Design of personalized creation model for cultural and creative products based on evolutionary adaptive network

Text generation image study

In the field of text-to-image generation, GANs have been widely adopted to facilitate multimodal synthesis by generating visual content conditioned on textual descriptions. However, conventional GANs often suffer from instability during training, which can result in mode collapse and suboptimal image quality. To address these challenges, researchers have proposed hierarchical generation architectures. Notably, the Stacked GAN introduced a two-stage generation process to sequentially refine image resolution and detail, with each stage operating independently (Tominaga & Seo, 2022). Building on this concept, Zhang et al. (2019) presented StackGAN++, which restructured the generation pipeline into a tree-like, end-to-end framework comprising multiple generators and discriminators for enhanced feature refinement. Expanding upon stacked architectures, attention-based GANs emerged to improve the semantic alignment between text and image. Among these, Attentional GAN by Xu et al. (2020) and Rich Feature GAN by Cheng et al. (2020) introduced attention mechanisms to selectively emphasize word-level semantics during image synthesis. Further advancements include Deep Fusion GANs proposed by Tao et al. (2022), Semantic GANs and Semantic-Spatial Aware GANs introduced by Liao et al. (2022), and Recurrent Affine Transformation GANs developed by Ye et al. (2024), all of which employed conditional affine transformations to achieve deeper integration of textual and visual modalities. Experimental evaluations demonstrate that even single-level architectures utilizing such transformations can produce photorealistic outputs that rival or surpass those generated by complex multi-stage models. However, the omission of word-level semantic granularity in these models often leads to a deficiency in fine-detailed image features.

To overcome this limitation, Tan et al. (2023) introduced a fine-grained semantic evaluation metric known as Semantic Similarity Distance, which inspired further development in contrastive pretraining methods. Based on this principle, CLIP model was proposed (Radford et al., 2021), laying the groundwork for the Parallel Deep Fusion GAN (PDF-GAN). PDF-GAN leverages multi-level semantic integration, employing a discriminator capable of simultaneously evaluating global coherence and local detail, thereby enhancing the fidelity and semantic richness of the generated images.

Image style transfer study

CariGANs, proposed by Cao, Liao & Yuan (2020) leverage neural networks in conjunction with principal component analysis to learn a mapping from facial keypoints in real photographs to those in caricature counterparts, thereby facilitating deformation based on structural correspondence. However, the exaggeration effect in this approach is rigidly anchored to the input image, constraining the generative diversity of resulting caricatures. Nonetheless, the tight coupling between stylization and deformation modules hinders its adaptability across varied application contexts (Hou et al., 2021).

Turja et al. (2022) further explored caricature deformation by predicting a deformation field from paired image data. Although effective in capturing artist-guided exaggeration, this method is heavily reliant on labor-intensive, high-quality supervision and treats caricature generation as a deterministic one-to-one mapping task, thus failing to produce diverse outputs. In contrast, Hou et al. (2021) incorporated latent encoding strategies to enrich the diversity of exaggeration patterns. While this approach offers potential for varied deformations, the absence of robust supervision introduces challenges in controllability and frequently leads to unstable or inconsistent distortions. To address these limitations, styleCariGAN (Jang et al., 2021) enhances shape exaggeration by embedding dedicated exaggeration blocks into pre-trained StyleGAN layers (Bermano et al., 2022). Despite its innovative architecture, the model predominantly emphasizes texture manipulation and struggles with achieving meaningful structural distortion, such as ironic or satirical deformation. Complementarily, Men et al. (2022) devised a segment-based, database-driven matching technique wherein user-assisted segmentation is employed to isolate hairstyles and facial regions. These segmented elements are then matched with stylized samples from a curated database, which are recombined to synthesize the final caricatured output (Men et al., 2022).

