Method and system for dynamically generating personalized advertisement based on AIGC

Abstract

The application discloses a kind of personalized advertisement dynamic generation method and system based on AIGC, it is related to digital advertising technical field.The method includes: obtaining the multidimensional data of target user, and multidimensional feature vector construction is carried out;Through the construction strategy of preset prompt word, structured generation prompt word is generated;The generation prompt word is input into AIGC model, and initial advertisement material is generated;Initial advertisement material is detected and quality is evaluated according to compliance;For background image, initial text box position is determined, and according to visual saliency, page self-adaptive adjustment is executed, and repositioning text box position is obtained;The initial advertisement material and the product main image data of product to be promoted are rendered into final personalized advertisement, and push to the target user.The application solves the problems of insufficient multidimensional feature fusion, text layout conflict, obvious synthesis trace and uncontrollable quality in advertisement generation in the prior art.

Description

Technical Field

[0001] This invention relates to the field of digital advertising technology, and in particular to a method and system for dynamically generating personalized advertisements based on AIGC. Background Technology

[0002] In today's digital marketing and e-commerce landscape, personalized advertising has become a core tool for improving conversion rates. Traditional ad creative generation relies primarily on human designers to create fixed posters or videos based on product characteristics. This process is not only time-consuming and costly, but also struggles to dynamically adapt to different users' interests, preferences, real-time emotions, and environments.

[0003] With the development of Artificial Intelligence Generated Content (AIGC) technology, especially the maturity of multimodal diffusion models, it has become possible to automatically generate images and text using text prompts. However, existing AIGC ad generation methods still have the following technical problems: 1) Mechanized prompt word construction: It is usually based on simple product tag splicing without deep integration of user multi-dimensional profiles (such as real-time emotions and behavioral preferences) and environmental context (such as weather and time), resulting in "one-size-fits-all" ads with poor emotional resonance.

[0004] 2) Image and text layout is prone to conflict: The background image generated by the AIGC model is unpredictable. Directly superimposing the product main image and text on the generated background can easily result in the text being submerged in the complex background and becoming unreadable, or the text obscuring the main product area, resulting in poor visual effects.

[0005] 3) Obvious composite traces: When blending the main product image with the generated background, flaws such as edge halo, jagged edges, and inconsistent lighting often occur. Furthermore, the background color is not adapted when rendering text, resulting in a lack of quality in the advertisement.

[0006] 4) Lagging compliance and quality control: AIGC models exhibit "illusion" phenomena, potentially generating images containing illegal content or with blurry quality, lacking a multi-dimensional quality assessment and interception mechanism beforehand.

[0007] Therefore, there is an urgent need for an AIGC personalized ad dynamic generation solution that can deeply integrate multi-dimensional features, automatically adjust the layout, and has a complete quality assessment and seamless synthesis mechanism. Summary of the Invention

[0008] This invention provides a method and system for dynamically generating personalized advertisements based on AIGC, relating to the field of digital advertising technology. This invention solves the problems of insufficient multi-dimensional feature fusion, easy conflicts in image and text layout, obvious synthesis traces, and uncontrollable quality in existing technologies for advertisement generation.

[0009] In a first aspect, embodiments of the present invention provide a method for dynamically generating personalized advertisements based on AIGC, the method comprising: Acquire historical interaction data and real-time environmental data of target users, as well as product feature data of the product to be promoted, and construct multi-dimensional feature vectors to generate comprehensive user feature vectors, environmental context feature vectors, and product feature vectors. Based on the comprehensive user feature vector, the environmental context feature vector, and the product feature vector, a structured generation prompt word is generated using a preset prompt word construction strategy. The generated prompts are input into a multimodal generation model constructed based on the AIGC algorithm to generate initial advertising materials containing background images and text content; The background images and text content of the initial advertising materials were subjected to compliance testing and quality assessment. If the detection passes and the quality assessment results meet the preset conditions, then for the background image of the initial advertising material, the initial text box position is determined, and an adaptive adjustment of the layout based on visual salience is performed to obtain the repositioned text box position; Based on the position of the repositioning text box, the initial advertising material and the main product image data of the product to be promoted are rendered and synthesized into a final personalized advertisement, which is then pushed to the target user.

[0010] The technical solution provided in this application has at least the following beneficial effects: By constructing user feature vectors encompassing demographics, behavior, interests, and emotions, and combining them with environmental context and product features, structured prompts are generated. This enables the AIGC model to produce personalized ads that highly match the user's current mood and environment, enhancing conversion potential. A comprehensive saliency map is introduced, not only avoiding high visual focus areas in the generated background but also incorporating the pre-defined product main image placement area as a high-saliency constraint into the map. The target loss function is used to optimize and relocate the text box, completely resolving the pain points of text and product occlusion and unreadable text. In the synthesis stage, semantic segmentation is used to extract a product mask to guide Poisson fusion, eliminating edge halos and jagged edges. After fusion, the text color is adaptively calculated based on the dominant color of the relocated text box area, and finally, alpha blending is performed to ensure high visual harmony in the final ad image. Before synthesis, dual compliance checks (image violation confidence and text violation confidence) are performed on the background image and text, and a weighted evaluation is conducted based on aesthetics, consistency, and clarity. A veto and threshold interception are implemented to ensure the safety and high quality of the ads.

[0011] In one optional implementation, the historical interaction data includes the target user's clickstream records, purchase history, search keywords, and basic attribute information; The real-time environmental data includes the time input data, location input data, and weather input data of the target user's environment; The product feature data includes the product text description and product main image data of the product to be promoted; The comprehensive user feature vector consists of demographic feature sub-vectors, behavioral feature sub-vectors, interest tag sub-vectors, and real-time emotional state sub-vectors.

[0012] In one optional implementation, based on the comprehensive user feature vector, the environmental context feature vector, and the product feature vector, structured generation prompts are generated using a preset prompt word construction strategy, including: Calculate the semantic similarity between the integrated user feature vector and the product feature vector in a preset tag library, select the Top-K keywords, and form a keyword set; Based on the environmental context feature vector and the real-time emotional state sub-vector in the user feature vector, an atmosphere control parameter generation model is used to obtain the atmosphere control parameter vector of the generated image. The keyword set and the style modifiers obtained by discretizing and mapping the atmosphere control parameter vector are filled into a preset prompt word template to generate structured prompt words.

[0013] In one alternative implementation, the multimodal generative model is constructed according to the AIGC algorithm, including: A text encoder is generated to generate a text semantic embedding vector based on the generated prompt words; Generate a visual encoder, which is used to generate a product visual feature vector based on the product main image data in the product feature data of the product to be promoted; The backbone module is used to generate a multi-scale latent feature map and output the predicted noise based on the noise latent variables at the current time and the fused latent features. The cross-attention module is used to generate text fusion features that integrate text semantic information based on the multi-scale latent feature map and the text semantic embedding vector. A noise scheduler is used to generate noise scheduling parameters that control the intensity of the denoising process based on a preset number of diffusion time steps. A text generation decoder is used to generate advertising copy based on the text semantic embedding vector and the product visual feature vector; The text-image alignment and control module is used to generate spatial control features based on the main product image data of the product to be promoted, to constrain the generation of the background image to ensure consistency with the product structure. A latent spatial encoder is used to generate a background image based on the denoised latent representation.

