Dynamic Transmission and Innovative Transformation of Cultural Heritage: Generative Artificial Intelligence Practices Based on Cultural Cognitive Models

Abstract

The rapid advancement of digital technologies is compelling the field of cultural heritage preservation to shift from static conservation toward dynamic transmission and innovative transformation. In response, this study proposes a Generative AI (GenAI) application approach based on a cultural cognitive model. First, a cognitive structure of cultural symbols is constructed based on symbolic interactionism, and grounded theory is applied to analyze how specific user groups interpret and internalize these symbols, thereby establishing a cultural cognition system. An enhanced Delphi method is then employed to synthesize expert judgments and develop a multi-level cultural-symbol dataset. The dataset is integrated into generated models through stable diffusion models and Low-Rank Adaptation (LoRA) to strengthen their capacity for recognizing and generating culturally significant features. The feasibility and effectiveness of the proposed model are evaluated through expert-based assessments. To further examine its generalizability, the study conducts a case application using Shanghai-style furniture design. The results demonstrate substantial improvements in output quality and alignment with design requirements. This research provides a reproducible methodology for the digital safeguarding and innovative development of cultural heritage, while expanding the application scenarios of AI technologies in the protection of intangible cultural heritage.

1. Introduction

With the intensification of globalization and the rapid development of digital technology, the protection and utilization of traditional cultural heritage are facing unprecedented challenges [1]. As their original social functions and cultural values are increasingly undermined by factors such as cultural homogenization [2,3] and economic structure, showing a trend of weakening and marginalization [4]. Consequently, cultural heritage urgently requires new modes of transmission—shifting from static preservation and formal replication to dynamic expression—in order to foster more socially embedded and interactive pathways of inheritance [5]. In China, a country which enriches cultural heritage resource and owes the responsibility to solving problems, how to realize dynamic inheritance [6] and innovative transformation [7] in digital era is especially an undoubtable urgency. In confronting these challenges, the digitalization and informatization of cultural heritage have become a key direction in the strategy of sustainable cultural development [8,9].

The rise in artificial intelligence (AI) technology has brought new opportunities for the preservation and revitalization of cultural heritage [10]. Many scholars and institutions have introduced AI into heritage preservation and utilization, achieving remarkable results [11,12]. Through techniques such as image recognition and natural language processing, AI can assist in documenting, preserving, restoring, and disseminating cultural heritage, thereby enhancing preservation efficiency and supporting the sustainable development of culture. Computer vision, for instance, has been widely applied in 3D reconstruction [13], restoration [14], and defect detection of cultural relics [15], enabling more accurate representations of their original forms. Meanwhile, machine learning and mega data technologies are also gradually applied in cultural heritage data analysis [16,17]. By constructing large-scale datasets, researchers can better understand correlations among cultural elements across different media and uncover shared underlying patterns. These technological tools have greatly expanded the modes of cultural heritage presentation, and enhanced its popularization and dissemination effect [18,19].

However, though existing research has provided significant support for cultural heritage preservation, several limitations remain [20]. First, recent studies predominantly focus on the physical structure or appearance features of cultural heritage, while ignoring its cultural and social attributes. This makes it hard to truly reflect people’s in-depth cognition of it. Second, most existing AI models are pretrained on common visual datasets like ImageNet and COCO. These datasets have cross-cultural open data sources but lack systematic labeling of regional cultural elements. As a result, cultural semantic gaps occur when the models capture the symbolic features of specific cultural heritage, making it hard for them to accurately grasp cultural uniqueness in cultural heritage preservation tasks. Meanwhile, the existing AI models, because of their reliance on open-source datasets, may be biased and discriminatory, polluting the AI models. Therefore, constructing cognitive models in specific cultural contexts and integrating them with existing general-purpose datasets can enhance the dataset’s cognitive ability for specific cultures. This approach improves the models’ ability to meet the specific needs of cultural heritage digitization, making it a field worthy of in-depth research.

In order to fill the above research gaps, this study, based on symbolic interactionism as the theoretical framework [21], employs grounded theory together with the Delphi method as its methodological foundation [22]. The research explores the cognitive hierarchy of specific groups on cultural heritage, effectively acquires the cognitive data of specific groups on specific cultural heritage, and organizes them into structured datasets, which are then used for the training of AI models. These datasets are then integrated with LoRA (Low-Rank Adaptation) fine-tuning, which enables efficient integration at low training costs. It also significantly enhances the models’ applicability in cultural heritage preservation and reuse. This study uses Shanghai-style furniture as a practical case, combining generated images with comparative analysis against current mainstream image generation platforms and utilizing IPA for quantitative evaluation. Results demonstrate the model’s superiority in cultural symbol restoration, stylistic consistency, and image quality. This research not only advances the theoretical application of symbolic interactionism in cultural heritage preservation but also provides novel pathways and methodologies for digital heritage conservation. It validates the approach’s effectiveness and applicability across relevant fields, propels deeper AI integration into broader cultural heritage domains, while expanding future potential for cultural preservation and innovation.

2. Literature Review

2.1. Digital Preservation and Utilization of Cultural Heritage

Cultural heritage (CH) includes both tangible and intangible heritage. The former refers to those cultural artifacts, monuments, or sites with historical, esthetic, archeological, scientific, ethnographic, or anthropological interests [23]; the latter includes the various practices, performances, expressions, knowledge, and skills recognized by specific groups, communities, or individuals as their cultural heritage, as well as the associated bearers of heritage and cultural spaces.

Digital technologies have emerged as an important tool for the preservation of cultural heritage during recent years. In 2011, the World Intellectual Property Organization (WIPO) and the International Council of Museums (ICOM) signed a Memorandum of Understanding (MOU) [24] to establish a shared framework for the access to and dissemination of digitized cultural objects. In this process of creating, using, and maintaining digital heritage, three basic measures are essential, namely digitization, access, and utilization. Early digital preservation focused on the digital storage of documents and archives [25], such as converting paper documents to electronic data for preservation and accessibility [26]. With the continuous advancement of 3D scanning and modeling technologies, the discipline of research is gradually expanding into the field of 3D reconstruction and virtual presentation [27], using technologies like 3D scanning [28] and virtual reality (VR) [29,30] to reconstruct and present the structure and appearance of cultural heritage [31,32], to provide an immersive experience for people.

With the maturity of computer technology, AI methods such as image recognition and computer vision have further promoted innovations in cultural heritage protection. AI has been applied to the automatic classification and management of cultural relics [33], supported by the development of large-scale datasets that allow AI systems to learn and identify visual features of cultural heritage [34,35]. In addition, deep learning is applied in image recognition and text generation, enables AI to automatically process and recognize subtle features in cultural heritage data, enhancing the efficiency of cultural heritage management and display. More recently, the natural language processing ability of AI has been gradually enhanced; some scholars have begun to utilize this technology to parse the symbolic meanings in cultural heritage, trying to understand the historical and social values behind it [36].