The aforementioned research underscores the growing maturity of cross-modal analysis and style transfer methodologies grounded in deep learning, particularly within the domain of artistic creation. Consequently, in the context of personalized cultural and creative product development, the application of deep learning and cross-modal generative technologies offers substantial potential to shorten production cycles and fulfill the demands of intelligent, user-centric customization. Building upon traditional text-to-image diffusion models, this study incorporates a style transfer module to accommodate more diverse generative requirements and employs a tailored planning algorithm to expand the range of design aesthetics. These advancements hold significant promise for the future of personalized product creation, marking a pivotal step toward efficient, adaptive, and culturally resonant design innovation.

Text vector and diffusion-based image generation

The image generation module serves as the central component of the proposed system, tasked with translating text-based semantic vectors and style preference vectors into coherent visual representations. In this work, the module employs a generation framework built upon a diffusion-based architecture—specifically, Stable Diffusion—which reconstructs images through an iterative denoising process, thereby enhancing both visual clarity and semantic alignment, while preserving considerable generative flexibility. Given the integral role of user input in the personalized design of cultural and creative products, Stable Diffusion is adopted as the backbone of the image generation module. This model, grounded in the latent diffusion model, achieves high-fidelity image synthesis with superior computational efficiency (Du et al., 2023). Unlike earlier diffusion approaches that operate directly in pixel space, Stable Diffusion performs transformations within a lower-dimensional latent space, substantially reducing the computational burden associated with training and inference. The overarching architecture is illustrated in Fig. 1.

Stable Diffusion uses a text encoder in a pre-trained CLIP model to extract the semantic representation of the input text. The input text $T$ is tokenized and fed into the Transformer, which outputs the embedding vector $\mathbf{c}\in {\mathbb{R}}^d$. This semantic vector is fed into the attention module of UNet as conditional information that guides the semantic direction of image generation. Its main direction is represented by Eq. (1):

(1) $$\mathbf{c}={\mathrm{C}\mathrm{L}\mathrm{I}\mathrm{P}}_{\mathrm{t}\mathrm{e}\mathrm{x}\mathrm{t}}(T).$$

While the other way of encoding results in the image $x\in {\mathbb{R}}^{H\times W\times 3}$. It is encoded by VAE as potential representation $z\in$ ${\mathbb{R}}^{h\times w\times c}$, where $h\ll H,w\gg W$, to reduce the arithmetic burden. The representation result of its main encoder z and decoder x is given in Eqs. (2) and (3):

(2) $$z=\mathrm{E}\mathrm{n}{\mathrm{c}}_{\varphi }\left(x\right)$$

(3) $$\hat{x}=\mathrm{D}\mathrm{e}{\mathrm{c}}_{\theta }(z).$$

After the feature extraction of the two-way network is completed, the forward process adds the latent variable $z$ to the Gaussian noise to construct the training samples:

(4) $$z_t=\sqrt{{\alpha }_t}{z}_0+\sqrt{1-{\alpha }_t}\cdot ϵ,ϵ\sim \mathcal{N}(0,I).$$

Let t denote the discrete time step, and α_t the noise attenuation factor determined by the scheduler. We denote by z_t the noisy latent representation at step t, by ε the Gaussian noise sample to be removed, and by ccc the textual condition embedding provided by the encoder. The conditional UNet network is trained to predict the noise component ε from (z_t, t, c), thereby learning to iteratively denoise the latent representation under text guidance:

(5) $${\widehat{ϵ}}_{\theta }\left(z_t,t,\mathbf{c}\right)=\mathrm{U}\mathrm{N}\mathrm{e}{\mathrm{t}}_{\theta }\left(z_t,t,\mathbf{c}\right).$$

The cross-modal attention mechanism is introduced in UNet for fusing textual semantics, and the time step $t$ is added to the network as an additional condition through time coding. So far the final loss function is obtained in the form of MSE whose diffusion loss is shown in Eq. (6):

(6) $${\mathcal{L}}_{\mathrm{d}\mathrm{i}\mathrm{f}\mathrm{f}\mathrm{u}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}}={\mathbb{E}}_{z_0,t,ϵ}\left[ϵ-{\widehat{ϵ}}_{\theta }{(z_t,t,\mathbf{c})}^2\right].$$

