A Visualization Method for English Textbook Materials Empowered by Generative Artificial Intelligence and Its Empirical Study

2.1 Overall technical framework

To achieve accurate and efficient visualization of English textbook materials and resolve core pain points such as imprecise semantic alignment, uncontrollable spatial layout, and insufficient text label fidelity, this paper proposes an end-to-end generative AI visualization technical framework. Based on multimodal content extraction and scene graph construction, this framework achieves precise alignment between textbook text semantics and generation targets through prompt engineering. It relies on a dual-stage fusion of code anchoring and diffusion generation to balance accuracy and visual aesthetics, and ultimately completes iterative optimization of generation results through a multi-agent self-reflection closed-loop, forming a complete technical link from text input to high-quality visual output. Three core innovation modules play key synergistic roles in the framework: the scene graph construction module transforms textbook text into structured spatial relationship representations, providing strong semantic constraints for the diffusion generation process to solve the problem of uncontrollable spatial layout; the code anchoring module builds a fidelity mechanism for text labels, ensuring absolute accuracy of textual information in generated images to meet the rigid demands of educational scenarios; the multi-agent self-reflection closed-loop module takes educational adaptability as an explicit optimization objective, achieving deep matching between generation quality and English teaching demands through multi-dimensional verification and iterative correction. Each module is closely connected and collaboratively linked, guaranteeing both the technical precision of image processing and the adaptability of educational scenarios, thereby constructing a dedicated visualization generation system adapted to the characteristics of English textbook materials. Figure 1 shows the overall technical framework diagram of generative English textbook material visualization.

Figure 1. Overall technical framework diagram of generative English textbook material visualization

2.2 Dependency-aware discrete diffusion scene graph construction

As the core bridge connecting English textbook text and visual images, the structural accuracy of the scene graph directly determines the spatial rationality and semantic alignment precision of the subsequently generated images. Figure 2 shows the Construction process of the dependency-aware discrete diffusion scene graph. To accurately capture entity associations and spatial dependencies in textbook text, the input English textbook text sequence is first preprocessed. The input text sequence is defined as $T=\left\{t_1, t_2, \ldots, t_L\right\}$, where $L$ is the length of the text sequence. Through Optical Character Recognition (OCR) and layout analysis technology, entity text blocks and their spatial coordinates are extracted, and then the core components of the scene graph are constructed: the entity node set $V=\left\{v_1, v_2, \ldots, v_N\right\}$ and the directed relation edge set $E=\left\{\left(v_i, r_{i j}, v_j\right)\right\}$, where $N$ is the number of entity nodes, and $r_{i j} \in \mathrm{R}$ represents the predefined spatial relationship type between entities $v_i$ and $v_j$, covering common spatial association forms to ensure that the scene graph can accurately map the spatial semantics in the textbook text.

Figure 2. Construction process of the dependency-aware discrete diffusion scene graph

To improve the robustness and accuracy of scene graph generation, this paper uses a discrete diffusion process to model the scene graph. Through the bidirectional process of forward corruption and reverse denoising, deep decoupling of graph structure and text semantics is achieved. The forward diffusion process contains K diffusion steps. In each step k, the original scene graph G is corrupted with noise at probability $\beta_k$ to generate a noisy graph $G_{\text {noise}}$. During the corruption process, the existence of some entity nodes and the types of relation edges are randomly masked, simulating the uncertainty of scene graphs caused by vague or ambiguous text semantics. The core of the reverse denoising process is learning the denoising network $p_\theta\left(G_{k-1} \mid G_k, \operatorname{enc}(T)\right)$. This network adopts a Graph Transformer architecture to capture the dependencies between entity nodes through a message-passing mechanism. Its message-passing formula is:

$h_i^{(l+1)}=A t t n\left(h_i^{(l)},\left\{h_j^{(l)} \mid\left(v_i, r, v_j\right) \in E\right\}\right)+M L P\left(h_i^{(l)}\right)$          (1)

where, $h_i^{(l)}$ represents the hidden state of entity node $v_i$ at the l-th layer of the Graph Transformer, Attn is the attention mechanism used to focus on the hidden states of adjacent nodes that have direct relation edges with the current node to achieve dependency-aware feature transmission, and Multilayer Perceptron (MLP) is used for nonlinear transformation of node hidden states to enhance feature expression capability. After iterative processing through $L_g$ layers of Graph Transformers, the probability map $\widehat{G}$ is output through the softmax function, where the probability values of each node and relation edge reflect their matching degree with the text semantics.

