The application of artificial intelligence-assisted technology in cultural and creative product design

Abstract

This study proposes a novel artificial intelligence (AI)-assisted design model that combines Variational Autoencoders (VAE) with reinforcement learning (RL) to enhance innovation and efficiency in cultural and creative product design. By introducing AI-driven decision support, the model streamlines the design workflow and significantly improves design quality. The study establishes a comprehensive framework and applies the model to four distinct design tasks, with extensive experiments validating its performance. Key factors, including creativity, cultural adaptability, and practical application, are evaluated through structured surveys and expert feedback. The results reveal that the VAE + RL model surpasses alternative approaches across multiple criteria. Highlights include a user satisfaction rate of 95%, a Structural Similarity Index (SSIM) score of 0.92, model accuracy of 93%, and a loss reduction to 0.07. These findings confirm the model’s superiority in generating high-quality designs and achieving high user satisfaction. Additionally, the model exhibits strong generalization capabilities and operational efficiency, offering valuable insights and data support for future advancements in cultural product design technology.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-024-82281-2.

Introduction

Research background and motivations

Globally, the cultural and creative industries are becoming increasingly significant drivers of economic growth^1–3. This industry spans not only traditional arts and design fields but also includes diverse sectors such as advertising, media, software, and gaming^4–6. Existing design practices predominantly depend on the experience and intuition of designers. While effective for certain straightforward creative tasks, these approaches exhibit considerable limitations when confronted with complex cultural and creative design challenges. For instance, traditional design processes often involve lengthy timelines, requiring repeated manual creation and modification, which hinders quick adaptation to market demands. Designers frequently face constraints imposed by their existing experience and styles, impeding their ability to overcome innovation barriers. Complex design tasks typically demand substantial time and human resources, posing a significant burden for small to medium-sized creative teams. With the rapid development of digital technology, particularly the widespread application of artificial intelligence (AI), the field of cultural and creative product design is facing unprecedented opportunities for transformation^7,8. AI technology, by providing new tools and methods, brings unique perspectives and endless possibilities to creative design^9–11. The introduction of AI models offers a promising solution to these issues by automating the generation of design solutions, optimizing decision-making processes, and enhancing both design efficiency and creativity. This shift can significantly reduce resource consumption and alleviate the constraints faced by traditional design methods.

However, despite AI’s innovative momentum in the creative industries, it also poses a series of challenges and issues^12–14. Designers often face difficulties in adopting and integrating these new technologies, including concerns about the potential impact of technological dependence on creativity and originality. Moreover, the fusion of AI with creative thinking requires deeper understanding and research^15–17. Additionally, the adaptability and cultural sensitivity of AI technology are areas of concern, as cultural products must not only meet market demands but also convey the values and significance of specific cultures^18–20.

This study focuses on the application of Variational Autoencoders (VAE) and reinforcement learning (RL) in the design field. These AI technologies have demonstrated strong capabilities in pattern recognition, data generation, and decision optimization. However, their potential in cultural and creative product design remains underexplored. By investigating how VAE and RL can support the creative process, this study seeks to enhance design innovation and responsiveness to market trends while evaluating the impact of AI on designers’ creative autonomy.

Research objectives

The main objectives of this study are as follows: (1) to analyze and evaluate the application and potential of AI technologies, specifically VAE and RL, in cultural and creative product design; (2) to construct an interdisciplinary framework that integrates theory and practice, aiming to deepen the understanding of the role of AI technologies in the design process and their impact on creative output; (3) to explore how designers adopt and accept these technologies and their specific influences on design practice; and (4) to empirically assess the effectiveness of AI-assisted design tools in enhancing design innovation, efficiency, and market adaptability. The study focuses on a diverse selection of cultural artifacts, including textiles, ceramics, furniture, and digital art products. This range spans from traditional craftsmanship to contemporary design, enabling a thorough evaluation of AI technologies’ applicability and performance across different types of cultural and creative product design. By encompassing these domains, the study provides a comprehensive assessment of how AI technologies can be integrated into various design contexts.

Literature review

Application of AI in creative design

In recent years, AI technology has made substantial advancements in the field of creative design. Researchers have explored various AI methods, including Generative Adversarial Networks (GANs), VAEs, and RL, to drive innovation in design processes.

For instance, Hosseini and Rajabipoor Meybodi²¹ proposed a comprehensive model for the sustainable development of Iran’s creative industries, leveraging digital transformation and explanatory structural modeling. Mateja and Heinzl²² analyzed existing computational cultural creative systems based on machine learning, offering practical guidance for designing such systems in diverse contexts. As et al.²³ introduced a graph-based machine learning system capable of handling three-dimensional structures. This approach extracted significant components as subgraphs and recombined them into new cultural product designs, demonstrating a structured and compositional methodology beyond conventional media like images or text. Mahmud et al.²⁴ explored cultural creative product studies within human-AI collaborative environments, emphasizing the potential for human-AI symbiosis as a future direction for AI integration in cultural industries. Xie²⁵ examined the appearance, decorative elements, and compositional features of intangible cultural heritage products, employing shape grammar to extract and transform heritage elements into preliminary design concepts. Fiebrink²⁶ presented teaching strategies grounded in educational research and creative computing practices to support students in achieving their goals in creative computing.

These technological advancements not only improve the efficiency of design processes but also facilitate the generation of novel ideas. For instance, AI can quickly generate design options, provide inspiration, and address complex design challenges. Research has shown how AI can analyze market trends and consumer preferences to propose design solutions tailored to market needs. Additionally, AI has demonstrated remarkable potential in artistic creation, enabling artists to produce innovative works that redefine traditional art forms and contribute to their evolution.

Applicability of VAEs

VAEs are powerful generative models capable of learning complex distributions in high-dimensional data. Their ability to handle such complexity has led to extensive applications in design generation, where they have demonstrated significant advantages in several areas. To assess the efficiency of various models in cultural and creative design tasks, a review of relevant studies was conducted. Mak et al.²⁷ showed that VAEs outperformed traditional models such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) in generating diverse design options. Similarly, Notin et al.²⁸ emphasized that the variational inference mechanism of VAEs enabled them to handle complex data generation tasks with superior efficiency. These findings collectively highlight the VAE model’s strengths in design tasks and its potential to advance creative design processes.

