Cross-Modal Image Enhancement and Visual-Semantic Fusion Empowered by Generative Artificial Intelligence for English Language Teaching

In response to the inadequacies of existing cross-modal generation and image enhancement techniques in accommodating the specialized semantic constraints and cognitive learning requirements of English language teaching scenarios, an end-to-end closed-loop cross-modal image enhancement and visual-semantic fusion framework is constructed. This framework is composed of four deeply coupled and progressively cascaded core modules: grammar-decoupled diffusion generation, multi-granularity visual-semantic fusion, pedagogical saliency-based adversarial augmentation, and semantic contradiction detection and repainting. Standardized English language teaching textual inputs are employed as the primary foundation, with optional reference textbook images incorporated for collaborative computation. Through a structured visual processing pipeline, discrete linguistic grammatical rules are transformed into continuously controllable visual generation features, fundamentally resolving the persistent issues of tense expression ambiguity and inaccurate spatial semantic modeling inherent to general-purpose generation models. The framework sequentially accomplishes grammatically precise initial instructional image generation, multi-scale feature fusion with textbook semantic consistency, cognitively oriented adaptive visual enhancement, and automated correction and restoration of fine-grained semantic biases. Each functional module is interconnected in a sequential manner and jointly optimized. Concurrently, through a training strategy combining stage-wise pre-training and global joint fine-tuning, efficient propagation of semantic features and gradient information throughout the complete processing pipeline is ensured. Ultimately, high-quality images characterized by grammatical accuracy, visual adaptability, and pedagogical practicality are generated, providing a systematic and generalizable technical solution for semantically driven cross-modal visual generation and enhancement tasks in educational contexts.

2.1 Grammar-decoupled conditional diffusion generation module

General-purpose latent diffusion models, which learn modal association features from large-scale image-text data, are capable of achieving high-quality visual image synthesis. However, these models are limited to learning statistical association patterns between textual and visual content, without the capacity to accurately model the tense logic and spatial topological relationships among objects that are inherent to English grammar. Consequently, semantic deficiencies such as action state distortion and spatial relation misalignment frequently emerge in the generated instructional images. To address this issue, a grammar-decoupled conditional generation mechanism is constructed based on a pre-trained diffusion model, through which linguistic rules are transformed into differentiable visual modulation signals. By means of a dual-branch architecture comprising tense feature dynamic modulation and syntactic structure attention constraints, grammatical prior knowledge is explicitly embedded during the diffusion denoising process, thereby overcoming the limitations of statistical semantic matching and achieving accurate instructional image generation under grammatical rule constraints. The network architecture of the grammar-decoupled conditional diffusion generation module is illustrated in Figure 1.

1.png

Figure 1. Network architecture of the grammar-decoupled conditional diffusion generation module

A lightweight tense encoder is designed to accomplish the visual mapping of linguistic tense features. The encoder is primarily composed of two stacked Transformer encoding units, with the one-hot encoding vector of verb tense taken as input. Through feature encoding, the visual representation patterns corresponding to different tenses are extracted, and two independent linear output heads are configured to achieve feature-decoupled modeling. From the first linear head, a motion blur kernel parameter vector is output, based on which a dynamically constructed two-dimensional motion convolution kernel of configurable size is built. Simultaneously, non-negativity and normalization constraints are imposed on the convolution kernel parameters to ensure the physical plausibility of the motion blur features and to accurately characterize the motion trajectory features of dynamic actions. From the second linear head, in conjunction with a Sigmoid activation function, an object integrity state vector along the channel dimension is output. The dimensionality of this vector is maintained consistently with that of the intermediate feature channels of the diffusion U-Net, serving to quantify the visual completeness features of objects under different tenses. During the progressive denoising process of the diffusion model, tense feature injection is performed after the shallow downsampling residual module of the U-Net encoder. First, the object completeness information of the feature maps is modulated through a channel-wise weighting operation, defined as:

$h^{\prime}=h \odot \omega$       (1)

where, h denotes the original network feature map, $\omega$ denotes the object integrity state vector, and h' denotes the tense-modulated feature map. On this basis, grouped depthwise separable convolution is applied to the feature maps using the dynamically generated motion blur kernel, imposing a unified motion blur constraint on all channel features. Tense information is thereby encoded into the low-level texture and edge features of the image, achieving differentiated visual representations characterized by dynamic blur for the progressive aspect and static completeness for the perfective aspect.