[0014] In one optional implementation, the generated prompt words are input into a multimodal generation model constructed based on the AIGC algorithm to generate initial advertising creative containing background images and text content, including: The generated prompt words are input into a multimodal generation model constructed based on the AIGC algorithm; Based on the generated prompt words, the generated text encoder is used to generate a text semantic embedding vector; Based on the product main image data in the product feature data of the promoted product, the generated visual encoder is used to generate a product visual feature vector; Based on the preset potential spatial distribution, the noise scheduler is used to generate initial latent noise variables; Based on the initial noise latent variable, the text semantic embedding vector, and the product main image data of the product to be promoted, the generated backbone module, the cross attention module, and the text-image alignment and control module are used to generate a denoised latent representation through iterative denoising. Based on the denoised latent representation, a background image is generated using the latent spatial encoder; Based on the text semantic embedding vector and the product visual feature vector, the text generation decoder is used to generate advertising copy; The initial advertising material is generated by combining the background image and the advertising copy.

[0015] In one optional implementation, based on the initial noise latent variable, the text semantic embedding vector, and the product main image data of the product to be promoted, the backbone generation module, the cross-attention module, and the text-image alignment and control module are used to iteratively denoise and generate a denoised latent representation, including: Based on the initial noise latent variables and the text semantic embedding vector, semantically enhanced latent features are generated using the generation backbone module and the cross-attention module; Based on the main image data of the product to be promoted, spatial control features are generated using the text-image alignment and control module. Based on the semantically enhanced latent features and the spatial control features, the generated backbone module is used to generate fused latent features; Based on the fused latent features, the backbone generation module is used to generate predictive noise; Based on the predicted noise, the initial noise latent variables, and the noise scheduler parameters, generate the potential representation for the next time step; The denoising steps are executed iteratively until the time step reaches zero, generating the denoised potential representation.

[0016] In one optional implementation, compliance checks and quality assessments are performed on the background image and text content of the initial advertising material, including: Based on the background image of the initial advertising material, an image compliance detection model is used to detect the compliance of the image and generate a violation confidence score. Based on the text content of the initial advertising material, a sensitive word filtering and semantic compliance detection model is used to perform text compliance detection and generate a text violation confidence score. Based on the background image of the initial advertising material, an aesthetic score is generated using a pre-trained aesthetic quality assessment model; Based on the background image and the generated prompt words, a pre-trained image-text consistency evaluation model is used to generate an image-text consistency score. Based on the background image of the initial advertising material, an image sharpness assessment algorithm is used to generate a sharpness score; Based on the image violation confidence level, the text violation confidence level, the aesthetic score, the image-text consistency score, and the sharpness score, a weighted fusion algorithm is used to perform a quality assessment and generate a comprehensive quality score as the quality assessment result.

[0017] In one optional implementation, if the detection passes and the quality assessment result meets preset conditions, then for the background image of the initial advertising material, the initial text box position is determined, and an adaptive adjustment based on visual salience is performed to obtain the repositioned text box position, including: If the image violation confidence level is less than or equal to the preset image security compliance threshold and the text violation confidence level is less than or equal to the preset text security compliance threshold, the detection passes and proceeds to the quality assessment result determination step. If the overall quality score of the quality assessment result is greater than or equal to the quality assessment threshold, then proceed to the layout adaptive adjustment step; Based on the background image of the initial advertising material, a comprehensive saliency map is generated using a pre-trained saliency detection model; Based on the ad layout template, determine the initial text box position and search range; Based on the comprehensive saliency map, the initial text box position, and the search range, a target loss function containing saliency cost and positional constraints is constructed. The target loss function is optimized using a sliding window or gradient descent algorithm to generate the optimal displacement vector; Based on the optimal displacement vector, the initial text box position is adjusted adaptively to obtain the repositioned text box position.

[0018] In one optional implementation, based on the repositioning text box position, the initial advertising material and the product main image data of the product to be promoted are rendered and synthesized into a final personalized advertisement, which is then pushed to the target user, including: Based on the main image data of the product to be promoted, a product mask is extracted using a semantic segmentation model. Based on the product mask, a Poisson fusion of the product and the background is performed to generate a fused base image. Based on the fused base image and the repositioned text box position, the dominant color of the background area is calculated to generate an adaptive text color; A text layer is generated based on the text content of the initial advertising material, the adaptive text color, and the position of the repositioned text box; Based on the fused base image and the text layer, perform Alpha blending to generate the final personalized advertisement; Based on the target user identifier and the final personalized advertisement, a push message package is constructed and sent to the target user.

[0019] Secondly, embodiments of the present invention provide a personalized advertising dynamic generation system based on AIGC, used to implement a personalized advertising dynamic generation method, the system comprising: The data acquisition unit is used to acquire historical interaction data and real-time environmental data of target users, as well as product feature data of the products to be promoted, and to construct multi-dimensional feature vectors to generate comprehensive user feature vectors, environmental context feature vectors, and product feature vectors. The prompt word engineering unit is used to generate structured prompt words based on the comprehensive user feature vector, the environmental context feature vector, and the product feature vector, using a preset prompt word construction strategy. An initial content generation unit is used to input the generated prompt words into a multimodal generation model constructed according to the AIGC algorithm to generate initial advertising materials containing background images and text content; The review unit is used to perform compliance checks and quality assessments on the background images and text content of the initial advertising materials; The intelligent layout unit is used to determine the initial text box position set according to the advertising template for the background image of the initial advertising material when the detection is passed and the quality assessment result meets the preset conditions, and to perform layout adaptive adjustment of the initial text box position according to the visual salience to obtain the repositioned text box position. The synthesis and distribution unit is used to render and synthesize the initial advertising material and the main image data of the product to be promoted into a final personalized advertisement based on the position of the repositioning text box, and push it to the target user.

[0020] A third aspect of this invention provides an electronic device, which includes: At least one processor, and a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by at least one processor, which are executed by at least one processor to enable the at least one processor to perform the method proposed in the first aspect of the embodiments of the present invention.

[0021] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in the first aspect of the present invention. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention; Figure 2 This is a flowchart of the steps of a personalized advertising dynamic generation method based on AIGC provided in an embodiment of the present invention; Figure 3 This is a functional unit diagram of a personalized advertising dynamic generation system based on AIGC provided in an embodiment of the present invention. Detailed Implementation

[0023] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0024] The present invention will be further described below with reference to the accompanying drawings.

[0025] Reference Figure 1 , Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention.

[0026] like Figure 1 As shown, the electronic device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen and an input unit, such as a keyboard. Optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk drive. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0027] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0028] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a data storage module, a network communication module, a user interface module, and an electronic program for a personalized advertising dynamic generation system based on AIGC.