The sampling strategy during the inference stage directly affects the reliability of the scene graph. Aiming at the possible linguistic ambiguity in English textbook text, this paper introduces the Monte Carlo Dropout sampling method to improve the robustness of spatial relationship constraints of the scene graph. This method extracts M candidate scene graphs from the posterior distribution and selects the optimal scene graph by calculating the entropy value of each candidate graph. The entropy calculation formula is:

$H(\widehat{G})=-\sum_{g \in \widehat{G}} p(g) \log p(g)$           (2)

where, p(g) is the probability value of nodes and relation edges in the candidate scene graph $\widehat{G}$. The smaller the entropy value $H(\widehat{G})$, the more determined the structure of the scene graph and the clearer the semantics, effectively avoiding scene graph deviations caused by language ambiguity. Finally, the scene graph with the smallest entropy value is selected as the spatial constraint condition for subsequent diffusion generation, ensuring that the entity relationships in the generated images are highly consistent with the textbook text descriptions.

The dependency-aware discrete diffusion scene graph construction method realizes precise modeling of graph structures through the discrete diffusion process, captures semantic associations between entities with the help of dependency-aware message passing of the Graph Transformer, and solves the language ambiguity problem combined with Monte Carlo Dropout sampling. The generated scene graph can not only accurately map the semantic information of English textbook text but also possesses strong robustness, providing reliable structured constraints for the subsequent code anchoring and diffusion generation modules, laying the foundation for solving the core pain point of uncontrollable spatial layout.

2.3 Text fidelity mechanism of code anchoring

Absolute accuracy of text labels is the core rigid requirement for the visualization of English textbook materials and also the main bottleneck restricting the application of existing diffusion generation methods in educational scenarios. To thoroughly solve the problems of text spelling errors and label misplacement that easily occur during the diffusion model generation process, this paper designs a code anchoring mechanism. By combining structured code generation with conditional constraints, a fidelity barrier for text labels is constructed to ensure that the English text in the generated images is completely consistent with the textbook text, while also taking into account the visual aesthetic characteristics of the images. This mechanism takes the scene graph and textbook text as input and solidifies the text labels and spatial structure constraints into a visual skeleton through executable code, providing precise text and geometric references for subsequent diffusion generation. Figure 3 shows the network architecture of code anchoring and three-way conditional guided diffusion generation.

Figure 3. Network architecture of code anchoring and three-way conditional guided diffusion generation

The core implementation flow of the code anchoring mechanism revolves around structured code generation and skeleton map construction. First, the previously generated scene graph and the original English textbook text are fed into a Large Language Model (LLM) to generate Python drawing code in a specific format. The generated code contains three core elements: geometric placeholder boxes, absolutely correct text labels, and relational coordinate constraints between entities. Among them, the design of the coordinates of the geometric placeholder boxes strictly follows the spatial relationships of entities in the scene graph to ensure that the positions of the text labels match the distribution of entities; the text labels are directly derived from the textbook text, and semantic verification by the LLM ensures no spelling errors; the relational coordinate constraints further solidify the spatial association between entities and text labels to avoid label misplacement during the subsequent generation process. Executing this Python code generates a low-resolution skeleton map with a resolution set to 256×256. This resolution can not only ensure the clarity and recognizability of text labels but also reduce the computational complexity of subsequent diffusion generation. The skeleton map adopts a fixed color mode, with a white background, light gray entity placeholder boxes, and black text labels. This design enhances the distinction between text, background, and entities, facilitating the subsequent conditional constraint module to capture text features.

