Generation method of generative adversarial network for planning resource design stylized layout

Abstract

The invention discloses a generation method of a generative adversarial network for planning a resource design stylized layout, and belongs to the technical field of artificial intelligence and machine learning, and the method comprises the following steps: S1, multi-source data collection: widely collecting a large number of existing planning resource samples; s2, data classification labeling; s3, constructing a generative adversarial network model; s4, performing model training; s5, style migration training; s6, optimizing the model; s7, generating a planning resource scheme; s8, optimizing the scheme; and S9, performing feedback iteration. According to the method, through integrating a technical path of multi-source design data acquisition, deep style feature extraction and generative adversarial network collaborative optimization, full-process automation from user demand input to high-quality planning resource layout generation is realized, and user personalized demands are met through a multi-scheme candidate and interactive adjustment function; and finally, constructing an incremental learning closed loop based on user feedback data, and continuously updating model parameters to adapt to an emerging design trend and an industry specification.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and machine learning technologies, and in particular to a method for generating a stylized layout of resource designs using a generative adversarial network. Background Art

[0002] In today's digital age, the demand for various planning resources continues to grow, and the requirements for design style and layout rationality are becoming increasingly stringent. Traditional design processes rely on manual creativity and experience, which is not only inefficient and unable to meet the needs of rapid iteration, but also easily influenced by the designer's subjective factors, making it difficult to achieve precise stylistic customization. Therefore, we propose a generative adversarial network method to generate stylized layouts for planning resource design to solve this problem. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for generating a stylized layout of resource designs using a generative adversarial network to solve the problems raised in the above background technology.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions: The generative adversarial network is used to generate a stylized layout of resource designs, including the following steps: S1. Multi-source data collection: extensively collect massive existing planning resource samples; S2. Data classification and annotation: Perform style classification and annotation on the collected samples, and divide the samples into different style categories based on the typical characteristics of the design style; S3. Build a Generative Adversarial Network Model: Build a generator network whose input is the style feature vector and layout constraints, and use the pre-trained style extraction model to extract features from the style examples; S4. Model training: Divide the preprocessed dataset into training set, validation set, and test set according to a certain ratio, and use the training set to initialize the generative adversarial network; S5. Style transfer training: Select sample pairs of different styles, convert the layout of one style to another, generate the converted layout through the generator and calculate the style transfer loss, which is added to the total loss function for training; S6. Model optimization: After model training is completed, the generator and discriminator are optimized to remove redundant network connections and parameters and compress the model size. S7. Planning resource solution generation: The user inputs the design requirements for planning resources through the interactive interface. The system uses a pre-trained style extraction model to extract the corresponding style features based on the design style selected by the user or the style example uploaded, and encodes them into a style feature vector. The style feature vector and layout constraint vector are input into the trained generator network. The generator generates a planning resource layout solution that meets the specified style and layout constraints based on the input information. S8, Scheme Optimization: Post-process the generated layout scheme to check the completeness and rationality of the layout, and perform fine-tuning and optimization based on design rules and aesthetic principles; S9. Feedback iteration: The generated resource planning layout is displayed to users, who provide ratings and feedback. Based on user feedback, the feedback information is converted into data labels that can be learned by the model, and the generative adversarial network model is fine-tuned and trained, and the model parameters are updated to achieve iterative optimization of the model.

[0005] Preferably, in said S1, the specific steps are as follows: S101. Multi-platform data crawling and integration: crawl various types of planning resource samples from public design communities, corporate case libraries, and open source data sets, covering mainstream styles such as flat, minimalist, and retro, and store them in unified PSD, Sketch, and PNG formats; S102. Multimodal data collection expansion: Supplementary collection of user hand-drawn sketches, style reference images, text descriptions, and voice commands to build a multimodal input sample library of images, text, and voice; S103. Data cleaning and format standardization: Filter low-resolution, unclear copyright or duplicate samples, unify image resolution and color space, and encode text in UTF-8 format to ensure data quality and consistency.

[0006] Preferably, in said S2, the specific steps are as follows: S201. Establishment of style classification standards: Define a style labeling system based on design theory and clarify the visual characteristics of each style; S202, Semi-automatic Labeling and Expert Review: Use the pre-trained style classification model to perform preliminary labeling of samples, and manual experts review and correct incorrect labels to ensure labeling accuracy; S203, Data enhancement and category balance: Perform enhancements such as rotation, color perturbation, and element replacement on scarce style samples to balance the amount of data in each category and prevent model training bias.