[0029] exist Figure 1In the electronic device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the electronic device of the present invention can be set in the electronic device. The electronic device calls the electronic program of the AIGC-based personalized advertising dynamic generation system stored in the memory 1005 through the processor 1001, and executes the AIGC-based personalized advertising dynamic generation method provided in the embodiment of the present invention.

[0030] Reference Figure 2 The present invention provides a method for dynamically generating personalized advertisements based on AIGC, the method comprising: S201: Obtain historical interaction data and real-time environmental data of the target user, as well as product feature data of the product to be promoted, and construct multi-dimensional feature vectors to generate comprehensive user feature vectors, environmental context feature vectors, and product feature vectors. S202: Based on the comprehensive user feature vector, the environmental context feature vector, and the product feature vector, a structured generation prompt word is generated using a preset prompt word construction strategy; S203: Input the generated prompt words into the multimodal generation model constructed according to the AIGC algorithm to generate initial advertising materials containing background images and text content; S204: Conduct compliance testing and quality assessment on the background images and text content of the initial advertising materials; S205: If the detection passes and the quality assessment result meets the preset conditions, then for the background image of the initial advertising material, determine the initial text box position and perform layout adaptive adjustment based on visual salience to obtain the repositioned text box position; S206: Based on the position of the repositioning text box, render and synthesize the initial advertising material and the main image data of the product to be promoted into a final personalized advertisement, and push it to the target user.

[0031] The technical solution provided in this application has at least the following beneficial effects: By constructing user feature vectors encompassing demographics, behavior, interests, and emotions, and combining them with environmental context and product features, structured prompts are generated. This enables the AIGC model to produce personalized ads that highly match the user's current mood and environment, enhancing conversion potential. A comprehensive saliency map is introduced, not only avoiding high visual focus areas in the generated background but also incorporating the pre-defined product main image placement area as a high-saliency constraint into the map. The target loss function is used to optimize and relocate the text box, completely resolving the pain points of text and product occlusion and unreadable text. In the synthesis stage, semantic segmentation is used to extract a product mask to guide Poisson fusion, eliminating edge halos and jagged edges. After fusion, the text color is adaptively calculated based on the dominant color of the relocated text box area, and finally, alpha blending is performed to ensure high visual harmony in the final ad image. Before synthesis, dual compliance checks (image violation confidence and text violation confidence) are performed on the background image and text, and a weighted evaluation is conducted based on aesthetics, consistency, and clarity. A veto and threshold interception are implemented to ensure the safety and high quality of the ads.

[0032] In one optional implementation, the historical interaction data includes the target user's clickstream records, purchase history, search keywords, and basic attribute information; The real-time environmental data includes the time input data, location input data, and weather input data of the target user's environment; The product feature data includes the product text description and product main image data of the product to be promoted; S2011: Obtain historical interaction data and real-time environmental data of the target users, as well as product feature data of the product to be promoted, and construct multi-dimensional feature vectors to generate comprehensive user feature vectors, environmental context feature vectors, and product feature vectors, including: S2012: Extract target user identifiers, historical interaction data, and real-time environmental data of target users, as well as product feature data of the products to be promoted, from the data management system; S2013: Clean and normalize historical interaction data, and map it into a comprehensive user feature vector of preset dimensions. The comprehensive user feature vector consists of demographic feature sub-vectors, behavioral feature sub-vectors, interest tag sub-vectors, and real-time sentiment state sub-vectors, as shown in the formula: In the formula, To integrate user feature vectors as the basis for determining "style preferences" and "emotional tone" in subsequent generated prompts, the system integrates global features of "who the user is," "what they did," "what they like," and "how they are feeling at the moment." This is a demographic feature vector describing a user's basic social attributes. It's typically a numerical value processed through one-hot encoding or embedding, such as age group (20-29 years old), gender (female), region (first-tier cities), income level, etc. It determines the basic audience fit of the advertising content; for example, high... (Older age) may suppress overly trendy or noisy visual elements in the generated prompts; These are the weighting coefficients for demographic characteristic dimensions; The behavioral feature sub-vector describes the user's historical interaction behavior, including normalized values ​​such as click-through rate, average browsing time, purchase frequency, and active time period preference, reflecting the user's activity level and conversion potential. Highly active users may generate more impactful and dynamic prompts. These are the weighting coefficients for the behavioral feature dimension; Interest tag sub-vectors describe users' long-term interest preferences. The interest distribution score is calculated based on the products the user has browsed and the topics they follow (such as "outdoor sports" and "technology and digital products"), which determines the most critical feature source of the advertising background elements (such as "mountains and rivers", "code", and "food"). These are the weighting coefficients for the interest feature dimension; The real-time emotional state sub-vector describes the user's current real-time emotional state. It is used to adjust the emotional tone of the generated image based on the emotional polarity (e.g., excitement, anxiety, calm) analyzed from the user's recent comments or browsing content. For example, "anxiety" may correspond to a warmer or redder tone, while "calm" corresponds to a cool tone. These are the weighting coefficients for the emotional feature dimension; Weighting coefficients are used to dynamically adjust the proportion of each dimension in the final profile, for example: Scenario Example A (New Product Promotion): To accurately reach the target audience, increase the... (Interest weight) and (Population weight), reduce (Emotional weighting), emphasizing the degree of match; Scenario Example B (Impulsive Purchase): In order to take advantage of the user's current emotions, amplify... (Emotional weight) and (Behavioral weighting), emphasizing atmosphere and sense of urgency; S2014: Maps discrete categorical and continuous variables in real-time environmental data to a unified environmental vector, generating an environmental context feature vector. The formula is as follows: In the formula, This is an environmental context feature vector, which is a comprehensive vector containing location, weather, and time information; Location input data in real-time environmental data is usually discrete geocoding or latitude and longitude coordinates, which determines the regional adaptability of advertising content (e.g., pushing down jackets to northern cities and T-shirts to southern cities). For embedding layer functions, neural network layer operations map discrete data into low-dimensional dense vectors, converting discrete "city IDs" into a set of semantically meaningful numerical values; The weather input data from the real-time environmental data is categorized as labels such as {sunny, cloudy, rain, snow, fog}, which directly affects the lighting and elements of the advertisement. For example, "rain" will trigger prompts to generate visual elements such as "wet ground, reflection, umbrella". For one-hot encoding functions, the operation of converting categorical variables into binary vectors, if... N This kind of weather results in a length of N A vector where only the position corresponding to the category is 1, and the rest are 0. N Total number of weather categories; The time input data in the real-time environment data is the current system timestamp, which is usually a continuous value (such as seconds) or the number of hours in a day (0-23), which determines the time context of the advertisement (such as breakfast, late-night promotion, holiday atmosphere). The sine and cosine coding functions are used to transform the periodic time using sine and cosine functions. After the SinCos transformation, the distance between 11 PM and midnight in the vector space becomes very close, which helps the model understand that "late at night" is a continuous time period. This is a vector concatenation symbol, used to join multiple vectors end-to-end into a single long vector. For example, location input data: discrete city IDs are input into the embedding layer, and the output dimension is typically 64-128 dimensions, with a numerical range of [-1, 1] or [0, 1]. Weather input data: Assuming a preset total number of weather categories N =10 (sunny, rainy, snowy, etc.), after OneHot encoding, the output is a 10-dimensional binary vector with values ​​{0,1}; Time input data: Sine and cosine encoding is used to map the hour hh(0-23) into two-dimensional features. This encoding ensures that 23:00 and 00:00 are close in the vector space. final The dimension is the sum of the above three, for example, 128 + 10 + 2 = 140 dimensions; S2015: Using pre-trained independent visual encoders and independent text encoders, extract and fuse the features of the product feature data of the product to be promoted to generate a product feature vector. The formula is as follows: In the formula, The product feature vector is a multi-dimensional real number array (e.g., 512-dimensional or 1024-dimensional), which includes both the product's "appearance" (visual features, such as color, shape, and style) and the product's "content" (textual features, such as function and brand tone). The main product image data in the product feature data for promoting products is usually a pre-processed RGB image tensor (such as a 3×224×224 pixel matrix) that provides intuitive information about the product's appearance and ensures that the generated advertising background is coordinated with the product image in terms of color and composition. For independent visual encoder functions, based on the feature extraction algorithm model of deep neural networks, such as Residual Network-50 (ResNet-50), Visual Geometry Group (VGG), Vision Transformer (ViT), or the image encoder part of Contrastive Language–Image Pre-training (CLIP), the input... It is transformed into feature vectors that computers can understand, and is responsible for "understanding" images and extracting deep visual features such as texture, edges, and object shapes; The product text description in the product feature data is a text string that typically includes the product name, key selling points, slogan, or attribute tags (such as "lightweight, breathable, running shoes"), providing a functional description of the product and ensuring that the generated advertising copy accurately conveys the product's value. As an independent text encoder function, it transforms the input text based on a natural language processing algorithm model, such as the text encoder part of Bidirectional Encoder Representations from Transformers (BERT), Word2Vec, or CLIP. It is transformed into a feature vector, which is responsible for "understanding" semantics and converting words into a numerical distribution that contains contextual meaning; Visual weighting coefficient; These are text weight coefficients; and The scalar parameters used to balance the importance of visual and textual features typically satisfy... And 0≤ ≤1; Visually-driven products (such as clothing, jewelry, and decorative paintings): Increase the size. Values ​​(e.g.) =0.8, =0.2), which means that when generating ads, more emphasis is placed on the aesthetics of product images, and the prompts will focus on visual styles such as "beautiful" and "fashionable". Function-driven products (such as pharmaceuticals, electronic products, and books): Increase Values ​​(e.g.) =0.3, =0.7), which means that when generating ads, more emphasis is placed on the accuracy of the text description, and the prompts will focus on semantic styles such as "technological, professional, and clear".

