Situated cognition theory

Situated cognition theory posits that knowledge acquisition is an activity deeply embedded in sociocultural environments (Goodwin, Reference Goodwin2000), rather than a simple information transmission process. This theory connects with cultural heritage transmission through three core points:

1. 1. Authentic contexts: Learning in real or simulated environments is more effective thrning (Lave and Wenger, Reference Lave and Wenger1991; Giamellaro, Reference Giamellaro2017; Li et al., Reference Li, Ch’ng and Cobb2023; Troussas et al., Reference Troussas, Papakostas, Krouska, Mylonas and Sgouropoulou2024);

2. 2. Social interaction: Learning is a social participation process that promotes understanding through interaction with others (Gasson and Waters, Reference Gasson and Waters2018; Nazlidou et al., Reference Nazlidou, Efkolidis, Kakoulis and Kyratsis2024; Cosentino et al., Reference Cosentino, Anton, Sharma, Gelsomini, Giannakos and Abrahamson2025);

3. 3. Practice participation: “Learning by doing” creates a deeper understanding than passive information reception (Danielsson and Linder, Reference Danielsson and Linder2009).

Based on situated cognition theory, recent research demonstrates that experiential learning has significant advantages in cultural heritage transmission. Rather (Reference Rather2020)’s research proves that compared to traditional “knowledge-based” approaches, “experiential” participation increases cultural memory retention by 47%. Yu et al. (Reference Yu, Wang, Cai, Zhang, Yu, Yang and Zhang2024) developed the MakeBronze system that enhances children’s understanding of bronze culture through hands-on experience, increasing participation willingness by 68%.

Co-creation theory

Co-creation theory originates from the service design field, emphasizing user participation in value creation processes (Ercsey, Reference Ercsey2017). Applied to cultural heritage, co-creation presents three key characteristics:

1. 1. Multi-party participation: Cultural transmission is no longer an exclusive activity of experts but a collaborative process among multiple parties (Paschen et al., Reference Paschen, Paschen, Pala and Kietzmann2020);

2. 2. Value co-construction: Participants jointly construct understanding of culture through interaction processes (Ji, Reference Ji2023);

3. 3. Innovation integration: Creating new cultural expressions by fusing traditional and modern elements (Sipiran, Reference Sipiran2024; Waidyaratne et al., Reference Waidyaratne, Noisri, Srisupakwong, Wijayasekara, Ratchatakorntrakoon, Bongsasilp, Vanichchanunt and Wuttisittikulkij2024).

Recent research demonstrates that co-creation paradigms have unique value in cultural heritage protection.

Stefanidi et al. (Reference Stefanidi, Partarakis, Zabulis, Adami, Ntoa and Papagiannakis2022) developed a virtual traditional craft platform that enhanced user understanding of traditional crafts through participatory design. Lee et al. (Reference Lee, Park and Song2019) promoted user exploration of cultural data through visual interactive systems, strengthening cultural identity.

Based on co-creation theory, and considering the integration of AI systems in contemporary digital heritage contexts, this research adopts the following core terminology:

“AI Collaborative Creation” refers to the process in which human users and AI systems jointly complete the creation of cultural content through multimodal interaction (textual descriptions, images, parameter adjustments), iterative refinement, and value co-construction. This term is distinct from “AI-assisted generation,”which emphasizes unidirectional content generation by AI, whereas AI collaborative creation emphasizes bidirectional interaction, mutual contribution, and shared creative agency between human and AI.This definition operationalizes co-creation theory in AI-enabled cultural heritage contexts, forming the theoretical foundation for our “Creation-as-Transmission” paradigm.

In summary, existing research has validated the respective values of situated cognition and co-creation theories, but lacks an integrative design framework that systematically combines both theories and is specifically designed to guide cultural heritage “learning and transmission through creation.”

Cultural content generation applications

Existing research primarily employs AIGC as a professional tool for the generation, restoration, and reconstruction of cultural content. These applications have achieved significant progress in technical implementation. For example, Yang (Reference Yang2024) combined Transformer and U-Net for Dunhuang mural restoration, achieving intelligent filling of damaged areas; Sun et al. (Reference Sun, Chi and Liu2024) applied Stable Diffusion to generate Tujia brocade patterns, providing creative inspiration for designers; Zeng (Reference Zeng2024) utilized Midjourney to reconstruct 3D models of architectural heritage from historical photographs. Additionally, Liu (Reference Liu2020) developed MoCapGPT to capture traditional craft movements, while Smirnova et al. (Reference Smirnova, Grafeeva and Tokman2024) developed ancient manuscript text recognition systems. These technologies have expanded the application boundaries of AIGC in the cultural heritage protection field.

To systematically evaluate the characteristics and limitations of these applications, Table 2 provides a comparative analysis from dimensions including user participation, cultural fidelity assurance, and application objectives.

Table 2.

Comparison of this study with existing AI cultural generation systems for intangible cultural heritage

Note:

• • User participation method: None = Pure technical tools; Limited = Parameter adjustment; “Deep participation (in this studyUser participation method: None = Pure technical tools; Limited = Parameter adjustment; “Deep participation (in this study).

• • Cultural fidelity guarantee: The main mechanisms by which each study maintains cultural characteristics.

• • Main limitations: Insufficiency from the perspectives of user participation and cultural inheritance.

• • Differences in this study: Highlighting the unique value of AIMagCreations.

