User-Generated Images and Artificial Intelligence Large Model-Driven Generation of Tourism Destination Promotional Content

2.1 Overall framework architecture

A closed-loop generation framework jointly driven by preference and space is constructed, in which high-precision controllable generation and autonomous optimization of tourism promotional imagery are achieved through three functionally coupled submodules. The overall architecture consists of a joint conditional encoding layer, a geographically semantically guided diffusion decoding layer, and a multi-dimensional preference-aligned reward layer, with all modules forming a complete bidirectional information flow pathway. Heterogeneous user aesthetic data and destination geospatial data are received by the joint conditional encoding layer, where adaptive fusion of heterogeneous information is performed through cross-modal feature learning, and dynamic conditional vectors that integrate both individualized aesthetic characteristics and objective geographic semantic constraints are output. The geographically semantically guided diffusion decoding layer takes the fused conditional vector as the core regulatory basis, and refined controllable denoising operations are executed within the latent space of the diffusion model, thereby achieving high-quality generation of tourism scene imagery. Generation quality is quantitatively evaluated by the multi-dimensional preference-aligned reward layer across multiple dimensions—image realism, geographic fidelity, and user preference matching—and differentiable reward supervision signals are generated to provide guidance for iterative optimization of network parameters. The overall architecture of the closed-loop generation framework jointly driven by preference and space is illustrated in Figure 1.

Two operational modes—training and inference—are distinguished within this framework, thereby enabling a balanced trade-off between optimization performance and inference efficiency. During the training phase, an end-to-end closed-loop learning mechanism encompassing encoding, generation, evaluation, and updating is formed through the three-layer modules, where quality evaluation signals are directly back-propagated to update the parameters of the joint conditional encoding layer. In this manner, generation evaluation is internally embedded within the iterative model optimization process, in contrast to the pipelined post-hoc filtering paradigm characteristic of conventional methods. During the inference phase, only the converged encoding layer and diffusion decoding layer are invoked, while the reward evaluation branch is discarded, and image generation is accomplished through single-step forward computation, thereby reducing inference overhead while maintaining generation quality. Through this tightly coupled architecture, the inherent limitations of static mapping in controllable generation are overcome, and autonomous evolution of the model's encoding strategy is realized via a dynamic feedback mechanism, effectively enhancing the controllability and adaptive capacity of content generation in complex cultural tourism scenarios.

1.png

Figure 1. Overall architecture of the closed-loop generation framework jointly driven by preference and space

2.2 Cross-modal symmetric cross-gating joint conditional encoding

The cross-modal symmetric cross-gating joint conditional encoding module proposed in this section is designed to accomplish unified representation and adaptive fusion of two types of heterogeneous information—user aesthetic preferences and destination geographic semantics—thereby providing dynamic and precise two-dimensional conditional constraints for the diffusion generation process. The architecture of the cross-modal symmetric cross-gating joint conditional encoding network is illustrated in Figure 2. A dual-branch parallel structure is adopted in the module to enable independent extraction of features from the two sources, through which standardized style representation and spatial semantic representation systems are respectively constructed. In the user preference feature branch, both low-level visual attributes and high-level aesthetic semantics are simultaneously mined from historical user-generated images and social text data, encompassing quantifiable visual features such as color distributions, texture roughness, and proportions of salient regions, as well as abstract stylistic features. The two types of features are dimensionally aligned through dedicated projection layers and uniformly mapped into a d-dimensional latent space, where a global user preference embedding is obtained via element-wise average fusion, with the representation form given by $\mathrm{E}_{\text {style}} \in \mathrm{R}^d$. In the geographic semantic feature branch, a scene topological graph structure is constructed based on the target geographic location information. An effective spatial range is demarcated with the target coordinates as the center, and point of interest landmarks and functional scenes within this region are treated as graph nodes, with node association weights defined jointly by spatial Euclidean distances and semantic category similarities. Structured spatial features are learned through a lightweight graph attention network, and a fixed-dimensional geographic semantic embedding $\mathrm{E}_{\text {geo}} \in \mathrm{R}^d$ is obtained via global average pooling compression, thereby preserving the spatial layout and scene semantic priors of the tourism destination in their entirety.