To accurately control the object spatial relationships corresponding to prepositions in the text, a structured syntactic graph is constructed from the input teaching text. Syntactic feature encoding is accomplished via a two-layer graph convolutional network, through which the dependency constraint relationships among words are extracted, yielding high-dimensional syntactic feature vectors for each word node. In conjunction with image token positional encoding and textual syntactic features, a cross-modal spatial relationship prior matrix is constructed to quantify the association strength between image visual regions and textual semantic units. The matrix is computed as follows:

$M_{s y n}(p, q)=M L P\left(\left[e_p^{\text {pos}}; h_q^{\text {syn}}\right]\right)$        (2)

where, $e_p^{\text {pos}}$ denotes the positional encoding of the p-th image visual token, h^syn_q denotes the syntactic feature of the q-th text token, and a multilayer perceptron is employed to fuse cross-modal features and output an association score, which is then normalized via a tanh activation function to yield a normalized bias weight. This prior matrix is incorporated into the cross-attention computation of the diffusion model as learnable weights, through which the cross-modal feature matching logic is corrected. The optimized formulation can be expressed as:

$\operatorname{Attention}(Q, K, V)=\operatorname{softmax}\left(\frac{\rho K^{\top}}{\sqrt{d}}+\lambda M_{\text {syn}}\right) V$        (3)

where, Q, K, and V denote the query, key, and value vectors of the attention mechanism, respectively; d denotes the feature dimensionality; and $\lambda$ denotes a learnable bias intensity coefficient. Through this mechanism, cross-regional associations constrained by syntactic rules are adaptively reinforced during the image layout generation stage, and object spatial positional relationships are regulated at the feature matching level, thereby effectively circumventing the spatial semantic confusion issues common to general-purpose models.

To simultaneously ensure the visual quality and grammatical semantic accuracy of generated images, a multi-constrained joint optimization loss function is constructed in this module. The standard denoising score matching loss of the diffusion model is adopted as the foundational optimization objective to guarantee the visual realism and textural fidelity of image generation. Concurrently, a differentiable scene graph parser is introduced to perform structured semantic parsing of generated images, from which image object relation triples are extracted. These triples are then matched against the prepositional spatial relation ground-truth triples corresponding to the teaching text, and a syntactic spatial relation regularization loss is formulated to quantify the cross-modal semantic matching deviation. The overall training loss function of the module is defined as:

$L_{\text {gen}}=L_{\text {denoise}}+\beta L_{s v n}$        (4)

where, $L_{\text {denoise}}$ denotes the denoising score matching loss, $L_{\text {syn}}$ denotes the syntactic spatial relation regularization loss, and $\beta$ denotes a loss balancing weight that coordinates the optimization priorities between visual generation quality and grammatical semantic accuracy. Through iterative optimization of the joint loss, the model is enabled to adhere rigorously to English grammatical rules during scene generation while preserving visual perceptual quality, ultimately producing a semantically accurate initial instructional generated image $L_{\text {gen}}$.

2.2 Multi-granularity visual-semantic fusion module

The initial instructional images generated through grammar-decoupled diffusion generation possess precise grammatical semantic features. However, significant discrepancies exist in visual style and scene element layout relative to standardized textbook images, and direct application to teaching scenarios is prone to issues such as visual system fragmentation and inadequate material adaptability. Conventional image fusion methods predominantly perform pixel-level superposition and style unification based on fixed weights, without incorporating task-specific semantic features to dynamically regulate the fusion process, thereby readily disrupting the established object spatial topological relationships and grammatical semantic information within the images. To address this issue, a multi-granularity cross-modal visual-semantic fusion architecture is constructed in this module. Through hierarchical feature modeling and a text-driven dynamic gating mechanism, adaptive fusion between generated images and reference textbook images is achieved, whereby the core pedagogical semantics are fully preserved while the visual presentation is unified, enabling co-optimization of visual adaptability and semantic fidelity. A schematic diagram of the multi-granularity visual-semantic fusion and alignment mechanism is illustrated in Figure 2.