Comparative analysis reveals that while existing systems demonstrate superior technical performance, they exhibit significant deficiencies in cultural transmission. First, most systems position AIGC as auxiliary tools for professionals, where the general public can only passively receive generated content, thereby constraining the widespread dissemination of cultural heritage. Second, research predominantly focuses on optimizing generation quality (e.g., FID, PSNR metrics) while neglecting users’ cultural cognition enhancement during interaction, lacking systematic evaluation of transmission effectiveness (Li et al., Reference Li, Wang, Deng, Pan and Chen2021). Furthermore, although systems are trained on cultural data, they lack effective constraints of cultural rules (Sun et al., Reference Sun, Chi and Liu2024). These limitations reflect a prevalent disconnect between “technology-oriented” approaches and “transmission needs” in current systems. In response, this research proposes a “user-centered” participatory paradigm that systematically addresses issues of insufficient user engagement and weak educational functionality through platform-based tools, cognition-oriented evaluation, and rule-based generation.

AI collaborative creation system research

Compared to the unidirectional technical support of cultural content generation applications, AI collaborative creation systems emphasize bidirectional interaction between users and intelligent systems, providing participatory pathways for cultural transmission. In recent years, such systems have demonstrated potential across multiple domains, yet their application in cultural heritage contexts still faces critical challenges. Table 3 systematically compares key characteristics of representative AI collaborative creation systems in cultural heritage applications, revealing the strengths and limitations of existing systems.

Table 3.

Comparison of this study with existing AI collaborative creation systems

Note:

• • Culture-specific constraints: Indicates whether the system incorporates built-in style/rule constraint mechanisms tailored to specific cultural heritage.

• • User threshold: Determined based on target user groups and required prior knowledge (high = professionals, medium = experienced users, low = general public).

• • Cultural fidelity mechanism: Technical methods extracted from original literature for preserving cultural characteristics.

• • Data source: Method descriptions and experimental settings from the original papers of respective systems.

The comparison in Table 3 reveals three core limitations of existing AI collaborative creation systems: (1) insufficient cultural adaptability: most systems (SmartPaint, ID.8) are not designed for specific cultural heritage; (2) polarized participation thresholds: professional systems (Reframer, DGBID) offer powerful functionality but restrict public participation; while MakeBronze (Yu et al., Reference Yu, Wang, Cai, Zhang, Yu, Yang and Zhang2024), despite lowering the usage threshold through physical interaction, lacks depth in cultural knowledge transmission; (3) imbalance between cultural fidelity and innovation: difficulty in achieving equilibrium between cultural constraints and creative freedom. In contrast, this research overcomes the cultural adaptation deficiencies of general models through LoRA-based lightweight fine-tuning, maintains cultural depth while lowering participation thresholds through multi-level guided design, and provides a novel solution for AI collaborative creation systems in cultural heritage applications by establishing a dynamic balancing mechanism between cultural constraints and innovative expression, alongside systematic evaluation of cultural transmission efficacy.

Basic cognitive status

The survey revealed that 43% (n = 208) of respondents had never been exposed to Cizhou kiln culture, with cognition showing significant stratification:

Historical background cognition rate: 86.3%.

Cultural value cognition rate: 59.6%.

Type characteristic cognition rate: 55.8%.

Representative work cognition rate: 48.1%.

These results indicate that even among those with basic cognition, a deep understanding of Cizhou kiln culture remains limited.

Transmission channel effectiveness

Table 4 data shows that although video resources are the main transmission channel (76.92%), retention rates are generally below 60%. Notably, interactive participation, despite having the lowest coverage (8.42%), achieved the highest user retention rate (82.60%), indicating that interactive experience has significant effects on maintaining user interest.

Table 4.

Usage and retention rates of different transmission channels

Participation barrier analysis

Through semi-structured interviews and open-ended responses, we identified three main types of participation barriers:

1. 1. Technical barriers: 78% of respondents mentioned that Cizhou kiln creation “requires professional training,” believing they lack necessary skills;

2. 2. Cognitive distance: 63% of respondents felt traditional culture is “distant from modern life,” struggling to find personal connections;

3. 3. Display formats: 57% of respondents considered existing display methods “too static,” lacking attractiveness and interactivity.

Framework overview and theoretical positioning

The framework integrates situated cognition theory and co-creation theory to address the core challenge of cultural heritage transmission: transforming users from passive recipients to active co-creators.

Synergistic integration: Rather than treating these theories separately, our framework achieves synergy by positioning co-creation as the method and situated cognition as the mechanism. Creative participation (co-creation) generates authentic contexts (situated cognition) where cultural learning naturally occurs. This integration addresses the research gap of systematic theoretical application in cultural heritage interaction design (Section “Research Gaps and Positioning of This Study”).

Three-layer architecture (Figure 1): The framework comprises (1) theoretical foundation: situated learning principles + co-creation paradigm; (2) core mechanisms: translating abstract principles into concrete design mechanisms; (3) outcomes: user role transformation.

Core design mechanisms

The framework operationalizes the design principles (Section “Design Insights and Principle”) through three interconnected mechanisms:

1. 1. Mechanism 1—Multi-level guided design: Implements a four-stage workflow-selection, generation, iteration, and materialization-that reduces cognitive load while maintaining cultural depth by decomposing complex cultural creation into manageable steps (Norman, Reference Norman2013).

2. 2. Mechanism 2—Cultural Embedding Architecture: Integrates cultural rules as interaction constraints and embeds knowledge cards within the workflow, enabling implicit learning through practice by transforming cultural knowledge from explicit instruction to embodied experience (Goodwin, Reference Goodwin2000).

3. 3. Mechanism 3—Social Transmission Loop: Establishes a creation-sharing-feedback cycle with cultural validation, transforming individual creation into socialized cultural dissemination through social participation that enhances cultural understanding (Lave and Wenger, Reference Lave and Wenger1991).

Framework validation: mapping challenges to solutions

To verify the applicability of the framework, we map three user research challenges to specific framework components:

Data collection

Using multi-source fusion strategies to ensure data comprehensiveness and authority:

1. 1. Field collection: Collaborating with Cizhou kiln intangible heritage inheritors for field research, using high- precision 3D scanners (0.1 mm accuracy) to obtain original artifact data;

2. 2. Literature digitization: Integrating authoritative literature such as “Cizhou Kiln Porcelain Pillows” line drawings, high-resolution photography (600dpi) of museum-housed Cizhou kiln artifacts;

3. 3. Network resource integration: Crawling public images from cultural heritage databases, with expert team review and screening.