Finally, after completing the corresponding training, for a given present description $T$, the condition vector $\mathbf{c}$ is generated, a latent variable $z_T\sim \mathcal{N}\left(0,I\right)$ is sampled from a Gaussian distribution, and the denoising is iterated step by step:

(7) $$z_{t-1}=\mathrm{D}\mathrm{e}\mathrm{n}\mathrm{o}\mathrm{i}\mathrm{s}\mathrm{e}\mathrm{S}\mathrm{t}\mathrm{e}\mathrm{p}\left(z_t,{\widehat{ϵ}}_{\theta }\left(z_t,t,\mathbf{c}\right)\right).$$

The final iteration is completed to get the reconstructed image as:

(8) $$\hat{x}=\mathrm{D}\mathrm{e}{\mathrm{c}}_{\theta }(z_0).$$

In this study, the image generation module functions as a foundational platform upon which adaptive control mechanisms and evolutionary strategies are integrated, enabling the synthesis of images that are diverse, semantically coherent, and stylistically distinct. As a central architectural element, it plays a critical role in facilitating personalized image generation, effectively aligning output aesthetics with user-defined preferences and contextual demands.

AdaIN-based image style transfer

The style adaptive module is designed to enable controllable and personalized modulation of image styles by incorporating external style representations—such as exemplar images or user-defined preference vectors—to align the visual output with specific cultural aesthetics or individual taste profiles (Gu & Ye, 2021). In this work, we employ AdaIN as the core mechanism for style regulation, owing to its efficiency and differentiability in facilitating image style transfer. AdaIN achieves style adaptation by normalizing the feature statistics of the content image—specifically, by adjusting its channel-wise mean and variance—and replacing them with those of the target style image. This operation alters the image’s textural and stylistic characteristics while preserving its underlying structural semantics. Formally, for an input image, AdaIN can be expressed as:

(9) $$\mathrm{A}\mathrm{d}\mathrm{a}\mathrm{I}\mathrm{N}(f_c,f_s)=\sigma (f_s)\cdot \left(\frac{f_c-\mu (f_c)}{\sigma (f_c)}\right)+\mu (f_s)$$where $f_c\in {\mathbb{R}}^{C\times H\times W}$: represents the feature map of the content image on a certain layer; $f_s\in {\mathbb{R}}^{C\times H\times W}$: represents the feature map of the style image; and for $\mu (\cdot ),\sigma (\cdot )$ represents the mean and standard deviation calculation function on the channel dimension. The generation of the style-transferred image is ultimately completed through a decoding process. In this study, the style adaptive module serves as a key mechanism for aligning visual output with users’ personalized aesthetic preferences. By enabling explicit style control, it ensures that the generated images not only maintain semantic fidelity to the input text but also embody stylistic characteristics that reflect cultural traditions or individual tastes. This capability constitutes one of the core components in achieving truly personalized cultural and creative expression.

The establishment for the evolutionary adaptive generative aesthetic network (EAGAN)

To enhance the adaptability of the generated multi-style images to a broader spectrum of product design requirements, this study introduces a genetic algorithm (GA)-based optimization strategy. This approach targets three critical elements within the model pipeline: the textual prompts in the diffusion model training phase, the style content vectors during style transfer, and the latent noise inputs in the final decoding stage. Serving as a pivotal mechanism for adaptive refinement, this module leverages population-based evolution—encompassing fitness evaluation, crossover, and mutation operations—to iteratively improve the quality, diversity, and personalization of the generated outputs.