To quantitatively evaluate the text fidelity effect, this paper defines a text label set and a Text Fidelity Index (TFI) to build a strict fidelity constraint system. Let the text label set be $L=\left\{l_1, l_2, \ldots, l_M\right\}$, where M is the total number of text labels that need to be retained in the textbook text, covering words, phrases, sentences, and other forms. The code anchoring mechanism ensures that for any label $l \in L$, the OCR result of the corresponding area in the skeleton map is completely consistent with l through syntax checking and execution verification, forming a hard fidelity constraint. The TFI is used to quantify the text fidelity performance of the final generated image, and its calculation formula is:

$T F I=1-\frac{1}{|L|} \sum_{l \in L} E D\left(l, O C R\left(\left.\hat{I}\right|_{\widehat{\Omega}_l}\right)\right)$          (3)

where, ED represents the normalized edit distance, used to measure the difference between the recognition result of the label area in the generated image and the original label. Its calculation formula is $E D(a, b)=\operatorname{edit}(a, b) / \max (\operatorname{len}(a), \operatorname{len}(b))$, where $\operatorname{edit}(a, b)$ is the edit distance between strings $a$ and $b$, and are the lengths of the two strings len(a) and len(b), respectively; $\hat{I}$ is the final generated image; $\widehat{\Omega}_l$ is the predicted area in the generated image corresponding to label $l$; and $O C R\left(\left.\hat{I}\right|_{\widehat{\Omega}_l}\right)$ is the OCR result of that area. The TFI ranges from [0, 1]; the closer the value is to 1, the higher the text fidelity. When TFI = 1, the text labels in the generated image are completely consistent with the original textbook text.

The code anchoring mechanism achieves a balance between text fidelity and visual aesthetics through the deep integration of ControlNet and the diffusion model. The generated skeleton map is used as the conditional input for ControlNet. ControlNet constrains the generation process of the diffusion model through additional conditional branches, forcing the text labels and geometric placeholder box structures in the skeleton map to remain unchanged during the generation process. Specifically, ControlNet downsamples the skeleton map to the same resolution as the latent space noise image of the diffusion model, converts the skeleton map features into dimensions compatible with the intermediate features of the diffusion model through zero convolution blocks, and then adds them element-wise to the output features of the U-Net encoder to achieve deep injection of conditional information. This fusion method not only retains the advantage of the diffusion model in generating high-quality visual images but also ensures the absolute accuracy of text labels through the rigid constraints of the skeleton map, stabilizing the TFI of the final generated images close to 1.0, thoroughly solving the text spelling error problem of existing methods, and meeting the core demands of English teaching scenarios.

2.4 Three-way conditional guided diffusion generation

Diffusion generation is the core link of English textbook material visualization, and its generation accuracy directly determines the semantic consistency, text fidelity, and spatial rationality of the final image. Single-condition guidance struggles to balance multi-dimensional constraint requirements. Therefore, this paper extends the U-Net denoiser of the base diffusion model, introducing three-way conditional collaborative constraints—text embedding, scene graph, and skeleton map—on the noise prediction process. Through the deep fusion of multi-source information, triple precise constraints of semantics, space, and text are realized. This not only retains the visual generation advantages of the diffusion model but also ensures that the generation results adapt to the core demands of English textbook visualization. The three-way conditions perform their respective duties and collaborate synergistically: text embedding provides global semantic and style priors, the scene graph strengthens entity spatial relationship constraints, and the skeleton map guarantees text label fidelity, jointly building a multi-dimensional, strongly constrained diffusion generation system.

Effective encoding of the three-way conditions is the basis for collaborative constraints. Aiming at the characteristics of different conditions, differentiated encoding strategies are designed to ensure that each condition's information can be accurately injected into the diffusion generation process. Text embedding is obtained through a pre-trained text encoder, converting the English textbook text sequence into a fixed-dimensional semantic vector, denoted as $e_{\text {text }}$. After being processed by the encoding function $\phi_{\text {text}}, \phi_{\text {text}}\left(e_{\text {text}}\right)$ is obtained. This embedding vector can capture the overall semantics and style tendency of the text, providing global semantic guidance for the generated image. Scene graph embedding is implemented using a Relational Graph Convolutional Network (R-GCN). By performing feature encoding on the entity nodes and relation edges of the scene graph, it captures the dependencies and spatial associations between entities. Its node feature update formula is:

$h_i^{(\text {graph})}=\sigma\left(\sum_{r \in R} \sum_{j \in N_i^r} \frac{1}{c_{i, r}} W_r h_j+W_0 h_i\right)$          (4)

where, $h_i^{(\text {graph})}$ is the graph embedding feature of entity node $v_i$, $\sigma$ is the Sigmoid activation function used to introduce nonlinear feature expression; $r$ is the predefined spatial relationship type, and $N_i^{V^{\prime}}$ represents the set of adjacent nodes having a relationship of type $r$ with node $v_i ; c_{i, r}$ is the normalization coefficient used to balance the contribution of different relationship types, calculated as $c_{i, r}=\sum_{j \in N i r} 1$, ensuring the stability of feature updates; $W_r$ is the relation-specific weight matrix, and $W_0$ is the node's own weight matrix, used to adjust the contribution degree of adjacent node features and the node's own features, respectively. The graph embedding features of all nodes are globally averaged pooled to obtain a fixed-dimensional scene graph global embedding, which is used for subsequent spatial constraints in diffusion generation. Skeleton map encoding is completed through the ControlNet branch. The low-resolution skeleton map is downsampled to the same resolution as the diffusion model's latent space noise image $\mathrm{Z}_t$. The geometric and text features of the skeleton map are converted into dimensions compatible with the U-Net encoder features through zero convolution blocks, and then added element-wise to the output features of the U-Net encoder to achieve deep injection of skeleton map conditional information, ensuring the stability of text labels and geometric structures.

The three-way conditions are deeply integrated with the U-Net denoiser through a cross-attention mechanism to construct a collaborative constrained noise prediction function, realizing efficient interaction and integration of multi-source information. The extended noise prediction function is as follows:

$\begin{gathered}\epsilon_\theta\left(\mathrm{z}_t, t, e_{\text {text}}, G, S\right)=\epsilon_\theta^{\text {base}}\left(\mathrm{z}_t, t\right)+\lambda_1 \cdot C A\left(\mathrm{z}_t, \phi_{\text {text}}\left(e_{\text {text}}\right)\right)+ \\ \lambda_2 \cdot C A\left(\mathrm{z}_t, \phi_{\text {graph}}(G)\right)+\lambda_3 \cdot C A\left(\mathrm{z}_t, \phi_{\text {sketch}}(S)\right)\end{gathered}$            (5)

where, $\epsilon_\theta$ is the extended noise prediction function, $\mathrm{z}_t$ is the latent space noise image at step $t$ in the diffusion process, and $t$ is the diffusion step; $\epsilon_\theta^{\text {base }}$ is the noise prediction result of the base diffusion model, ensuring the model retains basic visual generation capabilities; $C A$ is the cross-attention operation, used to establish associations between the latent space noise image and each conditional embedding feature, realizing precise matching of multi-source information; $\lambda_1, \lambda_2, \lambda_3$ are the weight coefficients of the three-way conditions, ranging from [0.1, 1.0], determined by cross-validation to obtain optimal values, adjusting the constraint strength of text embedding, scene graph, and skeleton map respectively, ensuring the three-way conditions work synergistically without interfering with each other; $\phi_{\text {sketch }}(S)$ is the encoding feature of the skeleton map, consistent with the output feature of the aforementioned ControlNet branch. This fusion method is not a simple feature superposition, but allows latent space features to interact deeply with each conditional feature through the cross-attention mechanism, enabling the generation process to respond simultaneously to semantic, spatial, and text constraints, improving the overall quality of the generated image.

To further strengthen spatial relationship constraints and ensure that the spatial associations of entities in the generated image are completely consistent with the scene graph, a spatial relationship consistency loss function is introduced. It is weighted summed with the original loss function of the diffusion model to form the final optimization objective. The spatial relationship consistency loss function formula is:

$\begin{array}{r}L_{\text {spatial}}=\sum_{\left(v_i, r_{i j}, v_j\right) \in \mathrm{E}} \text { CrossEntropy } \\ \left(\text { RelClassifier}\left(\text {bbox}_i, \text { bbox}_j\right), r_{i j}\right)\end{array}$           (6)

where, $L_{\text {spatial }}$ is the spatial relationship consistency loss value, $\left(v_i, r_{i j}, v_j\right)$ is a directed relation edge in the scene graph, $b b o x_i$ and $b b o x_j$ are the bounding boxes of entities $v_i$ and $v_j$ in the generated image respectively, extracted from the generated image via object detection algorithms; RelClassifier is a spatial relationship classification network that inputs the coordinate features of entity bounding boxes and outputs the predicted probability distribution of the spatial relationship between entities; Cross Entropy is the cross-entropy loss, used to measure the difference between the predicted relationship and the ground truth relationship $r_{i j}$ in the scene graph, penalizing generation results that violate spatial relationship constraints. This loss function acts on the intermediate feature layer of the U-Net denoiser through backpropagation, guiding the model to learn the spatial relationship rules in the scene graph, reducing entity spatial misplacement phenomena, further improving the spatial rationality of the generated image, and ensuring that the generation results are highly consistent with the spatial semantics described in the English textbook text. The final total loss of the model is:

$L_{\text {total}}=L_{\text {base }}+\alpha \cdot L_{\text {spatial }}$           (7)

where, $L_{\text {base}}$ is the original noise prediction loss of the diffusion model, and $\alpha$ is the loss weight used to adjust the strength of spatial relationship constraints. Experiments verified that setting to 0.3 achieves the optimal balance between spatial constraints and visual quality.

2.5 Multi-agent self-reflection closed-loop optimization

The initial visualized image obtained through three-way conditional guidance can only complete the basic constraints of semantic structure and text information. It is difficult for it to actively fit the cognitive laws of English teaching at different academic stages, nor can it autonomously correct residual detail deviations in the generation process. To this end, this paper constructs a multi-agent collaborative self-reflection iterative architecture, building a complete closed-loop process from generation, dual verification, strategic reflection to parameter revision. It incorporates both image objective quality evaluation and teaching scenario adaptation evaluation into the optimization objective system, so that the final output results not only conform to the general evaluation standards in the field of image processing but also match the usage requirements of actual classroom teaching. Figure 4 gives the multi-agent collaboration and self-reflection iterative optimization flow.

Figure 4. Multi-agent collaboration and self-reflection iterative optimization flow

The Semantic Verification Agent undertakes the consistency verification work between image content and the original text. This module receives three types of input data: the original textbook text sequence, the generated image to be evaluated, and the standardized scene graph. It completes comprehensive quantitative scoring from three dimensions: entity restoration, spatial relationship matching, and text information fidelity. The calculation formula for the comprehensive semantic alignment score is:

Score $_{\text{sem}}=\alpha \cdot$ ObjAcc $+\beta \cdot$ RelAcc $+\gamma \cdot$ SpellAcc           (8)

where, ObjAcc represents the object entity detection accuracy rate inside the image, used to measure the complete restoration level of the object content corresponding to the text; RelAcc is the F1 metric for entity spatial relationship classification, which can accurately reflect the degree of agreement between the entity arrangement structure and the semantics of the text description; SpellAcc directly reuses the TFI defined earlier to unify the text content evaluation system. The weight parameters $\alpha$, $\beta$, and $\gamma$ are all adaptively obtained through training on labeled datasets and can automatically adjust the proportion of each dimension evaluation according to different textbook text types such as dialogues and short essays. The study sets a fixed judgment threshold of 0.85. When the comprehensive semantic alignment score does not reach this value, the system automatically determines that the current image has problems of semantic misalignment or content missing, triggering the subsequent iterative correction process.

The Educational Adaptability Verification Agent focuses on the quantitative regulation of teaching application levels. Taking the generated image, the target audience grade identifier, and the preset cognitive load threshold as basic inputs, it relies on visual feature algorithms to sequentially calculate three core indicators image visual complexity $C_{v i s}$, text content information density $D_{\text {info}}$, and visual attention guidance efficiency $E_{\text {attn}}$. The module compares the measured indicator values with the preset standard thresholds corresponding to the grade level, and outputs a standardized dimensional adjustment scheme. This is used to achieve directional optimization of the visual style and content arrangement of the generated image, allowing the visual content to adapt to the cognitive acceptance ability of learners of different ages, completing the quantitative implementation of educational dimensional constraints, and making up for the deficiency of traditional generative models neglecting educational adaptability.

After completing the dual-agent evaluation, the system constructs an incremental prompt vector $\Delta p$ based on the two types of evaluation results. This vector is fused with the initial text prompt embedding vector, and the scene graph and skeleton map are synchronously adjusted according to the evaluation opinions to obtain updated structured constraints G' and S'. Then, the new constraint conditions are re-integrated into the three-way conditional diffusion generation module to carry out a new round of image generation. The entire iterative process sets unified convergence judgment rules. Only when the three constraint conditions of $Score_{\text {sem }} \geq 0.85, C_{\text {vis}} \leq \tau_{\text {vis}}$, and $D_{\text {info}} \leq \tau_{\text {info}}$ are met simultaneously, the iterative process officially terminates and outputs the optimal visual image. In practical application scenarios, this closed-loop optimization system only requires three to five iterations to reach a stable convergence state, which can not only fully complete detail deviation correction and teaching adaptation tuning but also effectively control the overall generation time consumption, achieving a two-way balance between optimization effect and computational efficiency.