2.png

Figure 2. Architecture of the cross-modal symmetric cross-gating joint conditional encoding network

To address the limitations of fixed-weight fusion and unidirectional feature modulation, which fail to adapt to dynamic scene constraints, a symmetric cross-gating fusion mechanism is designed to achieve bidirectional adaptive coupling of the two types of heterogeneous features. Through this mechanism, feature weights are mutually modulated by dual gating branches, and the optimal fusion proportion of the two information sources is learned in a data-driven manner. The specific computational expression is given by:

$v_{\text {cond}}=\sigma\left(W_s \mathrm{E}_{\text {style}}+b_s\right) \odot \mathrm{E}_{\text {geo}}+\sigma\left(W_g \mathrm{E}_{\text {geo}}+b_g\right) \odot \mathrm{E}_{\text {style}}$       (1)

where, $\sigma$ denotes the Sigmoid activation function, which generates continuously differentiable soft gating weights; $W_s$ and $W_g$ are learnable weight matrices; $b_s$ and $b_g$ are bias parameters for the respective branches; and $\odot$ represents the element-wise product operation. Unlike the single-constraint mapping logic of conventional unidirectional modulation methods, this mechanism enables mutual constraint and mutual enhancement between the two feature types, with gating weights being dynamically and adaptively adjusted according to the characteristics of the input scene. For strongly structured tourism scenes such as architectural landmarks, the gating response intensity of the geographic semantic branch is elevated, and spatial structure and landmark layout in the generated imagery are preferentially constrained. For weakly structured scenes such as natural landscapes, the user aesthetic branch assumes dominant weighting, and personalized artistic styles are fully rendered while basic scene semantic plausibility is maintained, thereby effectively accommodating the complex and variable generation constraint requirements of cultural tourism scenarios.

To achieve parameter-efficient injection of conditional information, a lightweight hypernetwork is constructed based on the fused dynamic conditional vector v_cond, through which fine-grained conditional modulation of the attention mechanism in the pre-trained diffusion model is performed. Incremental bias parameters for the attention layers are dynamically generated by the hypernetwork according to the input conditional features, and minor modifications are applied to the native key and value parameters. The final attention parameter update is formulated as:

$K^{\prime}=K+\Delta K, V^{\prime}=V+\Delta V$        (2)

where, $K$ and $V$ denote the intrinsic key and value matrices of the pre-trained U-Net, $\Delta K$ and $\Delta V$ denote the dynamic conditional biases predicted by the hypernetwork, and $K^{\prime}$ and $V^{\prime}$ denote the final attention parameters after the introduction of semantic constraints. All pre-trained parameters of the diffusion backbone network are frozen throughout the entire training process, and external conditional information is introduced solely through a small number of incremental bias parameters, thereby preserving, to the greatest extent possible, the image generation priors learned by the model from general-purpose visual datasets. Furthermore, conditional modulation is applied only to the intermediate-resolution feature layers of the U-Net, while the low-resolution layers are left to maintain global semantic perception capabilities and the high-resolution layers are left to preserve image detail reconstruction capabilities. In this manner, an optimal balance is achieved among controllable generation precision, personalized expressiveness, and parameter computational overhead.

2.3 Spatially aware controllable perturbation mechanism for latent diffusion

To enable refined grounding and regulation of the fused conditional vector within the diffusion latent space, a spatially aware controllable perturbation mechanism for latent diffusion is constructed, through which the model denoising process is collaboratively guided from three dimensions: global style rendering, local semantic binding, and structural constraints. The multi-scale controllable perturbation mechanism for spatially aware latent diffusion is illustrated in Figure 3. A hierarchical perturbation strategy is adopted in this mechanism to overcome the deficiency that single conditional encoding cannot simultaneously maintain global aesthetic consistency and local geographic structural authenticity, thereby enabling synchronous controllability of stylistic expression and spatial layout in tourism imagery. For the unified regulation of global visual style, a channel-wise adaptive style normalization method is employed to perform semantically driven affine transformations on the U-Net feature maps of the diffusion model, where feature distribution patterns are dynamically adjusted to accommodate user aesthetic preferences. The specific computation is formulated as:

$\widetilde{F}_l=\gamma_l \cdot \frac{F_l-\mu\left(F_l\right)}{\sigma\left(F_l\right)}+\beta_l$        (3)

where, $F_l$ and $\widetilde{F}_l$ denote the feature maps of the $l$-th U-Net layer before and after transformation, respectively, and $\mu\left(F_l\right)$ and $\sigma\left(F_l\right)$ denote the channel-wise mean and standard deviation of the feature maps. The scale parameter $\gamma_l$ and shift parameter $\beta_l$ are obtained by mapping the fused conditional vector $v_{\text {cond }}$ through a single-layer multilayer perceptron $\Phi_l$, and adaptive normalization parameters are dynamically generated according to user aesthetic features. This approach relies entirely on semantic conditions to drive feature reshaping, without requiring the introduction of external style reference images, thereby fundamentally distinguishing itself from the image transfer logic of conventional adaptive instance normalization. Through this mechanism, semantically controllable unified regulation of global tone, contrast, and texture style is essentially achieved.

