Advancing lacquerware design through human-AI collaboration with controllable diffusion models Authors: Jun Liang, Yixin Li, Zhileng Xiong, Qing Huang Published: 2025 ABSTRACT AbstractWe propose a human-in-the-loop AI-generated content (AIGC) system to advance lacquerware design innovation by merging artisan knowledge with controllable diffusion models. Traditional lacquerware design faces challenges in balancing heritage preservation and creative exploration, where artisans often struggle with limited inspiration sources and time-consuming manual processes. The proposed system tackles these problems by employing a closed-loop workflow that integrates AI-driven creation and human adjustment, which fosters iterative cooperation between artisans and the AI engine. The system architecture consists of a front-end interface, an AI engine operating on the back end, and a carefully assembled database of historical and modern lacquerware designs, which undergo preprocessing steps such as normalization, background removal, and detailed annotation to document stylistic and compositional features. Essentially, a diffusion model refined through LoRA adaptation for specialized domain control produces high-quality lacquerware designs without compromising their aesthetic coherence. Furthermore, the system integrates multi-modal control mechanisms, which comprise text-to-image prompts for stylistic direction and image-to-image operations with ControlNet to achieve structural accuracy, thereby granting artisans the capability to submit sketches or reference compositions that the AI subsequently augments with elaborate details. The interactive user interface includes an inspiration gallery, a canvas with dynamic properties, and adjustable parameter controls, granting artisans the ability to efficiently investigate a wide range of design alternatives. Our key achievements consist of an innovative implementation of controllable diffusion models in cultural heritage design, a lacquerware dataset with detailed annotations, and a fine-tuning method empirically proven to elevate generation quality. Experiments show the system’s capacity to generate culturally consistent designs while markedly decreasing the time required for artisans to conceptualize their work. This work holds importance due to its capacity to connect traditional craftsmanship with contemporary AI technologies, thereby promoting innovation while preserving the genuine qualities of lacquerware art. The system broadens availability of sophisticated design tools, thereby hastening creative processes and creating fresh opportunities for cross-disciplinary studies in AI-driven cultural heritage conservation.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-33119-y.Keywords: Lacquerware, AI-Generated content (AIGC), Human-in-the-Loop, Diffusion models, Controllable generation, ControlNet, Cultural heritage, Human-AI Co-Creation, Generative designSubject terms: Engineering, Mathematics and computing Introduction Lacquerware, as a traditional craft of profound cultural importance in East Asia, reflects centuries of artistic tradition and technical expertise. The intricate designs and labor-intensive production processes have made lacquerware both a cultural treasure and a challenging art form to sustain in contemporary contexts. Although recent initiatives in safeguarding cultural heritage have recorded conventional methods and patterns1, a pressing issue remains in the discipline: how to promote creativity without compromising the genuineness and cultural fidelity of this time-honored practice. The delicate balance between preservation and evolution has become increasingly difficult as artisan communities shrink and traditional apprenticeship systems decline2. The advent of artificial intelligence, especially concerning generative models, presents novel prospects for tackling these challenges. Diffusion models show outstanding performance in creating images3, and methods such as Low-Rank Adaptation (LoRA) make it possible to adjust these models effectively for specific applications4. These technological advances coincide with growing interest in human-AI collaboration frameworks, where human expertise guides and refines AI-generated outputs5. In artistic fields, such cooperative methods have proven effective in boosting human creativity instead of supplanting it6. Numerous efforts have been undertaken to implement artificial intelligence in the realm of conventional craftsmanship. For instance, researchers have explored AI-assisted design for porcelain7 and textile patterns8. Nevertheless, lacquerware poses distinct difficulties owing to its unique material attributes, multi-dimensional shapes, and intricate ornamentation methods. The existing approaches often fail to capture the nuanced aesthetics and cultural significance embedded in lacquerware designs9. We propose a Human-in-the-Loop AIGC system specifically tailored for lacquerware design innovation. The system integrates three essential elements: (1) a diffusion model optimized on a comprehensive lacquerware dataset employing LoRA technology, (2) multi-modal control mechanisms granting artisans the ability to direct the generation process by means of text, sketches, and reference images, and (3) an interactive interface supporting iterative refinement of AI-generated designs. This method contrasts with prior research in multiple key dimensions. Initially, the system is precisely tailored to address the distinct attributes of lacquerware, encompassing its material limitations and cultural dimensions. Second, it grants craftspeople diverse methods of control instead of depending exclusively on textual instructions. Third, the workflow emphasizes continuous human-AI interaction, where the AI serves as a creative partner rather than an automation tool. The proposed system contributes to both technical and cultural domains. From a technical standpoint, we illustrate how diffusion models can be successfully tailored for specific cultural heritage purposes by meticulous dataset curation and model adjustment. Combining diverse control methods overcomes the constraints of single-input interfaces, which frequently cannot adequately reflect the intricacy of artistic expression. From a cultural perspective, the system presents a pragmatic approach to addressing the difficulty of preserving traditional crafts in contemporary settings, equipping artisans with resources that broaden their artistic potential without compromising the craft’s historical integrity. Our method corresponds to contemporary views in cultural ecology10, stressing the necessity for adaptable conservation measures granting traditions the capacity to develop. Professional lacquerware artisans and quantitative metrics jointly affirm the efficacy of our approach. The findings indicate notable advancements in design efficiency and creative exploration relative to conventional manual methods. Artisans indicated the system aided them in overcoming creative obstacles and investigating design avenues they may not have pursued independently, without compromising their artistic autonomy and authority in decision-making. The remainder of this paper is organized as follows: "Related work" reviews related work in AI-assisted design and cultural heritage applications. "Background: lacquerware art, diffusion models, and human-AI co-creation" provides necessary background on lacquerware art and the technical foundations of our approach. "HITL-Lacquer: a human-in-the-loop AIGC framework for lacquerware design" details the HITL-Lacquer framework, followed by experimental setup in "Experimental setup" and results in "Experimental results and analysis" and "Results and empirical validation" discusses broader implications, with future work and conclusions in "Future work" and "Conclusion" respectively. Related work Artificial intelligence and traditional craftsmanship converge as a dynamic field of study, with multiple interrelated research trajectories. This section organizes prior work into three thematic clusters: AI-assisted cultural heritage preservation, human-AI co-creation frameworks, and specialized applications of generative models in craft design. AI in cultural heritage preservation Recent years have witnessed growing applications of AI in documenting and revitalizing traditional crafts. In porcelain design7, showed that diffusion models adapted with LoRA could produce culturally consistent patterns without altering the iconic blue-and-white visual style. Similarly8, applied AI-assisted techniques to reconstruct Ming Dynasty textile patterns, though their approach focused primarily on historical accuracy rather than creative innovation. These studies established the viability of AI for pattern generation but often treated cultural elements as static motifs rather than dynamic design components. The lacquerware field poses distinct difficulties setting it apart from alternative artistic practices. In contrast to porcelain’s planar surfaces, lacquerware designs must address multidimensional form factors and material interactions that influence visual perception9. emphasized the difficulty