Sustainable Design on Intangible Cultural Heritage: Miao Embroidery Pattern Generation and Application Based on Diffusion Models

Abstract

Miao embroidery holds significant cultural, economic, and aesthetic value. However, its transmission faces numerous challenges: a high learning threshold, a lack of interest among younger generations, and low production efficiency. These factors have created obstacles to its sustainable development. In the age of artificial intelligence (AI), generative AI is expected to improve the efficiency of pattern innovation and the adaptability of the embroidery industry. Therefore, this study proposes a Miao embroidery pattern generation and application method based on Stable Diffusion and low-rank adaptation (LoRA) fine-tuning. The process includes image preprocessing, data labeling, model training, pattern generation, and embroidery production. Combining objective indicators with subjective expert review, supplemented by feedback from local artisans, we systematically evaluated five representative Miao embroidery styles, focusing on generation quality and their social and business impact. The results demonstrate that the proposed model outperforms the original diffusion model in terms of pattern quality and style consistency, with optimal results obtained under a LoRA scale of 0.8–1.2 and diffusion steps of 20–40. Generated patterns were parameterized and successfully implemented in digital embroidery. This method uses AI technology to lower the skill threshold for embroidery training. Combined with digital embroidery machines, it reduces production costs, significantly improving productivity and increasing the income of embroiderers. This promotes broader participation in embroidery practice and supports the sustainable inheritance of Miao embroidery. It also provides a replicable technical path for the intelligent generation and sustainable design of intangible cultural heritage (ICH).

1. Introduction

1.1. Intangible Cultural Heritage: Miao Embroidery

Intangible cultural heritage (ICH) embodies the historical memory and cultural heritage of human society, serving as an important carrier of national cultural identity and artistic creation [1,2,3]. Under the impact of globalization and digitalization, the inheritance of ICH faces severe challenges [4,5]. The core issues in ICH skills, are the reduction in artisans [6,7] and the lack of market adaptability [8,9], which have made many endangered ICH projects urgently need new protection and development paths. Sustainable design, as an important research direction of ICH, emphasizes the coordinated development of culture, ecology, and society, providing a feasible strategy for the living inheritance of ICH [10,11]. Through sustainable design methods, ICH can not only adapt to the needs of modern society but also rejuvenate in the new era [12]. Miao embroidery, a primary decorative technique for Miao clothing, is prevalent in Guizhou and Hunan, with a history spanning over 2000 years [13]. As an important ICH of China, it embodies the unique historical culture of the Miao people and possesses extremely high artistic value and national identity. Culturally, it is deeply embedded in local traditions, such as ritual garments, festive attire, and symbolic motifs like butterflies, birds, and ancestral totems, which reflect the spiritual beliefs and historical narratives of the Miao communities [14,15]. Economically, Miao embroidery has played a key role in supporting local women’s livelihoods through the growing market for handmade crafts and cultural tourism, with embroidery markets becoming important sources of income in many rural communities [16]. In terms of aesthetic value, Miao embroidery exhibits remarkable visual diversity and craftsmanship, which demonstrates distinct structural patterns, colors, and techniques that inspire contemporary textile and fashion design innovation [17]. However, the inheritance of these techniques remains largely dependent on traditional master-apprentice systems, with a long and complex learning cycle, compounded by regional stylistic differences that hinder the standardization and scalability of transmission [18]. Therefore, Miao embroidery embodies profound cultural connotations, economic potential, and design aesthetics and should be effectively protected and sustainably developed in the context of contemporary society.

1.2. Diffusion Models: Principles, Cross-Domain Applications, and Relevance to ICH

With the rapid development of Generative AI, the digital protection and innovation of Miao embroidery have ushered in new technological opportunities. Among them, the diffusion model, as an emerging deep generative model, has shown significant advantages in the field of computer vision due to its powerful content generation capabilities [19,20,21]. A diffusion model is a deep generative model that is based on two stages, a forward diffusion stage and a reverse diffusion stage. In the forward diffusion stage, the input data are gradually perturbed over several steps by adding Gaussian noise. In the reverse stage, a model is tasked at recovering the original input data by learning to gradually reverse the diffusion process, step by step [22]. The core idea of the diffusion model is to generate high-quality new data samples by gradually adding noise to the data and learning to reverse this process [23]. This mechanism enables it to learn and reproduce complex visual styles effectively, with broad application potential in style transfer, creative design, and visual content generation. Currently, several representative diffusion models are available, including Stable Diffusion [24], DALL·E [25], and Imagen [26]. They have demonstrated excellent generative capabilities and broad application prospects. In the field of medicine, Schneuing, A. et al. [27] proposed the Special Euclidean group in three dimensions equivariant diffusion model to achieve structure-based drug design. Wang, Y. et al. [28] employed a generation method based on diffusion models and molecular dynamics to design antimicrobial peptides from scratch, thereby enhancing the diversity and novelty of the generated peptides. In the field of cultural relic restoration, Zhang, R. et al. [29] proposed an image colorization diffusion model, the Hybrid Loss Function Denoising Diffusion Probabilistic Model, which is suitable for ethnic historical costumes, and trained the loss function to improve the image coloring quality. In the field of materials science, Yang, Z. et al. [30] improved the generalization ability of the diffusion model through pre-training and fine-tuning, accelerating the discovery and application of high-performance solid electrolytes.