3.png

Figure 3. Multi-scale controllable perturbation mechanism for spatially aware latent diffusion

Global normalization operations are only capable of homogenized regulation of overall visual style and are unable to impose constraints on fine-grained geographic structures such as local landmarks, building contours, and scene boundaries in tourism scenes. To address this, a learnable spatial bias matrix is embedded within the cross-attention computation process of the diffusion model, through which a precise associative mapping between latent spatial positions and geographic semantic tokens is established. The optimized attention computation is formulated as:

$\operatorname{Attention}(Q, K, V)=\operatorname{Softmax}\left(\frac{Q K^T}{\sqrt{d}}+B\right) V$       (4)

where, $Q, K$, and $V$ denote the query, key, and value matrices of the attention mechanism, respectively; $d$ denotes the feature dimension; and $B$ denotes a spatial bias matrix of dimension $H W \times N$, with $H W$ representing the total number of latent spatial positions and N corresponding to the total number of geographic semantic tokens. The spatial bias matrix is generated by concatenating the fused conditional vector with two-dimensional sinusoidal positional encoding, followed by projection through a two-layer fully connected network, and independent modulation weights are assigned to each spatial position and each geographic semantic token category. Leveraging the temporal characteristics of diffusion denoising, a step-wise scheduling strategy is introduced, wherein spatial bias constraints are enabled only during medium-to-high noise iteration steps to govern the construction of the overall geographic structure of the scene, while bias constraints are disabled during low-noise iteration steps to preserve the model's capacity for reconstructing high-frequency image details. Through this approach, an effective balance between geographic structural fidelity and visual detail richness is achieved.

To further solidify the geographic authenticity of scene spatial layouts and compensate for the limitations of soft attention-based constraints, a geographic structural consistency loss function based on self-supervised features is constructed, through which strong supervisory constraints are imposed at each iteration step of the diffusion process. A clean image $\hat{x}_0$ at the current iteration stage is estimated in real time from the noisy latent variables via the inversion mechanism of the Denoising Diffusion Implicit Models (DDIM), and layout constraints are enforced by aligning the deep structural features of the generated image with those of the ground-truth scene. The loss function is defined as:

$L_{\text {geo}}=\left\|D I N O v 2\left(\hat{x}_0\right)-D I N O v 2\left(x_{\text {ref }}\right)\right\|_2$       (5)

where, $\mathrm{DINOv}_2()$ denotes the self-supervised feature extraction operation, and $x_{{ref}}$ denotes the mean feature baseline of ground-truth images of the target tourism destination, which is used to represent the intrinsic geographic structural prior of the scene. The pre-trained DINOv2 features exhibit strong spatial structure perception capabilities and robustness to variations in texture and color, enabling accurate capture of core geographic information such as scene layout and relative landmark positions. To balance structural constraints with generation diversity, a dynamic loss weight that decays linearly with noise intensity is introduced, with the weight coefficient given by $\lambda_{\text {geo}}=1-t / T$, where t denotes the current iteration step and T denotes the total number of denoising steps. A larger loss weight is assigned during high-noise stages to prioritize the correction of overall scene layout, while constraint intensity is weakened during low-noise stages to ensure sufficient creative freedom for the model.

The three-tier perturbation mechanism forms a progressively layered, soft–hard integrated latent space regulation system, through which refined control over the entire diffusion generation process is achieved. Channel-wise normalization accomplishes unified shaping of global aesthetic style, spatial bias attention enables precise binding of local geographic semantics, and structural consistency loss imposes closed-loop constraints on scene spatial logic at the feature level. Through the synergistic action of the three components, the latent-space evolution of the diffusion model is guided simultaneously by both user aesthetic preferences and destination geographic priors, fundamentally addressing the imbalance between stylistic rendering and real-scene structural fidelity in cultural tourism image generation.