2.png

Figure 2. Schematic diagram of the multi-granularity visual-semantic fusion and alignment mechanism

Multi-scale feature extraction of the dual input images is accomplished in this module based on a contrastive language-image pre-training visual encoder, through which visual representation information at different levels is fully exploited. By extracting output features from multiple layers of the encoder, a hierarchical feature set covering local texture details to global scene semantics is constructed, corresponding respectively to the multi-granularity feature representations of the generated image and the reference textbook image. Features at all levels are uniformly fed into a parameter-shared cross-modal Transformer fusion unit, ensuring consistency and stability in the fusion logic across different scales. To achieve text-semantic-guided dynamic fusion, a feature-wise linear modulation gating modulation structure is introduced, in which the global semantic embedding of the teaching text is adopted as the modulation reference to generate a spatial-dimensional hybrid mask and channel-wise affine transformation parameters, respectively. First, adaptive weighted fusion of the dual image features is accomplished through the spatial mask, with the specific computation defined as:

$\widetilde{F}_{\text {fused}}^l=\alpha^l \odot F_{\text {gen}}^l+\left(1-\alpha^l\right) \odot F_{\text {ref}}^l$       (5)

where, $\alpha^l$ denotes the spatial hybrid mask corresponding to the l-th layer feature, with values in the range [0, 1], used to adaptively allocate the fusion weights between the generated features and the reference textbook features; $F_{g e n}^l$ and $F_{r e f}^l$ denote the l-th layer features of the generated image and the textbook image, respectively. On this basis, semantic rectification and feature distribution optimization of the fused features are accomplished through adaptive instance normalization, with the computation process defined as:

$F_{\text {fused}}^l=\gamma^l \odot \frac{\widetilde{F}_{\text {fused}}^l-\mu}{\sigma}+\beta^l$       (6)

where, $\mu$ and $\sigma$ denote the channel-wise mean and standard deviation of the fused features, respectively, and $\gamma^l$ and $\beta^l$ denote the channel-wise scaling and bias parameters, used to adjust the semantic tendency and feature distribution of the fused features. The multi-scale optimized fused features are upsampled and reconstructed through a lightweight feature pyramid decoder, yielding a preliminary fused image.

To fundamentally prevent the fusion process from disrupting the key spatial semantic relationships essential to English language teaching, a differentiable scene graph alignment loss is introduced in this module to enforce high-level semantic constraints. Through a unified scene graph parser, the structured semantic graph of the teaching text and the visual scene graph of the fused image are respectively extracted. The sets of spatial prepositional relation edges in both graphs are filtered, and bidirectional optimal node matching is accomplished via the Hungarian algorithm, establishing precise correspondence between textual semantic relations and visual entity relations. Weighted constraints are imposed on the core spatial topological relationships in English language teaching to reinforce the model's capacity for preserving key semantics. The loss function is defined as follows:

$L_{s g}=\sum_{\left(r_t, r_f\right) \in \text { Match}} w_{r_t} \cdot C E\left(r_f, r_t\right)$       (7)

where, $r_t$ and $r_f$ denote the textual semantic relation and the image visual relation, respectively; $w_{r_t}$ denotes the semantic relation weight coefficient, with higher weight constraints assigned to prototypical spatial prepositional relations; and $C E$ denotes the cross-entropy loss function, used to quantify the classification deviation of matched semantic pairs. This loss is capable of approximating the scene graph edit distance in a differentiable form, precisely supervising semantic integrity throughout the fusion process.

A multi-constrained joint optimization strategy is adopted for the overall module to perform iterative parameter updates. Pixel-level L1 reconstruction loss and high-level perceptual loss are combined to ensure detail fidelity and visual naturalness of the fused image. An adversarial loss is employed to optimize overall style consistency, complemented by the scene graph alignment loss to constrain semantic fidelity. These multi-dimensional loss functions cooperate synergistically with complementary constraints, enabling the model to rigorously preserve the tense features and spatial semantic relationships established during the grammar-decoupled generation stage while adapting to the visual style of textbook images. Ultimately, a fused image $I_{\text {fused}}$ characterized by both visual adaptability and pedagogical semantic accuracy is produced.

2.3 Pedagogical saliency-driven adversarial augmentation module

Following multi-granularity semantic fusion, the processed images fully conform to English grammatical rules and textbook visual styles, with overall semantic structure and scene layout exhibiting pedagogical usability. However, the visual energy distribution in conventional fusion outputs tends toward uniformity, failing to align with the cognitive attention patterns of language learning. Conventional image enhancement methods predominantly employ globally uniform contrast and brightness adjustment strategies, with optimization objectives confined to general-purpose visual quality improvement, without the capacity to perform differentiated visual modulation according to the importance levels of teaching knowledge points. This results in insufficient visual saliency for key learning objects such as preposition-associated entities and core action regions, while secondary scene elements are prone to creating visual interference. To address these issues, learner cognitive visual patterns are integrated with an adversarial augmentation mechanism in this module, and a text-conditioned saliency-guided adaptive enhancement network is constructed. Pixel-level non-uniform visual optimization is driven by pedagogical tasks, achieving deep adaptation between image visual presentation and the attention mechanisms of language learning. The network architecture of the pedagogical saliency-driven adversarial augmentation module is illustrated in Figure 3.