conventional image generation methods face in reproducing the depth effects produced by traditional lacquer methods such as maki-e, which employs sprinkled metal powder for decoration. This highlights the necessity for tailored methods extending past conventional image generation. Human-AI collaborative design systems Human-in-the-loop AI has become increasingly prominent in creative fields, especially for overcoming the constraints of entirely autonomous generation6. noted principal advantages of these systems, such as their capacity to complement human creativity instead of supplanting it by means of repeated improvement processes. Their results correspond with data from11, which classified modes of engagement in generative artificial intelligence platforms and highlighted the necessity of preserving creative control. Practical implementations of these principles have taken various forms. In the realm of automotive interior design12, merged generative adversarial networks with cultural semantic modeling, yet their approach favored mass production rather than artisanal customization. More pertinent to craft settings13, investigated the collaboration between robotic systems and artisans in material fabrication, yielding critical findings on upholding craft genuineness in processes mediated by technology. Specialized generative models for craft design The technical core of contemporary AI-assisted design systems increasingly relies on diffusion models, which have surpassed earlier GAN-based approaches in both quality and controllability. The adaptation of these models for specialized domains has become an active research direction14. demonstrated the effectiveness of LoRA fine-tuning for ancient pottery patterns, achieving superior detail preservation compared to base models. Their work provided methodological foundations but did not address the interactive aspects crucial for design innovation. Control mechanisms represent another critical area of advancement. While15 surveyed various control techniques, their analysis revealed a gap in systems that combine multiple input modalities—a limitation particularly acute in craft contexts where artisans may express ideas through sketches, verbal descriptions, or existing references. The integration of ControlNet with diffusion models, as explored in16, offers promising directions for structural control but has not been systematically applied to traditional craft design. The proposed HITL-Lacquer system synthesizes insights from these diverse threads while addressing their limitations. Unlike prior work that treated cultural elements as fixed patterns, our approach enables dynamic exploration of lacquerware aesthetics through multi-modal control. The system extends beyond single-input interfaces by combining textual, structural, and referential guidance mechanisms. Compared to generic diffusion model applications, our specialized fine-tuning process and curated dataset specifically target the material and dimensional characteristics of lacquerware. These innovations collectively position the system as both a technical advancement and a practical tool for sustaining traditional craftsmanship in contemporary contexts. Background: lacquerware art, diffusion models, and human-AI co-creation To lay the groundwork for our proposed system, this section presents critical background knowledge spanning three interrelated areas: the artistic and technical dimensions of lacquerware, the foundations of diffusion models, and the theoretical structure of human-AI collaborative creation. Grasping these components is essential for recognizing how sophisticated generative technologies can be deliberately employed in traditional craft innovation without compromising cultural integrity. Lacquerware Art basics Lacquerware craftsmanship has existed for more than seven thousand years in East Asia, as archaeological findings indicate its beginnings in Neolithic China17. The unique material qualities of lacquer originate from the sap of Toxicodendron vernicifluum trees, which polymerizes upon contact with oxygen. This chemical process yields a resilient, lustrous finish that craftsmen have traditionally adorned with ornamental methods such as carving, embedding, and applying pigments. The material’s distinct properties, such as its viscosity, the manner in which it dries, and its reaction with pigments, determine the creative options and limitations that set lacquerware apart from other artistic forms18. Conventional manufacturing entails the deposition of multiple fine strata (generally 30–300), each necessitating exact environmental parameters to achieve adequate curing. This labor-intensive process yields objects possessing remarkable depth and luminosity, qualities conventional imaging finds difficult to reproduce faithfully. Regional differences exist in decorative techniques, as seen in Japanese maki-e (metal powder application), Chinese lacquer carving, and Vietnamese mother-of-pearl ornamentation, each reflecting unique artistic heritage19. The cultural importance of lacquerware goes beyond its tangible worth, acting as a vehicle for conveying artistic traditions, symbolic interpretations, and craft expertise between generations. Motifs frequently bear particular meanings, such as chrysanthemums denoting longevity in Chinese art, whereas cherry blossoms connote impermanence in Japanese traditions20. These cultural codes present both prospects and obstacles for AI-assisted design, as systems must identify and suitably adapt traditional visual language. Diffusion model fundamentals Diffusion models have become a robust framework for generative tasks, consisting of a forward procedure that incrementally introduces noise to data and a trained inverse procedure that produces samples from noise21. The forward process can be described by a differential equation that transforms the data distribution into a simple noise distribution : where controls the diffusion rate. The reverse process learns to invert this transformation through a neural network that predicts and removes noise at each step. Training involves optimizing a variational bound on the negative log-likelihood: Recent advances have improved the efficiency and controllability of these models through techniques like latent diffusion22 and guidance mechanisms23. For craft applications, the model’s ability to capture fine details and generate diverse variations makes it particularly suitable, though specialized adaptations are necessary to address domain-specific requirements. Human-AI co-creation concepts The paradigm of human-AI co-creation encompasses various interaction modes where human expertise and machine capabilities complement each other24. In design contexts, these interactions typically follow three patterns: human-directed workflows where AI executes specific tasks, AI-initiated processes where humans refine machine outputs, and iterative loops where both parties contribute dynamically25. Successful co-creation systems balance several competing demands: preserving human creative agency while benefiting from AI’s generative capacity, maintaining cultural authenticity while enabling innovation, and providing sufficient control without overwhelming users with technical complexity26. Psychological studies suggest that optimal collaboration occurs when AI acts as an “inspiration partner” rather than an autonomous creator, allowing humans to retain authorship while exploring expanded possibilities27. Image data preprocessing Effective application of generative models to lacquerware requires careful data preparation to capture the medium’s unique visual characteristics. Standard preprocessing includes geometric normalization to account for varying object shapes and sizes, followed by intensity normalization: where and represent the dataset’s mean and standard deviation. For lacquerware specifically, additional steps address challenges like specular highlights and transparency effects that conventional photography often misrepresents28. Data augmentation strategies must respect the integrity of decorative motifs—for example, avoiding arbitrary rotations that could disrupt culturally significant orientations. Image annotation techniques Accurate annotation enables models to learn the compositional principles underlying lacquerware designs. Multi-layer annotation schemes prove most effective, separating base forms, decorative patterns, and material effects into distinct channels29. Semantic segmentation captures regional styles (e.g., distinguishing Chinese carved lacquer from Japanese urushi), while keypoint annotation identifies characteristic features like the golden ratio proportions in Korean designs30. Semi-automatic approaches leverage pre-trained models for initial annotation, which domain experts then refine—a process particularly valuable for identifying subtle technical details like brushstroke directions or maki-e particle distributions31. These annotated datasets not only support model training but also facilitate style analysis and historical comparison, creating bridges between technical documentation and creative reinterpretation. HITL-Lacquer: a human-in-the-loop AIGC framework for lacquerware design The proposed HITL-Lacquer system establishes a novel framework for lacquerware design innovation through three core technical innovations: a domain-optimized diffusion model architecture, multi-modal human-AI interaction mechanisms, and a closed-loop refinement workflow. This segment outlines the system’s elements and their assembly into a unified design framework. Customized AIGC technology stack for lacquerware Our system is based on Stable Diffusion v2.132, adapted with Low-Rank Adaptation (LoRA) to focus on lacquerware aesthetics. The adaptation process introduces trainable low-rank matrices and into the attention layers of the base model, where the weight update is constrained by the rank : This approach grants effective adaptation without compromising the model’s broad competencies. For lacquerware-specific adaptation, we set based on empirical validation of detail preservation versus training efficiency. The training objective combines the standard diffusion loss with a perceptual loss term that emphasizes material characteristics: The perceptual loss is derived from a pre-trained VGG-19 network33 by contrasting generated and real samples across multiple feature scales to capture lacquer-specific textures and reflectance properties. ControlNet implementation34 delivers structural guidance by means of diverse conditioning mechanisms. Given an input sketch , the ControlNet branch processes spatial information through a trainable copy of the diffusion model’s encoder, producing conditioning features : These attributes adjust the primary diffusion trajectory by means of cross-attention layers, which permits exact alignment with compositions given by artisans while preserving stylistic consistency. The system accommodates a trio of conditioning modes: edge maps for preserving contour accuracy, depth maps for managing three-dimensional structure, and scribble inputs for creating imprecise conceptual drafts. Multi-modal human-AI collaborative workflow The interactive design process begins with artisan inputs through three parallel channels: textual prompts for stylistic direction, reference images for visual inspiration, and sketch canvases for structural specification. The text encoder handles natural language descriptions by employing a CLIP-based tokenizer35, featuring customized embeddings for lacquercraft terms (e.g., “tixi” carving method or “qiangjin” gold application). In image-based guidance, a hybrid encoding system merges global style transformation with local feature alignment. Given a reference image , the system extracts both high-level style codes and detail anchors : These representations condition the generation process through adaptive instance normalization36 for style transfer and spatial feature alignment for detail preservation. The sketch interface implements real-time processing through a lightweight edge detection module based on the HED architecture37. User strokes are converted to probabilistic edge maps that balance precision and flexibility: This formulation allows the system to interpret rough sketches while maintaining clean contours in the final output. The interface provides adjustable sensitivity controls, enabling artisans to determine how strictly the AI should adhere to their input lines. Iterative refinement and data management The closed-loop refinement system maintains a dynamic generation history represented as a directed acyclic graph , where vertices correspond to design states and edges record transformation operations. Each state encapsulates the full generation parameters, allowing non-linear backtracking and branching: Artisans navigate this history space through a visual timeline interface, comparing variants and merging desirable elements from different branches. The system automatically logs all interactions, creating a structured dataset of human preferences that informs future model updates. For data management, we implement a hierarchical storage system that organizes designs by stylistic attributes, technical parameters, and user annotations. Each design is indexed through multiple cross-referenced representations: This structure enables efficient retrieval during the design process, allowing artisans to quickly access relevant historical examples or similar concepts from the database. System architecture and implementation The complete system architecture, shown in Fig. 1, integrates these components through a modular pipeline. The frontend interface is implemented as a web application using React and Canvas API, communicating with backend services via RESTful APIs. The AI engine runs on NVIDIA A100 GPUs, with optimizations for real-time interaction including cached diffusion steps and progressive rendering. Enhanced HITL-AIGC System architecture for lacquerware design. Key performance optimizations include: Latent space caching for frequent operations. Latent space caching for frequent operations. Adaptive computation allocation based on interaction mode. Adaptive computation allocation based on interaction mode. Distributed processing for multi-variant generation. Distributed processing for multi-variant generation. Quantized inference for mobile compatibility. Quantized inference for mobile compatibility. The system achieves sub-second response times for sketch-based refinements and 2–3 s generation times for complete new designs, meeting usability requirements established through preliminary user studies. Technical validation and baseline performance Before user evaluation, we conducted extensive technical validation of the core generation capabilities. Quantitative metrics were computed on a held-out test set of 500 lacquerware designs, comparing our fine-tuned model against three baselines: vanilla Stable Diffusion, StyleGAN-338, and a conventional CNN-based autoencoder39. The results demonstrate significant improvements in both fidelity and controllability. Our model achieves a Fréchet Inception Distance (FID) score of 12.7 compared to 18.3 for vanilla Stable Diffusion, indicating better alignment with authentic lacquerware aesthetics. The LPIPS diversity metric40 shows a 22% increase in meaningful variation while maintaining stylistic coherence. For structural adherence, the ControlNet integration achieves 89% contour accuracy on sketch-to-design tasks, compared to 64% for text-only generation. This capability proves particularly valuable for preserving artisan intent in the conceptual stages while still allowing for creative elaboration by the AI. Experimental setup Dataset preparation and annotation The experimental assessment employs a thorough dataset of 3,200 lacquerware designs gathered from museum archives, contemporary displays, and artisan studios across China, Japan, and Vietnam. Every design is subjected to thorough preparatory procedures to guarantee uniformity and high standards. The preprocessing pipeline includes: Geometric normalization : All pictures are adjusted to 512 × 512 dimensions with aspect ratios preserved by means of smart padding41. Geometric normalization : All pictures are adjusted to 512 × 512 dimensions with aspect ratios preserved by means of smart padding41. Material-aware enhancement : Dedicated filters address the reflective properties of lacquerware by applying a specularity reduction algorithm. Material-aware enhancement : Dedicated filters address the reflective properties of lacquerware by applying a specularity reduction algorithm. 3.Multi-Layer Annotation: A team of five lacquerware experts annotates each design through: Multi-Layer Annotation: A team of five lacquerware experts annotates each design through: Semantic segmentation masks identifying 12 material regions (e.g., gold powder, mother-of-pearl). Semantic segmentation masks identifying 12 material regions (e.g., gold powder, mother-of-pearl). Structural decomposition into base form and decorative patterns. Structural decomposition into base form and decorative patterns. Cultural metadata including historical period, regional style, and symbolic motifs42. Cultural metadata including historical period, regional style, and symbolic motifs42. The final dataset is divided into training (2,560 designs), validation (320 designs), and test (320 designs) sets, with balanced distribution across styles and techniques. Participant selection, protocol, and methodological rationale The user study involved 12 participants, segmented into two cohorts to ensure a breadth of relevant expertise: Group A (Artisans): Comprised 6 professional lacquerware artisans (4 female, 2 male; age range 45–62, M = 53.5) with a mean of 21.3 years of professional experience (SD = 5.8). Group A (Artisans): Comprised 6 professional lacquerware artisans (4 female, 2 male; age range 45–62, M = 53.5) with a mean of 21.3 years of professional experience (SD = 5.8). Group B (Students): Comprised 6 final-year Master of Fine Arts students (3 female, 3 male; age range 22–25, M = 23.1) specializing in traditional craft design at a leading arts university. Group B (Students): Comprised 6 final-year Master of Fine Arts students (3 female, 3 male; age range 