[0007] Preferably, in S3, the specific steps are as follows: S301. Generator network architecture design: The input layer receives the style feature vector and layout constraint vector. The core structure is a U-Net generator combined with an attention mechanism to dynamically fuse style and layout information. The output layer generates layout parameters and style parameters. S302, Discriminator Network Construction: Using a multi-scale PatchGAN structure, it evaluates layout rationality, style consistency, and aesthetic scores respectively; S303, pre-trained style extraction model integration: load the CLIP-ViT model frozen weights, extract image style features, and encode them into a 128-dimensional vector input generator.

[0008] Preferably, in said S4, the specific steps are as follows: S401. Dataset partitioning strategy: The training set, validation set, and test set are divided into a ratio of 7:2:1 to ensure a uniform distribution of styles. The validation set is used for hyperparameter tuning, and the test set is used for final performance evaluation. S402, Adversarial Training and Loss Function Design: Distinguishing between real and generated layouts through adversarial loss, and jointly optimizing layout and style discrimination capabilities; S403, training strategy optimization: progressive training is adopted, training in stages from low resolution to high resolution, combined with gradient penalty to improve stability.

[0009] Preferably, in S5, the specific steps are as follows: S501, cross-style sample pair construction: select sample pairs with the same content but different styles, requiring the content elements to be consistent and only the style attributes to change; S502, style feature decoupling and transfer: Decouple content and style features through a variational autoencoder, inject the target style vector into the generator, and generate a cross-style layout solution; S503. Transfer loss calculation and fusion: Calculate the distance between the generated layout and the target style CLIP feature as the transfer loss, add the weighted value to the total loss function, and force the model to learn style-independent content retention capabilities.

[0010] Preferably, in said S6, the specific steps are as follows: S601, Model Pruning and Quantization: Channel pruning is used to remove redundant convolution kernels in the generator, and the weights are quantized to 8 bits, compressing the model volume by 60%; S602, Knowledge Distillation Lightweighting: Use the original large model to guide the lightweight small model, and transfer layout generation and style transfer knowledge through feature alignment loss; S603, Inference Acceleration Deployment: Integrate the TensorRT engine to optimize the computational graph and implement real-time inference on NVIDIA GPUs.

[0011] Preferably, in said S7, the specific steps are as follows: S701, User Needs Analysis and Feature Encoding: BERT extracts keywords and encodes them into layout constraint vectors. CLIP extracts style features of user-uploaded images and encodes them into 128-dimensional vectors. S702, generator reasoning and solution generation: input the style vector and layout constraints into the generator, and output layout parameters and style parameters; S703, multiple solution candidate output: generate differentiated solutions through noise injection for user selection or further adjustment.

[0012] Preferably, in said S8, the specific steps are as follows: S801, Layout integrity check: Detect element missing, overflow and overlap issues, and trigger automatic repair; S802. Fine-tune aesthetic rules: Optimize element spacing based on design principles and adjust color contrast to ≥4.5:1 to ensure accessible readability. S803. Dynamic responsive adaptation: Generate multi-resolution layouts based on the terminal device, prioritize vertical arrangement on mobile devices, and use a grid system for column display on PCs.

[0013] Preferably, in said S9, the specific steps are as follows: S901. User feedback data collection: record user ratings, manual adjustment records, and A / B test results; S902, Feedback Data Labeling: Convert user behavior into supervision signals and build incremental training datasets; S903, Online fine-tuning and version iteration: Inject new data into the training process every week, use the elastic weight consolidation algorithm to prevent catastrophic forgetting, and update the model version.