In the field of design, the application of diffusion models has been continuously expanding, covering multiple areas. In the field of interior design, Chen, J. et al. [31] constructed a unique dataset of interior decoration styles, proposed a new loss function that considers decoration style, and utilized the dataset to retrain the diffusion model. The proposed diffusion model can generate interior design images with specific decoration styles and spatial functions from text descriptions end-to-end, improving the design and decision-making efficiency of designers. In the field of architectural design, Cui, X. et al. [32] developed a latent diffusion model suitable for architectural design. By inputting architectural line drawings, it can generate high-quality urban design renderings with accurate layouts and details. Zhang, J. et al. [33] developed a diffusion model design method focusing on night store facades. By introducing the Low-Rank Adaptation of Large Language (LoRA) model [34], the accuracy and stability of the generated images were improved. In the field of game design, Xiang, W. et al. [35] applied text to image models, introduced the LoRA model based on the stable diffusion model, and modified the structure of the variable autoencoder to improve the efficiency of image generation. After obtaining the initial generation result, the local area of the generated image was personalized through mask operation. Through pre-generation and post-processing, the model can accurately generate semantically guided face images, effectively serving the character avatar design. Zhang, Y. et al. [36] introduced geometric and lighting conditions to ensure that the generated material is consistent with the given geometry and ambient light, thereby avoiding the shadow interference that occurs in traditional methods. At the same time, the text-guided generation method is combined with the diffusion model of geometry and lighting perception to optimize the material generation through Classifier Score Distillation loss. This method is suitable for generating high-quality Physically Based Rendering materials in game development, film production, and Augmented Reality/Virtual Reality scenes. In the field of product design, Yin, H. et al. [37] proposed a Midjourney product design method based on Artificial Intelligence Generated Content, which is equipped with a prompt formula and accompanying AIGC-based Midjourney Prompt Cards to significantly improve the efficiency of designers in the product development process. In the field of text design, Iluz, S. et al. [38] proposed a method for creating vector-format words as image illustrations automatically. Under the guidance of a pre-trained stable diffusion model, the outline of each letter was optimized to convey the desired concept. At the same time, an additional loss function was introduced to ensure the readability of the text and maintain the font style. The text image illustration demonstrates visual creativity and opens the possibility of semantic typography using large visual language models. In the field of clothing design, Chen, Y. et al. [39] proposed an intelligent fashion design generation method by building a LoRA model of prompt templates and specific attributes, combining fashion attribute knowledge into the generation process of a stable diffusion model. Zhang, Y. et al. [40] evaluated the effectiveness of Midjourney in fashion design and related commercial applications, demonstrating that this AI tool can assist fashion designers in creating visually expressive clothing and ready-to-wear products that meet established design standards and consumer needs. Overall, these previous references present how diffusion models, a form of generative AI, are being specifically adapted and applied across a wide range of design disciplines. The core theme is the customization of these models to meet the unique demands of each field. In conclusion, their methods portray a field that is rapidly advancing from the novelty of AI image generation to its practical, systematic, and specialized application. The overarching goal is to create more controllable, efficient, and high-quality generative tools that act as powerful assistants, augmenting the capabilities of professional designers across various industries.

1.3. Related Work on ICH Pattern Generation

In recent years, there has been an increasing number of studies on the generation of ICH patterns using diffusion models, including traditional craft patterns similar to Miao embroidery. For example, Wang, Y. et al. [41] developed a batik-style LoRA model through generative artificial intelligence, which achieved the digital preservation and innovation of traditional batik patterns. Daffa Izzuddin Wahid, R. et al. [42] used a multimodal model to generate batik image labels and fine-tuned the denoising module of LDM to achieve high-quality and customizable batik pattern generation. Peng, X. et al. [43] proposed a fine-grained semantic adapter that introduces multi-perspective information to guide the diffusion process, retain the elements of Chinese landscape painting, and achieve digital generation with a suitable style. Building upon these advancements, generative models have made significant progress in generating ICH patterns, particularly by learning the style of specific cultural elements, which enhances the cultural adaptability of generated images.

1.4. Objectives and Highlights

However, existing research focuses on flat works, such as batik and painting. Less attention has been paid to key needlework textures, three-dimensional sense, and process feasibility in embroidery, which makes it challenging to support embroidery ICH design tasks with highly handicraft characteristics effectively. At the same time, existing research has not yet focused on applying diffusion models to the generation and style modeling of Miao embroidery patterns. Moreover, there remains a lack of research on how generative AI technologies can support not only the design process but also the continuity of traditional knowledge and practices in local embroidery communities.

Hence, this study aims to answer the following three core issues: (1) How can a culturally rooted Miao embroidery image dataset be constructed to authentically represent the symbolic patterns, needlework textures, and regional styles of Miao embroidery? (2) How can the cultural semantic information and needlework textures of Miao embroidery be effectively integrated into a LoRA-based diffusion model to generate patterns that are culturally faithful and production-ready? (3) Can the generated Miao embroidery patterns be smoothly converted into machine-embroidered products under conventional parameterization processes and receive positive recognition from local artisans, thereby promoting the sustainable inheritance of Miao embroidery? We hypothesize that the proposed AI-assisted design framework can strengthen the cultural transmission and market adaptability of Miao embroidery, thereby supporting its sustainable inheritance.

Therefore, this study takes Miao embroidery as the research object and constructs a high-quality pattern generation model that integrates cultural characteristics and sustainable concepts under small sample conditions. It aims to provide traditional artisans with creative support tools through an artificial intelligence-assisted design method, supporting the sustainable inheritance of Miao embroidery.