22–25, M = 23.1) specializing in traditional craft design at a leading arts university. A sample size of N = 12 is conventional for formative usability and qualitative studies in Human-Computer Interaction (HCI), where the goal is to identify major design themes and interaction patterns rather than to generalize to a wider population. Research in this field has shown that samples in the range of 10–15 participants are typically sufficient to reach “thematic saturation,” the point at which new participants are unlikely to reveal new high-level insights. While a small sample might raise concerns about statistical power, a post-hoc power analysis confirmed the adequacy of our sample for the observed data. For our primary efficiency metric (time-per-design), the very large effect size (Cliff’s δ = -0.92) yielded statistical power exceeding 0.95, indicating that the detected difference was so substantial that it was reliably observable even with a modest sample size. To address potential confounding variables related to individual skill and proficiency, we deliberately employed a within-subjects (or repeated measures) experimental design. In this structure, each participant completed tasks under both conditions (Traditional Method and HITL-AIGC System), serving as their own control. This design is highly effective at isolating the effect of the intervention by factoring out baseline variations in individual speed, artistic style, and technical expertise, thereby ensuring that observed differences are attributable to the design method rather than pre-existing differences between participants. To mitigate order effects inherent in within-subjects designs, the sequence of the two conditions was counterbalanced using a Latin square design. Participants were randomly assigned to a starting condition via a computer-generated sequence. Prior to the HITL-AIGC task, each participant underwent a 30-minute familiarization protocol. This included a guided tutorial on the system’s interface and multi-modal controls, followed by a 15-minute sandbox period where they could freely experiment with the tool. Each participant then completed two 90-minute design tasks, one for each condition, based on identical design briefs. Evaluation metrics and protocols The evaluation model integrates both numerical and descriptive criteria spanning three distinct aspects. Creativity metrics The novelty of generated designs was quantified using a CLIP-based cosine similarity metric. To establish a baseline of existing designs, we compiled a comprehensive archival corpus of 50,000 unique historical and modern lacquerware patterns. This dataset was aggregated from the digital archives of several collaborating cultural institutions and private collections; due to institutional agreements, the specific provenance of this corpus cannot be publicly disclosed. For each design generated in our study, its visual features were embedded using the openai/clip-vit-base-patch32 model. This embedding was then compared against the pre-computed embeddings of all 50,000 archival patterns. The novelty score is defined as, where is the embedding of the generated design and is the set of historical embeddings. A higher score signifies greater visual dissimilarity from the known design archive43. The novelty of generated designs was quantified using a CLIP-based cosine similarity metric. To establish a baseline of existing designs, we compiled a comprehensive archival corpus of 50,000 unique historical and modern lacquerware patterns. This dataset was aggregated from the digital archives of several collaborating cultural institutions and private collections; due to institutional agreements, the specific provenance of this corpus cannot be publicly disclosed. For each design generated in our study, its visual features were embedded using the openai/clip-vit-base-patch32 model. This embedding was then compared against the pre-computed embeddings of all 50,000 archival patterns. The novelty score is defined as, where is the embedding of the generated design and is the set of historical embeddings. A higher score signifies greater visual dissimilarity from the known design archive43. Diversity: Measured through LPIPS distance between variants generated from the same prompt40. Diversity: Measured through LPIPS distance between variants generated from the same prompt40. Technical quality metrics Material Fidelity: Specialist assessments (1–7 scale) on the degree to which produced designs reflect lacquerware methods. Material Fidelity: Specialist assessments (1–7 scale) on the degree to which produced designs reflect lacquerware methods. Structural Soundness: Engineering analysis of 3D printable viability for generated forms. Structural Soundness: Engineering analysis of 3D printable viability for generated forms. Workflow metrics Task Completion Time: Recorded automatically by the system. Task Completion Time: Recorded automatically by the system. Interaction Frequency: Count of refinement iterations per design. Interaction Frequency: Count of refinement iterations per design. System Usability Scale (SUS): Standardized questionnaire administered post-task44. System Usability Scale (SUS): Standardized questionnaire administered post-task44. A panel of three independent lacquerware experts evaluates all outputs through blind assessment, with inter-rater reliability measured via Krippendorff’s alpha (α > 0.75 for all metrics). Implementation details The HITL-AIGC system implements several key optimizations for real-world usability: Model configuration Base model: Stable Diffusion v2.1 with 768-dimensional latent space. Base model: Stable Diffusion v2.1 with 768-dimensional latent space. LoRA rank: r = 64, trained for 12,000 steps (batch size 16). LoRA rank: r = 64, trained for 12,000 steps (batch size 16). ControlNet: Initialized with depth and edge conditioning models. ControlNet: Initialized with depth and edge conditioning models. Hardware setup Frontend: iPad Pro with Apple Pencil for sketch input. Frontend: iPad Pro with Apple Pencil for sketch input. Backend: NVIDIA A100 80GB GPUs with Triton inference server. Backend: NVIDIA A100 80GB GPUs with Triton inference server. Latency: <1.2s for sketch refinements, < 3.5s for full generations. Latency: <1.2s for sketch refinements, < 3.5s for full generations. Interface features Dynamic parameter sliders for: Style adherence (0-100%).Detail intensity (0-100%).Tradition/innovation balance. Dynamic parameter sliders for: Style adherence (0-100%).Detail intensity (0-100%).Tradition/innovation balance. Style adherence (0-100%). Style adherence (0-100%). Detail intensity (0-100%). Detail intensity (0-100%). Tradition/innovation balance. Tradition/innovation balance. Multi-touch canvas supporting simultaneous sketch and gesture inputs. Multi-touch canvas supporting simultaneous sketch and gesture inputs. Baseline comparisons The study compares HITL-Lacquer against three alternative approaches: Manual design: Traditional artisan workflow as current industry standard. Manual design: Traditional artisan workflow as current industry standard. AutoGen: Automated Stable Diffusion pipeline without human guidance32. AutoGen: Automated Stable Diffusion pipeline without human guidance32. PatternMix: CNN-based style transfer system adapted from45. PatternMix: CNN-based style transfer system adapted from45. Each baseline is evaluated on identical tasks using the same metrics, with appropriate adaptations to their respective workflows (e.g., PatternMix receives style reference images instead of text prompts). Ethical considerations The experimental protocol adheres to institutional review board guidelines, with: Informed consent obtained from all participants. Informed consent obtained from all participants. Anonymization of design outputs for expert evaluation. Anonymization of design outputs for expert evaluation. Compensation at fair market rates for artisan time. Compensation at fair market rates for artisan time. Cultural sensitivity review by heritage preservation specialists. Cultural sensitivity review by heritage preservation specialists. All generated designs are clearly marked as AI-assisted to maintain transparency about their provenance and creation process. Experimental results and analysis Quantitative performance evaluation Our analysis revealed significant improvements across key metrics when participants used the HITL-AIGC system compared to their traditional workflow. As our data includes ordinal ratings and did not conform to assumptions of normality, we employed the non-parametric Wilcoxon signed-rank test for all paired comparisons between the two conditions. To address the issue of multiple comparisons across our four primary metrics, we controlled the False Discovery Rate (FDR) using the Benjamini-Hochberg procedure, reporting the adjusted p-value (q-value) for each test. The results, summarized in Table 1, show that the system led to a substantial increase in both design diversity and novelty, with both findings remaining statistically significant after FDR correction. The effect sizes for these improvements, calculated using Cliff’s delta (δ), were large, indicating a pronounced and meaningful difference between the two conditions. The most dramatic gain was in efficiency; artisans were able to finalize their design concepts 58.5% faster with the system. This result was highly significant and corresponded to a very large effect size (δ = -0.92). Importantly, the expert-rated feasibility of the designs showed no significant difference between the two methods, with a negligible effect size. This suggests that the gains in creativity and speed did not compromise the practical viability of the final designs. Table 1Comparative performance metrics with detailed statistical analysis.MetricTraditional (Median)HITL-AIGC (Median)Test statistic (Z)p-valueq-value (FDR)Effect size (Cliff’s δ)95% CI for δDesign diversity0.520.77-3.180.0030.0060.81[0.55, 0.93]Novelty score4.25.8-2.310.0210.0280.74[0.41, 0.89]Feasibility rating6.15.9-0.820.4120.412-0.15[-0.48, 0.21]Time per design (min)8234-3.96< 0.001< 0.004-0.92[-0.98, -0.79] Comparative performance metrics with detailed statistical analysis. Multi-modal control effectiveness Analysis of interaction logs reveals distinct usage patterns across input modalities. As shown in Fig. 2, sketch-based methods constituted 63% of refinements, whereas text prompts directed 28% of iterations and reference images shaped 9%. This distribution indicates artisans’ preference for visual over verbal specification when collaborating with the AI system. Distribution of control modality usage during design refinement sessions. Each modality’s efficacy differs across design phases. Conceptualization: Text prompts excel at establishing broad stylistic directions (e.g., “Edo-period maki-e with seasonal flowers”). Conceptualization: Text prompts excel at establishing broad stylistic directions (e.g., “Edo-period maki-e with seasonal flowers”). Refinement : Sketch inputs are most beneficial for modifying compositional components, such as altering the placement of decorative motifs. Refinement : Sketch inputs are most beneficial for modifying compositional components, such as altering the placement of decorative motifs. Detailing : Reference images are optimal for making adjustments tailored to specific materials, such as achieving the density of gold powder. Detailing : Reference images are optimal for making adjustments tailored to specific materials, such as achieving the density of gold powder. The implementation of ControlNet shows notable efficacy in upholding structural coherence, with 89% of resultant designs retaining no less than 80% of the initial sketch outlines. This function grants craftsmen sovereign command over core arrangements while assigning complex elaborations to artificial intelligence. Ablation study of control modalities To isolate the unique contribution of each input modality, we conducted an ablation study. A subset of designs was generated using only a single control method at a time: text-only, sketch-only, or reference-only. These were then evaluated by our expert panel against designs created using the fully integrated multi-modal system. The results, presented in Table 2, demonstrate a clear synergistic effect. While each individual modality offers distinct advantages, the integrated HITL-AIGC system significantly outperforms any single component in both style accuracy and overall perceived quality. This suggests that the combination of textual, structural, and referential controls allows artisans to achieve a level of nuance and precision that is unattainable when using any single input method in isolation. Table 2Ablation study results for different control modalities.Control modalityStyle accuracy (%)Expert quality rating (Median, 1–7 Scale)Text-only81.20%4.5Sketch-only65.70%5Reference-only72.40%4.8HITL-AIGC (All Modalities)87.40%5.9 Ablation study results for different control modalities. User experience and adoption The system usability scores showed an average of 82.4 on the SUS scale (SD = 6.7), which suggests high levels of approval among both professional artisans and students. Qualitative feedback identifies multiple recurring themes. Creative amplification: “The system showed me combinations I wouldn’t have tried manually” (Participant A3). Creative amplification: “The system showed me combinations I wouldn’t have tried manually” (Participant A3). Workflow efficiency: “I could explore ten variations in the time it normally takes to draft one” (Participant B2). Workflow efficiency: “I could explore ten variations in the time it normally takes to draft one” (Participant B2). Skill augmentation: “It helped me attempt techniques beyond my training while keeping the results authentic” (Participant A5). Skill augmentation: “It helped me attempt techniques beyond my training while keeping the results authentic” (Participant A5). Nonetheless, a subset of individuals reported early unease with ceding some degree of autonomy, especially those veteran craftsmen with extensive hands-on experience. This suggests the importance of gradual onboarding when introducing AI tools to traditional craft contexts. Cultural coherence analysis To rigorously quantify cultural authenticity, we utilized a dedicated style classifier based on the ResNet-50 architecture. Crucially, to mitigate potential overfitting and ensure unbiased evaluation, this classifier was trained on an independent corpus of 10,000 lacquerware images—a dataset entirely distinct from the 3,200 images used for the diffusion model’s LoRA fine-tuning. This separation prevents data leakage and guarantees that the metric reflects genuine stylistic generalization rather than memorization. The classifier achieved a 96% agreement rate (top-1 accuracy) on a held-out validation set, establishing a reliable baseline for assessing the generated outputs. Using this validated model, we analyzed the AI-produced designs in relation to seven predefined lacquerware style classifications. Figure 3 displays the alignment of the distribution with human-generated artifacts, demonstrating that the system effectively preserves local stylistic traits (average style accuracy = 87.4%). Comparison of style distribution between AI-generated and traditional lacquerware designs. The system displays notable capabilities in: Preserving material-specific rendering of maki-e (94% accuracy). Preserving material-specific rendering of maki-e (94% accuracy). Preserving appropriate motif proportions in relation to the dimensions of the object (91% compliance). Preserving appropriate motif proportions in relation to the dimensions of the object (91% compliance). Honoring color pairings with cultural importance (89% accuracy). Honoring color pairings with cultural importance (89% accuracy). The findings establish the specialized LoRA adaptation effectively captures domain knowledge absent predefined rule-based limitations. Limitations and failure modes Although the system is mostly effective, it has a number of discernible shortcomings. Complex 3D Forms: Designs for highly sculptural objects (e.g., nested boxes) show 23% lower feasibility ratings. Complex 3D Forms: Designs for highly sculptural objects (e.g., nested boxes) show 23% lower feasibility ratings. Rare Techniques : Less conventional techniques such as tsuishu (carved red lacquer) show greater incidence of errors (37% compared to the typical 11%). Rare Techniques : Less conventional techniques such as tsuishu (carved red lacquer) show greater incidence of errors (37% compared to the typical 11%). Cultural Hybrids: Attempts to blend distinct regional styles sometimes produce aesthetically inconsistent results. Cultural Hybrids: Attempts to blend distinct regional styles sometimes produce aesthetically inconsistent results. Examination of unsuccessful instances indicates these problems arise chiefly from insufficient data in specific categories, not from inherent constraints of the model. The cyclical structure of the HITL process aids in addressing these issues, as craftsmen can adjust flawed results by hand during the improvement phase. Computational performance The system attains operational deployment velocities by means of multiple optimizations. Generation Time: 2.8s average for 512 × 512 images (A100 GPU). Generation Time: 2.8s average for 512 × 512 images (A100 GPU). Memory Usage: 8.2GB VRAM during active design sessions. Memory Usage: 8.2GB VRAM during active design sessions. Sketch Response: 0.9s latency for contour refinements. Sketch Response: 0.9s latency for contour refinements. These metrics indicate the practicality of deploying in real-world settings, including those with limited resources in workshop contexts. The progressive rendering method grants artisans the capability to initiate design evaluation within merely 1.2 s, as finer elements progressively improve during subsequent diffusion stages. Longitudinal skill development An unexpected benefit emerged in the system’s potential as a pedagogical tool. After engaging with the artificial intelligence tool, learners exhibited a 31% rise in pre/post-assessment results pertaining to conventional design tenets. Motif arrangement principles. Motif arrangement principles. Material technique