Despite these advancements, several limitations persist. First, current digitization practices often overlook the symbolic and social dimensions embedded in cultural heritage. Cultural heritage does not only exist physically, but also carries people’s collective memory of history, religion, social values, etc. [37]. However, existing digitization methods have difficulty capturing such deep cultural connotations, leading to the loss of cultural uniqueness in the process of preservation and reuse. In addition, many digitization studies mainly rely on large-scale open-source datasets and lack data support from specific cultural contexts. This makes it difficult for digitization results to accurately reflect the unique appearance and value of specific cultural heritage [38]. Secondly, current AI models mostly focus on technical realization and application effects while lacking in-depth excavation of the symbolic meaning of cultural heritage. This feature makes it difficult for the outcome to truly reflect people’s deeper understanding of cultural heritage, leading to the fact that the digitization process, while recording physical features, often neglects the symbolic meaning and social impact of cultural heritage in different cultural contexts, and fails to present the value of cultural heritage in a comprehensive way.

Although AI models fundamentally lack perceptual capabilities, current technological developments suggest that even future perceptual AI may neither interpret “culture” and “heritage” in conventional human terms nor fully grasp these concepts. Moreover, given the longstanding divergence between the values embedded in tangible and intangible cultural expressions, digital cultural heritage—currently grounded exclusively in human cognition—is expected to encounter substantial challenges in near future. The human and machine cognitive dimensions of cultural heritage will coexist, arousing the interference between these two conflicting values for the same cultural heritage [39].

Based on this, future research should focus on how to digitally capture the deeper cognition of human beings towards cultural heritage, especially the symbolic and social significance of cultural heritage. Methods like grounded theory, Mental Models Theory, and Distributed Cognition will be used, aiming to dig into the unique cognitive patterns of scholars, researchers, designers, and other specific groups of people about cultural heritage. These cognitive insights can then be converted into structured data and introduced into existing AI datasets, enabling AI systems to develop more human-aligned cognitive systems in the cultural heritage domain and thereby enhancing their role in heritage transmission and adaptive reuse.

2.2. Cultural Heritage in the Perspective of Symbolic Interactionism

Symbolic interactionism focuses on how individuals construct and convey meaning through symbols in social interactions [40,41], and how they form their understanding of themselves and the world through the interpretation of symbols [42]. Based on this, cultural heritage can be regarded as a “symbolic corpus” accumulated in the process of human social interaction [43], which is not only a material existence, but also an intermediation between collective cognition and social meaning transmission. The theory emphasizes that the values and meanings of cultural heritage are not inherent, but are continuously generated, shared, and recreated through the interaction of different cultural backgrounds and groups.

At the practical level, cultural cognition involves recognizing, interpreting, and reconstructing the meanings embedded in symbolic systems within culturally specific environments. In terms of technical implementation, this process involves feature extraction, semantic annotation, and symbolic association modeling of cultural elements. Specifically, by training models to analyze how different social groups construct collective meanings through symbolic systems, such cultural representations can be converted into structured data for AI models to be processed and understood.

Symbolic interactionism thus provides both a theoretical foundation for cultural heritage studies and methodological implications for AI model design and training. By treating cultural heritage as a symbolic corpus, artificial intelligence can recognize the external characteristics of cultural elements and deeply understand the social and cultural meanings behind these elements. The core ideas of symbolic interactionism [44], especially the emphasis on the continuous generation and re-creation of cultural heritage among different groups, offer guiding principles for the model. It prompts the model to integrate the cultural perspectives of specific groups during training. This approach helps AI gradually grasp the in-depth cognitive understanding of cultural heritage by groups, thereby incorporating humans’ symbolic comprehension of culture into the model’s judgment and generation processes. As a result, the model demonstrates higher cultural sensitivity and humanistic care in the preservation and application of cultural heritage [45,46].

2.3. The Application of Generative Artificial Intelligence (GenAI) in Cultural Heritage Area

Generative artificial intelligence (GenAI) has gradually started to be widely used in the field of cultural heritage in recent years, profoundly reshaping the practices and management of cultural heritage digitization [47]. GenAI offers analytical and decision-support tools [48] that enable the rapid aggregation and synthesis of large-scale datasets [49] and facilitate the reconstruction of cultural heritage objects and sites [50]. However, on the other hand, publicly accessible GenAI models such as GPT-5.1, Google Gemini, or DeepSeek V3.2 are vulnerable to data bias, which can lead to biased decision-making, inequality, and discrimination. In the database used by ChatGPT, some of the sources have been proven to have limitations in empirical studies in the fields of archeology [51], cultural heritage preservation [52], and design [53]. Its data sources are relatively homogeneous, with an over-reliance on common English texts and a lack of multilingual specialized literature support, which leads to a lack of standardization of terminology and even systematic errors and biases, which makes it inappropriate in the field of heritage studies.

Nowadays, transformer architectures [54] have driven the rapid development of GPT [55] and diffusion models [56] (e.g., stable diffusion model). The diffusion model significantly improves the quality of image generation through potential diffusion algorithms, contributing to its wide application in social science fields such as cultural heritage, archeology, and design. In cultural heritage preservation, AI has been widely used in the reconstruction and restoration of digitized cultural artifacts, especially in high-precision digital reconstruction of historical sites and artifacts through Generative Adversarial Networks (GANs) and diffusion model [57], as shown in the last diffusion model diagram in Figure 1. These technologies enable inaccessible or fragile cultural heritage to be preserved and presented virtually, which facilitates the dissemination and sharing of global cultural heritage [58]. In archeology, AI improves the efficiency and accuracy of archeological research by analyzing ancient site images, relic fragments, and documents to assist archeologists in site excavation, relic classification, and historical reconstruction [59]. In addition, the application of AI in design has also made significant progress, especially in generative art, architectural design, and product design, where AI can provide designers with innovative inspiration and generate works that meet specific esthetic requirements by learning historical art styles and design patterns [60]. These applications extend methodological approaches within the social sciences and foster interdisciplinary innovation, offering new technological support for heritage preservation, archeological exploration, and design creativity.

However, most of these applications rely on the training of targeted large models and the use of techniques such as fine-tuning [61], which requires professionals to utilize strong computational resources for training, as shown in Figure 2.

Low-Rank Adaptation (LoRA) of Large Language Models addresses this limitation by freezing the primary weights of the pretrained model and optimizing only a set of low-rank matrices. This enables efficient learning of cultural features under constrained computational conditions [62]. Its mathematical formulation is as follows:

$$W^{\prime }=W+∆W,∆W=AB$$

(1)

Here, A ∈ Rd × r and B ∈ Rr We believe that an additional explanation of the dashed line frame is not necessary. The dashed lines simply highlight key steps and components within the model, which are essential for understanding the workflow. As these elements are already clear in the figure, we feel that no further clarification is needed. × kB are low-rank matrices, where r is much smaller than d or k, thereby reducing computational costs. Only a small portion of parameters require fine-tuning that can adapt the model to specific tasks [62], which significantly lowers resource demands. In this way, by using fewer data and computational resources, researchers with limited AI usage and training experience can achieve fine-tuning of large diffusion models and yield more practical training outcomes [63]. This method has already been applied in fields such as cultural heritage preservation [64]. Experimental results demonstrate that LoRA enables effective fine-tuning using curated collections of cultural artifact images and descriptions, accurately generating images with historical characteristics that align with historical data and expert evaluations [65,66,67].