[0014] The beneficial effects of the present invention are: In the present invention, the generative adversarial network is used to generate a method for planning resource design stylized layouts, which solves the problems of data uniformity and insufficient style generalization capabilities of traditional methods through multi-source data collection and refined style classification and labeling technology. The system integrates massive design samples from multiple platforms, covering mainstream and niche styles such as flat, minimalist, and retro, and constructs a high-precision style labeling system through semi-automatic labeling and expert review mechanisms. Style features are extracted based on a pre-trained visual language model, combined with content and layout decoupling technology to ensure that the generated solution not only meets the user's specified style but also retains the core information structure. This method breaks through the style limitations of traditional template tools, can adapt to the design needs of multiple fields such as e-commerce, education, and culture, and significantly improves the diversity and professionalism of planning resources; In the present invention, the generative adversarial network is used to plan the generation method of the stylized layout of resource design. Through the multimodal fusion architecture and dynamic optimization strategy of the generative adversarial network, high-quality and diversified style transfer and layout generation are achieved. The generator network combines the attention mechanism with the pre-trained feature encoder to accurately align the style examples and layout constraints input by the user to generate a design scheme that is both aesthetically reasonable and functional. The discriminator uses a multi-scale evaluation module to jointly optimize from three dimensions: layout logic, style consistency, and visual appeal, to avoid common design defects such as element overlap and color conflict. This technology enables non-professional users to quickly generate schemes comparable to manual designs, significantly reducing corporate outsourcing costs and creation cycles; In the present invention, the generative adversarial network is used to plan the generation method of the stylized layout of resource design, and flexible style conversion and efficient resource adaptation are achieved through cross-style transfer training and lightweight model deployment technology. The system supports seamless migration of the same content to different style scenes, such as converting a science and technology PPT from a minimalist style to a cyberpunk style, while maintaining the integrity and readability of the content structure. Through model pruning, quantization and knowledge distillation technology, the generative network can run efficiently on mobile terminals and edge devices, supporting real-time layout generation and adjustment in offline environments. This capability makes it widely used in scenarios with high requirements for timeliness and portability, such as offline event planning and mobile advertising; In the present invention, the generative adversarial network is used to plan the generation method of the stylized layout of resource design, and a highly personalized design experience is achieved through user-driven interactive optimization and a multi-solution candidate mechanism. Users can input their requirements through natural language description, reference image upload or manual adjustment, and the system generates multiple sets of differentiated solutions and provides a visual comparison function. Combined with automated aesthetic rule verification and responsive adaptation technology, it dynamically optimizes element proportions, color contrast and cross-terminal display effects to ensure the visual consistency of the solution on multiple devices such as mobile phones, tablets, and large screens. This function is particularly suitable for complex design needs such as brand marketing and multinational companies that require multiple versions and multi-scene adaptation; In this invention, the generative adversarial network is used to generate a method for planning stylized layouts of resource designs. Through a closed feedback loop and incremental learning mechanism, the model achieves continuous evolution and long-term value enhancement. The system collects user ratings, modification behavior, and business indicator data in real time, converting them into learnable signals such as style preference weights and layout priorities, and injecting them into the model training process through an elastic weight consolidation algorithm. This mechanism enables the generation strategy to dynamically adapt to changes in user needs and iterative market trends, such as rapidly learning emerging design styles or updated industry standards. Compared to traditional static models, this system demonstrates greater adaptability and sustainable commercial value in long-term applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a schematic diagram of the three-dimensional structure of the method for generating stylized layouts of resource designs using a generative adversarial network proposed in the present invention. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0017] Reference Figure 1 , a generative adversarial network is used to generate a stylized layout of curated resource designs, comprising the following steps: S1. Multi-source data collection: extensively collect massive existing planning resource samples; S2. Data classification and annotation: Perform style classification and annotation on the collected samples, and divide the samples into different style categories based on the typical characteristics of the design style; S3. Build a Generative Adversarial Network Model: Build a generator network whose input is the style feature vector and layout constraints, and use the pre-trained style extraction model to extract features from the style examples; S4. Model training: Divide the preprocessed dataset into training set, validation set, and test set according to a certain ratio, and use the training set to initialize the generative adversarial network; S5. Style transfer training: Select sample pairs of different styles, convert the layout of one style to another, generate the converted layout through the generator and calculate the style transfer loss, which is added to the total loss function for training; S6. Model optimization: After model training is completed, the generator and discriminator are optimized to remove redundant network connections and parameters and compress the model size. S7. Planning resource solution generation: The user inputs the design requirements for planning resources through the interactive interface. The system uses a pre-trained style extraction model to extract the corresponding style features based on the design style selected by the user or the style example uploaded, and encodes them into a style feature vector. The style feature vector and layout constraint vector are input into the trained generator network. The generator generates a planning resource layout solution that meets the specified style and layout constraints based on the input information. S8, Scheme Optimization: Post-process the generated layout scheme to check the completeness and rationality of the layout, and perform fine-tuning and optimization based on design rules and aesthetic principles; S9. Feedback iteration: The generated resource planning layout is displayed to users, who provide ratings and feedback. Based on user feedback, the feedback information is converted into data labels that can be learned by the model, and the generative adversarial network model is fine-tuned and trained, and the model parameters are updated to achieve iterative optimization of the model.