The main highlights of this study include the following:

• A representative dataset of Miao embroidery images was collected, preprocessed, and labeled through fieldwork and expert interviews, providing culturally rooted input for model training and evaluation.
• An AI-assisted design method based on Stable Diffusion and LoRA fine-tuning was proposed to enable efficient and style-consistent pattern generation under limited data conditions.
• Based on feedback from artisans and inheritors, the proposed method has good business and social impacts in improving creative efficiency, enhancing the participation of artisans, and promoting vocational education and ICH inheritance.

The second section systematically introduces the primary methods used in the study, including image preprocessing, pattern generation, generated pattern quality evaluation, and embroidery production; the third section shows and analyzes the effect of generated patterns by different embroidery types and comprehensively evaluates the model output results by combining quantitative indicators and subjective evaluation. Two high-performing samples were selected for physical embroidery production. Finally, the whole research work is summarized, the limitations of the current research are identified, and future optimization directions and research prospects are proposed.

2. Method

The framework of this study is shown in Figure 1. First, the Miao embroidery image preprocessing step ensures dataset quality through image collection, data standardization, and augmentation to enhance image quality. This step was necessary to unify the format, enhance sample diversity, and improve model robustness. Second, the Miao embroidery pattern generation step includes data labeling, LoRA training, and pattern generation. These methods were chosen to achieve style fidelity under limited data conditions. Third, the generative pattern quality evaluation step applies Fréchet inception distance (FID) evaluation and expert evaluation, which guides the selection of embroidery patterns. Finally, the generated pattern embroidery production step performs pattern selection, digital design, and embroidery production. This approach links cultural data collection with physical output, ensuring authenticity and practicality.

2.1. Preprocessing of Sample Images

2.1.1. Miao Embroidery Images Collection

This study created a dataset for diffusion model training. A total of 329 Miao embroidery photos were collected from Miao embroidery inheritors, private collectors, and museums. These embroideries, dating from the late Qing Dynasty (1644–1911) to the contemporary era, represent a stage in which Miao embroidery styles had largely stabilized, and were selected for their clear and well-preserved patterns. These images came from Miao costumes and folk artifacts to guarantee the diversity of data. The primary stylistic differences in shape, color, and texture across the five embroidery types are summarized in

Table 1. Visual features comparison of five Miao embroidery types.

Miao Embroidery Types	Shape	Color	Texture
Xi embroidery	geometric	silver, black	metallic, tin gimp
Poxian embroidery	totemic	red, blue	smooth, delicate, silk thread
Dazi embroidery	curvilinear	green, rose red, white	dense circular, raised
Zhou embroidery	symmetrical	green, red	wrinkled, ribbon-textured
Tiebu embroidery	four-way layout	blue, white, yellow	layered fabric, stitched

.

Xi embroidery is made by processing tin sheets into narrow strips, coiling them in perpendicular directions on the ground fabric, and flattening them to secure. Poxian embroidery is a process of splitting silk threads into 2 to 16 strands and using these threads to cover paper-cut patterns for embroidery, producing an effect similar to satin stitch [44]. Dazi embroidery is characterized by knotted dots formed through thread looping, which is technically similar to the knot stitch [45]. Zhou embroidery uses couched braids to create pleats on a base fabric, which are then fastened by stitching. Tiebu embroidery, a form of appliqué, is a general technique of adding cut pieces of material to an existing ground fabric [46]. The visual features of these five embroidery types are shown in Figure 2.

The dataset comprises 40 images of Xi embroidery, 90 images of Poxian embroidery, 58 images of Dazi embroidery, 78 images of Zhou embroidery, and 63 images of Tiebu embroidery. Representative examples of each embroidery type are shown in Figure 3.

The image collection was collected through fieldwork in southeastern Guizhou, China, where Miao communities are densely concentrated. As shown in Figure 4, the five embroidery types were collected from multiple representative locations: Xi embroidery from Nanzhai Town, Jianhe County; Poxian embroidery from Shidong Town, Taijiang County; Dazi and Tiebu embroidery from Kaitang Town, Kaili City; and Zhou embroidery from Shidong Town, Taijiang County and Xijiang Town, Leishan County, with many motifs originating from historical Miao migrations and refined in the towns where they are still practiced today.

2.1.2. Standardization of Image Data

To ensure that the model can accurately learn the texture and structural features of Miao embroidery, a series of steps was applied. Specifically, we removed non-embroidery image background elements, corrected perspective distortion to obtain a frontal view, and aligned image orientations to ensure consistency across samples. These steps enhanced the accuracy of subsequent analysis and model training. The images were saved as PNG format, and their resolution was adjusted to 512 × 512 pixels. The preprocessed samples of each embroidery method are shown in Figure 5.

2.1.3. Data Augmentation and Quality Improvement

Miao embroidery, as a representative item of China’s ICH, faces challenges such as long learning cycles and limited artisan inheritance. These factors lead to a low number of handcrafted samples, resulting in insufficient data for model generalization. To enhance the diversity and robustness of the training data, we employed a data augmentation strategy (random cropping, scaling, randomly rotated (90°, 180°, 270°), and horizontally and vertically mirrored) that expands the number of images for each embroidery type to 500. This ensures that each image displays the detailed features and improves the diversity of the dataset.

2.2. Miao Embroidery Pattern Generation

2.2.1. Data Labeling

Under the guidance of five experienced Miao embroidery inheritors (corresponding to Experts 1–5 in Appendix A), who are approximately 60 years old and possess over 20 years of embroidery expertise, each image in the dataset was meticulously annotated based on five semantic dimensions: embroidery type, cultural motif, shape, color, and texture. Shape, color, and texture have been used as important visual characteristics in the analysis of embroidery [47]. In Miao embroidery, these elements also carry cultural meanings. A cultural motif captures implicit cultural meanings within the tradition [48]. The embroidery type was determined by the expert judgment of Miao embroidery inheritors, who also confirmed these dimensions.