The advantages of VAEs in creative design can be summarized as follows:

1. Large-Scale Data Handling: VAEs excel at processing extensive image datasets, extracting latent features, and generating novel creative samples.

2. Design Iteration: Thanks to their reconstruction capabilities, VAEs enable designers to iteratively modify generated samples, facilitating a more dynamic and flexible design process.

3. Task Versatility: Research demonstrates the effectiveness of VAEs in diverse design contexts, such as apparel design and product aesthetics.

4. Interpretability: VAEs offer strong interpretability, allowing designers to understand the relationships between generated samples and original inputs. This feedback supports informed decision-making and enhances the overall design process.

Role of RL

RL has garnered increasing attention in design generation, particularly for its ability to optimize design outcomes. Jang et al.²⁹ proposed an RL-based generative design process, where a reward function was employed to maximize the diversity of topological designs. This approach framed generative design as a sequential task of determining the optimal combination of design parameters relative to a reference design. Goodarzi et al.³⁰ outlined a three-step workflow—comprising photogrammetry, feature extraction, and discriminative feature analysis—aimed at decision support for cultural heritage using deep learning techniques. Additionally, Belhi et al.³¹ introduced a framework for automatic annotation and metadata completion, leveraging novel deep learning methods for classification and annotation. Their work also addressed the restoration of physically damaged artifacts using advanced image reconstruction methods through supervised and unsupervised learning approaches. RL models learn by interacting with their environment, enabling the iterative optimization of design choices to meet specific objectives and constraints. In creative design, RL has been employed to evaluate innovation, practicality, and aesthetic appeal. By defining appropriate reward mechanisms, RL can guide models to generate designs that align more closely with market demands. For instance, during the design evaluation phase, RL models can self-adjust based on user feedback and expert reviews, ensuring that the generated designs meet both creative and practical standards. Recent research demonstrates that combining RL with other generative models significantly enhances design quality and user satisfaction, further advancing the field of creative design.

Despite these advancements, challenges remain in effectively integrating interdisciplinary approaches from computer science, cognitive psychology, and art design to foster innovation in cultural and creative products^32–34. While existing studies explore various AI applications in design, there is limited in-depth research on the actual effects, efficiency, and cultural sensitivity of these technologies within specific cultural and creative product contexts. Furthermore, the influence of cultural and organizational backgrounds on the acceptance and adoption of these technologies is often overlooked. To address these gaps, this study seeks to empirically evaluate the role of AI-assisted design tools in promoting design innovation and market adaptability. It also examines designers’ attitudes toward these tools and their experiences during usage, offering insights into how RL can be further integrated to support creative industries.

Research methodology

AI models and their applications

Overview of AI models

VAEs are deep learning models used for generating new data samples, such as images or text. They consist of two main components: an encoder, which transforms input data into a compact latent representation, and a decoder, which reconstructs the original data from this representation. A defining feature of VAEs is their use of variational inference, a statistical method that helps the model learn the structure of the latent space, enabling it to generate diverse and novel data samples. Variational inference is a statistical technique that approximates complex probability distributions, addressing the computational challenges of calculating these distributions directly. In the VAE framework, variational inference estimates the true posterior distribution of the data, which is typically intractable. By introducing a simple prior distribution, usually a Gaussian distribution, it enables the identification of an optimal approximation to the true posterior. This process facilitates an effective mapping between the latent space and the data space, ensuring the model can generate meaningful and realistic outputs.

The VAE is particularly effective at generating diverse design options due to its use of variational inference. It learns from existing designs and generates new, creative variations. The model not only produces high-quality designs by optimizing the objective function, but also captures intricate patterns and variations in data through its latent representations. Latent representations refer to the compressed, abstract form of input data generated by the encoder, capturing the essential features of the data. In a VAE, these representations exist within a continuous, multidimensional space called the latent space. This space allows the model to generate novel data by making small adjustments to the latent variables, producing outputs that resemble the original data while introducing variations and creativity. The latent space facilitates continuous transformations of the data, enabling significant diversity and innovation in design options. This makes VAEs particularly suitable for cultural and creative tasks that require multiple design alternatives³⁵. Within the VAE framework, the latent space is a lower-dimensional, continuous space where the encoder maps input data. Each point in this space corresponds to a potential data sample. The design of the latent space is intentionally flexible, supporting smooth transitions between points. This flexibility enables the generation of diverse and creative outputs, fostering innovation and adaptability in the data produced by the model.

Based on this analysis, the following hypotheses regarding the application of AI in cultural and creative product design are proposed:

Hypothesis 1: The integration of VAE and RL generate more innovative, culturally relevant, and user-satisfactory cultural and creative product designs. This combined model is expected to better address complex creative design challenges compared to traditional design methods and other AI models.

Hypothesis 2:The VAE + RL model not only overcomes the limitations of traditional design methods but also enhances design efficiency, reduces design time, and improves overall design quality and market adaptability through dynamic optimization strategies in cultural and creative product design.

In cultural and creative product design, the VAE can automatically generate novel design schemes by learning the latent features of existing designs. This ability enables designers to explore new creative spaces and generate product designs that adhere to specific cultural aesthetics. The generative power of VAE is particularly well-suited for exploring innovative elements in cultural product design, especially in designs that combine traditional and modern elements.

On the other hand, RL optimizes decision-making processes by learning the optimal action strategy through interaction with the environment^36–38. By continuously refining design choices through feedback and iteration, RL can enhance the innovation, practicality, and market relevance of the generated designs. RL is a machine learning approach that allows an agent to learn optimal behavioral strategies by taking actions within an environment and receiving feedback, typically in the form of rewards or penalties, based on the outcomes of those actions. The objective of RL is to maximize the cumulative reward over time. This method is particularly effective for solving problems that require making sequential decisions, such as in games, robot navigation, and design optimization. Optimizing decision processes within the RL framework refers to the iterative process of identifying a strategy that enables the agent to make the most effective decisions for a given task. This optimization occurs progressively through repeated trials and learning, making it an ongoing process rather than a one-time event. An action strategy is the approach the agent uses to select actions in response to a specific state. The agent’s goal is to learn an optimal policy, which specifies the best action to take in any given state to maximize long-term cumulative rewards. A policy can be deterministic, where the agent consistently selects the same action for a given state, or stochastic, where the agent chooses actions based on defined probabilities. In the context of cultural and creative product design, RL can be used to optimize the decision-making process. For example, it can automatically adjust product designs to accommodate user preferences from different cultural backgrounds. By simulating user interactions with the design and optimizing the decisions made during these interactions, RL can help designers create products that are more appealing and culturally relevant.