Standing on previous research, the present study uses LoRA fine-tuning to incorporate group-specific cognitive structure into large AI models. This allows the models to approach cultural heritage interpretation from a human cognitive perspective, making their reasoning processes more aligned with human thought and cognitive systems. Consequently, the approach offers new methodological pathways for the preservation, transmission, and adaptive reuse of cultural heritage.

3. Materials and Methods

3.1. Case Study

Shanghai-style furniture emerged in the early 20th century in Shanghai, integrating traditional Chinese mahogany craftsmanship and Western Art Deco styles. Characterized by clean lines, symmetrical shapes, and the use of diverse materials, the style reflects Shanghai’s multiculturalism and modern esthetic values [68,69]. Because of its integration of Chinese and Western cultural elements, its stylistic evolution reflects broader social and cultural changes and enriches a symbol system including form, material, color, and decorative patterns. These characteristics make Shanghai-style furniture a suitable subject for cultural heritage digitization and an appropriate dataset for training and evaluating AI models based on cultural cognition.

To ensure the applicability, representativeness, and diversity of the sample, the data collection follows the principles of theoretical sampling and theoretical saturation. The research subject is practitioners in the field of cultural heritage digitization, so the interviewee should meet the following conditions: (i) have relevant experience in cultural heritage digitization or AI generative design (e.g., stable diffusion, LoRA, etc.); (ii) have participated in the research of cultural symbols, re-creation of cultural content, or related design practices, with verifiable research or application records, such as design proposals, thesis, lab reports, etc.; (iii) be able to fully understand the interview questions and clearly express personal experience and opinions.

The interviews were structured around the following themes: first, the interviewees were asked to describe their personal backgrounds including their careers, education, professional experiences, and practices in digitizing cultural heritage, describing their perception of specific cultural heritage (e.g., Shanghai-style furniture) and the use of cultural symbols. Then, the level of the interviewees’ perception of cultural heritage was explored, and how their core characteristics were defined; and finally, the interviewees reviewed the content of the interviews and summarized the key elements of the perception of specific cultural heritage.

From March to May 2025, 31 respondents (S01–S31) in total have completed formal interviews one after another, as shown in

Table 1. Interviewee background information.

ITEM	Value	Number	Percentage (%)
Sex	Male	13	41.93%
	Female	18	58.06%
Age (years)	20–29	12	38.71%
	30–39	11	35.48%
	40–49	5	16.13%
	50–59	2	6.45%
	60 and higher	1	3.22%
Professional background	Cultural heritage	10	32.25%
	Design	2	6.45%
	Artificial intelligence	5	16.13%
	Archeology	7	22.58%
	Sociology	7	22.58%
Educational background	Junior college and below	5	16.13%
	Undergraduate	7	22.58%
	Master	8	25.80%
	Doctor	11	35.48%

for basic information. In this study, the semi-structured narrative interview method was used to interview 31 respondents, and the interviews lasted an average of 40 min; all interviews were audio-recorded with the consent of the respondents and transcribed into text, obtaining a total of 156,901 words of transcribed text. After identifying the coded materials, this paper randomly selected 2/3 of the materials for bottom-up coding analysis with the help of Nvivo 12, and the remaining 1/3 of the materials were used for theoretical saturation testing.

3.2. Method

Theoretically framing symbolic interactionism and being rooted in grounded theory, this study constructs a cognitive indicator system by analyzing the layers of understanding of cultural symbols of cultural heritage by a group of practitioners. Expert opinions were subsequently integrated using the Delphi method to develop cultural symbol datasets specific to the selected heritage domain. Subsequently, LoRA fine-tuning technology was applied to integrate these cognitive datasets into the diffusion model to enhance the model’s ability to recognize and process cultural features. In the model development stage, the dynamic understanding and reuse ability of the AI system for cultural symbols is strengthened through the extraction of cultural feature vectors and the optimization training of the generative model. Ultimately, its effectiveness is verified through expert quantitative analysis. The specific research path in Figure 3 is as follows:

3.2.1. Cognitive Hierarchy Construction

To address AI’s shortcomings in cultural cognition, this study adopts procedural grounded theory to construct cognitive hierarchies. Grounded theory is a bottom-up research method [70], emphasizing extracting concepts through data analysis. It has evolved into the following three distinct schools since its inception [71]: classical grounded theory, grounded in positivism, emphasizes the objectivity and neutrality of research; constructive grounded theory integrates interpretivism and constructivism, focusing more on the interactive meaning-making between researchers and participants; procedural grounded theory, grounded in interpretivist philosophy, emphasizes the predictive and explanatory construction of theory through systematic coding steps. Its clear methodological structure and strong operationality make it particularly suitable for constructing complex, conceptually diverse cognitive systems. Compared to other grounded theory approaches, procedural grounded theory preserves the rich semantics of data while progressively refining a logically rigorous and structurally coherent theoretical system through the following three stages: open coding, axial coding, and selective coding. This aligns with the present study’s objective of constructing a group cognitive structure.

The symbolic interaction theory emphasizes that the meaning of cultural symbols is generated from social interaction [72], while the grounded theory provides a systematic methodology to explore the symbolic process of group cognition. Combining these two approaches enables the exploration of specific groups’ perceptions of cultural heritage from multidimensional and multilevel perspectives, gradually constructing their unique cognitive structures regarding cultural heritage. This methodological combination is particularly appropriate for exploratory cultural heritage research, as it effectively reveals group-specific cognitive patterns related to cultural symbols and supports the development of a dynamic framework for interpreting cultural heritage perceptions.

The implementation process includes the following key steps, as shown in Figure 4. First, performing open coding on the interview data of a specific group to identify the key concepts; second, performing axial coding on the basis of the initial concepts to establish categories and explore the logical relationship between the categories; and then using selective coding to further integrate the core categories and form a systematic cognitive hierarchy structure; finally, constructing a cultural cognitive hierarchy model based on the coding results to realize the integration of AI into the cognitive structure of specific groups under the cognitive structure of the big model.

3.2.2. Cognitive Data Labeling

When functioning as a theoretical model, the cultural cognitive hierarchy model can effectively reflect the cognitive hierarchy of cultural heritage among a specific group. However, it is still insufficient to meet the needs of cultural heritage digitization. The equivocality of cultural symbols and the differences in understanding in different contexts make it necessary to further match and calibrate the model with specific image datasets to ensure its effectiveness in practical applications. To address this need, the Delphi method is introduced to refine the model and enhance the accuracy of its correspondence with visual datasets. As a systematic and interactive method [73], Delphi method can predict expert consensus and therefore help make decisions. It can ensure the scientific validity and reliability of the model through multi-rounded anonymous feedback and opinion convergence [74]. In this study, experts in cultural heritage, professional practitioners, and researchers in artificial intelligence technologies participated in several rounds of evaluation [75], so as to transform the cognitive hierarchy of cultural heritage into the collective meanings of cultural symbols by utilizing expert consultation, as shown in Figure 5. Based on the cognitive hierarchy of cultural heritage constructed by the previous rooted theory analysis, this study transforms it into the multivariate expression structure of the dataset through multiple rounds of expert consultation and the dual annotation strategy of technology–culture. This process ensures the scientificity and consistency of the data and provides accurate corpus support for the intelligent reproduction and innovation of cultural heritage.