3.png

Figure 3. Network architecture of the pedagogical saliency-driven adversarial augmentation module

A text-conditioned pedagogical saliency prediction model is constructed in this study to achieve domain-specific visual attention distribution modeling for teaching scenarios. The model is built upon a cross-modally adapted SalGAN architecture, in which the conventional approach of relying solely on image pixel features for saliency detection is abandoned, and textual semantic priors are introduced to guide saliency region prediction. Global semantic features of the teaching text are extracted via a contrastive language-image pre-training text encoder, and the textual semantic information is progressively embedded into the multi-scale visual features of the saliency encoder through feature adaptive modulation layers, establishing a mapping relationship between textual grammatical semantics and visual attention regions. The model is pre-trained on a dedicated dataset comprising real learner eye-tracking data and expert pedagogical annotations, and is capable of outputting a standardized target attention probability distribution based on the input text and fused image. The computation can be expressed as:

$P_{\text {target}}=S\left(I_{\text {fused}}, T\right)$        (8)

where, $S$ denotes the pedagogical saliency prediction network, $I_{\text {fused}}$ denotes the input fused image, $T$ denotes the teaching text, and $P_{\text {target}}$ denotes the pixel-wise saliency probability map. This map is capable of precisely characterizing the effective attention regions in English learning scenarios, assigning high probability weights to core pedagogical targets such as preposition-interacting objects and action subjects, thereby providing precise cognitive supervisory signals for subsequent targeted enhancement.

A saliency-conditional adversarial augmentation network is constructed to achieve co-optimization of visual quality enhancement and cognitive attention alignment. The enhancement generator adopts a lightweight U-Net architecture integrated with spatially-adaptive normalization residual blocks. Channel-wise concatenation of the fused image and the target saliency probability map is performed as the joint input for feature decoding and image reconstruction. Through the spatially-adaptive normalization mechanism, differentiated enhancement requirements across saliency regions are accommodated, and a cognitively optimized enhanced image is produced. The discriminator employs a PatchGAN architecture, using paired features of images and saliency maps as discriminative criteria to distinguish, at the local patch level, between natural real image-text matching distributions and the visually enhanced distributions produced by the model, thereby constraining the local visual realism of the enhancement results. The overall adversarial optimization objective function is defined as:

$\min _{G_{\text {enh}}} \max _D \begin{gathered}E\left[\log D\left(I_{\text {fused}}, P_{\text {target}}\right)\right]+E\left[\log \left(1-D\left(I_{\text {enh}}, P_{\text {target}}\right)\right)\right]+\eta L_{\text {sal}}\end{gathered}$        (9)

where, $G_{e n h}$ and $D$ denote the enhancement generator and the discriminator, respectively; $E$ denotes the mathematical expectation operator; and $\eta$ denotes a weight balancing coefficient used to adjust the optimization priorities between the adversarial loss and the saliency regularization loss. The first two terms in the formulation constitute the classical adversarial training loss, responsible for constraining the visual naturalness and stylistic consistency of the enhanced images.

To further strengthen the optimization constraint at the cognitive level, a saliency distribution alignment loss is introduced in this study to quantify the deviation between the enhanced image and the standard pedagogical attention distribution. The specific formulation is expressed as:

$L_{\text {sal}}=\left\|S\left(I_{\text {enh}}, T\right)-P_{\text {target}}\right\|_2^2$       (10)

where, $S\left(I_{e n h}, T\right)$ denotes the real-time saliency distribution corresponding to the enhanced image. The overall deviation between the predicted distribution and the target distribution is minimized via the squared L2 norm. This loss is capable of precisely guiding the enhancement strategy at the pixel level, enabling the model to adaptively improve the clarity, contrast, and detail expressiveness of pedagogical core regions, while simultaneously attenuating the visual weight of redundant background information. Under the dual regulation of adversarial and cognitive constraints, the network is enabled to transcend the limitations of globally homogeneous enhancement, producing visual content that conforms to human cognitive habits in language learning, and ultimately yielding a high-pedagogical-validity enhanced image $I_{e n h}$.