Table 2. Labeled example of Poxian embroidery image across five features used for model training.

Original Image	Miao Embroidery Type	Cultural Motif	Shape	Color	Texture
	Poxian embroidery	Butterfly	Symmetrical, curved	Red ground fabric, blue and multicolor embroidery	Silk thread, satin stitch

illustrates one representative example from the full dataset.

Subsequently, an automatic image tagging tool (WD14 Tagger) was used to extract semantic labels from the embroidery dataset, assisting in constructing prompt text pairs for model training. Then, the automatically generated results were manually corrected based on expert advice. Label conflicts were resolved by removing mutually exclusive or inconsistent tags. Missing information, especially in the theme and texture dimensions, was supplemented by expert annotation. During this process, priority was given to expert-validated labels and culturally significant categories to ensure labeling accuracy and consistency.

2.2.2. LoRA Model Training

The dimension and the alpha parameter of the LoRA model were both set to 32, and only the low-rank matrix of the attention layer was optimized to improve the model’s adaptability to technical styles under small-sample conditions. The optimizer was AdamW8bit. The learning rates of U-Net and the text encoder were 0.0001 and 0.00001, respectively. The maximum number of training rounds was 10, and the batch size was 1. Mixed precision (bf16) and former acceleration were enabled during training, and potential features were cached to improve resource utilization efficiency. The training platform was configured with an Intel Core i5–12400F processor, an NVIDIA GeForce RTX 3060 Ti graphics card, and 16 GB of memory on Windows 11. The training for each type of embroidery LoRA model typically took approximately 1 h.

2.2.3. Embroidery Pattern Generation

After fine-tuning the model, the trained LoRA weights are loaded into the Stable Diffusion v1.5 inference framework by inputting positive and negative prompt words to constrain the generated content and style.

To explore the generation capabilities of the model under different prompt strategies, we designed two types of generation scenarios. In the first setting, only embroidery type words (e.g., “Xi embroidery”, “Poxian embroidery”) were used as positive prompts to control the overall stylistic characteristics of the generated patterns. In the second setting, cultural motif words (e.g., “butterfly”, “flower”, “dragon”) were combined with embroidery type prompts to test the model’s ability to integrate semantic content within a specific embroidery style. These two prompt strategies correspond to the experiments detailed in Section 3.1.1 and Section 3.1.2, respectively.

The parameter setup mainly includes LoRA weights and diffusion steps. The images generated under different parameter combinations are used for subsequent quantitative evaluation and subjective expert review. The impact of specific parameters on the generation effect will be analyzed in Section 3.2.

2.3. Generated Pattern Evaluation Method

To evaluate the quality and consistency of style in the generated pattern images, a comprehensive evaluation method is employed that integrates both objective metrics and subjective expert assessment.

Objectively, the FID is used to measure the similarity between generated images and authentic images in the feature space. FID evaluations were conducted on two steps: first, by comparing the generated results of five Miao embroidery types with their corresponding original images to assess the overall quality of style reproduction, and second, by using Tiebu embroidery as a representative case, evaluating FID scores across different generation parameters to analyze the impact of model tuning on image fidelity.

Subjectively, to comprehensively assess the stylistic fidelity and design adaptability of the generated Miao embroidery patterns, a panel of 10 experts was invited to score the images based on six weighted criteria: shape, color, texture, artistic value, creativity, and application potential. Shape, color, and texture follow the semantic dimensions described in Section 2.2.1, reflecting both visual accuracy and cultural fidelity. Artistic value and creativity are adapted from established indicators in design-quality evaluation. Application potential was proposed by the expert panel to evaluate a pattern’s adaptability in actual design, focusing on functional suitability across different materials, processes, and product forms. The expert panel consisted of five officially recognized Miao embroidery inheritors, two skilled Miao embroiderers, and three professional embroidery designers, ensuring a balanced representation of traditional expertise, technical execution, and contemporary design insight (see Appendix A for participant details).

A five-point Likert scale was used for scoring (1 = “very poor”, 2 = “poor”, 3 = “fair”, 4 = “good”, and 5 = “excellent”). For each embroidery type, eight images were randomly selected and evaluated independently by all experts. The overall score $\mathrm{s}$ for each type was calculated using a weighted average formula as follows:

$${\mathrm{s}}_{\mathrm{i}}={\displaystyle \frac{1}{10}}{\displaystyle \sum _{\mathrm{j}=1}^{10}}{\mathrm{s}}_{\mathrm{i}\mathrm{j}}$$

(1)

$$\mathrm{s}={\displaystyle \sum _{\mathrm{i}=1}^6}{\mathsf{\alpha }}_{\mathrm{i}}{\mathrm{s}}_{\mathrm{i}}\mathrm{with}{\displaystyle \sum _{\mathrm{i}=1}^6}{\mathsf{\alpha }}_{\mathrm{i}}=1$$

(2)

where ${\mathsf{\alpha }}_{\mathrm{i}}$ represents the coefficient of score for criterion $\mathrm{i}$, ${\mathrm{s}}_{\mathrm{i}\mathrm{j}}$ is the evaluation score by expert $\mathrm{j}$ on criterion $\mathrm{i}$, ${\mathrm{s}}_{\mathrm{i}}$ represents the average score for criterion $\mathrm{i}$, and $\mathrm{s}$ represents the final average score. The value of the ${\mathsf{\alpha }}_{\mathrm{i}}$ has been decided by 10 experts after a brainstorm discussion.