The generative English textbook material visualization method proposed in this paper has formed significant advantages in terms of technological innovation in image processing and adaptation to educational scenarios, possessing important technical value and application prospects. At the technical level, aiming at the core pain points of existing text-to-image generation technologies, this method constructs a dependency-aware discrete diffusion scene graph generation mechanism, realizing precise modeling of entity spatial relationships and compensating for the defect of insufficient spatial constraints in traditional methods; the code anchoring mechanism, through the deep integration of structured code generation and conditional constraints, builds an absolute fidelity barrier for text labels, solving the industry challenge of inaccurate text generation in diffusion models; the three-way conditional guided diffusion generation strategy realizes the collaborative constraints of multi-source information including text embedding, scene graph, and skeleton map, balancing image accuracy and visual aesthetics, enriching the innovative application of text-to-image generation technology in structured constraint scenarios. At the application level, this method can accurately adapt to the visualization demands of English textbook materials. The generated images not only meet the rigid requirements of text fidelity and spatial accuracy but also possess good educational adaptability. They can be directly applied to immersive English teaching to improve teaching efficiency and learning experience. Meanwhile, its core technical framework can be migrated to the visualization of textbook materials in other subjects such as Chinese and Science, possessing broad application potential. At the interdisciplinary level, based on the cross-integration of image processing and education, this paper constructs a complete research paradigm of "technological innovation-empirical verification-application landing," providing referenceable technical ideas and experimental design references for similar interdisciplinary research, and promoting the standardized application of generative AI technology in the field of educational visualization.

Although the proposed method demonstrates significant advantages in multiple experiments, there are still some limitations that cannot be ignored, reflecting the objectivity and rigor of the research. The scene graph parser is currently only fine-tuned on junior high school and high school English textbook corpora, and its generalization ability is limited by the domain corpus. When migrating to textbook materials in other subjects such as Science and History, due to differences in subject text semantic features and entity relationships, the scene graph generation accuracy may decrease, requiring further collection of domain-specific labeled data for optimization. In terms of generation efficiency, multi-round self-reflection iterations and scene graph Monte Carlo sampling inevitably increase end-to-end generation time. Although the dynamic early exit mechanism has achieved a certain degree of optimization, the generation speed is still difficult to meet real-time visualization demands and has limitations in large-scale textbook material batch generation scenarios. Educational adaptability optimization is still at a basic level, currently only focusing on general indicators such as cognitive load and attention guidance, without targeted design combined with specific teaching objectives such as English vocabulary memorization and grammar understanding, making it difficult to fully exert the supporting role of visualization technology in specific teaching links. In addition, the experimental dataset only covers English textbook materials from junior high school and high school sections, with limited sample size and coverage, making it difficult to fully verify the generalization performance of the method in different scenarios, which may affect the universality of the research conclusions.

Aiming at the above limitations and combining the research orientation in the field of image processing with the actual demands of educational scenarios, this paper proposes the following specific and feasible future research directions. To improve the cross-domain generalization ability of the scene graph parser, domain adaptive learning methods will be introduced. Through transfer learning and feature alignment of cross-disciplinary corpora, reliance on specific domain labeled data will be reduced to achieve accurate scene graph generation for textbook materials in different disciplines. To further improve generation efficiency, the deep integration of lightweight diffusion models and the self-reflection closed-loop will be explored, designing a more efficient iteration termination judgment mechanism, and combining model quantization and pruning technologies to reduce computational overhead, meeting real-time generation and batch generation demands. To deepen educational adaptability optimization, combining specific objectives of English teaching, a more refined educational adaptability indicator system will be designed to achieve deep matching between visualized images and specific teaching links such as vocabulary teaching and grammar teaching, enhancing the teaching value of the technology. Meanwhile, the coverage of the dataset will be expanded to include various material types such as primary school English, English picture books, and listening materials, further verifying the generalization of the method, and expanding application scenarios to promote the deep landing and innovative development of generative AI technology in the field of educational visualization.