3.2.3. Model Training and Application to Cultural Cognition Datasets

Building upon prior research, a text-image mapping dataset incorporating cultural cognitive structures was developed. To achieve digital representation of cultural heritage and enhance AI’s generative capabilities within specific cultural contexts, the model was systematically trained and optimized. Following the workflow of mainstream image-generation tasks, the study performed standardized preprocessing and parameter optimization to ensure scientific rigor and training efficiency.

A pre-trained stable diffusion model was selected as the foundation and introduced LoRA (Low-Rank Adaptation) technology for efficient fine-tuning. By incorporating low-rank matrices into the pre-trained model weights, LoRA effectively reduces training parameters and computational resource consumption, making it suitable for rapid cultural style adaptation under small-sample conditions. Compared to traditional fine-tuning methods, LoRA preserved the original model’s general capabilities and enhanced training stability. Training epochs and learning rates were set appropriately, and the Adam optimizer was employed to enhance model convergence. Simultaneously, they leveraged the powerful cross-modal alignment capabilities of the CLIP model. Cultural prompts were accurately mapped to the latent space to guide the model toward generating culturally coherent images that align with semantic contexts, as shown in Figure 6.

The model training conducted a subjective evaluation and screening of its performance. Experts in culture and design were invited to compare and assess candidate models alongside images generated by mainstream platforms, and to assign comprehensive scores based on dimensions such as artistic merit, cultural symbol consistency, and innovation. To enhance the structured nature of the evaluation, the Importance-Performance Analysis (IPA) method was adopted. Experts were organized to rate satisfaction and importance across various dimensions of the patterns generated by the models, constructing an IPA matrix to clearly reveal the strengths and weaknesses of each model in the user experience.

This systematic training and optimization approach not only enhances AI’s adaptability in the field of cultural heritage digitization but also provides a viable practical pathway for future cultural symbol generation tasks involving the generation and interpretation of cultural symbols.

3.3. Data Encoding and Analysis

3.3.1. Open Coding

To ensure the rigor and objectivity of the coding, the coders were asked to complete the coding independently and record the coding ideas. The feedback was subsequently revised through continuous comparison, discussion, and exchange. The final experiment yielded 78 concepts (A1~A78) and 68 basic categories. It should be noted that due to the large amount of data reorganization, adjustment, and abstraction, the ordering of the labels, conceptualizations, and preliminary categorization contents in this stage do not strictly correspond to each other.

3.3.2. Axical Coding

Based on the characteristics of the research problem and the object, the basic categories can be clustered into six more abstract main categories—the initial perceptual experience of cultural heritage: the symbolic significance of cultural heritage; the historical background and evolution context of cultural heritage; the social and esthetic value of cultural heritage, and the practical application of cultural heritage in design and communication, the critical reflection, and reuse of cultural heritage. These categories together form the cognitive hierarchy of cultural heritage; the results of the main axis coding are shown in

Table 2. Results of main axis coding for cultural heritage cognitive hierarchy.

Main Categories	Subcategories (Preliminary Categories)	Significance
Initial perceptual experience of cultural heritage	Visual Impressions and Perceptions	Initial contact and sensory impressions of cultural heritage, including visual, auditory, tactile, and other sensory aspects of identification
	Tactile and auditory perception
	External form and representation
	Direct experience of cultural heritage
	Initial perceptual response
The Symbolism of Cultural Heritage	Cultural Symbols and Symbolic Meanings	Understanding of the symbolic meaning and value of cultural heritage, and forming a preliminary decoding of cultural connotations
	The Cultural Context of Artistic Symbols
	Symbolic interpretation of cultural heritage
	Cultural Identity and Symbolic Expression
	Symbolic understanding of historical culture
Historical Background and Evolutionary Context of Cultural Heritage	Historical Development of Cultural Heritage	Understanding the historical background, evolutionary process, and social context of cultural heritage deepens our perception of its temporality and regionality.
	Historical timeline and regional context
	Regional culture and historical sites
	Historical Imprints in Cultural Change
The social and esthetic value of cultural heritage	The social significance of cultural heritage	Understanding the multiple values of cultural heritage from multiple perspectives, including social, esthetic, educational, and ecological perspectives.
	Historical and cultural educational function
	Esthetic Value and Artistic Expression
	Social Impact of Cultural Heritage
	The role of cultural heritage in social identity
The practical application of cultural heritage in design and communication	Cultural Heritage Design Innovation	The ability to translate an understanding of cultural heritage into practical activities such as design, communication, and education, reflecting the transformation of cognition into action.
	Cultural Relics Display and Exhibition Planning
	Cultural heritage education and dissemination
	Combination of creative industries and cultural heritage
The critical reflection and reuse of cultural heritage	Critics of the Reuse of Cultural heritage	Possess critical reflection skills and be able to judge and reflect on the process of cultural heritage reuse, the distribution of discourse power, and cultural politics.
	Analysis of Discourse Power in Cultural Heritage
	Cultural Identity and Political Discourse
	The controversy over authenticity and cultural heritage
	Ethical issues in cultural preservation

.

3.3.3. Selective Coding

A final stage of the rooted theory research, selective coding, aims to identify and establish the core categories through systematic integration of results of open coding and axial coding, thereby constructing a complete theoretical framework. After systematic analysis, the hierarchical progressive mechanism of cultural heritage cognition was established as the core category. This category reveals the gradual deepening of cognition in the field of cultural heritage, reflecting the comprehensive process from perceptual perception to rational interpretation, practical application, and reflection. Each hierarchy of cognition is interconnected, promoting the gradual improvement and transformation of cultural heritage cognition. This hierarchical structure not only helps to analyze the various dimensions of cultural heritage but also translates the collective experience of a specific group into structured cognitive data. As

Table 3. Hierarchical model of cultural cognition.

Core Categories	Axial Categories	Specific Cognitive Categories
Mechanism of hierarchical progression in the perception of cultural heritage	Perceptual Cognition	Visual Impression
		Sensory Perception
		Visceral Reaction
		Cultural Perception
		Eternal Feature
	Symbolic Cognition	Cultural Labels
		Symbolic Significance
		Semiotic Interpretation
		Symbolic Connotation
		Historical Symbol
	Historical Cognition	Historical Overview
		Time Characteristic
		Regional Culture
		Evolutionary Process
	Value Cognition	Cultural Value
		Educational Function
		Esthetic Experience
		Social Value
		Identification
	Applied Cognition	Design Utilization
		Exhibition Planning
		Teaching Translation
		Cultural Creativity
	Critical Cognition	Cultural Reproduction
		Discourse Construction
		Immigration Critics
		Authenticity Issues
		Ethical Reflection

shows, this framework provides a cognitive foundation for the AI model to support more accurate cultural heritage protection and dissemination.