Construct the initial population ${\mathcal{P}}_0=\left\{P_1,P_2,\dots ,P_N\right\}$, each individual is a set of generative configurations, where Pi is shown in Eq. (10):

(10) $$P_i=\left\{{\mathbf{t}}_i^{\mathrm{\prime }},{\mathbf{s}}_i,z_i\right\}.$$

Three of the parameters represent the text embedding vector, style vector and noise variable, and the individual assessment fitness function defined based on the above three elements is shown in Eq. (11):

(11) $$F\left(P_i\right)=\alpha \cdot \mathrm{C}\mathrm{L}\mathrm{I}\mathrm{P}\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}\left(x_i,T_i\right)+\beta \cdot \mathrm{S}\mathrm{t}\mathrm{y}\mathrm{l}\mathrm{e}\mathrm{S}\mathrm{i}\mathrm{m}\left(x_i,s_u\right)+\gamma \cdot \mathrm{P}\mathrm{r}\mathrm{e}\mathrm{f}\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}\left(x_i\right)$$where CLIPScore: semantic consistency; StyleSim: style match (with user preference vectors); PrefScore: from the rating prediction model. StyleSim is implemented as the cosine similarity between CLIP image embeddings of the generated output and the style reference. This embedding-based metric captures both visual appearance and higher-level stylistic cues, offering a more semantically grounded measure than handcrafted features or Gram-matrix correlations. Then it is selected, poorer as well as mutated accordingly to generate a new population for the next generation as follows:

(12) $${\mathcal{P}}_{t+1}=\mathrm{G}\mathrm{e}\mathrm{n}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{e}\left({\mathcal{P}}_t^{\mathrm{s}\mathrm{e}\mathrm{l}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{e}\mathrm{d}},\mathrm{C}\mathrm{r}\mathrm{o}\mathrm{s}\mathrm{s}\mathrm{o}\mathrm{v}\mathrm{e}\mathrm{r},\mathrm{M}\mathrm{u}\mathrm{t}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\right).$$

Keep iterating the algorithm until the corresponding requirement is finally satisfied. In this article, the overall structure flow of EAGAN, which is finally built by integrating the above text image generation, style transfer and adaptive iterative optimization, is shown in Fig. 2.

The proposed network, driven by user-provided textual input, integrates a diffusion-based generation mechanism, style transfer technology, and a genetic optimization strategy to construct a multi-module collaborative framework for the personalized creation of cultural and creative products. The process begins with the user articulating their creative intent via natural language, which is encoded into semantic vectors by the text encoding module. These vectors are then fed into a Stable Diffusion-based architecture, initiating image synthesis within a compact latent space. Simultaneously, user-supplied style references—whether in the form of exemplar images or abstract preference vectors—are incorporated through the Style transfer Module, which employs the AdaIN mechanism to modulate feature representations and infuse stylistic attributes. Following initial image generation, the system conducts an evaluation of the output using a multidimensional fitness function that accounts for semantic alignment, style coherence, and subjective aesthetic congruence. Based on this assessment, a genetic algorithm iteratively refines the textual prompts, style vectors, and latent noise parameters, enabling adaptive optimization across successive generations. Through multiple evolutionary cycles, the system progressively converges on outputs that exhibit high semantic fidelity, distinctive stylistic traits, and strong personalization. To achieve multimodal alignment, both textual and visual features are projected into a shared embedding space. Specifically, we adopt the CLIP encoder (ViT-L/14 variant), which jointly encodes text and images into semantically aligned vector representations through contrastive pretraining. Text prompts and user preference vectors are processed by the CLIP text encoder, while images and generated outputs are embedded using the CLIP vision encoder. Semantic consistency is enforced by measuring cosine similarity within this shared space, which also serves as the basis for optimization in the Evolutionary Module. This design ensures that the generated imagery aligns not only with the textual description but also with the encoded stylistic and preference cues, providing a robust multimodal bridge for both training and evaluation. Abstract concepts such as tension, fantasy, or urgency are not modeled through explicit rules but are implicitly captured by the CLIP text encoder, which aligns them with recurring visual patterns in large-scale training data. These embeddings guide the generator toward corresponding stylistic or compositional features, though the mapping remains indirect and may introduce ambiguity in certain cases.