2.4 Multi-reward guided closed-loop optimization of the conditional encoder

To achieve dynamic and iterative improvement in generation quality, a reinforcement learning mechanism is introduced to establish an end-to-end closed-loop optimization system, through which the conditional encoding strategy is adaptively updated under the reverse constraint of generation performance. Conventional optimization approaches for diffusion-based generation typically operate directly on the denoising network, where the step-wise latent-space evolution process tends to induce gradient oscillation and training redundancy. In this work, the complete tourism image generation pipeline is modeled as an episodic Markov decision process, through which the complex temporal optimization logic is effectively simplified. Within a single decision episode, the input user aesthetic features and geographic semantic features constitute the system state, the fused dynamic conditional vector serves as the agent's decision action, and the comprehensive quality score of the final generated image serves as the episodic supervisory reward. Through this modeling approach, the continuous optimization problem of multi-step denoising is transformed into a single-step conditional encoding decision problem, thereby significantly compressing the gradient computation dimension, suppressing gradient variance during training, and enhancing overall iterative stability. The multi-reward guided closed-loop optimization flow of the conditional encoder is illustrated in Figure 4.

4.png

Figure 4. Multi-reward guided closed-loop optimization flow of the conditional encoder

To comprehensively quantify the overall generation quality of cultural tourism imagery, a multi-dimensional fused reward function system is constructed, through which quantitative supervisory signals are established across three dimensions: image realism, geographic semantic matching, and user aesthetic alignment. Each reward component is first normalized to a unified numerical interval via min-max scaling, thereby eliminating weight bias caused by dimensional discrepancies. The realism reward is computed by evaluating the distributional discrepancy between generated images and real cultural tourism images using a pre-trained discriminator, with the formulation given by:

$R_{\text {real}}=\log \left(D\left(\hat{x}_0\right)\right)$       (6)

The geographic fidelity reward quantifies the degree of scene structural matching via semantic segmentation intersection over union, expressed as:

$R_{\text {geo}}=m \operatorname{IoU}\left(\operatorname{Seg}\left(\widehat{x}_0\right), \operatorname{Seg}_{\text {target}}\right)$       (7)

The user preference alignment reward measures the correspondence between generated content and user aesthetic style through feature cosine similarity, expressed as:

$R_{\text {pref}}=\cos \left(\right.$ Encoder$\left._{\text {img}}\left(\hat{x}_0\right), \mathrm{E}_{\text {style}}\right)$        (8)

The three reward components are combined through weighted fusion to obtain the final joint reward signal, with the overall formulation given by:

$R=\lambda_1 R_{\text {real}}+\lambda_2 R_{\text {geo}}+\lambda_3 R_{\text {pref}}$        (9)

where, $\lambda_1, \lambda_2$, and $\lambda_3$ denote the balancing weights for each reward component, respectively. A user preference prediction model is pre-trained in an offline manner, and its parameters remain fixed throughout the closed-loop training process, enabling the generation of stable and reliable aesthetic matching supervisory signals while avoiding reward fluctuation issues caused by dynamic training.

Based on the constructed joint reward signal, a policy gradient algorithm with baseline subtraction is employed to iteratively update the parameters of the conditional encoding network, with the optimization objective defined as the maximization of the expected reward. The policy gradient is computed as:

$\nabla_\theta J(\theta)=E_{\hat{x}_0 \sim p_\theta\left(\cdot \mid v_{\text {cond }}\right)}\left[\left(R\left(\hat{x}_0\right)-b\right) \cdot \nabla_\theta \log p_\theta\left(\hat{x}_0 \mid v_{\text {cond}}\right)\right]$        (10)

where, $\theta$ denotes all trainable parameters of the joint conditional encoding layer, and b denotes the baseline parameter computed as the mean reward over historical iterations. Through baseline deviation correction, gradient estimation bias is further reduced and training convergence efficiency is enhanced. The reparameterization trick is introduced to facilitate gradient computation in the continuous action space, where the stochastic sampling process of the conditional vector is transformed into a deterministic mapping superimposed with controllable noise, thereby effectively resolving the difficulty of gradient estimation in continuous decision spaces.