3.3.4. Theoretical Saturation Test

This study strictly adhered to grounded theory requirements and theoretical saturation guidelines [76]. The researchers compiled, coded, and analyzed data immediately after each interview, continuously iterating between data collection, coding, analysis, theory construction, and theory comparison. By the 31st respondent, no new concepts, categories, or relationships emerged. To ensure that the theory was saturated and the sample was sufficiently rich, the research still conducted 10 more interviews. However, these also yielded no new concepts, categories, or relationships. Additionally, 30% of the materials were re-coded after a 5-month interval, and the Kappa coefficient and coding consistency percentage were calculated to evaluate the reliability and consistency of the coding. The results showed a Kappa coefficient of 0.82 and a coding consistency percentage of 90.6%, which met the consistency standards. Therefore, the theoretical testing in this study has met the saturation requirements.

3.4. Image Dataset Construction

3.4.1. Image Acquisition and Cleaning of Shanghai-Style Furniture

With the authorization from Shanghai Jinze Craft Society, an on-site photographic collection of its Shanghai-style furniture was conducted, obtaining more than 700 photos. After preliminary screening, 303 clear and complete image samples were selected, covering the following three categories: full-form chair images; characteristic structural views of typical components; and decorative pattern illustrations.

To ensure data quality for subsequent image model training, all images were standardized through the following procedures: (1) each image contained only one single category of furniture and underwent background removal to ensure a pure white background; (2) image resolution was uniformly standardized to 512 × 512 pixels; (3) sample quantities across the three categories were balanced to avoid class imbalance; (4) a total of 143 morphological diagrams required complete furniture appearance representation; 87 structural feature diagrams emphasized woodworking structures and material texture details; 73 ornamental diagrams ensured clear pattern discernibility. These curated image samples form the foundation of the digital pattern database used in this study, as illustrated in Figure 7.

3.4.2. Classification of Shanghai-Style Furniture Image Dataset

To further refine the classification of image samples, the dataset was classified according to the major stylistic schools of Shanghai-style furniture. Drawing on existing scholarship regarding the historical development and stylistic evolution of Shanghai-style furniture, samples within each category were systematically identified and quantified. This annotation system exhibits high rigor and systematicity, not only covering the major style of Shanghai-style furniture and their typical characteristics but also aligning with authoritative research on the developmental evolution of Shanghai-style furniture, as illustrated in Figure 8.

Shanghai-style furniture encompasses major styles including British Neoclassicism, French Rococo, Eclectic (Victorian), European Baroque, and American Art Deco. This classification reflects its inheritance of Ming and Qing classical furniture traditions alongside the fusion of modern Chinese and Western design characteristics. Within each style category, its typical structural forms and decorative pattern elements were further subdivided and annotated.

Neoclassical Shanghai-style furniture emphasizes clean lines and balanced proportions, retaining the restrained elegance of traditional mortise-and-tenon craftmanship. French Rococo style excels in its voluptuous, flowing curves and intricate, ornate carvings, exemplifying the luxurious and romantic artistic flair of 18th-century French court furniture. Eclectic furniture draws heavily from Western classical designs, incorporating Rococo and Baroque curved elements into Ming- and Qing-inspired forms, while simplifying ornamental excess to highlight the natural grain of rosewood. European Baroque furniture features heavy, ornate forms and exaggerated decorations, pursuing dramatic visual impact and majestic grandeur. Art Deco style, characterized by innovative, fashionable forms and clean, sharp lines, incorporates industrial-era design concepts. It introduced novel furniture types, such as sofas, dressing tables, wardrobes, and swivel chairs, which are rare in traditional Chinese contexts, alongside new materials like glass and metal, becoming a quintessential fashionable furniture movement in 20th-century Shanghai.

This classification system fully integrates the eclectic cultural attributes of Shanghai-style furniture with academic summaries of its stylistic characteristics, as

Table 4. Stylistic classification and characteristics of Shanghai-style furniture.

Stylistic Schools	Morphological Images	Structural Feature Images	Decorative Pattern Images
Neoclassical Style	Furniture has symmetrical shapes and regular structures, with an overall elegant and restrained style	Common conical grooved legs, straight legs, and structural lines are straight and concise	Simple classical patterns such as plant wreaths and oval frames, with coordinated proportions between decoration and structure
French Rococo Style	Curves are smooth, shapes are light and gorgeous, with common shell-like and scroll-like decorative shapes	Decorative components such as curved chair legs (S-shaped legs), beast hoof feet, and carved flower holders.	Decorative patterns such as scrolls, roses, and bows, with exquisite and complex carvings and soft lines
Eclectic Style	Combines multiple Western classical elements, with mixed forms and grand or changeable shapes	Combinations of various European decorative leg types, such as curved legs, cylindrical legs, and dental plates	Integrates decorative motifs such as Gothic, Baroque, and Rococo. Typical patterns include flowers and leaves, flying apsaras, mythological patterns, etc.
Art Deco	Strong sense of geometry, simple lines, neat volume, and a modern sense	Use of metal parts, glass, straight legs, and new types of plates, such as plywood	Abstract geometric patterns, radial, stepped, and parallel line compositions, highlighting modernist design concepts
Chinese–Western Fusion Style	Chinese frames embed Western functional elements, with coordinated proportions and diverse styles	Western-supporting components are combined within Chinese frames, such as claw feet and appliqué legs	Integrates traditional Chinese patterns (such as Shou character, peony, auspicious cloud) with Western flowers and scroll patterns to form a mixed decoration system

shows. Through this multi-layered annotation approach, logical clarity and systematic completeness can be ensured in image classification. It highlights traditional craftsmanship esthetics while reflecting the fusion of Chinese and Western design features, demonstrating the annotation system’s professionalism and cultural adaptability. Each style category exhibits distinct formal, structural, and ornamental characteristics reflective of their historical and esthetic lineage.

3.4.3. Expert Group

Eighteen experts were invited from related fields like cultural heritage, design science, artificial intelligence, and cultural and creative industries to ensure a balanced coverage of diverse areas. The expert team members include cultural heritage scholars with experience in cultural heritage preservation, non-heritage research, and in-depth understanding of cultural symbols, craftsmanship, and historical context; design experts who focus on the inheritance and innovation of cultural symbols, and are concerned with cultural heritage re-creation and digitalization adaptability; AI researchers who are familiar with deep learning, computer vision, and generative AI technologies, and can ensure that the dataset can be used for model training; museum and cultural creation industry experts with practical experience in the cultural content industry, concerned with the social value of cultural heritage and communication models. The construction of the expert panel ensures interdisciplinary cross-fertilization and provides a multi-dimensional academic perspective for this study.