The application of these two models in cultural and creative product design acts as a bridge, merging traditional creative processes with modern AI technologies. It provides designers with unprecedented tools to explore new creative possibilities and optimize design decisions^39–41. This integration enhances not only design efficiency and innovation but also the alignment between products and user cultures. The framework depicted in Fig. 1 illustrates this approach for the design of cultural and creative products.

Fig. 1.

AI-assisted process for cultural and creative product design.

Data collection

Data collection is the first step in the process of cultural and creative product design using AI-assisted technology. At this stage, a substantial amount of data about cultural and creative products is collected, including images, textual descriptions, user reviews, and more. The diversity and richness of this data are essential for subsequent model training and design generation. The experimental dataset utilized in this study is diverse, encompassing a broad range of design works^42–44. It includes collections from modern art museums, showcasing various contemporary art forms and design concepts, from avant-garde pieces to modern design elements, reflecting the diversity and evolving creativity within the contemporary art community. Additionally, the dataset incorporates cultural and artistic collections from the National Museum of China, showcasing a rich heritage of traditional culture, including ancient ceramics, textiles, traditional furniture, and crafts. This collection underscores China’s extensive cultural history and artisanal techniques. To ensure the dataset’s comprehensiveness and representativeness, a variety of cultural and creative products were selected, covering both classical and modern designs. The data encompass the following categories of design works:

Ceramics: This category includes both ancient and modern ceramic works, illustrating craftsmanship across various historical periods and styles.

Textiles: The dataset features traditional brocade and embroidery alongside modern textile designs, showcasing the evolution and innovation in fabric design.

Furniture: A wide range of furniture designs, from classical to modern, is covered, demonstrating traditional craftsmanship alongside contemporary innovations.

Digital Media Artworks: This category includes interactive art, digital animation, and virtual reality designs, highlighting the integration of modern digital technologies into artistic creation.

Examples of images from these categories of museum collections are illustrated in Fig. 2.

Fig. 2.

Museum collections (a. ceramics; b. textiles; c. furniture; d. digital media artworks).

To further enrich the dataset, active designers in the cultural and creative industries were invited to contribute original works. These submissions include design sketches, completed design images, and accompanying descriptions that provide insight into the creative concepts. The design sketches offer a glimpse into the entire creative process, from the initial concept to the final product, while the completed images showcase the final outcomes. The accompanying descriptions and concept explanations provide a comprehensive understanding of the inspiration and design philosophy behind each work, clarifying the designers’ creative intentions and conceptual frameworks. Examples of design cases are illustrated in Fig. 3.

Fig. 3.

Designer works (a. ceramics; b. textiles; c. furniture).

Data preprocessing

In the VAE model, image data from cultural and creative products undergo preprocessing to generate visual features for training. Textual description data are converted into embedding vectors using natural language processing (NLP) techniques. These vectors are then integrated with image features and fed into the RL model. This approach allows the model to combine visual information from images with semantic information from text, thereby producing design solutions that align more closely with practical requirements.

During preprocessing, image data are subjected to feature extraction using CNNs, resulting in image embedding vectors with a dimensionality of 512. Textual data are processed using pretrained language models (such as Word2Vec or BERT) to generate 300-dimensional text embeddings. Despite the differences in the dimensionality of these embeddings, effective integration occurs during the model fusion phase. Textual descriptions are encoded using pretrained language models, which convert text into semantic vector representations, capturing the semantic relationships within the text. For instance, the BERT model generates rich semantic embeddings related to images through its contextual understanding. These text embeddings are then fused with the image features to ensure the model can effectively utilize both modalities. In the model, the image feature vectors and text embedding vectors are integrated in the feature fusion layer using techniques such as simple concatenation or weighted averaging. The resulting integrated vector encapsulates the joint feature representation of both image and text, providing the model with a multimodal input. This joint representation comprehensively reflects both the visual and textual information involved in the design task.

The output vectors from the model are decoded through the decoder module. The image component is decoded into new image designs via a deconvolutional layer, while the text component generates corresponding design-related textual descriptions. This process ensures the model can simultaneously produce visual designs and their textual explanations. After the image and text vectors are concatenated, the output represents a fusion of multimodal features. These outputs are not simple reproductions but rather integrated representations of both data sources. In the design generation process, the output captures the relevance and complementarity of visual and linguistic information. For example, the model may generate design elements that semantically match the input textual descriptions or extract visually appealing details from image features. Consequently, the output not only produces designs that are culturally relevant but also fosters personalized innovation based on the textual descriptions. The collected data are preprocessed through cleaning, standardization, classification, and other steps to ensure data quality and consistency, thereby enhancing the effectiveness and reliability of subsequent model training.

Model training and evaluation methods

The implementation of the AI system integrates VAE and RL techniques. The process involves several critical steps in data processing and design generation:

1. Data Preprocessing: Design sketches and completed images are standardized to ensure consistency and data quality. The preprocessing steps include noise removal, missing value imputation, and data scaling, all of which create a uniform dataset for training.

2. Model Training: The VAE model generates design samples, while the RL model optimizes these samples through feedback mechanisms. A substantial amount of design data is used to enhance the generative capabilities of the VAE and the optimization performance of the RL model, aiming to improve both design quality and user satisfaction.

3. Design Generation and Optimization: The trained models produce a series of design samples that not only replicate successful historical designs but also introduce creative elements to push traditional design boundaries. These generated designs cover a broad range of cultural and creative products, including textiles, ceramics, furniture, and digital media works. The RL model then optimizes these designs to ensure innovation, aesthetic appeal, and practicality, aligning with modern market demands.