selection. Material technique selection. Cultural symbolism application. Cultural symbolism application. This indicates that the system can render implicit design knowledge more transparent and available to learners, which may help resolve difficulties in craft education. Pilot study: a bridge to 3D-aware design generation To specifically address the critical transition from planar ideation to volumetric fabrication, we conducted a targeted pilot study exploring the integration of depth-conditioned diffusion with orthogonal projection mapping. This investigation aims to validate whether the aesthetic coherence achieved in the 2D latent space can be preserved when designs are mapped onto the curved geometries typical of tixi boxes and meiping vases. Methodological Protocol. The pilot workflow extends the HITL-Lacquer system by incorporating a Depth-Aware ControlNet branch coupled with a custom UV-mapping pipeline. We selected a representative subset of ten high-complexity designs (five maki-e compositions and five carved lacquer patterns) and projected them onto standardized 3D meshes of traditional lacquerware vessels. The process involved three critical steps: Canonical view acquisition: We rendered normalized depth maps of the target 3D meshes using an orthogonal projection camera. This eliminates perspective distortion inherent in standard photography, ensuring that the generative model perceives the “true” geometric proportions of the substrate. Canonical view acquisition: We rendered normalized depth maps of the target 3D meshes using an orthogonal projection camera. This eliminates perspective distortion inherent in standard photography, ensuring that the generative model perceives the “true” geometric proportions of the substrate. Geometry-conditioned generation: The diffusion model, conditioned on these depth maps, synthesized surface textures. Crucially, this conditioning forces the model to modulate the density of high-frequency details—such as gold powder distributions in maki-e—according to the surface curvature gradient. This mimics the artisan’s technique of bokashi (gradation), where ornamentation is often rarefied at the peripheral curvature to enhance volumetric perception. Geometry-conditioned generation: The diffusion model, conditioned on these depth maps, synthesized surface textures. Crucially, this conditioning forces the model to modulate the density of high-frequency details—such as gold powder distributions in maki-e—according to the surface curvature gradient. This mimics the artisan’s technique of bokashi (gradation), where ornamentation is often rarefied at the peripheral curvature to enhance volumetric perception. Inverse projection mapping: The generated 2D manifolds were back-projected onto the 3D meshes using a cylindrical mapping algorithm for vases and a planar-box mapping for tiered containers. Inverse projection mapping: The generated 2D manifolds were back-projected onto the 3D meshes using a cylindrical mapping algorithm for vases and a planar-box mapping for tiered containers. Quantitative assessment of geometric consistency To rigorously evaluate the fidelity of this translation, we introduced the Geometric Consistency Metric (GCM). GCM quantifies the geodesic distortion of semantic keypoints (e.g., the center of a chrysanthemum motif or the eye of a crane) when mapped from the planar generation space to the curved mesh surface. It is defined as: wheredenotes the geodesic distance between keypoints on the 3D mesh surface, and represents the Euclidean distance in the generated 2D latent image. A lower GCM indicates superior structural preservation. Results and empirical validation The pilot study yielded a mean GCM of 0.152 (normalized scale 0–1), significantly outperforming a baseline “naive wrapping” approach (mean GCM = 0.425). As illustrated in Table 3, the depth-conditioned workflow successfully maintained semantic integrity across convex surfaces, with motif distortion remaining below the perceptual threshold for 80% of the test cases. However, we observed minor localized artifacts in regions of extreme topology, specifically singularity distortions at the poles of spherical forms. Table 3Quantitative evaluation of geometric consistency in pilot study.Geometry typeTest object (Model)Baseline (Naive Wrap) GCMOurs (Depth-Conditioned) GCMImprovement (%)Artifact observationConvex PrimitiveSphere (Tea Caddy)0.452 0.151 + 66.6%Polar singularity (minor)CylindricalBrush Pot (Cylinder)0.284 0.089 + 68.7%Seam alignment validCompound curveMeiping Vase0.553 0.178 + 67.8%Neck curvature distortionPlanar/TieredJubako (Box)0.051 0.032 + 37.3%Minimal distortionComplex organicStatue base0.785 0.310 + 60.5%Occlusion shadowsMEAN SCORE–0.425 0.152 + 64.2% –Bold values indicate the best performance for each metric. Quantitative evaluation of geometric consistency in pilot study. Bold values indicate the best performance for each metric. Despite these topological constraints, this pilot study empirically validates that the stylistic grammars captured by our LoRA model possess sufficient semantic robustness to be lifted into 3D environments. By leveraging depth maps as a structural prior, the system effectively mitigates the “texture-geometry dissonance” often observed in naive AIGC applications, establishing a verified pathway for future integration with neural rendering techniques like NeRF or Gaussian Splatting. Discussion: implications for creative practice and cultural heritage Limitations of the HITL AIGC system for lacquerware design Although the suggested system shows encouraging outcomes, a number of technical and practical limitations warrant attention. The existing approach faces challenges with designs demanding accurate multi-dimensional visualization, especially for intricate functional items such as tiered containers or curved receptacles. This constraint arises due to the two-dimensional structure of the diffusion model, incapable of entirely grasping the interaction between surface ornamentation and three-dimensional shape central to numerous lacquerware objects46. Moreover, the system shows diminished capability in producing designs that blend various distinct methods concurrently (e.g., carved lacquer with gold inlay), as the model often favors one method over others in combined results. While the pilot study (Pilot study: a bridge to 3D-aware design generation) demonstrated the viability of depth-conditioned mapping for standard geometries, the current system still encounters hurdles with highly irregular topologies. The intrinsic two-dimensional structure of the underlying latent diffusion model cannot natively encode the complex self-occlusions found in intricate carved lacquer (tixi) works or the undercut geometries of elaborate handle attachments. Although our GCM analysis confirms that semantic coherence is maintained for convex forms, concave geometries occasionally suffer from projection shadows where the generative model fails to infer hidden surface details, necessitating manual intervention during the UV unwrapping phase. The training data distribution also introduces subtle biases in generation quality. Designs that echo prevalent historical motifs attain greater fidelity compared to those venturing into atypical styles, which underscores the difficulty of reconciling innovation with convention in AI-aided design. This phenomenon aligns with observations in other AI-assisted craft systems47, where models often favor dominant styles in the training corpus. The current annotation schema, while comprehensive, may inadvertently reinforce these biases by categorizing designs according to established historical classifications rather than more fluid stylistic dimensions. Furthermore, our evaluation of cultural coherence, while grounded in expert assessment and a style classification model, highlights a broader challenge in the digital humanities: the difficulty of quantitatively measuring cultural authenticity. Our current approach relies on expert consensus and classification against known styles, which are valid and widely used methods. However, these are ultimately proxies for the complex, intangible qualities that define cultural identity in craft. We acknowledge that our metrics do not fully capture the nuanced rules of motif combination, symbolic meaning, or the appropriateness of a design for a specific cultural context. This limitation suggests a need for more sophisticated evaluation frameworks that can move beyond stylistic similarity to assess deeper semantic and cultural appropriateness. Similarly, our use of a CLIP-based similarity score to measure novelty warrants careful interpretation. This metric effectively quantifies a design’s visual and semantic distance from a large corpus of historical examples, providing an objective measure of dissimilarity. However, we recognize that true “innovation” in a cultural context is not merely about being different, but about being different in a meaningful way that respects and extends a tradition. The CLIP score, in its current form, cannot distinguish between a culturally resonant innovation and an aesthetically incongruous outlier. For