3.4.4. Expert Interviews and Multiple Rounds of Views Convergence

Three rounds of the Delphi method were conducted to achieve expert consensus and ensure the reliability of the dataset annotations. First, the experts conducted structured consultations on the different cultural cognitive hierarchies of each picture, provided standardized short word labels based on the cultural heritage cognitive structure, and proposed preliminary style classification. Feedback was collected via an open-ended questionnaire and integrated into the corresponding categories in the cultural heritage cognitive hierarchy. Cluster analysis was then conducted to identify areas of consensus and divergence. In the second round of interviews, the research team summarized the core labels based on the feedback from the first round and calculated the consistency of experts’ opinions (Kendall-W) to determine the degree of convergence of experts’ opinions through quantitative analysis.

For images with high divergence, the research team provided the categorization opinions of the experts from the previous round and invited the experts to reevaluate and adjust based on these categorizations. In the third round, experts conducted a final review of the cultural cognition labels refined in the first two rounds to ensure that all labels conformed to the cultural cognitive structure and remained standardized in their descriptions. When the consistency coefficient reached a set threshold (W > 0.75), the research team identified the final set of cultural labels and used them as the core component of the dataset.

3.4.5. Textual Dataset Construction

During the data organization phase, the research team further screened the labels provided by experts, removing redundant or ambiguous tags while merging similar labels from different experts to ensure data usability and consistency. The final curated label set adopted short-word formats. Based on empirical experience in training generative models, effective dataset labels can be categorized into the following three types: trigger words, feature labels, and distractor labels as shown at

Table 5. Label set categories.

Category	Introduction
Trigger word	The trigger word is usually a special character, such as pinyin. Concepts that cannot be understood in the image features to be learned by the model correspond to this word. Inputting the trigger word will cause image features related to the trigger word to appear in the generated image, such as dragon and phoenix carvings, inlaid studs, etc.
Feature label	Feature labels are labels that are strongly associated with trigger words, such as curved leg, Victorian carving, wooden door, sofa, closet, etc. More is not necessarily better; only easily recognizable features are necessary, such as sofa, closet, coffee table, etc.
Disturbance Label	Interference labels are screen elements like white backgrounds, clutter, etc., which are non-furniture related. These elements need to have labels applied to additionally.
Negative Label	Negative labels are tags established to further enhance the effectiveness of model training, such as poor quality and damage. Their role is to prevent the model from generating undesirable output results.

. All image samples were assigned standardized labels and organized according to a cultural cognitive hierarchy, as shown in

Table 6. Sample structured annotation of furniture images for cultural cognition hierarchies.

Image	Image Label
	Trigger Word	Feature Labels	Disturbance Label	Negative Label
	Art Deco Furniture	chair; streamlined; geometric; functional; oval solid wood armrest; roller shape; leather upholstery; brown leather	white background	low resolution, crack, stain, blurry texture, noise, overexposure, blurry edge, deformed structure, blurry detail, scratch
	Art Deco Furniture	dining cart; geometric lines; modernism; glass countertop; concise lines; open structure; wooden cabinet; metal accessories; metal support rod; curved metal handle; roller	white background	low resolution, crack, stain, blurry texture, noise, overexposure, blurry edge, deformed structure, blurry detail, scratch
	Chinese–Western Fusion Style Furniture	dressing cabinet; large round mirror; wood carving decoration; geometric lines; line relief; cloud and crane pattern; solid wood cabinet; square cabinet feet	white background	low resolution, crack, stain, blurry texture, noise, overexposure, blurry edge, deformed structure, blurry detail, scratch

.

4. Results

4.1. Environment and Parameter Configuration

After constructing the Shanghai-style furniture image and symbol database and organizing it into a training set, this study conducted model training for specific cultural recognition in a low-resource environment. Stable-diffusion/reaisticVisionV60B1 v5lVAE, safetensors foundational model was selected as the pre-training framework, with NVIDIA RTX 5000 GPU (16 GB VRAM), Intel i9 10th generation CPU, and 128 GB RAM computational environment. Key parameters during training were systematically adjusted to ensure stable convergence under constrained computational resources, including repetitions, training epochs, batches, learning rate, and optimizer settings, etc. Training was initiated under this configuration, ultimately yielding a set of LoRA fine-tuned models tailored for Shanghai-style furniture. This establishes the technical foundation for subsequent generative experiments and cultural symbol reproduction, as shown in

Table 7. Training configuration.

Classification	Parameters	Classification	Parameters
model_train_type	Sd-lora	max_train_epochs	15
pretrained_model	Stable-diffusion/reaisticVisionV60B1_v51VAE.safetensors	train_batch size	1
resolution	512 × 512	unet_lr	1 × 10−4
output_name	Shanghai-style furniture	text_encoder_lr	1 × 10−5
network_dim	64	Ir_scheduler	cosine_with_restarts
network_alpha	32	Ir_scheduler_num_cycles	1
Save the model every N epoch.	2	optimizer_type	AdamW8bit

.

4.2. Training and Testing of Stylized LoRA Models

Following model training, five LoRA models was obtained, namely Shanghai-style furniture 000001000.safetensors—Shanghai-style furniture.safetensors. These models were imported into the Stable Diffusion platform for evaluation. Model performance was examined systematically using the X/Y/Z parameter matrices to assess whether the generated outputs accurately reproduced the core stylistic attributes of Shanghai-style furniture.

During the testing, the X-axis represented model category while the Y-axis indicated weight parameter (Strength). When Y values fell within 0.0 to 0.4, generated images displayed low stylistic relevance to Shanghai-style furniture, scarcely reflecting typical symbolic features. As Y weights increased to the 0.4 to 0.8 range, models began gradually incorporating distinctive elements of Shanghai-style furniture. Models integrated in the fourth and eighth round can clearly present typical forms and decorative symbols of Shanghai-style furniture, as shown in Figure 9.

However, different models revealed significant differences in generalization ability and stability: some showed insufficient generalization—images derived from out-of-training-set prompts exhibit a deficiency in the requisite cultural consistency; the others exhibited overfitting, resulting in limited diversity in generated outputs. Based on a synthesis of quantitative metrics and expert blind reviews, the fourth model was selected as the core model for subsequent experiments and applications due to its optimal balance between cultural symbol fidelity, stylistic stability, and generalization capability.

4.3. Superiority Evaluation of Generated Images

To further validate the effectiveness of the cultural cognition model proposed in enhancing cultural-symbol understanding, a comparative analysis was conducted against other mainstream open-source generative models. Additionally, Importance-Performance Analysis (IPA) method was introduced to establish a relevant evaluation system and obtain more structured quantitative evaluation results [77]. The study discarded traditional objective evaluation methods such as Fréchet Inception Distance (FID) and Inception Score (IS) [78]. The primary reason for this approach is that the study focuses on identifying shortcomings in AI’s cultural cognition capabilities and proposing optimization strategies. Cultural cognition is inherently an abstract process, rooted in human interpretive processes rather than being measurable through conventional deep-learning metrics. Therefore, combination of expert subjective evaluation and quantitative assessment was deemed more appropriate for evaluating the model’s performance in cultural cognition.