The specific process is illustrated in Fig. 4.

Fig. 4.

Design generation workflow.

• 4.Result Evaluation: The AI-generated design samples are evaluated based on structural similarity, user satisfaction, and market performance. The real-world effectiveness of the designs is validated by comparing them with market data and user feedback, leading to necessary model adjustments and improvements.
• 5.Model Improvement: Based on the evaluation outcomes, iterative refinements are made to the model. These adjustments include fine-tuning the VAE and RL model parameters, optimizing data processing techniques, and refining the training process to further enhance the system’s performance and adaptability.

This comprehensive approach to design data processing and AI model application ensures a thorough evaluation of the system and provides valuable insights into how AI technology can foster innovation across various design tasks.

Parameter settings and model optimization

To ensure the quality and consistency of input data, comprehensive preprocessing procedures are implemented. Initially, all design sketches and images are standardized to a uniform size before being input into the model. This step is crucial for facilitating the comparison and analysis of design samples from various sources on the same scale. All images are adjusted to a consistent resolution. To address variability in color information, color space conversion is performed, standardizing all images to the RGB color space. This helps mitigate discrepancies arising from different image formats and color standards.

Subsequently, data noise is managed using techniques such as Gaussian filtering and median filtering to eliminate random noise and texture distortion from the images. Incomplete or blurry design sketches are repaired and supplemented using interpolation methods, ensuring their suitability for model training. During the cleaning process, special attention is given to preserving the complex patterns and details characteristic of cultural and creative designs. Edge detection algorithms are employed to maintain these details, ensuring the quality and richness of the generated designs. Additionally, to enhance the model’s generalization capabilities, data augmentation techniques such as rotations, flips, and scaling are utilized. These techniques allow the model to accommodate variations in design angles and styles, improving the diversity and innovation of the generated results.

In the process of building and using VAE and RL models, appropriate parameter settings and continuous model optimization are crucial. For VAE, the dimensionality of the latent space determines the complexity and abstraction level of the learned space^45–47. The learning rate affects the speed and stability of model training. A learning rate that is too high may lead to instability during training, while a learning rate that is too low may slow down progress. A larger batch size can improve memory utilization and training stability but may increase computational costs. In VAE, balancing the reconstruction loss with the Kullback-Leibler (KL) divergence is essential. Adjusting these weights can affect the model’s emphasis on data reconstruction accuracy versus the regularity of the latent space. The KL divergence is a metric used to quantify the difference between two probability distributions. It measures the information loss incurred when approximating one distribution by another. The KL divergence is expressed in Eq. (1):

1

In Eq. (1), P and Q represent the two probability distributions, and x denotes the sample points. The KL divergence quantifies the relative entropy or information discrepancy between the actual distribution P and the reference distribution Q. For RL, a higher discount factor makes the model prioritize future rewards more heavily^48,49. In the early stages of training, a higher exploration rate helps the model explore more possible strategies, with the exploration rate gradually reduced as training progresses to focus more on utilizing the learned strategies. The design of the reward function is also crucial for RL models, and it must be precisely defined to ensure that the model optimizes in line with the design objectives.

In the context of cultural and creative product design, this study adopts a personalized cross-validation approach to adapt to the diversity of these designs. This method not only considers traditional statistical effectiveness but also incorporates creativity evaluation metrics, such as originality and aesthetic value. In this way, the model must perform well in data fitting while also meeting specific standards in innovation and design uniqueness. Let D be the entire dataset, and k be the number of groups; the innovation-weighted cross-validation error is defined as shown in Eq. (2):

2

Equation (2) describes the innovation-weighted cross-validation error, where is the testing error for the i-th iteration, is the innovation score, and λ is the weight of the innovation factor. Hyperparameter tuning employs a dynamic adjustment strategy, which modifies parameters not only based on performance feedback but also dynamically according to the creave requirements of the design tasks. In the field of creativity, model distillation aims to simplify models and improve efficiency while “distilling” the creative essence from complex models into simpler ones. This method preserves and reinforces innovative features through specially designed loss functions, enabling lightweight models to still generate high-quality creative designs. Assuming is the output of the teacher model and is the output of the student model, the distillation loss for innovation preservation is represented by Eq. (3):

3

.

In Eq. (3), is the innovation preservation loss, and β is the weight of the innovation feature. Early stopping is designed to be context-aware, meaning that the condition for stopping training includes not only a decrease in loss but also the quality of the creative output. Training is halted when the model begins to repeatedly generate similar designs or when the innovation metric starts to decline, ensuring diversity and novelty in design.

Results and discussion

Experimental environment and evaluation process

Appendices 1 and 2 present the experimental environment utilized in this study, detailing the hardware and software configurations. The parameter settings for the VAE and RL models are provided in Appendices 3 and 4.

The experiment evaluates the performance of the proposed VAE and RL combination in cultural and creative product design using several metrics, including model accuracy, generation quality, user satisfaction, and computational efficiency. These results are compared with those of traditional single models, which have not integrated multiple techniques or methods. These single models include fundamental approaches such as GANs, VAEs, and RL.

In evaluating the quality of the generated designs, both quantitative and qualitative measures are applied. Quantitatively, the Structural Similarity Index (SSIM) is employed to assess the similarity between AI-generated designs and reference designs from historical data, providing an objective measure of output accuracy and quality. SSIM evaluates image similarity based on brightness, contrast, and structural features. Additionally, Contrast Fidelity, Texture Fidelity, and Color Fidelity are used to assess visual contrast, texture detail, and color reproduction accuracy, respectively, by comparing the generated designs to real-world counterparts. Texture Fidelity can be analyzed by extracting texture features, such as using Gabor filters, and calculating the similarity between these extracted features. The SSIM formula is expressed in Eq. (4):

4

In Eq. (4), and are the means of images x and y, and are their variances, represents the covariance between the two images, and and are constants to prevent division by zero. Contrast Fidelity, which assesses how well the contrast is preserved in the generated image, is defined as Eq. (5):

5

In Eq. (5), is the standard deviation of the generated image, and is the standard deviation of the reference image. A Contrast Fidelity value close to 1 indicates strong contrast preservation.