In terms of the parameter update paradigm, a localized fine-tuning optimization strategy is adopted, wherein only the parameters of the symmetric cross-gating fusion module, feature projection layers, and hypernetwork are updated via back-propagation, while the parameters of the diffusion backbone network, feature extraction networks, and reward evaluation models remain frozen throughout training. Through this optimization strategy, the general image generation priors of the pre-trained diffusion model are fully preserved, and only the fusion and injection strategies for heterogeneous conditions are targeted for optimization. In this manner, sustained iterative improvement in generation quality is achieved while parameter efficiency and training stability are simultaneously maintained, ultimately establishing an endogenous closed-loop optimization paradigm encompassing encoding, generation, evaluation, and updating.

The closed-loop generation framework jointly driven by preference and space proposed in this work exhibits excellent model compatibility and cross-scenario generalization capability. A modularly decoupled architecture is adopted in the overall framework, wherein the conditional encoding, spatial perturbation regulation, and closed-loop optimization submodules are independent of the diffusion backbone network, enabling adaptation to various mainstream pre-trained diffusion foundations without requiring large-scale restructuring or fine-tuning of the main model architecture. At the level of cultural tourism application scenarios, the framework is capable of adaptively accommodating diverse regional landscapes and scene structural characteristics. Both precise constraint of spatial proportions and layout logic in highly structured landmark buildings and flexible adaptation to personalized stylistic creation requirements in weakly structured scenes—such as natural landscapes and urban streetscapes—are achieved, demonstrating universal regulatory capacity across diversified tourism scenes. Furthermore, the model dynamically adjusts the fusion weights of dual-source features according to the aesthetic characteristics of different users, accommodating differentiated user-generated content creation styles such as fresh, vintage, and textured aesthetics. From the perspective of technical transferability, the heterogeneous feature adaptive fusion and feedback-driven iterative optimization mechanisms constructed in this work are not limited to cultural tourism image generation scenarios but can be extended to multiple cross-modal generation tasks, including controllable remote sensing image generation and customized scene-oriented visual creation, thereby offering substantial general research value and broad extension potential.

Although the proposed method achieves synergistic optimization of geographic fidelity and preference matching in conventional tourism scenarios, certain performance limitations persist in extreme and niche scenes. For niche cultural tourism scenes with scarce sample accumulation, the limited scene prior information prevents the model from accurately capturing exclusive spatial structures and semantic features, and the adaptive regulation precision of the gating fusion mechanism is correspondingly diminished, resulting in slight deviations in the geographic detail authenticity of generated content. When users exhibit extreme aesthetic preferences characterized by highly artistic or surreal styles, a trade-off conflict arises between the demand for intensive stylistic rendering and the constraint of geographic structural fidelity, and the model struggles to achieve absolute equilibrium between the two optimization objectives, frequently giving rise to issues such as localized weakening of scene structure and fine-grained semantic deviations. Furthermore, under extreme lighting conditions—including backlighting, overexposure, and nighttime scenes—the distribution of low-level visual features in the imagery undergoes shifts, and the regulatory stability of the latent-space constraint mechanism is reduced, leading to unnatural light transitions and localized detail distortions in the generated imagery. Consequently, there remains room for improvement in the model's robustness to complex and extreme environments.

This study offers novel theoretical insights and technical paradigm support for the field of controllable diffusion generation. Existing controllable generation research has largely relied on static conditional inputs and open-loop generation logic. In contrast, through the construction of a closed-loop encoding-side optimization system in this work, it has been verified that multi-dimensional quality feedback signals can endogenously drive the autonomous iterative evolution of conditional encoding strategies, thereby breaking through the performance bottleneck of static mapping generation paradigms and advancing the theoretical framework of controllable generation with multi-heterogeneous-condition collaborative constraints and multi-objective joint optimization. The hierarchical spatially aware perturbation and bidirectional adaptive feature fusion strategies also provide novel technical pathways for resolving multi-source condition conflicts and achieving fine-grained scene semantic control. Future research can further integrate three-dimensional geographic information data to strengthen precise constraints on the three-dimensional spatial structure of tourism scenes, and incorporate multi-round human–computer interaction mechanisms to enable dynamic iterative updating of user aesthetic preferences. Additionally, the closed-loop collaborative optimization paradigm can be extended to multi-dimensional digital content generation tasks—such as cultural tourism short videos and panoramic imagery—thereby further promoting the refinement and deployment of intelligent content generation technology systems for smart cultural tourism.