[0033] In one optional implementation, based on the comprehensive user feature vector, the environmental context feature vector, and the product feature vector, structured generation prompts are generated using a preset prompt word construction strategy, including: S2021: Calculate the semantic similarity between the integrated user feature vector and the product feature vector in the preset tag library, select the Top-K keywords, and form a keyword set, using the following formula: In the formula, The set of keywords consisting of the Top-K keywords, such as {“basketball shoes”, “dynamic”, “red”, “street”}, serves as the core semantic unit for subsequent prompts. It is the result of mapping back from the vector space to natural language. This set combines the product’s own attributes and the style of user preferences; Top-K is the total number of keywords. The feature vectors of the input preset tag library belong to the product feature vectors. or user feature vector ; A pre-defined tag dictionary containing tags k Pre-trained word embedding vectors It contains various words that the AIGC model can understand, such as object names (car, shoes, flowers), style descriptions (oil painting, minimalism, cyberpunk), and environment descriptions (rainy day, forest, space). For tags k The pre-trained word embedding vectors are used to map natural language words into a high-dimensional vector space using pre-trained models (such as Word2Vec, Global Vectors for Word Representation (GloVe), or CLIP's text encoder). In this space, semantically similar words are close to each other (for example, the distance between "cat" and "dog" is closer than the distance between "cat" and "car"). kCandidate label variables, strings, such as "skyscraper" or "warm lighting"; The cosine similarity function measures the directional similarity between two vectors. It calculates a value between -1 and 1 (or 0 and 1). The closer the value is to 1, the better the similarity between the feature vectors of the input predefined label library and the candidate labels. k The closer their meanings are; To filter the top with the highest similarity K The operation function for each tag directly returns the top similarity rankings. K The tag set filters out noisy words with low matching degree. For example, although "running" and "basketball" are both sports, if it is basketball shoes, only the top 3 words with the highest similarity are taken to avoid the prompt words being too long and causing the generation effect to deteriorate. K To filter hyperparameters, values ​​are typically set between 3 and 5; The union operator, meaning "to combine", is used in the formula. It will traverse { , }, respectively based on product feature vectors or user feature vector Calculate the cosine similarity with the tag dictionary and select the top [values]. K Each tag is used to merge the two batches of keywords into a single keyword set. S2022: Based on the environmental context feature vector and the real-time emotional state sub-vector in the user feature vector, use a pre-trained atmosphere control parameter generation model to obtain the atmosphere control parameter vector of the generated image. The atmosphere control parameter generation model is constructed based on the Conditional Multi-Layer Perceptron (Conditioned MLP) algorithm, and includes: The sentiment branch, consisting of several cascaded fully connected layers and a ReLU activation function, is responsible for processing the real-time sentiment state sub-vector in the user feature vector. The environment branch, consisting of a fully connected mapping layer, is used to map high-dimensional environment context feature vectors to a low-dimensional space consistent with the dimension of the atmosphere parameters, and is responsible for processing the environment context feature vectors. Weighted fusion mechanism: The outputs of the emotion branch and the environment branch are weighted and summed element by element through style weight coefficients, and the final output is an atmosphere control parameter vector; The formula is: In the formula, This is an atmosphere control parameter vector containing values ​​such as {light intensity, color temperature, saturation}, used to insert specific technical modifiers (such as "volume light" or "high contrast") into the generated prompts. The real-time emotional state sub-vector determines the emotional tone of the advertisement. For example, for the emotion of "anxiety", the model may generate a warmer and more reassuring atmosphere, while for the emotion of "excitement", it will generate a high-contrast and high-brightness atmosphere. The weight matrix of the neural network that generates the atmosphere control parameters is used to linearly transform the real-time emotional state sub-vectors from the emotion space to the atmosphere parameter space, defining the mapping relationship of "which atmosphere change corresponds to a certain emotion" (for example, the model may learn that "excitement" corresponds to "high saturation"). The bias term of the neural network that generates the atmosphere control parameters is used to adjust the baseline of the atmosphere parameters so that a basic atmosphere output can be maintained even if the emotion vector is zero. Style weighting coefficients for the sentiment branch; To activate the model, a nonlinear factor is introduced to enhance its expressive power and ensure that the ambient parameters (such as light intensity and saturation) are non-negative, since image parameters cannot be negative. The ReLU function sets all calculation results less than 0 to zero, which is consistent with physical logic. This is a feature vector representing the environmental context. The linear projection matrix for the environment branch is used to project the normalized high-dimensional environment vector onto the same dimensional space as the atmosphere parameter vector, thus solving the dimension mismatch problem. The normalization function is a function that scales a vector to a standard range (usually [0, 1]). When data contains different units of measurement (e.g., time is 0-23, and weather is categorically coded), directly adding the values ​​will cause the data with the larger unit of measurement to dominate the results. After normalization, the influence of weather and time is brought to the same level, ensuring fair integration. The style weighting coefficient for the environment branch; and A hyperparameter that balances the importance of "subjective emotion" and "objective environment," and satisfies... + =1, for example: Scenario A (Emotional Marketing): For impulse-buying products (such as blind boxes and games), increase the intensity of marketing. , reduce This means that we should follow the user's mood; even if it is daytime, if the user is negative, we should generate dark-toned ads. Scenario B (Scenario Marketing): For functional products (such as umbrellas, takeout), increase the intensity of marketing. , reduce This means that it follows the environment, regardless of the user's mood. If the real-time environment is raining, it generates ads that match the rainy environment. S2023: Fill the keyword set and the style modifiers obtained by discretizing and mapping the atmosphere control parameter vector into the preset prompt word template to generate structured prompt words; In this embodiment, since the atmosphere control parameter vector is a continuous numerical feature (such as light intensity 0.8), it cannot be directly used as text input. It is necessary to convert the numerical values ​​in the vector into natural language modifiers through preset mapping rules (such as threshold judgment or lookup table). For example, the value of the light intensity parameter greater than 0.7 is mapped to "bright light" or "volume light", and the value of the saturation parameter greater than 0.8 is mapped to "vibrant color" or "high contrast". After this conversion, the style modifiers obtained by mapping are filled into the prompt word template together with the keyword set to ensure that the generated prompt words can be correctly processed by the text encoder of the multimodal generation model. Mapping rule example: If the light intensity component is greater than 0.7, then the text modifier "bright light" or "volume light" will be generated; If the color temperature component is less than 0.3, the generated text modifiers will be "cool tone" or "tech feel". If the saturation component is greater than 0.8, then the text modifiers "vibrant colors" or "high contrast" will be generated. Finally, fill in the template, for example, to generate prompts such as: "A background image about basketball shoes, with a dynamic style, vibrant colors, and bright light elements."