Color Fidelity is assessed using the CIEDE2000 color difference formula, as shown in Eq. (6).

6

In Eq. (6), L^∗, a^∗, and b^∗ represent the color components in the CIE Lab color space.

For the qualitative assessment, a comprehensive set of evaluation criteria is used to ensure the design works meet high standards. These criteria include innovation, aesthetic appeal, cultural relevance, cultural adaptability, and design practicality. Innovation refers to the uniqueness and creativity of the design. Experts evaluated the level of innovation using a scoring rubric, employing a scale from 1 to 10, where 1 indicates a lack of innovation and 10 represents high creativity. The scale is defined as follows. For example: 1–3 points indicate a lack of innovation, with designs closely resembling common, repetitive patterns; 4–6 points reflect some innovation but insufficient distinction; 7–8 points demonstrate strong innovation with notable uniqueness; and 9–10 points represent highly innovative designs showcasing ground-breaking ideas.

Aesthetic appeal focuses on the visual impact of the design, including elements such as color combinations, shapes, and compositions. Both users and experts rate the designs based on their aesthetic quality, ensuring objectivity and consistency in the evaluation process. Cultural relevance assesses how well the design aligns with a specific culture. Experts score designs based on their effectiveness in cultural communication, ensuring that the designs not only have aesthetic appeal but also resonate with cultural significance. Cultural adaptability evaluates how effectively the design integrates into a specific cultural context. Experts score based on the cultural elements embodied in the design, ensuring its relevance to the target culture. Finally, design practicality concerns the feasibility of the design in practical applications, assessing functionality, usability, and market acceptance.

To gather expert feedback, professionals from various fields within the cultural and creative industries—such as designers, cultural researchers, and market analysts—are invited to participate. Prior to the evaluation, experts undergo training on the evaluation criteria and scoring system, ensuring they fully understand the requirements and meaning behind each indicator. Experts use standardized review forms to score the designs, ensuring a systematic and consistent evaluation process. Simultaneously, user feedback is collected through structured questionnaires, covering aspects such as visual appeal, cultural communication effectiveness, and design practicality. Users rate each aspect using a Likert scale, supplemented with open-ended questions for deeper insights (e.g., 1 indicating strong disagreement and 5 indicating strong agreement).

During the analysis phase, the collected scoring data are subjected to statistical analysis to compute the mean and standard deviation for each criterion. Analysis of variance (ANOVA) is used to assess significant differences between the various designs. Additionally, thematic analysis is performed on users’ open-ended feedback to identify key patterns and themes, providing valuable insights. This evaluation framework helps to better understand design performance within the cultural and creative industries, offering theoretical foundations and practical recommendations for future design practices.

User interaction data, such as actions like viewing, selecting, and modifying designs, provides insights into user engagement with the designs. For computational efficiency, the focus is on assessing the runtime and resource consumption of the model during the design generation process. Time measurements for each model’s design generation are recorded under identical hardware conditions, allowing for a comparative analysis of computational efficiency across different models. This data is then compared with the performance of traditional single models, providing a thorough evaluation of the advantages of combining VAE and RL models in the creation of cultural and creative products.

Performance evaluation

Model generation results

The results of several design cases generated by the proposed model are presented in Fig. 5.

Fig. 5.

Samples of generated design cases (a. ceramics; b. textiles; c. furniture).

These examples demonstrate that the generated designs successfully retain the original design elements while integrating innovative features to enhance their creativity and imaginative appeal. For instance, in Fig. 5a, the Jingdezhen ceramics incorporate additional creative patterns and flowing elements inspired by the original museum collection, resulting in a more dynamic and visually engaging interpretation of the porcelain. In Fig. 5b, the textiles reinterpret traditional Qing Dynasty garments by introducing distinctive colors and intricate embroidery, blending traditional craftsmanship with modern stylistic elements to align with contemporary aesthetic preferences. Similarly, Fig. 5c reimagines an ancient chair by incorporating modern design elements, creating a visually innovative piece that harmonizes the ingenuity of ancient Chinese design with modern aesthetic standards, rendering it highly appealing to contemporary audiences.

Quantitative evaluation results of generated products

The evaluation results for model accuracy are shown in Fig. 6.

Fig. 6.

Evaluation of cultural element recognition accuracy by different models.

In Fig. 6, the VAE + RL model demonstrates exceptional performance across all metrics, particularly excelling in accuracy and F1 score compared to other models. The accuracy of the VAE + RL model reaches 94.5%, significantly surpassing other single models such as VAE (92.3%), RL (88.7%), and GAN (87.0%). These findings underscore the effectiveness of combining VAE with RL in capturing both global and local features during complex design generation tasks, resulting in enhanced overall performance. In comparison, GPT and Llama-3 models also exhibit commendable performance, with accuracies of 90.4% and 91.2%, respectively. While these models excel in text generation tasks, they face slight limitations in handling cross-modal tasks, particularly those requiring seamless integration of visual and textual data, when compared to the VAE + RL model. Notably, GPT and Llama-3 models achieve recall scores of 91.7% and 90.8%, respectively, reflecting their strong capabilities in recognizing design elements. However, the VAE + RL model outperforms in design diversity and innovation, achieving an F1 score of 93.4%, further reinforcing its superiority in terms of design quality and user satisfaction.

Despite the widespread use of the GAN model in various generation tasks, its performance in this study is relatively suboptimal, achieving an F1 score of only 87.5%, which is lower than that of other models. This result underscores the GAN model’s limitations in managing design complexity and innovation, as it struggles to balance generation quality with design creativity as effectively as the VAE + RL model. These findings highlight that the VAE + RL model excels not only in accuracy and recall but also outperforms other models in overall performance, particularly in generating high-quality, innovative design samples. While GPT and Llama-3 models showcase exceptional abilities in text generation, their effectiveness diminishes when addressing tasks that demand design innovation and the integration of multimodal data. This contrast underscores the substantial advantage of combining VAE and RL techniques, which enhances both the generation and optimization of designs in the cultural and creative product design domain.

Figure 7 provides a visual comparison of the generated quality among different models, illustrating the superior outcomes achieved by the VAE + RL approach.

Fig. 7.

Comparison of generated quality by different models.