[0018] In this embodiment, in S1, the specific steps are as follows: S101. Multi-platform data crawling and integration: crawl various types of planning resource samples from public design communities, corporate case libraries, and open source data sets, covering mainstream styles such as flat, minimalist, and retro, and store them in unified PSD, Sketch, and PNG formats; S102. Multimodal data collection expansion: Supplementary collection of user hand-drawn sketches, style reference images, text descriptions, and voice commands to build a multimodal input sample library of images, text, and voice; S103. Data cleaning and format standardization: Filter low-resolution, unclear copyright or duplicate samples, unify image resolution and color space, and encode text in UTF-8 format to ensure data quality and consistency.

[0019] In this embodiment, in S2, the specific steps are as follows: S201. Establishment of style classification standards: Define a style labeling system based on design theory and clarify the visual characteristics of each style; S202, Semi-automatic Labeling and Expert Review: Use the pre-trained style classification model to perform preliminary labeling of samples, and manual experts review and correct incorrect labels to ensure labeling accuracy; S203, Data enhancement and category balance: Perform enhancements such as rotation, color perturbation, and element replacement on scarce style samples to balance the amount of data in each category and prevent model training bias.

[0020] In this embodiment, in S3, the specific steps are as follows: S301. Generator network architecture design: The input layer receives the style feature vector and layout constraint vector. The core structure is a U-Net generator combined with an attention mechanism to dynamically fuse style and layout information. The output layer generates layout parameters and style parameters. S302, Discriminator Network Construction: Using a multi-scale PatchGAN structure, it evaluates layout rationality, style consistency, and aesthetic scores respectively; S303, pre-trained style extraction model integration: load the CLIP-ViT model frozen weights, extract image style features, and encode them into a 128-dimensional vector input generator.

[0021] In this embodiment, in S4, the specific steps are as follows: S401. Dataset partitioning strategy: The training set, validation set, and test set are divided into a ratio of 7:2:1 to ensure a uniform distribution of styles. The validation set is used for hyperparameter tuning, and the test set is used for final performance evaluation. S402, Adversarial Training and Loss Function Design: Distinguishing between real and generated layouts through adversarial loss, and jointly optimizing layout and style discrimination capabilities; S403, training strategy optimization: progressive training is adopted, training in stages from low resolution to high resolution, combined with gradient penalty to improve stability.

[0022] In this embodiment, in S5, the specific steps are as follows: S501, cross-style sample pair construction: select sample pairs with the same content but different styles, requiring the content elements to be consistent and only the style attributes to change; S502, style feature decoupling and transfer: Decouple content and style features through a variational autoencoder, inject the target style vector into the generator, and generate a cross-style layout solution; S503. Transfer loss calculation and fusion: Calculate the distance between the generated layout and the target style CLIP feature as the transfer loss, add the weighted value to the total loss function, and force the model to learn style-independent content retention capabilities.

[0023] In this embodiment, in S6, the specific steps are as follows: S601, Model Pruning and Quantization: Channel pruning is used to remove redundant convolution kernels in the generator, and the weights are quantized to 8 bits, compressing the model volume by 60%; S602, Knowledge Distillation Lightweighting: Use the original large model to guide the lightweight small model, and transfer layout generation and style transfer knowledge through feature alignment loss; S603, Inference Acceleration Deployment: Integrate the TensorRT engine to optimize the computational graph and implement real-time inference on NVIDIA GPUs.

[0024] In this embodiment, in S7, the specific steps are as follows: S701, User Needs Analysis and Feature Encoding: BERT extracts keywords and encodes them into layout constraint vectors. CLIP extracts style features of user-uploaded images and encodes them into 128-dimensional vectors. S702, generator reasoning and solution generation: input the style vector and layout constraints into the generator, and output layout parameters and style parameters; S703, multiple solution candidate output: generate differentiated solutions through noise injection for user selection or further adjustment.

[0025] In this embodiment, in S8, the specific steps are as follows: S801, Layout integrity check: Detect element missing, overflow and overlap issues, and trigger automatic repair; S802. Fine-tune aesthetic rules: Optimize element spacing based on design principles and adjust color contrast to ≥4.5:1 to ensure accessible readability. S803. Dynamic responsive adaptation: Generate multi-resolution layouts based on the terminal device, prioritize vertical arrangement on mobile devices, and use a grid system for column display on PCs.