[0034] In one alternative implementation, the multimodal generative model is constructed according to the AIGC algorithm, including: A text encoder is generated to generate a text semantic embedding vector based on the generated prompt words; Generate a visual encoder, which is used to generate a product visual feature vector based on the product main image data in the product feature data of the product to be promoted; The backbone module is used to generate a multi-scale latent feature map and output the predicted noise based on the noise latent variables at the current time and the fused latent features. The cross-attention module is used to generate text fusion features that integrate text semantic information based on the multi-scale latent feature map and the text semantic embedding vector. A noise scheduler is used to generate noise scheduling parameters that control the intensity of the denoising process based on a preset number of diffusion time steps. A text generation decoder is used to generate advertising copy based on the text semantic embedding vector and the product visual feature vector; The text-image alignment and control module is used to generate spatial control features based on the main product image data of the product to be promoted, to constrain the generation of the background image to ensure consistency with the product structure. A latent spatial encoder is used to generate a background image based on the denoised latent representation; The output of the text encoder is connected to the first input of the cross-attention module and the first input of the text generation decoder, respectively, and is used to input the generated text semantic embedding vector into the cross-attention module as a semantic condition for image generation, and into the text generation decoder as a semantic context for copywriting generation. The output of the visual encoder is connected to the second input of the text generator decoder, and is used to input the generated product visual feature vector into the text generator decoder as visual reference information; the main image data of the product to be promoted is directly input into the text-image alignment and control module as a spatial control signal. The output of the noise scheduler is connected to the first input of the backbone generation module, and is used to provide the backbone generation module with the initial noise latent variable and the noise scheduling parameters corresponding to each time step in the iterative denoising process. The cross-attention module is embedded in or connected in parallel with the generation backbone module. Its second input end is connected to the output end of the generation text encoder. It is used to receive the feature map extracted by the generation backbone module as a query, receive the text semantic embedding vector as a key and value, generate features that fuse text information, and feed them back to the generation backbone module. The text-image alignment and control module is connected to the backbone generation module and is used to inject the spatial control features generated based on the product main image data of the product to be promoted into the corresponding level of the backbone generation module, so as to constrain the background image structure generated by the backbone generation module to be consistent with the product features. The output of the backbone generation module is connected to the input of the latent spatial encoder, and is used to input the denoised latent representation after iterative denoising to the latent spatial encoder to decode and generate a pixel-level background image. The text generation decoder is configured to generate advertising copy that matches the background image based on the received text semantic embedding vector and the product visual feature vector.

[0035] In one optional implementation, the generated prompt words are input into a multimodal generation model constructed based on the AIGC algorithm to generate initial advertising creative containing background images and text content, including: S2031: Input the generated prompt words into the multimodal generation model constructed according to the AIGC algorithm; S2032: Based on the generated prompt words, use the generated text encoder to generate a 768-dimensional text semantic embedding vector; S2033: Based on the product main image data in the product feature data of the promoted product, use the aforementioned visual encoder to generate a 768-dimensional product visual feature vector. S2034: Based on the preset potential spatial distribution, use the noise scheduler to generate initial noise latent variables; In this embodiment, the diffusion time step is 50, and an initial noise latent variable is generated with a shape of 4×64×64 (corresponding to the latent spatial resolution). S2035: Based on the initial noise latent variable, the text semantic embedding vector, and the product main image data of the product to be promoted, the generated backbone module, the cross attention module, and the text-image alignment and control module are used to generate a denoised latent representation through iterative denoising. S2036: Based on the denoised latent representation, use the latent spatial encoder to generate a 3×512×512 pixel-level background image; S2037: Based on the text semantic embedding vector and the product visual feature vector, use the text generation decoder to generate advertising copy, for example, the advertising copy is "On a rainy night street, red combat boots light up the whole scene, taking control at every step!"; S2038: Combine the background image and the advertising copy to generate initial advertising materials.