2.4. Generated Embroidery Production

To validate the feasibility and adaptability of the generated patterns for production processes, we followed expert recommendations to select the highest-quality images produced by our method. These images are digitally converted and parameterized to embroidery design through specialized software (GiS BasePac 10, ZSK STICKMASCHINEN GMBH, Krefeld, Germany). Subsequently, physical production is conducted using a JF-0215-495 industrial embroidery machine (ZSK STICKMASCHINEN GMBH, Krefeld, Germany). Although specific workflows may vary depending on the digital embroidery machine used, the general production process can be divided into three main steps: (1) importing the generated design files into the embroidery machine, (2) adjusting machine parameters such as stitch type, density, and color sequence, and (3) executing the embroidery process. Throughout this process, appropriate threads and base fabrics corresponding to the specific embroidery styles are meticulously chosen to effectively retain the original cultural textures and intricate details of each pattern.

3. Results and Discussion

This section presents the experimental results of the Miao embroidery pattern generation method based on the latent diffusion model framework. Specifically, Section 3.1 shows the pattern generation performance of five types of Miao embroidery after fine-tuning the basic stable diffusion and LoRA models; Section 3.2 compares and analyzes the quality differences and parameter optimization trends of generated images under different models through the FID indicator; Section 3.3 combines expert subjective scores to evaluate the style consistency and conversion potential of generated images from the two dimensions of cultural restoration and design performance; Section 3.4 further compares the performance of five mainstream deep generation models in the Miao embroidery pattern generation task; Section 3.5 uses digital embroidery physical production as a carrier to verify the feasibility and innovative value of this method in practical applications.

3.1. Pattern Generation Results Using Stable Diffusion and LoRA Models

3.1.1. Generated Patterns Based on Embroidery Types

Under the same generation parameter settings, the generation time for each pattern ranged from 10.3 to 10.9 s, with only minor fluctuations, indicating that the model exhibited relatively stable efficiency across different embroidery styles, as shown in

Table 3. Time comparison of the diffusion model in generating patterns of different Miao embroidery types.

Miao Embroidery Types	Average Time Consumption of the Generation of Pattern (s)
Xi embroidery	10.9
Poxian embroidery	10.5
Dazi embroidery	10.4
Zhou embroidery	10.8
Tiebu embroidery	10.3

.

In addition, in order to more intuitively demonstrate the model’s generation effect on different Miao embroidery types, Figure 6 shows a representative original image sample and its corresponding generated pattern. From the generation effect, the model exhibited high consistency in pattern structure, central color control, and other aspects and can better restore the typical style characteristics of each type. Geometric types, such as Xi embroidery and Tiebu embroidery, exhibited clear styles and better model-fitting performance. Poxian embroidery also showed relatively stable results, with strong color fidelity and recognizable motif composition. In contrast, types with more complex texture levels, such as Dazi embroidery and Zhou embroidery, produced relatively fewer stable results. However, the basic outlines and color-matching rules were preserved, indicating that further improvements can be made in local accuracy and detail refinement. The generation results primarily demonstrate that the stable diffusion with the LoRA model has a strong ability to control the reconstruction of traditional pattern styles, providing strong support for the digital generation of Miao embroidery patterns.

3.1.2. Generated Patterns Based on Embroidery Types and Cultural Motifs

This section further investigates the model’s generation performance under culturally symbolic prompts. Consistent with the results in Section 3.1.1, the generation time for each pattern remains stable at approximately 10 s under the same parameter settings. To further evaluate the pattern generation ability of the LoRA model for each embroidery method under the guidance of concrete semantics, a generation experiment combining “specific embroidery type + specific prompt words” was designed here. The selected prompt word, “butterfly,” originates from the widely circulated “butterfly mother” pattern in Miao embroidery. It has a high degree of cultural symbolism and pattern stability. It not only carries a profound national cultural connotations but also provides a unified semantic anchor for comparing cross-technique styles. To ensure the comparability of the generation results, the same generation parameter settings were uniformly used during the experiment. Except for the corresponding prompt words used for each embroidery method, the other variables were kept consistent to eliminate the interference of non-model factors on the style expression.

From the results in Figure 7, although the generated patterns are all centered around the theme of butterflies, the embroidery models all show significant style differences and visual characteristics in terms of configuration language, pattern layout, color expression, and pattern details. Among them, the butterfly patterns generated by the Xi embroidery and Poxian embroidery models exhibit strong structural symmetry and cultural feature retention capabilities. In contrast, the Zhou embroidery is more decorative in terms of color and texture expression. The Tiebu embroidery and Dazi embroidery models, on the other hand, are outstanding in terms of pattern creativity and formal diversity. This part of the experiment verifies that the LoRA fine-tuning model has good semantic response capabilities while retaining the core characteristics of the technique style, providing a feasible solution for the semantic-driven generation of ICH patterns.

3.2. Pattern Generation Quality Analysis

The quantitative results of FID analysis for five embroidery types are listed in

Table 4. Quantitative Comparison of FID Scores across Miao embroidery types.

Miao Embroidery Types	FID Scores
	Stable Diffusion v1.5	Proposed Model
Xi embroidery	308.8767	108.4913
Poxian embroidery	220.9007	140.9676
Zhou embroidery	192.5459	149.6429
Dazi embroidery	207.3201	159.3247
Tiebu embroidery	278.9957	179.8882

. The proposed model achieves a significant FID reduction across all embroidery types, demonstrating strong style adaptability and generalization performance. Among them, the FID of Xi embroidery drops from 308.88 to 108.49, which is the largest drop, indicating that the model has a significant advantage in processing patterns with rich details and complex structures. The FID of Poxian embroidery decreases from 220.90 to 140.97, and the FID of Zhou embroidery decreases from 192.55 to 149.64, both showing good generation consistency.