[0026] In this embodiment, in S9, the specific steps are as follows: S901. User feedback data collection: record user ratings, manual adjustment records, and A / B test results; S902, Feedback Data Labeling: Convert user behavior into supervision signals and build incremental training datasets; S903, Online fine-tuning and version iteration: Inject new data into the training process every week, use the elastic weight consolidation algorithm to prevent catastrophic forgetting, and update the model version.

[0027] In this embodiment, during the specific implementation, by integrating the technical paths of multi-source design data collection, deep style feature extraction and generative adversarial network collaborative optimization, the entire process from user demand input to high-quality planning resource layout generation is automated. The system first collects multimodal data from design communities, corporate case libraries, and user interaction behaviors, and constructs a high-precision style annotation dataset through cleaning, classification, and enhancement; then, through a generative adversarial network architecture, it encodes the user's input text descriptions, reference images, or voice commands into style feature vectors, generates an initial plan based on layout constraints, and evaluates the layout rationality, style consistency, and aesthetic performance based on a multi-scale discriminator; then, through cross-style transfer training and dynamic optimization strategies, it achieves seamless conversion of the same content in different style scenarios, while utilizing model lightweighting and edge deployment technologies to adapt to multi-terminal real-time generation needs; in the output stage, the system combines automated aesthetic rule verification with responsive adaptation technology to optimize element proportions and cross-device display effects, and meets user personalized needs through multiple plan candidates and interactive adjustment functions; finally, it constructs an incremental learning closed loop based on user feedback data, continuously updates model parameters to adapt to emerging design trends and industry standards, and forms a complete intelligent design ecosystem from data-driven generation to user feedback iteration, significantly improving the efficiency, diversity, and commercial value of planning resource creation.

[0028] The above is a detailed introduction to the method for generating a stylized layout of resource design using a generative adversarial network provided by the present invention. Specific embodiments are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. It should be pointed out that for those skilled in the art, without departing from the principles of the present invention, the present invention can also be improved and modified in several ways, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A method for generating stylized layouts of resource designs using generative adversarial networks, characterized in that: The following steps are involved: S1. Multi-source data collection: extensively collect massive existing planning resource samples; S2. Data classification and annotation: Perform style classification and annotation on the collected samples, and divide the samples into different style categories based on the typical characteristics of the design style; S3. Build a Generative Adversarial Network Model: Build a generator network whose input is the style feature vector and layout constraints, and use the pre-trained style extraction model to extract features from the style examples; S4. Model training: Divide the preprocessed dataset into training set, validation set, and test set according to a certain ratio, and use the training set to initialize the generative adversarial network; S5. Style transfer training: Select sample pairs of different styles, convert the layout of one style to another, generate the converted layout through the generator and calculate the style transfer loss, which is added to the total loss function for training; S6. Model optimization: After model training is completed, the generator and discriminator are optimized to remove redundant network connections and parameters and compress the model size. S7. Planning resource solution generation: The user inputs the design requirements for planning resources through the interactive interface. The system uses a pre-trained style extraction model to extract the corresponding style features based on the design style selected by the user or the style example uploaded, and encodes them into a style feature vector. The style feature vector and layout constraint vector are input into the trained generator network. The generator generates a planning resource layout solution that meets the specified style and layout constraints based on the input information. S8, Scheme Optimization: Post-process the generated layout scheme to check the completeness and rationality of the layout, and perform fine-tuning and optimization based on design rules and aesthetic principles; S9. Feedback iteration: The generated resource planning layout is displayed to users, who provide ratings and feedback. Based on user feedback, the feedback information is converted into data labels that can be learned by the model, and the generative adversarial network model is fine-tuned and trained, and the model parameters are updated to achieve iterative optimization of the model.

2. The method for generating a stylized layout of a resource design using a generative adversarial network according to claim 1, characterized in that: In S1, the specific steps are as follows: S101. Multi-platform data crawling and integration: crawl various types of planning resource samples from public design communities, corporate case libraries, and open source data sets, covering mainstream styles such as flat, minimalist, and retro, and store them in unified PSD, Sketch, and PNG formats; S102. Multimodal data collection expansion: Supplementary collection of user hand-drawn sketches, style reference images, text descriptions, and voice commands to build a multimodal input sample library of images, text, and voice; S103. Data cleaning and format standardization: Filter low-resolution, unclear copyright or duplicate samples, unify image resolution and color space, and encode text in UTF-8 format to ensure data quality and consistency.