[0036] In one optional implementation, based on the initial noise latent variable, the text semantic embedding vector, and the product main image data of the product to be promoted, the backbone generation module, the cross-attention module, and the text-image alignment and control module are used to iteratively denoise and generate a denoised latent representation, including: S20351: Based on the initial noise latent variables and the text semantic embedding vector, use the generation backbone module and the cross attention module to generate semantically enhanced latent features; In this embodiment, at each time step of the iterative denoising process, the backbone generation module first performs convolution processing on the input noise latent variables to extract features. Subsequently, the cross-attention module uses the text semantic embedding vector as a condition to calculate the association between image features and text words, injecting semantic information such as "basketball shoes," "rainy night," and "neon" into the feature map to generate potential features with preliminary semantic content. The formula is as follows: In the formula, For potential features to enhance semantics, the feature map already contains information about objects and atmosphere described in the text; For time stept The noise latent variable is the initial noise latent variable in the first iteration; t For time step index; This generates the convolutional layer functions in the backbone module, which are used to extract the spatial structural features of the image; It is a text semantic embedding vector containing style and object descriptions from the prompt words; This is the cross-attention calculation function for the cross-attention module. It uses image features as the query and text features as the key and value, and outputs a feature map that integrates text information. S20352: Based on the product main image data of the product to be promoted, use the text-image alignment and control module to generate spatial control features; In this embodiment, in order to ensure that the generated background can be perfectly integrated with the main product image (such as consistent perspective angle and reasonable lighting logic), the text-image alignment and control module (such as adding conditional control to text-to-image diffusion models (ControlNet) network) directly receives the main product image data of the product to be promoted, and extracts the spatial structure information such as the edge, depth or posture of the product through a preset edge detection or depth estimation algorithm, and converts it into the corresponding control map or spatial control features; For example, directly receive 3×224×224 product main image data, use the Canny edge detection algorithm to extract the shoe outline, and generate a 1×64×64 spatial control feature map; S20353: Based on the semantically enhanced latent features and the spatial control features, the generation backbone module is used to generate fused latent features; In this embodiment, the backbone module receives semantically enhanced potential features and spatial control features, and superimposes the product's structural constraints onto the semantic features through residual connection or feature addition, forming a fusion feature that conforms to the text description and has an accurate spatial structure. S20354: Based on the fused latent features, use the backbone generation module to generate predictive noise; In this embodiment, the backbone generation module performs final upsampling and processing on the fused latent features, and outputs a tensor with the exact same shape as the input noise. This tensor represents the model's prediction of the noise components contained in the current latent variables. S20355: Generate the potential representation for the next time step based on the predicted noise, the initial noise latent variable, and the noise scheduler parameters; In this embodiment, the parameters provided by the noise scheduler are used in conjunction with the initial noise latent variables. Based on the diffusion sampling formula (such as Denoising Diffusion Probabilistic Models (DDPM) or Denoising Diffusion Implicit Models (DDIM)), the latent variables of the previous time step are calculated as the latent representation of the next time step. This step is essentially to remove some of the predicted noise to make the image clearer. S20356: Perform the denoising steps iteratively until the time step is zero, generating the denoised potential representation.

[0037] In one optional implementation, compliance checks and quality assessments are performed on the background image and text content of the initial advertising material, including: S2041: Based on the background image of the initial advertising material, use a security and compliance detection model to perform image compliance detection and generate an image violation confidence score; In this embodiment, a trained binary or multi-label classification security compliance detection model (such as a workplace inappropriate content detector or a violence / pornography recognition model) is used to scan the background image and determine whether it contains sensitive, illegal, or harmful information (such as bloody, nude, or disturbing elements). S2042: Based on the text content of the initial advertising material, use a sensitive word filtering and semantic compliance detection model to perform text compliance detection and generate a text violation confidence score; S2043: Based on the background image of the initial advertising material, generate an aesthetic score using a pre-trained aesthetic quality assessment model; In this embodiment, an aesthetic quality assessment model (such as a neural image assessment model) trained based on human subjective rating data is used to score the composition, color matching, and lighting effects of a background image, and to predict its attractiveness to the public. S2044: Based on the background image and the generated prompt words, use a pre-trained image-text consistency evaluation model to generate an image-text consistency score; In this embodiment, in order to ensure that the generated background image is faithful to the user's intent, the feature vectors of the background image and the generated prompt words are calculated using the image-text consistency evaluation model, and the cosine similarity between the two is calculated as the image-text consistency score. S2045: Based on the background image of the initial advertising material, generate a sharpness score using an image sharpness evaluation algorithm; In this embodiment, the key indicator for evaluating whether the background image is clear and free of blur is to calculate the gradient or variance of the Laplacian operator of the image, thus preventing the generation model from producing blurry ghost images. S2046: Based on the image violation confidence level, the text violation confidence level, the aesthetic score, the image-text consistency score, and the sharpness score, a weighted fusion algorithm is used to perform a quality assessment, generating a comprehensive quality score as the quality assessment result. The formula is: In the formula, The final indicator used for comprehensive quality scoring and quantitative evaluation of the overall quality of advertising materials is usually [0,1] or [0, 100]. A value of 0 indicates that the material is non-compliant or of extremely poor quality, while the closer the value is to 1 (or 100), the higher the quality of the material and the more it meets the publication standards. Image violation confidence level; The preset image security compliance threshold; Confidence level for text violation; A preset text security compliance threshold is set to prevent the generation of text containing false advertising or sensitive words; Rate the aesthetics; The score represents the consistency between the text and the image. Rate the clarity; Weighting coefficients for aesthetic scoring; This is the weighting coefficient for the consistency score between text and images; Weighting coefficients for clarity scoring; As a normalization function, it maps data of different dimensions to a mathematical transformation function with a unified scale, eliminates the influence of dimensions, and prevents indicators with a large value range (such as sharpness variance) from masking indicators with a small value range (such as aesthetic scores of 1-10).