4.3.1. Comparison and Evaluation of Mainstream Platforms

To evaluate the relative performance of our research model, a panel of fifteen experts in cultural heritage and design were invited to review. A benchmarking analysis was conducted against leading commercial generative systems, including OpenAI’s DALL·E 3 and Midjourney (v5). Through collective discussion, the panel identified three sets of prompt phrases for generating Shanghai-style furniture that align with cognitive patterns and are commonly used in image generation. Corresponding image sets were generated on each platform for blind review alongside outputs from the model. Experts evaluated images from different sources across six cognitive dimensions within the core cognitive model structure, without prior knowledge of each image’s origin, ensuring impartiality in the comparison.

Comparative analysis reveals notable limitations among mainstream general-purpose models, due to their lack of training specific to particular cultural styles, although Midjourney generates visually exquisite and realistic images of Shanghai-style furniture. However, the forms and arrangements of certain typical patterns do not fully align with authentic Shanghai style; DALL·E’s outputs occasionally display stylistic inconsistencies or misapplication of cultural elements. Doubao, as a Chinese image generation model, demonstrates a relatively stronger performance in capturing cultural characteristics in its Shanghai-style home furnishings. However, it sometimes exhibits structural errors, overly prominent patterns, and poor stylistic consistency. In contrast, the fine-tuned model developed in this study performs significantly better in terms of cultural symbol richness and stylistic coherence. The generated patterns exhibit a high degree of alignment with the expected design requirements, as shown in Figure 10.

While large models like Midjourney retain advantages in image clarity and diversity, the approach conducted in this study achieves comparable visual quality while maintaining cultural accuracy. Expert evaluations confirm that the proposed model’s performance outperforms culturally untrained general-purpose models in key dimensions including cultural style alignment. This result validates the effectiveness of integrating cultural cognition structure into generative models and enhances the model’s potential for increasing image quality and cultural consistency.

4.3.2. Importance-Performance Analysis

To further structure the evaluation, the Importance-Performance Analysis (IPA) method was employed to further analyze the expert scoring results.

IPA was proposed by Martilla and James in 1977 [79] to measure the relative importance of product/service quality attributes. It compares the perceived importance and perceived satisfaction of evaluation indicators within a system, plotting perceived importance on the x-axis and perceived satisfaction on the y-axis. Using the mean value of the importance variable as the x-axis reference line and the mean of the satisfaction variable as the y-axis reference line, the plane is divided into four quadrants. Each factor is then mapped to its respective quadrants based on their mean values, forming an importance-performance matrix that helps prioritize product improvements. This matrix serves as a critical basis for strategic design and resource allocation.

In this study, IPA is employed to systematically analyze user satisfaction with patterns generated by the cultural cognition model for Shanghai-style furniture. This analysis further validates the model’s value in promoting the inheritance and dissemination of Shanghai-style furniture, a Chinese intangible cultural heritage.

• Experiment Process

An on-site experiment was designed to evaluate generative models enhanced with cultural cognition. A total of 125 professionals from fields including design, fine arts, cultural communication, intangible cultural heritage preservation, and traditional crafts were recruited to participate in the survey. Under researcher supervision, participants completed Shanghai-style furniture image generation tasks using personal computers equipped with Intel i9 processors and NVIDIA RTX 5000 (16 GB) graphics cards. Immediately after completing the tasks, participants were asked to fill out questionnaires to ensure the immediacy and validity of experimental data. This method enabled researchers to observe participants’ operational processes in real time, thereby guaranteeing the scientific rigor and credibility of experimental outcomes.

• Questionnaire design

The questionnaire design integrates cultural cognition models with the Unified Theory of Acceptance and Use of Technology (UTAUT2) [80] and the System Usability Scale [81]. This combination was to systematically evaluate users’ technical acceptance, system usability, and cultural cognition consistency during model fine-tuning. Based on these theoretical foundations, a multi-dimensional user satisfaction evaluation system was constructed, as shown in Figure 11, and developed a questionnaire designed using a Likert 5-point scale. This design enables a comprehensive examination of user experience and satisfaction across multiple evaluative dimensions.

• Experiment

The experiment was conducted in July, with 125 questionnaires distributed and 123 valid responses collected. Prior to the formal survey, the research team provided participants with a detailed explanation of the study objectives and model applications, introducing the introduction to LoRA model, Shanghai-style furniture culture, and related concepts.

• Analysis

Statistical analysis of the questionnaires was performed using SPSS 26.0. The KMO value was 0.869 for the satisfaction section and 0.914 for the importance section, with Bartlett’s sphericity test yielding 0.000 for both. This indicates strong construct validity of the questionnaire. In reliability analysis, Cronbach’s alpha coefficient for satisfaction was 0.796 and 0.964 for importance, both exceeding 0.7, indicating good internal consistency and reliability.

The average satisfaction and importance scores were then calculated for each dimension, and an Importance-Performance Analysis (IPA) matrix was then constructed. As shown in Figure 12, the overall average satisfaction score was 3.93, while the average importance score was 3.75. The horizontal axis represents attribute importance, and the vertical axis represents satisfaction. Through the IPA matrix, the model’s performance across different cultural cognitive domains is clearly categorized into four quadrants: Strengths (Quadrant I), Maintenance (Quadrant II), Opportunities (Quadrant III), and Improvements (Quadrant IV).

IPA reveals a different distribution of the model’s performance across different cultural cognition dimensions. Dimensions are positioned in the Advantage and Maintenance quadrants, indicating that the model has achieved high user recognition in areas such as perceptual cognition, symbolic cognition, historical cognition, and performance expectations. Particularly regarding cultural characteristics associated with Shanghai-style furniture, the model performed accurate identification and reproduction, receiving positive evaluations for its generation efficiency and learnability. Overall, these strengths should be maintained and further consolidated in future applications.

Factors in the Opportunity Quadrant—including critical cognition, effort expectations, hedonistic motivation, and consistency—show relatively lower importance and satisfaction, yet still hold optimization potential in areas like system interface style and entertainment interactions. These findings suggest these domains may warrant further exploration and refinement in subsequent research.

The Improvement Quadrant reveals several critical issues, including interpretable presentation, insufficient contextual transfer operations, excessive complexity, and cumbersome operational paths. These issues directly involve core cognitive domains such as sensory perception, authenticity concerns, design application, and symbolic interpretation.

To address these weaknesses, future improvements should focus on expanding the model’s application scenarios, enhancing its universality and adaptability across diverse domains. Simplifying user interfaces could lower accessibility barriers, while integrating lightweight or efficient inference techniques to accelerate generation speeds would further optimize the overall operational experience. These enhancements are expected to strengthen the model’s performance across cultural cognition hierarchies, thereby providing more comprehensive support for the digital preservation and innovative transformation of Shanghai-style furniture.