Figure 7 provides a comparative analysis of generative quality across different models, emphasizing the superiority of the VAE + RL model across all evaluated metrics. The VAE + RL model achieves a SSIM of 0.92, a Contrast Fidelity of 0.95, a Texture Fidelity of 0.94, and a Digital Color Fidelity of 0.93. These results demonstrate that designs generated by the VAE + RL model not only closely align with real-world structural characteristics but also exhibit exceptional fidelity in contrast, texture, and color. This superior performance is attributed to the synergy between the VAE and RL. The VAE effectively captures diverse latent representations from existing designs, while RL optimizes these designs iteratively through interaction with feedback mechanisms, producing high-quality outputs.

In comparison, GPT and Llama-3 models also deliver strong generative performance, particularly in structural similarity and contrast fidelity. GPT achieves an SSIM of 0.89 and a Contrast Fidelity of 0.91, while Llama-3 records an SSIM of 0.88 and a Contrast Fidelity of 0.90. These models, although primarily developed for natural language generation, display robust generative capabilities that extend to design tasks. However, their performance in Texture Fidelity and Digital Color Fidelity is slightly lower, reflecting their limited specialization in image-based generation compared to the VAE + RL approach.

When used independently, VAE and RL achieve moderate generative quality. The VAE records an SSIM of 0.87, a Contrast Fidelity of 0.90, a Texture Fidelity of 0.88, and a Digital Color Fidelity of 0.85. Similarly, RL achieves an SSIM of 0.85, a Contrast Fidelity of 0.88, a Texture Fidelity of 0.85, and a Digital Color Fidelity of 0.84. While both models produce coherent designs individually, their outputs lack the polished quality achieved through their combined use. In contrast, the GAN model underperforms across all metrics, with an SSIM of 0.83, a Contrast Fidelity of 0.86, a Texture Fidelity of 0.83, and a Digital Color Fidelity of 0.81. While GANs are known for their ability to generate diverse designs, they often face challenges such as mode collapse and training instability, which negatively impact their consistency and overall quality. These results underscore the substantial benefits of integrating VAE and RL, particularly for tasks demanding high-fidelity outputs and innovative design capabilities. The combination proves to be a robust approach in the domain of cultural and creative product design, offering superior performance and adaptability compared to traditional single-model methods.

Qualitative evaluation results of generated products

Figure 8 illustrates the qualitative evaluation scores assigned by experts and users for various design works based on the defined metrics, with scores ranging from 1 (very poor quality) to 10 (excellent quality).

Fig. 8.

Qualitative evaluation results.

The scoring results presented in Fig. 8 show that:

1. Design A, which includes three ceramic products, received high ratings for innovativeness and aesthetic appeal, with particular emphasis on design practicality, resulting in a user satisfaction rate of 95.2%. This design was recognized for its ability to combine traditional ceramic techniques with modern creative elements, ensuring both visual appeal and functionality.

2. Design B, featuring three improved versions of Qing Dynasty garments, excelled in cultural relevance and cultural adaptability, scoring 9. This score reflects the design’s success in integrating traditional cultural elements with contemporary aspects. Although its aesthetic appeal was slightly lower than that of Design A, its overall performance remained strong, with a user satisfaction rate of 90.3%, indicating its effectiveness in blending cultural heritage with modern aesthetics.

3. Design C, which includes three enhanced furniture pieces (chairs), achieved a score of 9 for both innovativeness and cultural relevance, reflecting the design’s successful fusion of traditional cultural elements with modern design innovation. The user satisfaction rate for this design was 90.7%, underscoring the appeal of the furniture in merging design innovation with cultural expression.

The results indicate variations in performance across the different design types concerning innovativeness, cultural adaptability, and user satisfaction. Design A stands out for its strong performance in practicality and aesthetic appeal, while Design B demonstrates its value in merging cultural heritage with modern enhancements. Design C further highlights the potential of integrating cultural relevance with innovative design, offering valuable insights into the field of furniture design.

Model efficiency evaluation

Figure 9 illustrates the resource consumption and runtime of different models.

Fig. 9.

Comparison of resource consumption and runtime.

Figure 9 compares the resource consumption and runtime of various models, offering insights into their computational efficiency and suitability for real-time design generation:

1. The VAE + RL model exhibits exceptional performance with an average training time of 5 h and an average inference time of 0.2 s. This high efficiency makes it ideal for practical applications where rapid generation of high-quality designs is essential, providing a competitive advantage in environments requiring fast turnaround times.

2. The GPT and Llama-3 models demonstrate strong generative capabilities but are relatively less efficient in terms of resource consumption and runtime. GPT requires an average of 8 h for training and 0.4 s for inference, while Llama-3 needs 9 h for training and 0.35 s for inference. These longer training times reflect the higher computational demands of these models, which may limit their effectiveness in real-time design applications where speed and efficiency are critical.

3. The VAE and GAN models show moderate resource consumption, with training times of 6 h and 7 h, respectively, and inference times of 0.25 s and 0.3 s. These models offer a balance between generative quality and resource efficiency, making them suitable for applications requiring both high-quality outputs and optimized resource use.

4. The RL model, while efficient in inference, has the longest training time, requiring 8 h. Its inference time of 0.35 s is comparable to other models, but the extended training period is needed to optimize design decisions effectively.

In conclusion, the VAE + RL model emerges as the most efficient and effective choice for cultural and creative product design, offering both superior generative quality and resource efficiency. In contrast, the GPT and Llama-3 models are more computationally demanding and may be better suited for scenarios with ample computational resources or more stringent quality requirements. This analysis highlights the strengths and limitations of each model, helping guide their optimal application in different contexts.

Statistical analysis results

To further validate these findings, an ANOVA test is performed, and Table 1 provides the adjusted performance statistics for each model.

Table 1.

Statistical analysis of different model performances.