This comprehensive framework embodies a synergistic and iterative co-evolution across four key dimensions—text, image, style, and optimization—offering a robust, scalable, and user-adaptive solution for creative ideation and customized content generation.

The proposed model, EAGAN, is built upon the Stable Diffusion architecture and designed to enable personalized image generation guided by semantic and stylistic preferences. The input data consisted of paired text prompts and cultural image content, sourced from benchmark datasets relevant to cultural and creative product design. During data preprocessing, all images were resized to a consistent resolution, and the corresponding text prompts were tokenized and encoded using a pre-trained semantic text encoder. Image normalization was applied to ensure consistency across visual inputs. Style references were also collected and standardized to feed into the AdaIN-based modulation module for effective style conditioning. The evolutionary component required encoding user preferences into cue phrases and latent vectors, which were initialized randomly and then iteratively evolved through a genetic algorithm.

Dataset and experiment setup

Given that this study encompasses both text-to-image generation and style transfer, we employ subsets from the WikiArt, BAM, and LAION datasets to support comprehensive evaluation. The WikiArt dataset comprises over 80,000 paintings produced by artists worldwide, encompassing more than 20 canonical artistic styles such as Impressionism, Cubism, and Realism, and includes rich metadata on styles and creators. For this research, we selected 100 representative images from each of the ten most prominent styles, constructing a 1,000-image subset to assess the model’s style control capabilities (Eichler, Eichler & Del Pino, 2023). In contrast, the BAM dataset (Froehlich & Koeppl, 2024) emphasizes modern design and illustration, featuring approximately 65,000 art images with clearly annotated style labels. Its categories—such as digital painting, graphic design, and hand-drawn line art—align more closely with the practical aesthetic requirements of the cultural and creative industries. Similarly, we extracted 100 samples from each of ten styles within BAM and integrated them with the WikiArt subset to establish a consolidated benchmark dataset for evaluating style adaptability and control.

In order to evaluate the model’s performance on text–image semantic consistency, we further utilized the LAION-5B dataset (Schuhmann et al., 2022). From this large-scale collection, we extracted a dedicated evaluation subset consisting of 1,000 high-quality image–text pairs. The curation process combined keyword-based filtering with manual inspection: candidate samples were first selected through thematic keyword queries, and then carefully reviewed to ensure that the associated textual descriptions not only captured the central thematic content but also included stylistic indicators. To maintain consistency and reduce noise across samples, several preprocessing steps were applied. All images were standardized to a fixed resolution of 512 × 512 pixels, and samples exhibiting low visual quality, vague or incomplete textual descriptions, or clear semantic mismatches between text and image were systematically excluded. This multi-stage filtering procedure ensured that only coherent and representative pairs were retained.

The final evaluation set was structured into two complementary subsets:

Stylistic expressiveness subset (sourced from WikiArt and BAM), focusing on artistic attributes, visual style, and expressive qualities.

Semantic alignment subset (sourced from LAION), emphasizing accurate correspondence and fidelity between textual descriptions and visual content.

Together, these curated subsets enable a robust, multi-dimensional evaluation protocol. By balancing stylistic control with semantic fidelity, the dataset supports comprehensive validation and comparative analysis of the proposed model’s performance. Importantly, compared with the raw large-scale datasets, the curated subsets underwent stricter quality control and noise reduction, thereby enhancing the reliability and stability of the evaluation outcomes.

To ensure reproducibility and minimize evaluation bias, the dataset was divided into training and validation subsets following an 80/20 split, stratified across both the stylistic (WikiArt/BAM) and semantic (LAION) partitions to preserve representativeness. The validation set was kept entirely disjoint from training data to avoid information leakage. For user preference modeling, stylistic indicators were encoded using categorical embeddings, while semantic alignment cues were represented by normalized dense vectors derived from CLIP text embeddings; these preference vectors served as control signals during model conditioning. All experiments were conducted with fixed random seeds to guarantee reproducibility, and repeated runs yielded stable results with minimal variance. To mitigate potential evaluation bias, sample selection was balanced across categories, manual curation was cross-validated by multiple annotators, and reported metrics represent averaged performance across subsets rather than relying on isolated cases.