The findings of this study demonstrate the integration of LoRA fine-tuning with optimized cultural cognition hierarchy. The combination significantly enhances the diffusion model’s performance in cultural cognition augmentation. The generated cultural heritage images not only align with cultural logic but also satisfy visual requirements, offering broad applicability and advantages for the digital preservation of cultural heritage and cultural innovation design.

Previous research findings also support these conclusions. Wang and Zhou successfully achieved the digital preservation and innovative interpretation of traditional Blue Clamp-Resist Dyeing by integrating LoRA fine-tuned models [82]. Zhou et al. proposed an innovative low-computational-power generation method, using LoRA to fine-tune diffusion models for generating novel kite designs, thereby promoting the revival and innovation of kite-making craftsmanship [83]. These studies indicate that LoRA technology not only has significant advantages in reducing computational resource requirements but also effectively facilitates the innovation and inheritance of intangible cultural heritage.

5. Discussion

5.1. Theoretical Significance

This study makes a theoretical contribution to the application of generative artificial intelligence (AI) technology in the dynamic transmission of intangible cultural heritage. It breaks through the limitations of the museum-like static display caused by the object restoration [84] model of traditional digital preservation through the mediation of cultural cognition model. Based on symbolic interactionism, a cultural cognition hierarchy model was constructed based on group perceptual structures, and a diffusion model was fine-tuned with LoRA technology. This integration enables AI to engage more effectively with the semantic vitality of culture in specific social contexts and its communication adaptability. The framework also reinforces ethical constraints on Al ensuring that cultural symbols are neither distorted nor misappropriated during generation, thereby preventing cultural bias and discriminatory outcomes. It provides a theoretically interpretable, technically reproducible, and practically adaptable solution for the modern representation of intangible cultural heritage.

In addition, this study responds to the academic criticisms of generative AI in terms of its ability to understand cultural contexts [85]. For the first time, cultural cognitive structures derived from human groups are directly transformed into model-training inputs, realizing an effective interface between collective human understanding and AI generative logic. This innovation not only expands the boundaries of cultural heritage digitization research but also introduces a new research paradigm for the interdisciplinary application of AI in humanities and social sciences. By supporting the dissemination and innovation of intangible cultural heritage technically, the study provides systematic theoretical support for realizing both presence and regeneration of cultural heritage in the digital era.

5.2. Practical Significance

This study proposed a replicable and low-threshold cultural heritage digitization and protection path, which provides a practical technical solution for the innovative transformation of intangible cultural heritage. By constructing a cultural cognition model and integrating it with LoRA fine-tuning technology, this study breaks through the previous longstanding technical barriers in cultural heritage digitization that are highly dependent on AI expertise and high-performance computing environments [86]. The approach significantly reduces computational resource requirements, alleviates hardware dependencies, and enables cultural heritage work to be carried out effectively under limited-resource conditions, thereby supporting the sustainable development of this field.

This empowers cultural heritage researchers, even those without AI development experience, to independently carry out high-quality generative cultural content production. Driven by cultural cognition as the core, this generative AI approach not only improves the quality of cultural heritage generation in digital scenarios but also creates new opportunities for expanding the application of intangible cultural heritage content. These opportunities cover the fields of education, public communication, and immersive display, etc. In addition, this study achieves the controllable application of generative artificial intelligence through local deployment, thereby mitigating risks associated with uploading cultural data to external platforms [85]. Localized deployment safeguards cultural ownership, restricts data usage to authorized contexts, and enhances credibility through human-in-the-loop verification mechanisms. This significantly reduces the risk of reliance on external platforms, ensuring the authentic reproduction and respect of cultural symbols and heritage content, and enhancing the security and reliability of digital heritage workflows.

The method exhibits strong versatility and generalizability, making it adaptable to different types of cultural heritage digitization projects across different regions. The data-labeling system and model training process constructed in the research has strong modularization characteristics, enabling rapid adaption and reuse in diverse cultural heritage scenarios. For museums, cultural centers, non-genetic heritage bases, and universities, this research offers a new paradigm for them to realize in-depth presentation, contextual interpretation, and public dissemination of cultural content under limited resources, while laying a foundation for integrating cultural heritage with science and technology education, and realizing the innovative transformation of cultural heritage.

In sum, this study not only responds to the contemporary demand of “making cultural heritage come alive” but also provides practical and easy-to-operate digital technology solutions for the protection, research, and dissemination of cultural heritage, which has important practical value and wide application prospects.

6. Conclusions

This study proposes a generalized, low-threshold, and replicable pathway for the digital protection and innovative transformation of cultural heritage through generative artificial intelligence. It focuses on the construction and application of cultural cognition models. This study responds to criticisms from some scholars. Those criticisms point out that existing cultural heritage digitalization pathways overly focus on the reproduction of physical characteristics, neglecting the in-depth exploration of cultural semantics and social meanings [87]. Based on the cultural cognitive structure of symbolic interactionism, this research offers a new theoretical and methodology perspective. By integrating cultural symbols and cognitive datasets into generative model, the approach moves beyond surface-level reproduction and enables the capture and regeneration of underlying cultural and social meanings. This methodological innovation demonstrates strong adaptability and universality, especially for the preservation and dissemination of intangible cultural heritage across diverse cultural contexts.

Compared with traditional AI model training methods [88], LoRA can significantly reduce the demand for computational resources. This enables more cultural institutions and researchers with limited resources to participate in the digital preservation of cultural heritage. This technological innovation aligns with academic calls for lowering barriers to Al adoption and promoting inclusive cultural-technological participation [89]. The approach also provides new methodological insights for educational communication, public exhibition, and cultural-creative development, opening up fresh possibilities for the digital activation of intangible cultural heritage.

Nonetheless, the study still has several limitations. First, the current dataset focuses on a single cultural style, limiting the validation of model generalizability across broader applications and diverse cultural types. Second, although the cultural cognition hierarchy is constructed based on a large number of expert opinions, it is still affected by subjective factors like semantic ambiguity and cross-cultural understanding, and more interdisciplinary perspectives need to be introduced to enhance the stability and adaptability of the cognitive system. Furthermore, while LoRA significantly reduces training costs, its performance in high-complexity scenarios or multimodal-data integration still requires improvement.

Future research can expand the model training datasets to cover more ICH (intangible cultural heritage) categories and diverse regional cultural styles, so as to verify the cross-cultural adaptability of the method. Meanwhile, combining multi-model data structures (e.g., voice, action, video, etc.) could further enrich the means of cultural semantic modeling and promote the expansion of culturally generated content towards greater interactivity and immersion. In addition, developing an open cultural knowledge collaboration platform and encouraging cultural workers, AI researchers and the public to participate in data construction and algorithm optimization will help to realize the cultural co-creation mechanism empowered by AI.

Overall, this study provides a systematic concept and practical pathway for the digital transformation process of cultural heritage from understanding to generation. It not only represents an innovative response of cultural heritage science in the context of AI era but also lays an important foundation for the promoting of the integration of the protection of cultural diversity with technological advancement.