2.4 Semantic contradiction detection and localized repainting correction module

The aforementioned generation, fusion, and augmentation modules effectively ensure the macro-level grammatical structure, visual style adaptability, and cognitive attention rationality of instructional images. However, semantic deviations at the pixel level and in fine-grained entity attributes remain difficult to avoid. Stochastic perturbations in the latent space during image generation are prone to causing inconsistencies between detailed attributes—such as color, quantity, and material—and the textual instructional descriptions. Such subtle semantic contradictions cannot be corrected through macro-level semantic constraints or global visual optimization. Moreover, conventional cross-modal editing methods rely on manual localization of defective regions, lacking an automated detection and correction pipeline. To address this issue, a fine-grained image-text alignment-driven semantic error correction mechanism is constructed in this module. Through fully automatic contradiction region detection, dynamic mask construction, and semantically constrained diffusion-based repainting strategies, precise correction of detailed semantic deviations in instructional images is achieved, thereby completing the closed-loop optimization capability of the overall visual processing pipeline. A flowchart of the semantic contradiction detection and fine-grained diffusion-based repainting process is illustrated in Figure 4.

4.png

Figure 4. Flowchart of semantic contradiction detection and fine-grained diffusion-based repainting

Fine-grained matching between local image regions and textual semantic phrases is accomplished in this module based on the Q-Former architecture, enabling pixel-level localization of semantic contradiction regions. Multiple sets of learnable visual query tokens are output by the image encoder, with each set independently corresponding to a local semantic region of the image, forming a global visual-semantic representation sequence. Simultaneously, syntactic dependency parsing is performed on the input teaching text, from which basic semantic units such as noun phrases, prepositional phrases, and attribute modifier phrases are extracted, and independent semantic embeddings for each phrase are generated through the text encoding branch. To quantify the degree of local image-text matching, one-to-one alignment between visual tokens and text phrase features is performed, cosine similarity is computed, and the maximum value is taken as the single-region alignment score, thereby characterizing the degree of correspondence between local visual content and textual semantics. A semantic contradiction heat mask is constructed based on the pixel-level alignment scores, with the computation defined as:

$M_{\text {incon}}(x, y)=\operatorname{ReLU}\left(\tau-S_{\text {align}}(x, y)\right)$        (11)

where, $S_{\text {align}}(x, y)$ denotes the full-resolution image-text alignment score map, and $\tau$ denotes a preset semantic matching threshold used to distinguish between well-matched regions and semantically mismatched regions. The rectified linear unit activation function is capable of filtering out positive matching deviations, retaining only pixel responses where semantic contradictions exist. To eliminate single-pixel noise interference and ensure the spatial continuity of defective regions, a conditional random field is introduced in this study to perform smooth optimization of the initial heat mask, precisely localizing defective regions such as entity attribute errors and fine-grained semantic mismatches, thereby providing precise mask constraints for subsequent localized repainting.

Based on the generated semantic contradiction mask, a latent-space diffusion-based inpainting model is adopted in this study to accomplish local image correction, whereby semantic repainting of defective regions is achieved while preserving the valid semantic content and visual structure of the original image. The pixel-space contradiction mask is downsampled to the latent-space resolution of the diffusion model, yielding a latent-adaptive mask. During each iteration of the reverse diffusion sampling process, partitioned updates of the latent variables are performed, with the fusion formulation expressed as:

$z_{t-1}=M_{\text {incon}}^z \odot \tilde{z}_{t-1}+\left(1-M_{\text {incon}}^z\right) \odot z_{t-1}^{\text {known}}$        (12)

where, $M_{\text {incon}}^Z$ denotes the latent-space semantic contradiction mask, $\tilde{z}_{t-1}$ denotes the corrected latent variable predicted by the model based on text semantics, and $z_{t-1}^{\text {known}}$ denotes the effective latent variables of the original image after noise addition. This computation enables dynamic repainting of defective regions while preserving intact regions unchanged, avoiding the disruption of already optimized grammatical structures and visual features through global reconstruction. To further constrain the semantic accuracy of the repainted content, a semantic gradient guidance mechanism is introduced in the post-sampling stage, through which latent features are optimized via gradient ascent to maximize the similarity between the visual features of the masked region and the target text phrase. The optimization process can be expressed as:

$z_t \leftarrow z_t+\delta \cdot \nabla_{z_t} \log \operatorname{sim}\left(v_{\text {masked}}\left(z_t\right), t_{\text {phrase}}\right)$        (13)

where, $\delta$ denotes the gradient update step size, $v_{\text {masked}}$ denotes the local visual token features decoded from the masked region, and $t_{\text {phrase}}$ denotes the semantic embedding of the corresponding textual modifier phrase. Concurrently, a semantic correction loss is introduced to perform end-to-end fine-tuning of the denoising module of the diffusion model, continuously enhancing the model's capacity for fine-grained semantic restoration.