Through these methods, 13 TB of raw data were obtained. Using the perceptual image hashing algorithm (ImageHash) to remove duplicate images (similarity 95%), manually eliminating low-resolution ( 1024 × 1024), damaged, or obviously restored samples, finally retaining 6,000 high-quality pattern images.

Data annotation and structuring

Multi-dimensional annotation of pattern data to construct a structured database:

The annotation process adopts expert review mechanisms to ensure annotation quality and cultural accuracy. The final database contains 6,000 high-quality samples covering the main Cizhou kiln vessel types and pattern categories. Cizhou Kiln data annotation statistics are presented in Table 6.

Table 6.

Cizhou Kiln data annotation statistics

Multi-scale image perceptual compression

To optimize computational efficiency while maintaining pattern details, we designed a multi-scale image perceptual compression algorithm based on variational autoencoder (VAE). The algorithm progressively compresses images through multiple scales, reducing computational complexity from original space to 1/64 when downsampling factor =8.

Key technical specifications:

1. (1) Architecture: VAE encoder-decoder with multi-scale processing

2. (2) Loss function: MSE reconstruction loss + KL divergence regularization

3. (3) Performance: Single inference time 12.3 2.1 s (NVIDIA RTX 3090 GPU)

4. (4) Compression ratio: 1/64 while preserving cultural pattern details

Detailed algorithm implementation is provided in Supplementary Appendix E1.

Aesthetic template-based composition guidance

To maintain typical Cizhou kiln composition characteristics (such as “three-seven composition”), we designed a template-based composition guidance mechanism that constrains the generation process through cultural rule templates.

Key design principles:

1. (1) Template representation: Binary masks encoding composition rules _.

2. (2) Adaptive denoising: Adjustable denoise strength controls creative freedom

3. (3) Cultural validation: Automatic verification of “three-seven ratio” and symmetry rules.

This mechanism ensures generated content follows traditional composition rules while preserving creative space. Implementation details are provided in Supplementary Appendix E2.

Multimodal style transfer

Combining IP-Adapter and LoRA technologies for efficient style transfer, addressing the lack of cultural specificity in general models: Core technical components:

1. 1. IP-Adapter module: Achieves multimodal feature fusion through decoupled cross-attention mechanisms, enabling the model to simultaneously process text descriptions and reference images. Uses CLIP ViT–L/14 as image encoder, cross-attention weights set to 0.8, with an additive attention mechanism for feature fusion;

2. 2. LoRA fine-tuning strategy: Uses low–rank decomposition method (r = 64) for efficient adaptation of pre-trained models, significantly reducing computational costs while maintaining style consistency. Specific configurations include learning rate 2e–4 (cosine scheduling), 30,000 training steps, batch size 16, using AdamW optimizer ( 1 = 0.9, 2 = 0.999, weight_decay = 0.01).

Technical implementation configuration: Based on Stable Diffusion–XL v1.0 (resolution 1024 1024) as base model, trained on 4 NVIDIA A100 GPU (80GB), 128GB RAM hardware environment, using CUDA 11.8 and PyTorch 2.0.1. These parameter configurations achieve optimal style transfer effects while maintaining computational efficiency, particularly LoRAIs low-rank constraint (r = 64) significantly reducing computational complexity while maintaining Cizhou kiln characteristics.

Performance evaluation results: Quantitative assessment shows this method reduces training time from traditional full–parameter fine–tuning’s 48.7 hours to 12.3 hours (74.7% reduction) while achieving precise cultural characteristic preservation. All code and pre–trained models will be publicly released after publication to ensure research reproducibility.

Cultural constraint generation algorithm

To ensure generated content conforms to Cizhou kiln cultural characteristics, we designed a multi-condition constrained generation process that integrates cultural rules throughout the generation pipeline:

Constraint mechanisms:

1. 1. Composition constraints: Validation of “three-seven ratio” (subject-to-background ).

2. 2. Style constraints: Verification of black-white contrast characteristics.

3. 3. Semantic constraints: Cultural meaning validation of pattern elements the generation process employs iterative refinement with periodic cultural validation (every check_interval steps) to ensure both technical quality and cultural authenticity. The complete algorithm workflow is detailed in Supplementary Appendix E3.

Algorithm performance evaluation

Algorithm performance was evaluated through comparative experiments, with results shown in Table 7.

Table 7.

Pattern generation performance evaluation

Notes: (1) FID (Fréchet Inception Distance): Lower is better, 30 indicates excellent performance. (2) PSNR (Peak Signal-toNoise Ratio): Unit is indicates high-quality reconstruction. (3) Compositional accuracy: The conformity to “threeseven composition” rules independently annotated by 3 experts, Fleiss’ Kappa = 0.83. (4) Cultural feature fidelity: Weighted average of three sub-dimensions including black-white contrast, symmetry, and decorative semantics. (5) Baseline model: Stable Diffusion-XL v1.0 original model (not optimized for Cizhou kiln).

Experimental results show that by introducing cultural feature constraint mechanisms, the system maintains traditional Cizhou kiln composition rules (92.7% accuracy) and cultural feature fidelity (89.4% accuracy) while maintaining fast generation speed (average 15.0 seconds).

Interaction framework design

The core innovation of AIMagCreations lies in the “Creation-as-Transmission” interaction paradigm, achieved through a four-stage AI collaborative creation framework: selection, generation, iteration, and materialization (Figure 3).

1. 1. Selection phase: Users explore cultural elements through three entry points:

   • • Theme selection: Choose from six traditional themes (prosperity, academic achievement, etc.);

   • • Form selection: Browse 10 classic vessel types, learning historical backgrounds;

   • • Pattern elements: Decorative element library organized by symbolic meaning.