3. The method for generating a stylized layout of a resource design using a generative adversarial network according to claim 1, characterized in that: In S2, the specific steps are as follows: S201. Establishment of style classification standards: Define a style labeling system based on design theory and clarify the visual characteristics of each style; S202, Semi-automatic Labeling and Expert Review: Use the pre-trained style classification model to perform preliminary labeling of samples, and manual experts review and correct incorrect labels to ensure labeling accuracy; S203, Data enhancement and category balance: Perform enhancements such as rotation, color perturbation, and element replacement on scarce style samples to balance the amount of data in each category and prevent model training bias.

4. The method for generating a stylized layout of a resource design using a generative adversarial network according to claim 1, characterized in that: In S3, the specific steps are as follows: S301, generator network architecture design: the input layer receives the style feature vector and the layout constraint vector, the core structure is a U-Net generator combined with an attention mechanism, dynamically integrating style and layout information, and the output layer generates layout parameters and style parameters; S302, Discriminator Network Construction: Using a multi-scale PatchGAN structure, it evaluates layout rationality, style consistency, and aesthetic scores respectively; S303, pre-trained style extraction model integration: load the CLIP-ViT model frozen weights, extract image style features, and encode them into a 128-dimensional vector input generator.

5. The method for generating a stylized layout of a resource design using a generative adversarial network according to claim 1, characterized in that: In said S4, the specific steps are as follows: S401. Dataset partitioning strategy: The training set, validation set, and test set are divided into a ratio of 7:2:1 to ensure a uniform distribution of styles. The validation set is used for hyperparameter tuning, and the test set is used for final performance evaluation. S402, Adversarial Training and Loss Function Design: Distinguishing between real and generated layouts through adversarial loss, and jointly optimizing layout and style discrimination capabilities; S403, training strategy optimization: progressive training is adopted, training in stages from low resolution to high resolution, combined with gradient penalty to improve stability.

6. The method for generating a stylized layout of a resource design using a generative adversarial network according to claim 1, characterized in that: In said S5, the specific steps are as follows: S501, cross-style sample pair construction: select sample pairs with the same content but different styles, requiring the content elements to be consistent and only the style attributes to change; S502, style feature decoupling and transfer: Decouple content and style features through a variational autoencoder, inject the target style vector into the generator, and generate a cross-style layout solution; S503. Transfer loss calculation and fusion: Calculate the distance between the generated layout and the target style CLIP feature as the transfer loss, add the weighted value to the total loss function, and force the model to learn style-independent content retention capabilities.

7. The method for generating a stylized layout of a resource design using a generative adversarial network according to claim 1, characterized in that: In S6, the specific steps are as follows: S601, model pruning and quantization: channel pruning is used to remove redundant convolution kernels in the generator, the weights are quantized to 8 bits, and the model volume is compressed by 60%; S602, Knowledge Distillation Lightweighting: Use the original large model to guide the lightweight small model, and transfer layout generation and style transfer knowledge through feature alignment loss; S603, Inference Acceleration Deployment: Integrate the TensorRT engine to optimize the computational graph and implement real-time inference on NVIDIA GPUs.

8. The method for generating a stylized layout of a resource design using a generative adversarial network according to claim 1, characterized in that: In said S7, the specific steps are as follows: S701, User Needs Analysis and Feature Encoding: BERT extracts keywords and encodes them into layout constraint vectors. CLIP extracts style features of user-uploaded images and encodes them into 128-dimensional vectors. S702, generator reasoning and solution generation: input the style vector and layout constraints into the generator, and output layout parameters and style parameters; S703, multiple solution candidate output: generate differentiated solutions through noise injection for user selection or further adjustment.

9. The method for generating a stylized layout of a resource design using a generative adversarial network according to claim 1, characterized in that: In said S8, the specific steps are as follows: S801, Layout integrity check: Detect element missing, overflow and overlap issues, and trigger automatic repair; S802. Fine-tune aesthetic rules: Optimize element spacing based on design principles and adjust color contrast to ≥4.5:1 to ensure accessible readability. S803. Dynamic responsive adaptation: Generate multi-resolution layouts based on the terminal device, prioritize vertical arrangement on mobile devices, and use a grid system for column display on PCs.

10. The method for generating a stylized layout of a resource design using a generative adversarial network according to claim 1, characterized in that: In said S9, the specific steps are as follows: S901. User feedback data collection: record user ratings, manual adjustment records, and A / B test results; S902, Feedback Data Labeling: Convert user behavior into supervision signals and build incremental training datasets; S903, Online fine-tuning and version iteration: Inject new data into the training process every week, use the elastic weight consolidation algorithm to prevent catastrophic forgetting, and update the model version.

Images