[0038] In one optional implementation, if the detection passes and the quality assessment result meets preset conditions, then for the background image of the initial advertising material, the initial text box position is determined, and an adaptive adjustment based on visual salience is performed to obtain the repositioned text box position, including: S2051: If the image violation confidence level is less than or equal to the preset image security compliance threshold and the text violation confidence level is less than or equal to the preset text security compliance threshold, the detection passes and proceeds to the quality assessment result determination step. S2052: If the overall quality score of the quality assessment result is greater than or equal to the quality assessment threshold, then proceed to the layout adaptive adjustment step; S2053: Based on the background image of the initial advertising material, a comprehensive saliency map is generated using a pre-trained saliency detection model; In this embodiment, a pre-trained saliency detection model (such as a depth-gazing model, a boundary-aware salient target detection model, or a saliency detection model based on spectral residuals) is used to analyze the background image and predict the region that the human eye is most interested in (heatmap), i.e., the visual saliency map. High saliency regions usually include high-contrast regions, complex texture regions, or salient elements such as generated faces in the background. Text should avoid these regions as much as possible to ensure readability and visual balance (Note: At this time, the product main image has not yet been synthesized, and saliency detection is only for the generated pure background image). At the same time, in order to prevent the text from being obscured by the product main image during subsequent synthesis, or the text obscuring the product main image, the preset product main image placement area is regarded as a "high saliency" region that must be avoided. The visual saliency maps are merged to generate a comprehensive saliency map to ensure that the text box is not positioned in the product's preset placement area. S2054: Determine the initial text box position and search range based on the advertisement layout template; In this embodiment, the initial coordinates and size of the text defined in the advertisement template are read. At the same time, a maximum allowable search radius or offset range is defined to limit the adjustment range of the text box and prevent it from deviating too far from the design or moving out of the screen. S2055: Based on the comprehensive saliency map and the initial text box position and search range, construct a target loss function that includes saliency cost and positional constraints; In this embodiment, to find the optimal position, an energy function is constructed. This function consists of two parts: first, the saliency cost, which is the sum of the saliency of the text box's covered area (the smaller the better, indicating less occlusion); and second, the position constraint cost, which is the distance between the current position and the initial position (the smaller the better, indicating that the original design intention is maintained). By weighted summation, the requirements of "avoiding the main subject" and "maintaining the layout" are balanced. The formula is as follows: In the formula, The objective loss function is a scalar value used to evaluate the movement of the text box. The overall advantages and disadvantages are determined by the value; a smaller value indicates a better position. To position the text box horizontally ( x Displacement distance (in pixels) on the axis. For the text box in the vertical direction ( y Displacement distance (in pixels) on the axis. For the comprehensive saliency plot on the coordinates The image combines the visual saliency of the background with the high saliency constraint of the product placement area. The coordinates of the comprehensive significance plot; The significance weighting coefficient is used to control the importance of the objective of "avoiding visual subjects". It is usually set to a high value (such as 0.8) to ensure that products or faces are avoided first. This is a positional weighting coefficient used to control the importance of the goal of "maintaining layout design". It is usually set to a low value (such as 0.2) as a regularization term to prevent text boxes from moving to the edge of the image or completely deviating from the design aesthetic in order to find areas of extremely low salience. The candidate text box area is a set of pixels representing the specific coverage area after the text box is moved, starting from the initial text box position. Add displacement and the original width and height of the text box Joint decision; This is the initial text box position value in the horizontal direction (x-axis). The initial text box position is in the vertical direction ( y Values ​​on the axis; This is the original width of the text box; The original height of the text box; The pixel coordinates of the candidate text box area; The horizontal coordinate (x-axis) of the image pixels represents the position of the pixel in the candidate text box region along the width of the image. The vertical coordinate (ordinate) of the image pixels represents the position of the pixel in the candidate text box region along the image height direction. S2056: Use a sliding window or gradient descent algorithm to optimize the target loss function and generate the optimal displacement vector; S2057: Based on the optimal displacement vector, perform layout adaptive adjustment on the initial text box position to obtain the repositioned text box position.

[0039] In one optional implementation, based on the repositioning text box position, the initial advertising material and the product main image data of the product to be promoted are rendered and synthesized into a final personalized advertisement, which is then pushed to the target user, including: S2061: Based on the main image data of the product to be promoted, extract the product mask using a semantic segmentation model, and perform Poisson fusion of the product and the background based on the product mask to generate a fused base image; In this embodiment, to ensure the coordination of product lighting and color tone with the background, the Poisson fusion algorithm is first used to seamlessly fuse the product main image into the generated background image. Since the Poisson fusion algorithm relies on an accurate foreground mask to determine the fusion region, and the input product main image data is usually an RGB image, a pre-trained semantic segmentation model (such as the Segment Anything Model (SAM)) is first used to analyze the product main image data and extract an accurate product main mask. Then, based on the mask, the fusion region is determined, the Poisson fusion operation is performed, and a fused base image with a unified visual effect is generated to avoid halo or jagged edges on the product. S2062: Based on the fused base image and the repositioned text box position, calculate the dominant color of the background area and generate an adaptive text color; In this embodiment, in order to ensure that the text is sufficiently readable on the fused image, the average pixel color (or average brightness) of the corresponding area of ​​the repositioned text box in the fused base image is extracted, and the optimal text rendering color is calculated according to the contrast principle (such as dark background with light text, and light background with dark text). S2063: Generate a text layer based on the text content of the initial advertising material, the adaptive text color, and the position of the repositioned text box; In this embodiment, the text content is rasterized into an image layer (usually with an alpha transparency channel) based on the text box coordinates, font size and adaptive color determined by the repositioned text box position, using a font rendering engine. S2064: Based on the fused base image and the text layer, perform Alpha blending to generate the final personalized advertisement; S2065: Construct a push message package based on the target user identifier and the final personalized advertisement, and send the push message package to the target user; In this embodiment, the final advertisement image is bound to the target user's identifier to generate a push request, and the advertisement is sent to the user's mobile device or information stream through a message queue or advertisement delivery interface.

[0040] This invention also provides a personalized advertising dynamic generation system based on AIGC, referring to... Figure 3 The diagram shows a functional unit diagram of a personalized advertising dynamic generation system 300 based on AIGC according to the present invention. The system may include the following units: The data acquisition unit 301 is used to acquire the target user's historical interaction data and real-time environmental data, as well as the product feature data of the product to be promoted, and to construct multi-dimensional feature vectors to generate comprehensive user feature vectors, environmental context feature vectors, and product feature vectors. The prompt word engineering unit 302 is used to generate structured prompt words based on the comprehensive user feature vector, the environmental context feature vector, and the product feature vector, through a preset prompt word construction strategy. The initial content generation unit 303 is used to input the generated prompt words into a multimodal generation model constructed according to the AIGC algorithm to generate initial advertising materials containing background images and text content; The review unit 304 is used to perform compliance testing and quality assessment on the background images and text content of the initial advertising materials; The intelligent layout unit 305 is used to determine the initial text box position set according to the advertising template for the background image of the initial advertising material when the detection is passed and the quality assessment result meets the preset conditions, and to perform layout adaptive adjustment of the initial text box position according to the visual salience to obtain the repositioned text box position. The synthesis and distribution unit 306 is used to render and synthesize the initial advertising material and the main image data of the product to be promoted into a final personalized advertisement based on the position of the repositioning text box, and push it to the target user.

[0041] According to the same inventive concept, another embodiment of the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; The processor, when executing a program stored in memory, implements the AIGC-based personalized advertising dynamic generation method of the present invention.

[0042] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus. The communication interface is used for communication between the aforementioned terminal and other devices. The memory can include Random Access Memory (RAM) or Non-Volatile Memory (NVM), such as at least one disk storage device. Optionally, the memory can also be at least one storage device located remotely from the aforementioned processor.

[0043] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0044] Furthermore, to achieve the above objectives, embodiments of the present invention also propose a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the AIGC-based personalized advertising dynamic generation method of the present invention.

[0045] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products implemented on one or more computer-usable hardware devices (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0046] The embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (apparatus), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0047] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0048] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0049] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. "And / or" indicates that either one or both can be chosen. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes the element.