2. 2. Generation phase: Transform multimodal inputs into Cizhou kiln style designs:

   • • Multimodal input: Support text descriptions, uploaded images, or both;

   • • Cultural constraints: Automatically apply Cizhou kiln composition rules;

   • • Visual feedback: Real–time visualization maintains user engagement.

3. 3. Iteration phase: Support guided refinement:

   • • Parameter adjustment: Modify specific aspects within culturally appropriate ranges;

   • • Guided adaptation: System suggests culturally appropriate modifications;

   • • Comparative view: Compare with historical examples.

4. 4. Materialization phase: Transform digital designs into physical objects:

   • • 3D visualization: Map to 3D vessel models;

   • • Physical output: Transfer to ceramic vessels through digital printing or transfer paper techniques;

   • • Creation story: Generate narratives describing cultural elements in the design.

Figure 3.

“Selection-generation-iteration-materialization” AI collaborative creation framework.

User interface design

The AIMagCreations interface implements the four-stage framework through progressive disclosure methods,maintaining creative freedom while guiding users (Figure 4). The interface design particularly emphasizes balance between cognitive load management and cultural guidance, adopting a three-layer information architecture: (1) interactive layer, providing direct operation interfaces; (2) cultural knowledge layer, providing cultural backgrounds through contextual prompts and embedded information cards; (3) creation guidance layer, providing timely creation suggestions. This multi-layer architectural design responds to Norman (Reference Norman2013)’s “knowledge in the world” design concept, embedding cultural knowledge not only in user cognition but also in interface elements.

Figure 4.

AIMagCreations user interface flow.

Key interface design features include:

1. 1. Progressive disclosure architecture with cognitive load management staged information architecture: Implements a four-phase workflow design following Nielson’s progressive disclosure principles, with each interface optimized for specific cognitive tasks to minimize extraneous cognitive load (Sweller, Reference Sweller, Mestre and Ross2011).

Hierarchical task decomposition: Employs modular interface segregation where complex cultural design tasks are decomposed into manageable sub-tasks, supporting users’ working memory limitations.

Contextual scaffolding: Integrates just-in-time cultural knowledge delivery through embedded information cards and contextual tooltips, reducing split-attention effects.

1. 2. Multimodal interaction design with cultural constraint integration

Input modality convergence: Implements parallel input channels (textual, visual, template-based) within a unified interaction space, supporting diverse user preferences and expertise levels real-time visual feedback loop: Employs immediate preview mechanisms in the central workspace to maintain user agency and reduce gulf of evaluation (Norman, Reference Norman2013).

Culturally informed affordances: Integrates traditional aesthetic constraints as interface affordances rather than restrictions, making cultural rules discoverable through interaction design.

1. 3. Distributed cognition framework for cultural knowledge

External memory support: Implements Norman’s “knowledge in the world” paradigm by embedding cultural semantics directly into interface elements, reducing reliance on user’s prior cultural knowledge.

Semantic layering architecture: Employs three-tier information architecture: (1) direct manipulation layer, (2) cultural context layer, and (3) generative guidance layer, supporting both novice and expert users.

Cultural sensemaking support: Provides comparative visualization tools and historical exemplar references to support users’ cultural pattern recognition and appropriation processes.

1. 4. Adaptive automation with human agency preservation

Mixed-initiative design: Balances AI-driven cultural constraint application with user creative control through transparent system suggestions and override capabilities.

Granular control mechanisms: Implements progressive complexity options (primary patterns, auxiliary patterns, decorative elements) allowing users to engage at their preferred level of detail.

Explainable AI integration: Makes algorithmic cultural adaptations visible through interface feedback, supporting user understanding and trust calibration.

1. 5. Social cognitive design for cultural transmission

Community-integrated workflow: Seamlessly integrates individual creation with social sharing mechanisms, supporting cultural knowledge propagation through peer learning.

Digital-physical bridging: Implements tangible interaction endpoints (QR codes, physical output options) that extend digital cultural creation into material culture practices.

Collaborative affordances: Provides social comparison tools and community integration points that support collective cultural meaning-making processes.

Design rationale integration: This interface design operationalizes Activity Theory principles by treating cultural ceramic design as a mediated activity system where traditional knowledge, digital tools, and user creativity form an integrated socio-technical ensemble. The progressive disclosure approach specifically addresses the challenge of cultural HCI identified by Winschiers-Theophilus et al. (Reference Winschiers-Theophilus, Zaman, Maasz, Muashekele, Mbinge, Jengan, Stanley and Ab Hamid2022) - maintaining cultural authenticity while enabling creative appropriation through technology mediation.

Figure 5 complete digital–physical cultural transmission ecosystem demonstrating three interconnected loops. (a) AI collaborative creation mode enables cultural learning through theme selection, multimodal input processing, and intelligent pattern rendering. The ecosystem implements a complete digital-physical loop by seamlessly connecting virtual creation with physical manifestation, supporting both individual cultural learning and collective cultural preservation. This comprehensive approach transforms traditional unidirectional cultural transmission into a participatory, cyclical experience that reinforces cultural connections through both digital engagement and tangible ownership. (b) Physical generation mode materializes digital designs into tangible ceramic products through 3D preview, cultural information display, digital asset preservation via QR code, and physical production ordering.

Figure 5.

Complete digital-physical cultural transmission ecosystem.

Interaction loop construction

To maximize cultural transmission effects, three interconnected interaction loops were designed:

1. 1. Digital-physical loop: Connecting virtual creation with physical objects through ceramic transfer technology;

2. 2. Creation-sharing loop: Promoting social sharing of creation processes and final works;

3. 3. Learning-creation loop: Embedding cultural knowledge in creation processes for implicit learning.

These interconnected loops transform traditional unidirectional cultural transmission into cyclical participatory experiences, continuously reinforcing cultural connections.