To comprehensively evaluate the performance of the proposed model in cultural and creative image generation, this study constructs a multidimensional evaluation framework encompassing four key dimensions: image quality, semantic consistency, style control, and personalized adaptability.

For image quality, two widely adopted metrics are employed: FID and Inception Score (IS). FID quantifies the distributional distance between generated and real images in the feature space, with lower values indicating higher realism. IS simultaneously considers image diversity and clarity, offering a complementary perspective on generative fidelity.

In terms of semantic consistency, the study uses CLIPScore, which computes the cosine similarity between the CLIP-encoded embeddings of the input text and the generated image. This metric reflects the model’s ability to accurately capture and visually represent user intent as expressed through natural language input.

To assess style control ability, LPIPS distance and Gram Matrix similarity are employed. LPIPS evaluates perceptual differences in visual style, while Gram Matrix similarity captures correlations of feature activations related to texture, brushstroke, and color. Since style transfer tasks preserve semantic structure and only modulate visual appearance, IS is excluded from this component of the evaluation.

For personalized adaptability, the model’s performance is gauged by its responsiveness to varying user preferences through iterative optimization, reflected in improved scores across the above dimensions under evolving input constraints.

To benchmark the effectiveness of the proposed framework, a suite of representative models is selected for comparison:

• Stable Diffusion (Du et al., 2023): a high-fidelity, text-driven baseline with strong semantic preservation capabilities.

• VQGAN+CLIP (Crowson et al., 2022): a hybrid model known for its abstract, expressive generation through joint optimization of image and text embeddings.

• StyleGAN-NADA (Gal et al., 2022): a semantically guided style transfer model, well-suited for direct comparison with the proposed style transfer module.

These models encompass the prevailing paradigms in generative AI and provide diverse, challenging baselines for comparative analysis. The experimental environment used for training and evaluation—including hardware specifications, software versions, and configuration details—is summarized in Table 1.

The genetic algorithm was configured with a population size of 30, a mutation rate of 0.2, and a crossover rate of 0.5, employing a tournament selection strategy. The number of iterations was fixed at 20, as preliminary convergence analysis showed that performance improvements largely stabilized beyond this point. To assess the robustness of our results, all experiments were repeated with three random seeds, and we report the average performance along with the corresponding standard deviation. This procedure ensures that the improvements observed in FID, CLIPScore, LPIPS, and Style Similarity are not due to seed-specific effects. The results show stable trends across runs, with small variances, indicating that the gains of EAGAN over baselines are statistically reliable.

Method comparison and result analysis

Following the completion of model construction, we conducted comprehensive testing using the two constructed datasets. The performance of the proposed model was evaluated across the primary metrics outlined earlier. The corresponding results—covering image quality, semantic consistency, and style control—are visually presented in Figs. 3 and 4, respectively, demonstrating the effectiveness of the model under varying input conditions and across multiple evaluation dimensions.

Figure 3 presents the comparative results of image generation performance across various methods on the LAION-5B dataset. Evaluated along three key metrics—FID, IS, and CLIPScore—the proposed EAGAN model demonstrates superior performance across all dimensions. Specifically, EAGAN achieves the lowest FID score (21.3), indicating that its generated images exhibit the highest resemblance to real image distributions. It also records the highest CLIPScore (0.78), evidencing strong semantic alignment between the generated visuals and their corresponding textual descriptions. In terms of generative diversity and visual fidelity, EAGAN attains an IS of 8.5, outperforming baseline models and validating the effectiveness of its integrated evolutionary optimization and style regulation mechanisms.