this reason, the novelty score was intentionally triangulated with expert feasibility and coherence ratings to provide a more holistic assessment. The tension between computational novelty and human-judged innovation remains a key area for future research. Potential application scenarios beyond lacquerware design The central structure of the system implies potential application to additional conventional crafts sharing analogous design features. Metalwork featuring elaborate surface designs, including cloisonné or repoussé, mirrors lacquerware’s focus on planar ornamentation applied to volumetric shapes48. The multi-modal control mechanisms could potentially translate well to these domains, with adjustments to the material-specific components of the diffusion model. Similarly, the approach shows promise for textile design applications where pattern continuity across seams and structural elements presents analogous challenges to lacquerware’s curved surfaces. In addition to conventional crafts, the human-AI collaboration model yields valuable perspectives for modern design disciplines grappling with analogous conflicts between uniformity and innovation. Industrial design processes requiring both technical precision and aesthetic innovation, such as automotive interior detailing or consumer electronics casing, could benefit from adapted versions of the iterative refinement workflow49. The system’s capacity to uphold stylistic harmony while generating varied adaptations indicates specific relevance for branding and identity design scenarios, where uniformity in appearance across different implementations is critical. Ethical issues in using cultural heritage data The development and deployment of such systems raise important questions about cultural ownership and representation. Although our dataset consists of publicly available museum collections and designs contributed by artisans, the alteration of these works by AI generation produces derivative forms whose cultural authenticity demands meticulous examination50. The existing framework currently establishes elementary attribution monitoring by connecting produced designs to their principal reference origins, yet more comprehensive provenance approaches could be required as the technology gains broader acceptance. A notably delicate matter pertains to the possible commodification of AI-created designs derived from conventional motifs. Numerous lacquerware designs hold profound cultural meaning for their originating communities, which necessitates attention to proper application settings and equitable distribution of advantages51. The current implementation includes style tagging to identify culturally sensitive elements, but additional safeguards may be needed to prevent inappropriate adaptations or decontextualized usage. These challenges mirror broader debates in the digital humanities about balancing open access with cultural respect52. Impact of the AIGC system on traditional lacquerware craftsmanship Incorporating AI tools into conventional craft methodologies invariably influences the dissemination and transformation of artisanal expertise. Although certain professionals voiced apprehensions regarding possible deskilling, our findings indicate outcomes that are more complex in nature. The system seems to alter rather than diminish skill demands, as artisans acquire fresh expertise in directing and improving AI results without losing their essential technical proficiencies53. This phenomenon resembles historical transitions where new tools transformed rather than replaced traditional skills, such as the adoption of power tools in woodworking. The system’s influence on creative decision-making processes warrants particular attention. Certain craftspeople described shifts in their thought processes, drawing upon the AI’s unforeseen proposals as catalysts for creativity instead of relying solely on innate inspiration. This transformation reflects mental adjustments seen in additional AI-supported artistic fields54, where individuals modify their processes to include possibilities generated by machines. Although predominantly favorable, these alterations could exert subsequent influences on the progression of design traditions across successive cohorts of practitioners. Contribution of the system to long-term lacquerware cultural heritage preservation Apart from direct design uses, the system’s organized depiction of lacquerware beauty serves as important archival material for safeguarding cultural heritage. The trained model captures design principles and stylistic connections which might otherwise remain unspoken knowledge possessed by expert craftsmen55. This digitally encoded repository of information may aid forthcoming preservation initiatives by establishing benchmarks for genuine reconstruction projects or assisting in the reassembly of impaired artifacts from the past. The iterative design history captured by the system also creates a novel resource for studying creative processes. Unlike traditional craft documentation that typically records only final products, the system preserves the complete trajectory from initial concept to finished design56. These records could inform both art historical research and craft pedagogy, revealing patterns in how artisans explore compositional variations or resolve design challenges. Over time, aggregated usage data might identify emerging stylistic trends or shifting cultural preferences within the lacquerware community. Future work A number of prospective avenues arise for advancing the functionalities and uses of the HITL-Lacquer framework. Expanding the model’s three-dimensional awareness constitutes a crucial subsequent advancement, given that existing two-dimensional production restricts the framework’s applicability for intricate operational artifacts. The adoption of neural radiance fields (NeRFs) or diffusion-based 3D synthesis would grant artisans the capacity to perceive designs within spatial environments, more effectively addressing the curvature and structural dynamics specific to lacquerware forms. Initial experiments with depth-conditioned generation indicate this extension is viable yet necessitates precise adjustments to preserve material consistency across different perspectives. To address the challenge of objectively measuring cultural authenticity, a future research direction will focus on developing a quantitative framework for cultural coherence. Moving beyond simple style classification, this framework would integrate multiple data streams. For instance, we could employ object detection models to analyze the frequency and co-occurrence of symbolic motifs, comparing them against established patterns in historical corpora. We could also develop metrics for color palette analysis that measure adherence to traditional color harmonies specific to different regions and periods. By combining these quantitative indicators with our existing expert evaluation protocol, we aim to create a more robust, multi-dimensional measure of cultural authenticity that balances computational objectivity with nuanced human judgment. Enhancing cross-cultural style blending presents an additional advantageous opportunity. Although the existing framework performs admirably in conserving local distinctiveness, the deliberate creation of blended forms may cultivate novel expressions that honor conventional roots. Creating a disentangled latent space for stylistic features, including motif density, color palettes, and compositional rules, would grant artisans the capacity to investigate interpolations between disparate traditions (for instance, merging Japanese maki-e and Vietnamese mother-of-pearl inlay) while preventing aesthetically incongruent results. This ability would correspond with current tendencies in cultural interaction while reducing dangers of unsuitable adoption. Long-term adaptation through continuous learning could address the system’s current limitations with rare techniques. Adopting a memory-augmented architecture permits the progressive addition of novel designs and artisan input while preserving essential core competencies against catastrophic forgetting. Methods such as elastic weight consolidation or generative replay can stabilize performance on established styles while absorbing new patterns from continuous workshop application. This adjustment would reflect the natural development of artisanal customs, where new ideas steadily become part of the process without disrupting the consistency of technique. Physical prototyping integration serves as a tangible method for connecting digital and physical processes. Integrating the AIGC system with robotic lacquer application or CNC carving systems would make it possible to quickly test generated designs in physical materials. This process necessitates methods for converting 2D designs into 3D manufacturing specifications while addressing material properties such as brushstroke depiction dependent on viscosity or dimensional changes during solidification. A two-way interaction between physical prototypes and digital adjustments can markedly improve design viability while