Statistical analysis methods

All statistical analyses were conducted using established procedures with appropriate assumption testing and effect size reporting.

Descriptive statistics: Central tendency and dispersion were summarized using means, standard deviations, and confidence intervals. Normality was assessed via Shapiro–Wilk tests, and outliers were identified using boxplots ( IQR criterion).

Between-group comparisons: Independent samples t-tests compared experimental and control groups after verifying assumptions (normality and homogeneity of variance via Levene’s test). When assumptions were violated, Welch’s t-test or non-parametric Mann–Whitney U tests were employed as appropriate.

Repeated measures analysis: Within-subjects comparisons used repeated measures ANOVA with Greenhouse–Geisser correction for sphericity violations (assessed via Mauchly’s test). Post-hoc pairwise comparisons employed Bonferroni correction.

Multiple testing correction: To control family-wise error rates, Bonferroni correction was applied systematically: adjusted / number of tests.

Effect sizes: Cohen’s d quantified standardized mean differences (small: , medium: , large: ). Pearson correlation coefficients (r) assessed association strength (weak: , moderate: , strong: ).

Regression analysis: Hierarchical multiple regression examined predictive relationships, with thorough diagnostic testing (linearity, independence, homoscedasticity, normality of residuals, multicollinearity via VIF). Standardized beta coefficients enabled cross-predictor comparisons.

Software and reproducibility: Analyses employed SPSS Statistics 27.0 and R version 4.2.1 (packages: ez, emmeans, car, lmtest, QuantPsyc). All random processes used specified seeds for reproducibility. Significance level was set at unless otherwise noted.

Experimental design

Twenty experts from interaction design, cultural heritage, and education fields (average age 40.35 2.54 years, 50% female) participated in the evaluation. The experiment used standardized tasks and the System Usability Scale (SUS), with procedures including:

1. 1. Pre-experiment preparation (10 minutes): Sign informed consent forms, complete background survey questionnaires, watch system introduction;

2. 2. System experience (40 minutes): Complete guided tasks (10 minutes), free exploration (10 minutes), deep creation (20 minutes);

3. 3. Evaluation feedback (30 minutes): Complete SUS questionnaire, participate in in-depth interviews.

Results analysis

Expert evaluation results (Table 9) show the system has good usability, specifically demonstrated in:

1. 1. Quantitative evaluation results: System achieved average SUS score of 78.625 (SD = 7.411), significantly higher than “good” usability threshold of 70 points (t(19) = 5.23, p 0.001, Cohen’s d = 1.17);

2. 2. Function completion: 88% of experts independently completed all tasks (M duration = 18.5 min, SD = 3.2), with key operation error rate below 5%;

3. 3. Expert feedback themes: Through thematic analysis of interview data, three core feedback categories emerged:

   • • Interface design advantages (mention rate 72%): Interaction flow follows intuition, visual guidance reduces learning costs;

   • • Technical efficacy recognition (mention rate 65%): Real–time generation speed ( 15s) and high style transfer fidelity;

   • • Improvement suggestions (frequency 3 times): Add preset templates (5 people), optimize loading prompts (3 people).

Table 9.

SUS scores in each dimension

Notes: Each dimension scored on 5-point scale, overall SUS score on 100-point scale Q2, Q4, Q6, Q8, Q10 are reverse items, scores reverse-calculated Overall SUS Score: 78.63 7.41 (range: 65–92.5, n = 20).

Experimental design

To validate the critical role of the LoRA module in Cizhou kiln pattern style transfer, three control experiments were designed:

1. 1. Complete model: Contains complete VAE encoder, IP-Adapter, and LoRA modules;

2. 2. Without LoRA model: Removes LoRA module, retaining only base Stable Diffusion-XL model;

3. 3. Full parameter fine-tuning model: Full parameter fine-tuning of UNet network using the same training dataset.

For each configuration, random sampling from 600 test images for evaluation, with 20 professional evaluators (average age 38.45 3.21 years) conducting technical metrics and subjective experience assessments. Technical metrics results and subjective experience assessments results are presented in Table 10 and Table 11.

Table 10.

Three model technical metrics comparison

Table 11.

User experience score comparison (mean SD)

Assumption testing. Prior to between-group comparisons, Levene’s test assessed homogeneity of variance across model configurations. For FID comparisons, the test revealed unequal variances between complete and w/o LoRA models (F(1,1198) = 4.23, p = 0.040), necessitating Welch’s t-test correction. Similarly, PSNR comparisons showed variance heterogeneity (F(1,1198) = 3.87, p = 0.049). Comparisons with the Full Fine-tuning model met variance homogeneity assumptions (FID: F(1,1198) = 0.89, p = 0.346; PSNR: F(1,1198) = 0.65, p = 0.420), permitting standard t-tests. For subjective evaluations (Table 11), Levene’s tests confirmed homogeneous variances across all six dimensions (all F(2,57) 2.34, all p 0.105), satisfying ANOVA prerequisites.

Key findings include:

The ablation study validates the critical role of the LoRA module from both technical performance and user perception perspectives:

1. (1) Technical performance improvements

   1. 1. Style fidelity: Removing the LoRA module caused FID to surge from 21.7 to , ), demonstrating LoRA’s indispensable role in capturing Cizhou kiln-specific stylistic features;

   2. 2. Detail preservation: The complete model achieved significantly higher PSNR (38.6 dB) than the w/o LoRA model ( ), confirming that low-rank constraints ( ) effectively mitigate detail degradation during style transfer;

   3. 3. Computational efficiency: Compared to full fine-tuning, LoRA reduced training time by vs. 48.7 h) and memory footprint by (5.3 GB vs. 15.8 GB, significantly enhancing system deployability while maintaining generation quality.