Figure 4 illustrates the comparative performance of various methods in style transfer tasks on the BAM and WikiArt datasets. The proposed EAGAN model demonstrates consistently superior results across all evaluated metrics. Notably, EAGAN achieves the lowest FID score (20.5) and lowest LPIPS distance (0.15), indicating that its generated images maintain high visual quality and preserve intricate details while undergoing style transformation. In addition, the model attains the highest Style Similarity score (0.81), substantially outperforming StyleGAN-NADA (0.73) and other baseline approaches. These results highlight EAGAN’s enhanced capacity for style expression and alignment, effectively validating the impact of its evolutionary optimization strategy and adaptive style regulation module.

Ablation experiment and application analysis

After benchmarking EAGAN against existing models, we further evaluated its evolutionary optimization performance, specifically the impact of the GA module, through a series of ablation experiments. The results of these analyses are presented in Figs. 5 and 6, which highlight the performance differences with and without GA integration. These comparisons reveal the critical role of the evolutionary component in enhancing semantic fidelity, style precision, and overall image quality, thereby confirming its effectiveness in driving adaptive and personalized generation.

Figure 5 demonstrates the impact of the evolutionary optimization module on model performance, with AGAN representing the ablation version of the model without GA integration. The comparison clearly shows that the full EAGAN model surpasses AGAN across all three key metrics—FID, IS, and CLIPScore. In particular, EAGAN achieves a notably lower FID, indicating a substantial reduction in the distributional gap between generated and real images. Simultaneously, the higher CLIPScore reflects enhanced semantic alignment between textual prompts and generated visuals. These improvements validate the effectiveness of the GA module in adaptively exploring and refining the parameter space, thus reinforcing its crucial role in optimizing image quality and semantic fidelity.

Figure 6 presents the results of the ablation experiments focused on the style transfer task using the BAM and WikiArt datasets. When compared to AGAN—the variant lacking the evolutionary optimization module—the complete EAGAN model demonstrates superior performance across all evaluated metrics. Specifically, EAGAN records lower FID and LPIPS values, indicating the generation of higher-quality images with smoother and more coherent style transitions. Furthermore, its Style Similarity score approaches 0.85, significantly outperforming AGAN and underscoring the effectiveness of the genetic algorithm in enhancing style alignment. These findings affirm the pivotal role of the evolutionary optimization module in strengthening both the consistency of stylistic expression and the overall visual fidelity of the generated results.

Figure 7 illustrates the impact of varying the number of GA iteration rounds on model performance across both datasets. The results reveal that as the number of GA iterations increases, EAGAN consistently improves in terms of FID and LPIPS, reflecting enhanced image quality and better style fidelity through evolutionary refinement. On the LAION-5B dataset, both CLIPScore and IS reach their peak between 15 and 25 iterations, after which a slight decline suggests the emergence of a search equilibrium. In contrast, on the BAM + WikiArt datasets, Style Similarity continues to rise steadily with more iterations until it stabilizes, indicating a progressive enhancement in stylistic alignment.

Overall, these findings highlight the significant influence of GA iteration count on generative performance. Based on the observed trends, the model achieves its optimal multidimensional results at 20 iterations, which is selected as the default configuration for this study.

Following the quantitative analysis, subjective human perception—an essential factor in evaluating personalized generative models—was assessed through a user study. Volunteers were asked to rate, on a 10-point scale, the visual quality and stylistic relevance of text-based images generated under varying GA iteration counts. As shown in Fig. 8, user scores exhibit a clear upward trajectory as the number of GA iterations increases, followed by a plateau. At the 5th iteration, the outputs were relatively unrefined, yielding an average score of 7.7, suggesting that early-stage evolutionary optimization was insufficient. However, between the 15th and 20th iterations, marked improvements in semantic accuracy and stylistic expression were observed, with the 20th iteration achieving the peak average score of 8.8. Beyond this point, user ratings fluctuate slightly but remain consistently high, indicating convergence in perceived quality. These results confirm that the GA module significantly enhances the aesthetic appeal, style alignment, and personalization of generated images. Moreover, the convergence trend supports the conclusion that a moderate iteration count effectively balances visual quality with computational efficiency for practical deployment scenarios.