology(ControlNet),to break through the bottlenecks of existing AI models in cultural semantic disconnection and structural disorder.By constructing a multi-dimensional database of 709 Kraak porcelain patterns and using LoRA to achieve efficient fine-tuning of parameters,the style characteristics of blue and white porcelain were retained.Introduce the ControlNet line drawing constraint mechanism to precisely control the lighting layout of the main patterns,secondary patterns and additional patterns.Experiments show that the generated patterns can stably present typical structures such as central symmetry and octagonal light,while maintaining the aesthetic characteristics of blue and white porcelain.In the restoration task,the models of LoRA and ControlNet were combined to complete the patterns on the damaged porcelain plates,and the effectiveness was verified through subjective and objective indicators.Quantitative analysis shows that after the introduction of ControlNet,both the structural similarity and the peak signal-to-noise ratio have increased,significantly optimizing the coherence and detail reproduction of the generated images.The ablation experiment further confirmed that ControlNet plays a key role in controlling the shape,size and transition of secondary patterns of the opening,solving the problem of uneven layout when only LoRA is used.This study has constructed a technical paradigm for the digital regeneration of cultural heritage.Through cross-modal semantic embedding,it realizes the controllable expression of the cultural genes of patterns,providing methodological support for the living inheritance and innovative design of traditional craftsmanship.In the future,we will deepen the cross-cultural symbol deconstruction algorithm,explore small-sample incremental learning,and integrate the experience of art historians with AI technology to promote the leap of cultural heritage protection from“technology-driven”to“cultural interpretation”.
作者
任玉洁
鲁浩天
刘驰
Ren Yujie;Lu Haotian;Liu Chi
出处
《艺术设计研究》
北大核心
2025年第3期5-13,135,共10页
Art & Design Research
基金
国家自然科学基金项目“面向复杂开放交互环境中AI图像合成的防御技术研究”(项目编号:62402009)的阶段性成果。
作者简介
任玉洁,澳门城市大学创新设计学院副教授博士;鲁浩天,澳门城市大学创新设计学院硕士研究生;刘驰,澳门城市大学数据科学学院助理教授博士。