2. (2) User perception validation

   1. 1. Visual aesthetics: The complete model scored significantly higher ( =6.0 0.8) than the w/o LoRA model ( =3.9 1.2, p 0.001), with no significant difference from full fine-tuning ( =5.8 0.9, =0.18), indicating LoRA achieves comparable visual quality to full-parameter approaches;

   2. 2. Cultural element recognition: The complete model received the highest rating ( =6.3 0.7), significantly outperforming both the w/o LoRA model ( =3.7 1.1, ) and full fine-tuning ( =5.7 0.8, p 0.01), suggesting LoRA’s cultural constraint mechanism surpasses generic training strategies;

   3. 3. Overall cultural fidelity: Across three dimensions-style consistency, detail preservation, and cultural recognition-the complete model averaged 6.13, exceeding the w/o LoRA model by 57.4% ( 0.001) and full fine-tuning by 6.6% ( 0.05).

3. (3) Metric-perception consistency

   Objective technical metrics demonstrated strong alignment with subjective perceptual evaluations. Correlation analyses revealed that FID scores exhibited a strong negative correlation with visual aesthetics ( = − 0.82, ), while composition accuracy showed a strong positive correlation with cultural element recognition ( =0.79, 0.001). This metric-perception consistency demonstrates that: (a) the proposed cultural constraint mechanism not only optimizes computational objectives but also aligns closely with human aesthetic judgment and cultural cognition; (b) the LoRA module achieves optimal balance across efficiency, quality, and cultural fidelity through precise cultural feature modeling.

Experimental design

We conducted controlled experiments with 100 participants aged years (average age years, balanced gender ratio), Participants were randomly assigned to experimental ( ) or control ( ) groups using computer-generated random number sequences. The experimental scenarios are shown in Figure 6 and Figure 7. Randomization employed a stratified block design to balance key variables including gender, age, and cultural background. A priori statistical power analysis (GPower 3.1.9.7) indicated that the sample size of provides power to detect anticipated effects at . Detailed randomization procedures, power analysis, baseline equivalence verification, and sensitivity analyses are provided in Supplementary Appendix F. Comparing AIMagCreations with traditional learning methods (lectures and videos):and exploring the relationship between creative behaviors and cultural cognition.

1. 1. Experimental design: Between-subjects design, random assignment to two groups:

   • • Experimental group (n = 50): System function video tour (8 minutes), free creation experience (20 minutes), complete evaluation questionnaire (2 minutes);

   • • Control group (n = 50): Cultural explanation video (8 minutes), themed lecture (20 minutes), complete evaluation questionnaire (2 minutes).

2. 2. Measurement indicators:

   • • Knowledge acquisition: Five standardized questions evaluating cognitive levels (Supplementary Appendix B3 Questionnaire 1);

   • • Cultural interest: Four scales measuring learning enjoyment and other indicators (Supplementary Appendix B3 Questionnaire 2);

   • • Behavioral indicators: Sharing behavior, exploration patterns, and subsequent information seeking.

   • • Behavioral indicators: Creative engagement indicators, interaction pattern indicators, socialized transmission indicators.

1. 3. Procedure: Participants completed pre-test, exposure to one condition, post-test.

Figure 6.

Control group (explanation picture of Cizhou Kiln).

Figure 7.

Experimental group (experimental process diagram).

Assumption testing: Independent samples t-tests require verification of variance homogeneity across groups. Levene’s test results for cultural cognition (Table 12) revealed that three questions violated the homogeneity assumption: SQ2 (F(1,98) = 4.89, p = 0.029), SQ4 (F(1,98) = 8.21, p = 0.005), and SQ5 (F(1,98) = 9.45, p = 0.003), necessitating Welch’s t-test with adjusted degrees of freedom. Questions SQ1 and SQ3 met the assumption (F(1,98) = 2.87, p = 0.093 for both), permitting standard t-tests. For cultural interest dimensions (Table 13), all four scales satisfied variance homogeneity (all F(1,98) 3.42, all p 0.067), allowing standard t-test procedures. Additionally, Shapiro–Wilk tests confirmed normality for all dependent variables (all p 0.05), validating parametric test appropriateness.

Table 12.

Cizhou Kiln knowledge cognition t-test analysis results

* p < 0.05, **p < 0.01.

Table 13.

Cizhou Kiln knowledge cognition t-test analysis results

*p < 0.05, **p < 0.01.

Notes: (1) Questionnaire items are detailed in Supplementary Appendix B3. (2) Independent samples t-test was employed, two-tailed test. (3) Effect size calculation: Cohen’s pooled. (4) All p-values were Bonferroni corrected (correction factor = 5).

Results analysis

Table 12 data analysis shows the experimental group achieved significantly higher accuracy rates on all five knowledge questions compared to the control group, particularly on deep understanding question SQ5 with the most significant difference (94.00% 23.99% vs 54.00% 50.35%, p 0.001, Cohen’s d = 1.02). According to (Cohen, Reference Cohen2013) standards, d 0.8 indicates large effect size, demonstrating this difference has substantial educational significance. Similarly, Table 11 shows the experimental group achieved significantly higher scores on all four interest dimensions, with all dimensions showing effect sizes exceeding 1.5, indicating the “Creation-as-Transmission” mode has powerful and consistent effects in stimulating cultural interest.

Further analysis shows the experimental group’s cultural cognitive accuracy positively correlates with creation completion degree (r = 0.76, p 0.001), supporting our theoretical hypothesis that creative activities as cognitive mediators can effectively promote cultural learning. This finding echoes (Kolb, Reference Kolb2014)’s experiential learning theory but further reveals the unique value of creative participation in cultural learning.

Correlation analysis between creative behaviors and cognitive effects

To elucidate the mechanism underlying the “Creation-as-Transmission” paradigm, we conducted a multi-dimensional correlation analysis of creative behaviors and cultural cognition performance among experimental group participants ( =50).

1. (1) Descriptive statistics

Table 14 presents the creative behavior characteristics of experimental group users.