In comparison, the improvements of Dazi embroidery and Tiebu embroidery are relatively small, with FID scores decreasing from 207.32 to 159.32 and from 278.99 to 179.89, respectively. However, the LoRA fine-tuning model still outperforms the original model in these two types, verifying its adaptability in multi-style scenarios.

In summary, the model proposed in this paper outperforms Stable Diffusion v1.5 in the five types of Miao embroidery styles, significantly reduces the distribution difference between the generated images and the original images, and provides an effective path to improve the quality of the automatic generation of ICH patterns.

To optimize the model parameter configuration and improve pattern generation quality, this study takes Tiebu embroidery as a representative case to analyze the influence of different generation parameters (sampling step and LoRA scaling factor) on FID scores. For the sampling step setting, the range was set from 5 to 50 with an interval of 5, balancing pattern quality and generation efficiency. Preliminary experiments showed that when the step count is below 5, the outputs tend to be blurry and structurally ambiguous, failing to meet the basic requirements of FID evaluation. Conversely, increasing steps beyond 50 results in a single pattern generation time exceeding 30 s, significantly affecting experimental efficiency and resource usage. The chosen range of 5–50 thus captures the most relevant quality transition zone while maintaining experimental feasibility. After fixing the step range, the LoRA scaling factor was varied from 0.2 to 2.0 in increments of 0.2. This range encompasses weak to strong style control, allowing for the evaluation of generation performance under varying degrees of stylistic influence. These results suggest that the parameter space defined by a scale from 0.2 to 2.0 and steps from 5 to 50 encompasses a practical control range, ensuring comparability in both the style expression and structural clarity of the generated images. This provides a robust foundation for subsequent FID analysis and parameter optimization. Based on this configuration, the study visualizes and analyzes the FID performance of Tiebu embroidery under different control settings to explore variations in generation quality.

Figure 8 shows the change in FID scores between the generated images and the original images of the LoRA fine-tuning model for Tiebu embroidery under different parameter combinations. Thirty images were generated for each set of parameters and compared with the original Tiebu embroidery image set for evaluation. The results show the FID scores follow a clear trend with parameter changes. In the low step range (≤15), the FID score remains high, and the generated images are not fully optimized. In contrast, with a medium step range (20–40) and a medium LoRA scale (0.8–1.2), most parameter combinations result in lower FID values, indicating this as the optimal range for Tiebu embroidery style generation. When the LoRA scale is within the range of 1.4–1.6, the pattern quality shows some fluctuations, but the low FID value can still be maintained within the step range of 25–35, indicating that moderately enhancing the model’s style expression ability can be achieved while maintaining stability. However, when the scale is ≥ 1.8 or the step is ≥ 40, the FID value generally increases. This may be attributed to overfitting or a style shift, causing the generated pattern to deviate from the original distribution, which can lead to problems such as pattern distortion and texture confusion.

Overall, a medium to slightly high scale (0.8–1.2) combined with a medium step number (20–40) constitutes a high-quality parameter range for Tiebu embroidery pattern generation, providing a practical reference for subsequent model optimization. This parameter range also proved effective across the other four embroidery types, suggesting its general applicability for high-quality generation in multi-style embroidery tasks.

3.3. Expert Evaluation of Pattern Generation Results

The evaluation weights for each dimension were determined by a panel of domain experts based on professional relevance. As shown in

Table 5. Expert evaluation of generated embroidery patterns (eight samples per embroidery).

Embroidery Types	$\mathbf{Average}\mathbf{Score}\mathbf{for}\mathbf{Each}\mathbf{Criterion}\left({\mathbf{s}}_{\mathbf{i}}\right)$						Final $\mathbf{Average}\mathbf{Score}\left(\mathbf{s}\right)$
	${\mathsf{\alpha }}_1$ = 0.2 (Shape)	${\mathsf{\alpha }}_2$ = 0.15 (Color)	${\mathsf{\alpha }}_3$ = 0.1 (Texture)	${\mathsf{\alpha }}_4$ = 0.2 (Artistic)	${\mathsf{\alpha }}_5$ = 0.25 (Creativity)	${\mathsf{\alpha }}_6$ = 0.1 (Application)
Xi embroidery	3.95	4.61	4.01	3.85	3.54	4.11	3.95
Poxian embroidery	4.04	3.92	3.88	3.87	3.95	3.54	3.90
Dazi embroidery	3.03	3.42	3.36	3.07	3.24	3.17	3.20
Zhou embroidery	2.78	3.26	3.08	2.76	3.10	2.88	2.97
Tiebu embroidery	3.46	3.47	4.12	3.96	4.13	4.23	3.87

, Xi Embroidery exhibited the most consistent and balanced performance across all evaluation dimensions. It achieved the highest scores in color (4.61) and texture (4.01), indicating that the generated patterns retained strong visual fidelity and structural clarity. With a weighted average of 3.95, it ranked first overall. Poxian Embroidery followed closely, with a total score of 3.90. The model demonstrated high performance in form (4.04), artistic value (3.87), and creativity (3.95), reflecting the strong stylistic distinctiveness and visual richness of this embroidery style. However, its relatively lower application score (3.54) suggests room for improvement in the practicality or usability of the generated outputs. Tiebu Embroidery achieved a weighted score of 3.87, showing notable strength in texture (4.12), artistic value (3.96), creativity (4.13), and application (4.23). These results indicate a high potential for real-world design integration and user acceptance. In contrast, Dazi Embroidery and Zhou Embroidery received lower overall evaluations, with scores of 3.20 and 2.97, respectively. Both styles underperformed in artistic and application dimensions, with Zhou Embroidery scoring especially low in form (2.78) and artistic value (2.76). These results reflect the model’s limited ability to reproduce the finer structural or symbolic features present in these more intricate traditional styles.