Metric	Traditional method	VAE	RL	VAE + RL	GPT	Llama-3
Mean	72.24	86.84	84.33	90.50	89.00	87.00
Standard deviation	3.62	3.02	2.44	2.00	3.00	4.00
Minimum	67.65	82.35	80.73	87.00	85.00	83.00
25th percentile	68.95	84.75	83.24	89.00	87.00	85.00
Median	72.60	87.06	84.22	90.00	89.00	87.00
75th percentile	73.69	88.14	85.30	92.00	91.00	89.00
Maximum	77.90	91.26	89.40	94.00	93.00	92.00

The data in Table 1 clearly demonstrates the superior performance of the VAE + RL model in cultural and creative product design. The traditional design methods, with a mean score of 72.24, are significantly outperformed by more modern approaches, showing their limited capability in addressing complex and innovative design tasks. These traditional methods fall short in both efficiency and design innovation when compared to the newer, data-driven models. The VAE model shows a significant improvement over traditional methods, with a mean score of 86.84, reflecting its ability to generate innovative designs due to its strong generative capabilities. However, while VAE excels in design creativity, the RL model—scoring 84.33—is particularly strong in optimizing decision-making processes. While the RL model shows promising results, its design innovation falls slightly behind the VAE model, as expected given its focus on optimizing rather than generating designs. The VAE + RL model achieves the highest mean score of 90.50, underscoring the synergy between the generative power of VAE and the decision-making capabilities of RL. This combination not only fosters higher levels of design innovation but also enhances user satisfaction. Additionally, the standard deviation of the VAE + RL model is 2.00, indicating stable performance across various design tasks, ensuring consistent high-quality outcomes. The wide score range (from 87.00 to 94.00) further emphasizes the model’s reliability in delivering superior results across different types of cultural and creative designs.

Although the GPT model (mean score: 89.00) and the Llama-3 model (mean score: 87.00) perform well in innovation, their strengths are primarily in text generation and handling complex design issues. When it comes to decision optimization and overall design processing, both models fall behind the VAE + RL model. GPT and Llama-3, though excelling in text generation, lack the integrated design optimization capabilities that the VAE + RL model offers. In conclusion, the VAE + RL model stands out as the most effective approach for cultural and creative product design, not only enhancing design innovation but also improving the efficiency and stability of the design process. This combination leads in terms of design quality, user satisfaction, and design optimization, offering valuable insights and robust technical support for future advancements in the field. Compared to the individual VAE, RL, GPT, and Llama-3 models, VAE + RL delivers a more comprehensive and effective solution for cultural and creative product design.

Turning test results

The Turing test conducted to assess the intelligence of the VAE + RL model involves determining whether the designs generated by the model can be perceived as indistinguishable from those created by human designers. This test provides direct insight into the model’s generative capabilities and its ability to mimic human creativity.

The VAE + RL model is initially employed to generate a series of design schemes, ensuring a broad range of diversity and creativity. These designs incorporated various styles and themes to simulate real-world cultural and creative design tasks. The generated designs are then mixed with those created by human designers to form a comprehensive evaluation set. To maintain fairness in the evaluation process, all designs are anonymized, removing any identifiable markers that could indicate whether the design was generated by the VAE + RL model or by a human designer. A double-blind assessment is conducted with a panel consisting of 20 design experts and 20 ordinary users. Each evaluator is provided with a set of designs, which included both VAE + RL model-generated and human-created designs. Evaluators are tasked with determining whether each design is created by a human, with their judgments based on factors such as innovation, practicality, and artistic value. After collecting the evaluations, the proportion of model-generated designs that are incorrectly identified as human-created was calculated. The classification results from each evaluator are then aggregated, and a confusion matrix is used, along with accuracy metrics, to assess the intelligence level of the model.

Table 2 presents the feedback based on differ rent design types and evaluator groups. Each test includes 20 design samples from both the VAE + RL model and human designers.

Table 2.

Turing test results.

Turing test results	Design source	Number of times judged as human-created (experts)	Number of times judged as human-created (general users)	Expert accuracy (%)	General user accuracy (%)
Modern art	VAE + RL	12	17	60.0	85.0
	Human designers	18	20	90.0	100.0
Traditional craft	VAE + RL	13	17	65.0	85.0
	Human designers	19	18	95.0	90.0
Digital illustration	VAE + RL	17	19	85.0	95.0
	Human designers	20	20	100.0	100.0
Product design	VAE + RL	15	16	75.0	80.0
	Human designers	18	17	90.0	85.0

Table 2 illustrates the performance variations of the VAE + RL model across different design types. The model’s designs in modern art and digital illustration are often perceived as human-created, demonstrating strong performance in these areas. In contrast, the VAE + RL model shows weaker performance in traditional craft and product design, with notably lower accuracy rates in product design, as evaluated by both experts and general users.

In the fields of modern art and digital illustration, the VAE + RL model performs notably well. For modern art, experts recognize 12 of the model-generated designs as resembling human creations, resulting in an accuracy rate of 60.0%. General users assess 17 designs with a higher accuracy rate of 85.0%. In the digital illustration category, experts evaluate 17 designs with an accuracy rate of 85.0%, while general users assess 19 designs, achieving a 95.0% accuracy rate. Conversely, the model is less effective in traditional craft and product design. For traditional craft, experts identify 13 designs with an accuracy rate of 65.0%, while general users assess 17 designs, yielding an 85.0% accuracy rate. In product design, experts evaluate 15 designs with a 75.0% accuracy rate, while general users assess 16 designs with an 80.0% accuracy rate. Overall, design experts typically achieve higher accuracy rates compared to general users, reflecting their more precise evaluation of the designs. General users perform better in modern art and digital illustration but demonstrate lower accuracy rates in traditional craft and product design. This discrepancy could be attributed to differences in design experience and sensitivity to details.

Discussion

The outstanding performance of the VAE + RL model in design quality and user satisfaction underscores its effectiveness in generating high-quality and attractive product designs. Jang et al.²⁹ highlight that combining generative models with decision optimization methods can significantly improve the quality and diversity of design solutions. This finding aligns with that viewpoint, as both SSIM and user satisfaction indicators were higher than those of other models. Furthermore, the high rating of design innovation by users emphasizes the critical role of innovation in the cultural and creative industries⁵⁰, reflecting Chen⁵¹ who identifies innovation as a key driver of success in cultural products. Despite the VAE + RL model’s impressive performance, its resource consumption and runtime limitations are important considerations. Zhan et al.⁵² note that highly complex model optimization can result in significant increases in computational costs. This suggests that while the model excels in design quality and user satisfaction, its computational demands could hinder its broader adoption, particularly in resource-constrained environments.