1. (2) Correlation analysis

Table 14.

Descriptive statistics of creative behaviors in experimental group ( = 50)

Notes: Mean SD = mean standard deviation; range = minimum–maximum; median = median (50th percentile); multimodal input type is a categorical variable, with mode and its proportion reported; functional exploration breadth = (number of functions actually used/total number of system functions) 100%; social dissemination indicators are binary variables, with incidence rates reported.

Pearson correlation analysis (Table 15) revealed significant associations between creative behavior and cultural cognition accuracy.

1. (3) Multiple regression analysis: Predictive model of creative engagement on cognitive enhancement.

Table 15.

Correlation analysis between creative behavior and cultural cognition accuracy ( = 50)

Notes: * , ** , *** .

1. Values in the table represent Pearson correlation coefficients (r), ranging from −1 to +1, with positive values indicating positive correlations. 2. Correlation strength criteria: indicates weak correlation, 0.3–0.6 indicates moderate correlation, and indicates strong correlation. 3. Significance levels: * (significant), ** (highly significant), *** (extremely significant). 4. SQ1–SQ5 represent the five questions in the cultural cognition test, with SQ5 being a higher-order question assessing deep understanding. 5. All correlation coefficients are positive, indicating a positive relationship between creative behavior and cognitive performance. 6. The correlation coefficient between iteration frequency and SQ5 ( ) is the highest value in the table, suggesting that iterative refinement has the most pronounced effect on promoting deep understanding.

To quantify the independent contribution of each creative behavior to cognitive outcomes, we constructed a hierarchical multiple regression model (Table 16):

Table 16.

Hierarchical regression analysis of cultural cognition accuracy ( = 50)

Notes: * , ** , *** .

Alsaqqaf and Li (Reference Alsaqqaf and Li2023): Springer, Cham, Switzerland; Cohen (Reference Cohen2013): Abingdon, Oxon, UK; Kolb, D. A. (Reference Kolb2014): Upper Saddle River, New Jersey, USA; Lave J and Wenger E (Reference Lave and Wenger1991): Cambridge, UK; Lawton et al. (Reference Lawton, Ibarrola, Ventura and Grace2023): Sydney, NSW, Australia; Lee et al. (Reference Lee, Park and Song2019): Springer, Cham; Liu (Reference Liu2020): Springer, Cham; Polidoro et al. (Reference Polidoro, Liu and Craig2024): New York, NY, USA; Shang (Reference Shang2015): Beijing, China; Tian (Reference Tian1985): Shanghai, China; Troussas et al. (Reference Troussas, Papakostas, Krouska, Mylonas and Sgouropoulou2024): New York, NY, USA; Waidyaratne et al. (Reference Waidyaratne, Noisri, Srisupakwong, Wijayasekara, Ratchatakorntrakoon, Bongsasilp, Vanichchanunt and Wuttisittikulkij2024), Pattaya, Thailand; Zhang (Reference Zhang2000): Beijing, China; Zhang et al. (Reference Zhang, Yao, Wu, Lin, Liu, Yan and Ying2022): New Orleans, LA, USA.

1. Hierarchical regression analysis: Variable groups were added sequentially to observe changes in explanatory power. 2. standardized regression coefficient, representing the independent effect after controlling for other variables; larger indicates stronger influence. 3. standard error, reflecting estimation precision; smaller SE indicates more reliable estimation. proportion of variance explained by the model; indicates that of cognitive variance is explained. 5. additional explanatory power contributed by newly added variables; indicates that Step 2 explains an additional overall significance test of the model; larger F value indicates more effective model. 7. Significance levels: . “–” Indicates that the variable was not included in the current model. 9. All variables were standardized; therefore, values are directly comparable in terms of effect size. 10. The coefficient for iteration frequency in Model is the largest among all predictors, indicating that it makes the most significant independent contribution to cognitive enhancement.

Interpretation of regression models:

Model 1 (basic engagement): Creative duration alone explained 38.4% of the variance in cognitive outcomes, demonstrating the foundational role of time investment.

Model 2 (deep interaction): With the addition of iteration frequency and cultural element exploration, significantly increased to . Iteration frequency exhibited the largest standardized regression coefficient ( ), indicating that each one standard deviation increase in iteration frequency corresponded to a 0.48 standard deviation improvement in cognitive accuracy, highlighting the central value of the “trial-error-reflection-optimization” cycle.

Model 3 (full model): After incorporating social dissemination variables, reached . Work sharing behavior ( ) and knowledge card viewing ( ), though demonstrating smaller effect sizes, remained significant, suggesting that social interaction and active learning further reinforce cognitive outcomes.

The final model revealed the multi-layered mechanism of “creation as communication”:

Cultural cognition time investment iterative exploration element breadth + social sharing active learning.

This model demonstrates that “deep engagement” in creation (iteration + exploration, contributing of the effect) has 3.6 times greater impact on cultural cognition than “superficial exposure” (time investment alone, contributing of the effect), providing quantitative evidence for the “creation as communication” paradigm.

AI collaborative creation experience research

Experimental design

To understand user AI collaborative creation experiences in depth, 60 multidisciplinary undergraduate students (average age 20.85 1.75 years) participated in experiments using mixed research methods (Figure 8):

1. 1. Creation task: Participants completed Cizhou kiln plate creation within 60 minutes using “task exploration + free creation” mode;

2. 2. Data collection:

   • • Quantitative measurement: Using MICSI scale (18-item 7-point scale) to evaluate collaborative creation experience (Supplementary Appendix B3 MICSI Scale);

   • • Qualitative methods: Screen recording, retrospective think-aloud, and semi-structured interviews;

Figure 8.

Experimental process case.

Analysis methods: Combining quantitative statistics and thematic analysis for comprehensive user experience evaluation.