Overall, the evaluation suggests that the generative model performs best when applied to embroidery styles characterized by strong structural regularity and visual distinctiveness. In contrast, styles with higher semantic abstraction or complex craftsmanship present greater challenges.

3.4. Comparison with Other Generative Models

To further verify the superior performance of the proposed method in a broader range of generative model systems, this paper compares and analyzes the pattern generation capabilities of mainstream deep generative models under the same prompt conditions, focusing on evaluating the style adaptability and generation quality of different models in the Miao embroidery pattern task.

This study selected the current mainstream text-generated pattern models, such as Stable Diffusion v1.5, Stable Diffusion XL, DALL·E 3, and Midjourney, as comparison objects and carried out a quality comparison analysis of pattern generation under the unified prompt words “Poxian embroidery, red tone, butterfly, silk thread.” As shown in Figure 9, the generation results of each model, obtained under the same semantic input, are presented.

To further quantify the generation results, this paper calculated the average FID value of 30 patterns generated by each model, as shown in

Table 6. Quantitative Comparison of FID Scores for Butterfly Pattern across Generative Models.

FID Scores
Stable Diffusion v1.5	Stable Diffusion XL	DALL·E 3	Midjourney	Proposed Model
222.3312	303.1107	329.0787	266.9262	177.8229

. The results show that our method achieved the best FID score (177.8229), significantly outperforming the other models. This indicates that the proposed model offers superior generation quality and stronger style consistency in traditional embroidery pattern synthesis.

As shown in

Table 7. Expert ratings and weighted evaluation results of pattern effects of different generative models.

Generative Models	$\mathbf{Average}\mathbf{Score}\mathbf{for}\mathbf{Each}\mathbf{Criterion}\left({\mathbf{s}}_{\mathbf{i}}\right)$						Final $\mathbf{Average}\mathbf{Score}\left(\mathbf{s}\right)$
	${\mathsf{\alpha }}_1$ = 0.2 (Shape)	${\mathsf{\alpha }}_2$ = 0.15 (Color)	${\mathsf{\alpha }}_3$ = 0.1 (Texture)	${\mathsf{\alpha }}_4$ = 0.2 (Artistic)	${\mathsf{\alpha }}_5$ = 0.25 (Creativity)	${\mathsf{\alpha }}_6$ = 0.1 (Application)
Stable Diffusion v1.5	2.01	1.63	2.26	2.12	2.18	1.85	2.03
Stable Diffusion XL	1.55	1.43	2.25	2.14	2.82	1.90	2.07
DALL·E 3	1.92	2.28	2.35	2.71	1.68	3.62	2.29
Midjourney	2.52	3.88	3.12	4.65	4.68	3.25	3.82
Proposed model	4.60	4.75	4.66	3.92	4.58	4.03	4.43

, among the evaluated models, the proposed model achieved the highest weighted average score (4.43), outperforming all other models in every evaluation indicator. It demonstrated clear advantages in form, color, and texture, as well as in creativity and practical application. Midjourney also showed strong performance, particularly in artistic and creative aspects, with a weighted score of 3.82. In comparison, DALL·E 3 displayed moderate overall results, while SD v1.5 and SD XL obtained the lowest scores across most dimensions, indicating relatively limited effectiveness for this specific pattern generation task. These findings highlight the superior capability of the proposed model in capturing both the visual and cultural characteristics required for high-quality, style-consistent Miao embroidery pattern generation. The comprehensive expert evaluation further confirms the robustness and applicability of the proposed approach for practical ICH digitalization and creative design tasks.

3.5. Innovative Pathways: From Pattern Generation to Digital Embroidery Application

To verify the feasibility of the aforementioned generative model in actual design applications and explore the path from digital generation to the physical expression of traditional patterns, this study selected the two Miao embroidery types, Xi embroidery and Poxian embroidery, as representatives. It conducted digital embroidery transformation experiments on generative images to create physical embroidery products that can be worn or displayed.

Figure 10 shows the generated patterns and photos of real embroidery products, where contour extraction, line path planning, and stitch type assignment were performed according to the specific requirements of each embroidery type. Material selection during fabrication was also tailored to stylistic features: silver-toned metallic threads (Fujix, Kyoto, Japan) were used for Xi embroidery to replicate the dense patch structure and metallic texture characteristic of traditional tin foil patterns; for Poxian embroidery, polyester silk threads (New brothread, Guangzhou, China) were employed to simulate the sheen and delicacy of fine silk, highlighting the layered and refined aesthetic of the original hand-stitched work.

4. Impacts on Business and Society

In addition to demonstrating technical feasibility, these embroidery samples received positive feedback from experienced inheritors. They not only demonstrate the practical application potential of AI-assisted design in traditional crafts but also reveal its dual role in sustainable development and cultural propagation.

4.1. Business Impact

Many local artisans participate in the current Miao embroidery markets, turning their traditional skills into commercial products with sustainable income potential. Figure 11 shows a scene of a booth of Miao embroidery products in the local market. The various embroidery patterns in the booth reflect the adaptability and market vitality of traditional patterns of embroidery products in the contemporary context. This form of engagement shows that ICH skills not only have cultural value but also contain sustainable economic potential. In the context of rural revitalization and ICH protection policies, integrating digital tools into craft industries is encouraged to stimulate local economies and improve artisans’ livelihoods [49].