The advantages of the VAE + RL model go beyond traditional evaluation metrics, demonstrating adaptability to market demands. As Vuong and Mai⁵³ assert, integrating generative models with decision optimization more effectively meets market needs for innovation and personalization. This integration enables the VAE + RL model to produce design solutions that better align with user expectations, thereby strengthening its competitive position in the market. However, given the model’s high computational demands, further research should focus on optimizing computational efficiency. Techniques such as model pruning and quantization could help reduce resource consumption while maintaining core performance. Choudhary et al.⁵⁴ suggest that model compression techniques are effective in lowering computational costs and enhancing practical application efficiency. Additionally, strategies like distributed computing and parallel processing could reduce model training and inference times, improving computational efficiency. Beyond the cultural and creative design sector, the design optimization capabilities of the VAE + RL model hold potential for various other domains. Applications could extend to fields like architectural design and product development. Future research should explore these cross-domain applications, assess the model’s performance in different design tasks, and develop targeted optimization strategies to enhance its utility.

A comparison with similar studies underscores the advantages of the proposed research methodology and its outcomes. For example, Liu et al.⁵⁵ explored the application of AI in the labor market, emphasizing data-driven decision support based on statistical analysis. In contrast, the current research not only provides data-driven design decision support but also generates diverse design solutions through VAEs and enhances design efficiency and user satisfaction through RL. Furthermore, this approach is specifically tailored to the cultural and creative design domain, with the specificity of the application scenario and the creativity of the generated solutions representing key advantages of the model.

Similarly, Li et al.⁵⁶ focused on AI-supported industrial perception, addressing sensor and data processing challenges in intelligent manufacturing. While their research emphasizes hardware integration and industrial optimization, the present study focuses on optimizing design processes in the cultural and creative sector, substantially improving design quality and user experience. For instance, user satisfaction in this study reached 95%, while existing models in industrial perception often overlook user feedback on design solutions. Additionally, the model addresses gaps in existing research by incorporating cultural adaptability and diversity generation assessments. Furthermore, Zhu⁵⁷ proposed an adaptive agent decision-making model based on deep RL, applied to decision optimization in the logistics sector. While both studies utilize RL frameworks, the current research distinguishes itself by integrating RL with VAEs. This not only optimizes design decisions but also leverages generative models to enhance the diversity of design solutions for cultural and creative products. Additionally, the model emphasizes the generation of optimized solutions based on user feedback, marking a significant departure from the logistics focus on efficiency and path optimization. As such, this approach is particularly suited for the design innovation domain, demonstrating greater adaptability and practical significance.

The comparative analysis above highlights the specificity and innovativeness of the proposed research methodology in the context of cultural and creative design, offering valuable insights for future studies in related fields. This research introduces a generative optimization model that combines VAE and RL, achieving both theoretical and practical advancements in the domain of cultural and creative product design. The experimental results and comparative analysis clearly demonstrate the method’s superiority in design quality, diversity, and user satisfaction. More importantly, this study presents a new paradigm for AI-assisted design within the cultural and creative industries. Unlike traditional design support systems, the proposed approach not only facilitates decision-making but also generates design solutions. The integration of generative models shifts the design process from “selection optimization” to “solution creation,” establishing a foundation for tackling more complex and diversified design challenges in the future. In terms of application prospects, the model developed in this study extends beyond the cultural and creative sector and holds significant cross-domain potential. For instance, its capabilities in generation and optimization can be applied to industrial design, educational content creation, and other areas, thus expanding the possibilities for AI-driven intelligent design. This cross-domain adaptability underscores the model’s versatility and provides substantial opportunities for future research and development.

Conclusion

Research contribution

The primary contribution of this study is the successful integration of VAEs and RL technologies in the field of cultural and creative product design, leading to the development of a novel AI-assisted design model. Experimental validation has shown that this model significantly enhances design quality, creativity, and user satisfaction. Specifically, design solutions generated by this model improve user satisfaction by 18% and innovation evaluation by 26% compared to traditional methods. In quantitative assessments, the model demonstrates exceptional performance, achieving an accuracy of 94.5%, a recall rate of 92.3%, and an F1 score of 93.4%, surpassing the performance of individual models. The experimental results indicate that the model not only replicates historical designs but also surpasses traditional design limitations by incorporating innovative elements. Furthermore, this study introduces a comprehensive range of quantitative metrics, including SSIM, contrast fidelity, texture fidelity, and color fidelity, along with user satisfaction and design innovation scores. These metrics provide robust evaluation tools for future AI design research, significantly enhancing the ability to assess and refine AI-assisted design models.

Future work and research limitations

While this study demonstrates the significant advantages of the VAE + RL model in cultural and creative product design, several limitations and opportunities for future research are apparent. The model’s performance heavily depends on high-quality, diverse datasets, and limitations in data diversity and scale could affect the model’s generalization and applicability to new, unseen design challenges. Additionally, the model’s high computational demands may restrict its application in resource-constrained environments. Future research should focus on developing more efficient model training and optimization techniques, including the exploration of lightweight models and advanced distillation methods, to reduce computational resource requirements and enhance adaptability. Moreover, increasing the model’s interactivity by integrating user feedback directly into the design process could enhance the personalization of design solutions and increase user satisfaction. These advancements will help extend the model’s application to broader design contexts and improve its effectiveness and practicality in real-world design tasks.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

Jing Liang: Conceptualization, methodology, software, validation, formal analysis, investigation, resources, data curation, writing—original draft preparation, writing—review and editing, visualization, supervision, project administration, funding acquisition.

Data availability

The datasets used and/or analysed during the current study available from the corresponding author Jing Liang on reasonable request via e-mail 18995763382@163.com.

Competing interests

The authors declare no competing interests.

Ethics statement

The studies involving human participants were reviewed and approved by Fashion Media Academy, Jiangxi Institute of Fashion Technology Ethics Committee (Approval Number: 2022.605966). The participants provided their written informed consent to participate in this study. All methods were performed in accordance with relevant guidelines and regulations.