Figure 9 demonstrates the system’s four multimodal input methods and their corresponding outputs. Each row shows: (1) Generation method: the interaction mode used; (2) Input content: user-provided information or media; (3) Output content (Patterns): generated 2D Cizhou kiln-style decorative patterns in black-and-white format; (4) Output content (Entity): 3D visualization of the pattern mapped onto ceramic plates. The four methods include: Theme Selection – users choose from traditional cultural themes (e.g., abundance every year) to generate culturally appropriate motifs; Text-to-Image Generation - natural language descriptions are converted into patterns (e.g., “A little boy is by the river, holding a lotus leaf. There are fish in the river”); Image-to-Image Generation - uploaded photographs (e.g., a bluebird) are transformed into Cizhou kiln style while preserving compositional structure; Free Drawing - hand-drawn sketches are refined into traditional decorative patterns. All methods maintain high cultural fidelity (89.4%) and generate quality (FID = 21.7), demonstrating the system’s capability to accommodate diverse user skill levels and creative preferences.

Figure 9.

Experimental cases of different generation methods.

Results analysis

MICSI Scale Scores (Figure 10):

Figure 10.

Research results on co-creation experience: (a) MICSI subscale score and (b) MICSI exploratory score.

Quantitative analysis shows the system performs positively in overall user experience, with average scores across dimensions reaching 5.554 (SD = 0.674), specifically:

1. 1. Core function evaluation: System achieved an effective threshold ( 5.0 points) in 77.8% of evaluation dimensions, with core functions performing exceptionally (M = 5.790, SD = 0.710);

Dimensional differences: Agency cognition dimension scored significantly lower (M = 4.533, SD = 1.641), reflecting need for improved collaborative capabilities;

1. 2. User feedback: Image generation freedom (67%) and operation diversity (23%) constitute the main advantages, while system usability (43%), prompt experience requirements (27%), and generation latency (30%) become focus areas.

Based on 125 interview codings, qualitative analysis shows improvement needs concentrated in three dimensions:

1. 1. Function enhancement (52 mentions): Need for more templates and customization options;

2. 2. Interface optimization (38 mentions): Visual elements and workflows need improvement;

3. 3. Creation support (35 mentions): Need for more creative guidance and example demonstrations.

90% of users recognized the system’s creative experience, but low scores in the agency cognition dimension indicate users tend to view the system as a tool rather than a collaborative partner, providing important directions for future system optimization.

Validity considerations and directions for improvement

Theoretical dialogue with experiential learning paradigm

Rather (Reference Rather2020) demonstrated that experiential engagement enhanced cultural memory retention by 47%, establishing the theoretical foundation for experiential learning in cultural heritage dissemination. Building upon this foundation, our study advances the field through three theoretical dimensions:

1. 1. The deepening of the paradigm from experience to creation: While Rather’s research focused on immersive experiences with users as “observers,” this study advances to active creation with users as “producers.” This transformation manifests in two aspects: (1) At the cognitive level, the paradigm shifts from declarative knowledge (“knowing what”) to procedural knowledge (“knowing how”), with the substantial difference in deep comprehension questions (94% vs. 54%) demonstrating the unique advantage of creative activities in promoting higher-order cognition; (2) At the motivational level, creative activities generate deeper intrinsic motivation through “artifact-self” identity connection, evidenced by the voluntary sharing rate (82.6%) significantly exceeding the memory retention rate (47%) in Rather’s study.

2. 2. Quantitative breakthrough in mechanistic evidence:This study provides the first quantification of causal pathways between creative behaviors and cognitive enhancement. Multiple regression analysis reveals: (1) Iteration frequency emerged as the strongest predictor ( ); (2) Cultural element exploration quantity ( ) validates the applicability of levels-of-processing theory (Craik and Lockhart, Reference Craik and Lockhart1972); (3) The complete model explains of cognitive variance, providing a reproducible predictive model for understanding how experience transforms into learning.

3. 3. The closed loop from individual experience to social communication:This study extends individual learning to socialized dissemination through the “create-share-feedback” loop. The finding that 64% of users expressed interest in further learning traditional crafts indicates that creative experiences catalyze transformation from transient interest to sustained commitment.

Critical comparison with Yu et al.’s MakeBronze system

Yu et al.’s (Reference Yu, Wang, Cai, Zhang, Yu, Yang and Zhang2024) MakeBronze system represents pioneering work in human-AI co-creation for cultural heritage, enhancing children’s understanding of bronze culture through physical interaction (lost-wax casting experience) and increasing participation willingness by 68%.

1. 1. Differences in target user groups and design strategies: MakeBronze reduces cognitive load for children by simplifying cultural elements (object constraints), making it suitable for foundational cultural education. In contrast, our study targets adult users, maintaining cultural integrity (composition rule accuracy: 92.7%; cultural fidelity: 89.4%) while achieving usability through multi-level guidance design. This comparison demonstrates that the creation paradigm can accommodate users at different cognitive levels by adjusting constraint intensity.

   Regarding evaluation dimensions, MakeBronze primarily measures affective indicators such as willingness and satisfaction. Our study advances this by incorporating objective cognitive accuracy measurements and establishing a predictive model from behavioral engagement to cognitive outcomes through behavior-cognition correlation analysis ( =0.731), providing quantitative foundations for system optimization.

2. 2. Complementarity of technical implementation pathways: MakeBronze employs a mixed reality approach combining physical interaction with AR visualization, emphasizing haptic feedback and spatial experience. Our study adopts a fully digital pathway with AI generation. Together, these studies demonstrate that the effectiveness of cultural heritage co-creation systems depends not on specific technical pathways, but on adherence to core design principles: (1) progressive guidance that lowers participation barriers; (2) constraint mechanisms that preserve cultural authenticity; (3) iterative design that promotes deep engagement. The “creating-as-disseminating” framework (Figure 1) and implementation guidelines (Table 15) proposed in this study distill these cross-technical design principles.