Based on fieldwork, the current business path of Miao embroidery can be mainly divided into the following three models:

• A pure manual model, with inheritors as the core, completely retaining the traditional process, often seen in the high-end art or collectibles market;
• The combination of manual and machine models, combining traditional skills with mechanical assistance, suitable for cultural and creative products;
• A pure mechanical model, mainly using an embroidery machine for mass production, serving the tourist souvenirs, mass consumer goods, and other markets.

The AI-assisted pattern generation method proposed in this study is mainly applicable to the second and third types of models. In these two models, the efficiency and diversity of pattern design are crucial to product development. However, nowadays, many young artisans primarily use existing patterns as references, leading to homogenized outputs. To this end, the proposed model in this study generates Miao embroidery patterns with consistency in style, good aesthetic quality, certain innovation, and suitability for actual embroidery production. This approach can significantly lower the technical barrier, improve creation efficiency, help artisans more efficiently transform creativity into patterns, and enhance their market participation and economic benefits. Compared with previous studies, our approach shares similarities with the work of Wang, Y. in increasing public engagement and pattern innovation [41]. More importantly, this study’s approach also increased artisans’ incomes. Feedback from local sellers indicated that they achieved strong sales results through three sales models, distinguishing between pure handmade, semi-handmade, and machine-made products and pricing them from high to low. This provides an effective strategy for meeting the needs of different customer groups.

4.2. Social Impact

The traditional Miao embroidery products are mainly handmade crafts and rely heavily on the individual experience of the artisans. However, in the current community practice, artisans are usually engaged in basic and repetitive tasks during the fabrication of embroidery products. They often use traditional patterns as the base for their embroidery. This fact makes many embroiderers perceived more as executors than as originators in the traditional Miao embroidery production, making it difficult to express their individual style and design intentions, thus limiting the scope for their creative expression in cultural production. This phenomenon has also affected the appeal and participation motivation of traditional crafts among young groups to a certain extent.

To address the problems of labor-intensive work with low added value in traditional embroidery, this study used AI technology in an actual community environment. We invited local Miao embroidery sellers and artisans to participate in practical experiments and provide feedback on the generated results. Although some artisans were unfamiliar and reserved about the technology in the early stage, in the actual operation process, the AI technology was gradually recognized for its efficient output, diverse styles, and pattern modifiability. Many artisans reported during the trial that the generative AI patterns provide more variability on the basis of maintaining the traditional style, reducing the dependence on traditional patterns and increasing their enthusiasm for participating in the initial design work.

The introduction of AI tools aims toward a collaborative mechanism at the design level, supporting option generation, parameterized refinement, and the translation of designs into digitizable formats, thereby expanding artisans’ involvement in creative activities. Especially in vocational education and ICH inheritance training, the pattern generation ability of AI tools can provide structured learning support for young artisans, helping them to further develop their creative capacities and broaden their role from primarily embroidery, the traditional pattern, to actively contributing as cultural participants.

In summary, AI-assisted pattern generation not only improves the efficiency of pattern design but also provides traditional artisans with the opportunity to express their individual style and actively participate in the creative design process. The application of this AI technology at the community level will help break the solidified labor division structure in traditional crafts, promote the transition of ICH inheritance from “skill reproduction” to “personal expression”, and enhance its vitality and participation in the context of contemporary society. This method also reflects the national digital heritage strategy, which advocates integrating advanced digital tools into community-based heritage education and practice [50]. Moreover, AI technology has contributed to higher income levels and attracted more local people to participate in embroidery practice. It also provides strong support for the sustainable inheritance of Miao embroidery.

5. Conclusions

This study proposes an AI-assisted method for generating Miao embroidery patterns and constructs a culturally rooted and production-oriented workflow to promote the sustainable inheritance of ICH. The results showed that (1) the small-sample Miao embroidery image dataset we constructed can truly present the pattern semantics, embroidery texture, and regional style, providing reliable input for model training and evaluation; (2) after incorporating cultural semantics and pattern features into the diffusion model fine-tuned by LoRA, the quality of five representative Miao embroidery generated patterns is superior to the original diffusion model and other generated models in terms of style consistency and cultural fidelity; (3) the generated patterns can be successfully parameterized and proofed for machine embroidery through a digital process, improving the production efficiency and replicability of Miao embroidery products.

Furthermore, our proposed AI technology can effectively assist artisans and embroidery designers in the design phase, enabling them to create more efficiently. This allows for the rapid generation of specific patterns and the smooth translation of designs into a production-ready digital format. Feedback from local sellers and artisans indicates that this approach significantly lowers the barrier to entry for young artisans to learn Miao embroidery and creates new income opportunities for the local community. These findings support our hypothesis that the AI-assisted design method can enhance the cultural communication and market adaptability of Miao embroidery, thereby providing a viable and replicable path for its sustainable development.

However, there are still certain limitations in converting the patterns generated by the diffusion model into actual manufacturing. The digital embroidery process is affected by factors such as stitch control, color level, and material characteristics, which makes some complex generated patterns need to be further adjusted during the real product manufacture process. Future research will further optimize the quality of AI pattern generation, enhance the model’s ability to comprehend the logic of the embroidery process, and improve the simulation of various stitches, stitch arrangements, 3D effects, and gloss effects to match the production needs of digital embroidery better. At the same time, we will explore the optimization of the entire process from AI generation to actual manufacturing and promote the in-depth application of AI in ICH protection, intelligent manufacturing, and sustainable design.