upholding artisan control over the ultimate outcome. Community-centered knowledge sharing mechanisms would amplify the system’s cultural preservation impact. Creating a federated framework for artisan networks could support collaborative design exploration while adhering to regional intellectual property standards. Differential privacy methods can support style acquisition from decentralized datasets without necessitating the consolidation of a central dataset. This method would foster region-specific creativity and establish a dynamic repository of progressing craft techniques, thereby tackling obstacles related to preservation and accessibility in conventional arts education. Cognitive science investigations into human-AI creative dynamics could refine interaction paradigms. Eye-tracking research conducted in design sessions can uncover the visual evaluation processes of artisans when examining AI-generated suggestions, which contributes to optimizing interfaces for directing visual attention. Neurophysiological assessments can measure differences in cognitive demand between conventional and AI-supported workflows, thereby pinpointing ways to lessen decision fatigue during extended creative tasks. These insights would advance general human-computer interaction principles for skill-based domains where expertise shapes tool engagement patterns. Ethical framework development remains imperative as adoption grows. Establishing protocols for attribution of AI-assisted works in exhibitions and sales would clarify authorship boundaries in this hybrid creation space. Computational traceability features, such as blockchain-based design lineage tracking, could document human and machine contributions while supporting fair compensation models. Partnerships with experts in cultural history and indigenous knowledge systems are necessary to guarantee that system applications adhere to the values and benefit-sharing requirements of source communities. Scalability adaptations for diverse workshop environments would broaden accessibility. Compact mobile applications employing refined diffusion models could assist rural craftspeople with restricted internet access, whereas server-hosted iterations may deliver sophisticated functionalities for organizational clients. Modular design grants the capacity for adaptation to distinct regional practices, such as the addition of local symbolism dictionaries or modules tailored to specific techniques. This adaptability would guarantee the technology supports rather than homogenizes the varied global traditions of lacquerware. These directions together progress toward a vision in which artificial intelligence serves as both a medium for preservation and a driver of innovation, a twofold function critical for keeping traditional crafts viable in modern settings. Subsequent versions could investigate generative approaches for supplementary elements such as reconstructing historical records or forecasting deterioration in antiquities, thereby aligning technological advancements with preservation goals. The ultimate measure of success will lie not in technical sophistication alone, but in how effectively the tools empower artisans to extend their craft’s legacy while retaining its cultural soul. Conclusion The HITL-Lacquer system illustrates that artificial intelligence can act as a substantive partner in the innovation of traditional crafts if its development prioritizes domain-specific needs and preserves human creative autonomy. The framework effectively connects lacquerware’s cultural legacy with modern design demands by merging controllable diffusion models and multi-modal interaction mechanisms. The system’s capacity to produce culturally consistent adaptations while retaining artisan oversight resolves a pivotal issue in heritage conservation, fostering progress without diminishing genuineness. Advances in model adaptation techniques, specifically via LoRA fine-tuning and ControlNet application, achieve exact correspondence with the material and stylistic attributes of lacquerware. These developments address shortcomings of general generative models, which frequently do not grasp the subtle interaction of method, structure, and cultural meaning embedded in conventional artisanal practices. The closed-loop refinement workflow further distinguishes the approach by keeping human oversight during the creative process, which guarantees that final designs align with artisan intent instead of algorithmic bias. Apart from short-term efficiency improvements, the system’s foremost impact could reside in its capacity to transform vocational training and the dissemination of cultural knowledge. The tools render implicit design knowledge more accessible and explorable, creating novel avenues for skill acquisition while honoring the master-apprentice traditions that have preserved lacquerware for thousands of years. The technology’s capacity to record artistic choices additionally grants novel possibilities for examining and safeguarding non-material elements of artisanal traditions, which conventional preservation techniques frequently fail to capture. The moral issues posed by this research, spanning cultural depiction to financial effects on craft-based populations, underscore the necessity for continuous discourse as these innovations advance. Upcoming advancements should harmonize technological capacities with deliberate regulatory structures safeguarding indigenous knowledge practices while fostering suitable progress. The system’s architecture establishes a basis for these dialogues by emphasizing transparency, attribution, and user autonomy at each phase. As evidenced by empirical assessment, the HITL-Lacquer method attains functional efficacy while preserving artisanal quality beyond mechanized replication. Artisans retain their role as creative directors while gaining expanded capacity to explore compositional possibilities. This balanced dynamic suggests a viable model for human-AI collaboration in other heritage domains facing similar pressures of preservation and renewal. The technology’s ultimate value will be measured not by its standalone outputs, but by how it empowers practitioners to extend their craft’s legacy in culturally grounded yet forward-looking ways. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements The authors would like to thank their colleagues, who distributed the questionnaires to students, and the students who took the time to complete the questionnaires online. Author contributions Author Contribution StatementJL and YL conceptualized the research. YL supervised the project and was the corresponding author. ZX was responsible for the methodology, developing the HITL-AIGC system, implementing the software, and performing the model training. ZX conducted the primary experiments and data analysis, with QH assisting in the experimental process and data collection. JL provided essential domain expertise for the curation and annotation of the lacquerware dataset. YL and ZX wrote the main text of the manuscript. ZX prepared Figs. 1, 2 and 3; Table 1. All authors reviewed and approved the final manuscript. Funding This work was supported by the National Art Fund project “Coromandel Lacquer Screen Creation Talent Training” (Project No: 2025-2-05-077-608). Data availability All data generated or analyzed during this study are included in this published article. The raw datasets generated during the current study are not publicly available due to restrictions imposed by our institution. However, the data are available from the corresponding author, Yixin Li, upon reasonable request. Declarations Competing interests The authors declare no competing interests. Ethical approval All experimental protocols involving human participants were conducted in strict adherence to relevant ethical guidelines and regulations, including the Declaration of Helsinki. Informed consent was obtained from all participants prior to their involvement in this study. The authors confirm that all experiments were conducted in accordance with relevant guidelines, regulations, and ethical standards. Approval for the experimental protocol(s) was obtained from the Ethics Committee of Hunan University of Technology and Business. Prior to submission, the manuscript’s Methods section was rigorously reviewed to ensure un-ambiguous disclosure of compliance and approval details. Furthermore, our pilot study on 3D-aware generation confirms that the stylistic fidelity achieved in the 2D domain can be successfully translated to volumetric forms through depth-conditioned control, establishing a foundational link between AI-assisted ideation and digital fabrication workflows. Informed consent Informed consent was obtained from all subjects involved in this study. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement All data generated or analyzed during this study are included in this published article. The raw datasets generated during the current study are not publicly available due to restrictions imposed by our institution. However, the data are available from the corresponding author, Yixin Li, upon reasonable request.