The four functional modules proposed in this work form a complete processing pipeline characterized by progressive cascading and deep coupling. The overall data flow follows an optimization logic driven by the teaching text, sequentially accomplishing grammatically precise generation, textbook semantic fusion, cognitive attention enhancement, and fine-grained semantic error correction. Model training is conducted using a strategy of stage-wise pre-training combined with global joint fine-tuning. The module-specific loss functions jointly constrain the network parameter updates, ensuring stable propagation and lossless iteration of linguistic priors, spatial semantic relationships, and cognitive visual features throughout the pipeline. The entire architecture achieves full-process modeling from discrete English grammatical rules to continuously controllable visual generation, ultimately producing a high-quality image $I_{\text {out}}$ characterized by fully aligned fine-grained semantics and full suitability for teaching scenarios, thereby providing a comprehensive technical paradigm for the precise generation and intelligent optimization of educational cross-modal visual content.

The syntactic graph bias attention mechanism introduced in this work exhibits excellent interpretability, effectively ameliorating the uncontrollability and implicit black-box characteristics inherent in the attention matching processes of conventional text-to-image models. This mechanism constructs a structured prior matrix based on textual syntactic dependency structures, transforming abstract prepositional spatial topological relations into learnable attention bias constraints, and precisely modulating the feature association strength between image visual regions and textual semantic units. Compared with cross-modal matching approaches that rely solely on statistical data-driven learning, explicit syntactic constraints are capable of guiding the model to establish stable object spatial correspondences during the scene layout stage, fundamentally reducing generation defects arising from semantic confusion across various spatial prepositions. Simultaneously, the pedagogical saliency augmentation mechanism exhibits differentiated optimization effects across different preposition knowledge points. While the improvement in visual discriminability for basic spatial relations is steady and significant, the cognitively oriented non-uniform enhancement strategy, in complex prepositional scenarios involving occlusion and nesting logic, is capable of prominently highlighting the core regions of entity interactions while suppressing interference from irrelevant background information. The auxiliary gain for challenging grammar learning is particularly pronounced, thereby validating the necessity of task-specific image enhancement strategies.

The proposed framework demonstrates stable performance advantages in routine English language teaching image-text processing tasks; however, certain performance limitations persist in complex semantic scenarios. When confronted with overly lengthy teaching texts involving multiple entities and nested grammatical structures, the corresponding syntactic graph topologies become more intricate, rendering the graph convolutional encoding process susceptible to feature coupling and relational redundancy, which in turn leads to a degradation in the model's parsing accuracy for multi-level spatial semantics. Constrained by the sample distribution of the dataset, the model's generalization capability for modeling low-frequency, rare prepositional relations remains relatively weak, rendering the precise visual expression of various niche spatial semantics difficult to achieve. Furthermore, the performance of the semantic contradiction detection module is significantly influenced by the threshold parameter. When the threshold value is set too high, numerous fine-grained attribute defects are filtered out, resulting in missed detections of semantic errors. Conversely, when the threshold is set too low, redundant contradiction masks are generated, triggering unnecessary localized repainting perturbations that disrupt the already optimized visual and semantic structures of the image. This parameter sensitivity imposes certain constraints on the general deployment of the model.

The core innovations and academic contributions of this work are concentrated in the fields of cross-modal image processing and visual generation, with English Language Teaching scenarios serving merely as a highly constrained and standardized validation vehicle and application context. All functional modules throughout the paper are centered around core computer vision technologies, with conditional diffusion generation, multi-scale cross-modal fusion, cognitive saliency modeling, and semantically constrained image inpainting serving as the technical kernel. Linguistic rules are employed solely as computable priors for achieving precisely controllable regulation of visual tasks. This work does not focus on research into language teaching theories or pedagogical strategies. Rather, through the construction of a mapping pathway from linguistic semantics to visual features, the industry-wide problems of insufficient semantic controllability and poor task adaptability in general-purpose image generation and enhancement methods are addressed. A novel technical paradigm and feasible optimization directions are thereby provided for general computer vision tasks such as semantically driven controllable image generation, task-oriented visual enhancement, and fine-grained image-text alignment correction.