[0050] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for dynamically generating personalized advertisements based on AIGC, characterized in that, The method includes: Acquire historical interaction data and real-time environmental data of target users, as well as product feature data of the product to be promoted, and construct multi-dimensional feature vectors to generate comprehensive user feature vectors, environmental context feature vectors, and product feature vectors. Based on the comprehensive user feature vector, the environmental context feature vector, and the product feature vector, a structured generation prompt word is generated using a preset prompt word construction strategy. The generated prompts are input into a multimodal generation model constructed based on the AIGC algorithm to generate initial advertising materials containing background images and text content; The background images and text content of the initial advertising materials were subjected to compliance testing and quality assessment. If the detection passes and the quality assessment results meet the preset conditions, then for the background image of the initial advertising material, the initial text box position is determined, and an adaptive adjustment of the layout based on visual salience is performed to obtain the repositioned text box position; Based on the position of the repositioning text box, the initial advertising material and the main product image data of the product to be promoted are rendered and synthesized into a final personalized advertisement, which is then pushed to the target user.

2. The method for dynamically generating personalized advertisements based on AIGC according to claim 1, characterized in that, The historical interaction data includes the target user's clickstream records, purchase history, search keywords, and basic attribute information; The real-time environmental data includes the time input data, location input data, and weather input data of the target user's environment; The product feature data includes the product text description and product main image data of the product to be promoted; The comprehensive user feature vector consists of demographic feature sub-vectors, behavioral feature sub-vectors, interest tag sub-vectors, and real-time emotional state sub-vectors.

3. The method for dynamically generating personalized advertisements based on AIGC according to claim 2, characterized in that, Based on the comprehensive user feature vector, the environmental context feature vector, and the product feature vector, structured generation prompts are generated using a preset prompt word construction strategy, including: Calculate the semantic similarity between the integrated user feature vector and the product feature vector in a preset tag library, select the Top-K keywords, and form a keyword set; Based on the environmental context feature vector and the real-time emotional state sub-vector in the user feature vector, an atmosphere control parameter generation model is used to obtain the atmosphere control parameter vector of the generated image. The keyword set and the style modifiers obtained by discretizing and mapping the atmosphere control parameter vector are filled into a preset prompt word template to generate structured prompt words.

4. The method for dynamically generating personalized advertisements based on AIGC according to claim 3, characterized in that, The multimodal generation model is constructed based on the AIGC algorithm and includes: A text encoder is generated to generate a text semantic embedding vector based on the generated prompt words; Generate a visual encoder, which is used to generate a product visual feature vector based on the product main image data in the product feature data of the product to be promoted; The backbone module is used to generate a multi-scale latent feature map and output the predicted noise based on the noise latent variables at the current time and the fused latent features. The cross-attention module is used to generate text fusion features that integrate text semantic information based on the multi-scale latent feature map and the text semantic embedding vector. A noise scheduler is used to generate noise scheduling parameters that control the intensity of the denoising process based on a preset number of diffusion time steps. A text generation decoder is used to generate advertising copy based on the text semantic embedding vector and the product visual feature vector; The text-image alignment and control module is used to generate spatial control features based on the main product image data of the product to be promoted, to constrain the generation of the background image to ensure consistency with the product structure. A latent spatial encoder is used to generate a background image based on the denoised latent representation.

5. The method for dynamically generating personalized advertisements based on AIGC according to claim 4, characterized in that, The generated prompts are input into a multimodal generation model constructed based on the AIGC algorithm to generate initial advertising creatives containing background images and text content, including: The generated prompt words are input into a multimodal generation model constructed based on the AIGC algorithm; Based on the generated prompt words, the generated text encoder is used to generate a text semantic embedding vector; Based on the product main image data in the product feature data of the promoted product, the generated visual encoder is used to generate a product visual feature vector; Based on the preset potential spatial distribution, the noise scheduler is used to generate initial latent noise variables; Based on the initial noise latent variable, the text semantic embedding vector, and the product main image data of the product to be promoted, the generated backbone module, the cross attention module, and the text-image alignment and control module are used to generate a denoised latent representation through iterative denoising. Based on the denoised latent representation, a background image is generated using the latent spatial encoder; Based on the text semantic embedding vector and the product visual feature vector, the text generation decoder is used to generate advertising copy; The initial advertising material is generated by combining the background image and the advertising copy.

6. The method for dynamically generating personalized advertisements based on AIGC according to claim 5, characterized in that, Based on the initial noise latent variables, the text semantic embedding vector, and the product main image data of the product to be promoted, the backbone generation module, the cross-attention module, and the text-image alignment and control module are used to generate a denoised latent representation through iterative denoising, including: Based on the initial noise latent variables and the text semantic embedding vector, semantically enhanced latent features are generated using the generation backbone module and the cross-attention module; Based on the main image data of the product to be promoted, spatial control features are generated using the text-image alignment and control module. Based on the semantically enhanced latent features and the spatial control features, the generated backbone module is used to generate fused latent features; Based on the fused latent features, the backbone generation module is used to generate predictive noise; Based on the predicted noise, the initial noise latent variables, and the noise scheduler parameters, generate the potential representation for the next time step; The denoising steps are executed iteratively until the time step reaches zero, generating the denoised potential representation.

7. The method for dynamically generating personalized advertisements based on AIGC according to claim 6, characterized in that, The background images and text content of the initial advertising materials undergo compliance testing and quality assessment, including: Based on the background image of the initial advertising material, an image compliance detection model is used to detect image compliance and generate an image violation confidence score. Based on the text content of the initial advertising material, a sensitive word filtering and semantic compliance detection model is used to perform text compliance detection and generate a text violation confidence score. Based on the background image of the initial advertising material, an aesthetic score is generated using a pre-trained aesthetic quality assessment model; Based on the background image and the generated prompt words, a pre-trained image-text consistency evaluation model is used to generate an image-text consistency score. Based on the background image of the initial advertising material, an image sharpness assessment algorithm is used to generate a sharpness score; Based on the image violation confidence level, the text violation confidence level, the aesthetic score, the image-text consistency score, and the sharpness score, a weighted fusion algorithm is used to perform a quality assessment and generate a comprehensive quality score as the quality assessment result.

8. The method for dynamically generating personalized advertisements based on AIGC according to claim 7, characterized in that, If the detection passes and the quality assessment result meets the preset conditions, then for the background image of the initial advertising material, the initial text box position is determined, and an adaptive adjustment based on visual salience is performed to obtain the repositioned text box position, including: If the image violation confidence level is less than or equal to the preset image security compliance threshold and the text violation confidence level is less than or equal to the preset text security compliance threshold, the detection passes and proceeds to the quality assessment result determination step. If the overall quality score of the quality assessment result is greater than or equal to the quality assessment threshold, then proceed to the layout adaptive adjustment step; Based on the background image of the initial advertising material, a comprehensive saliency map is generated using a pre-trained saliency detection model; Based on the ad layout template, determine the initial text box position and search range; Based on the comprehensive saliency map, the initial text box position, and the search range, a target loss function containing saliency cost and positional constraints is constructed. The target loss function is optimized using a sliding window or gradient descent algorithm to generate the optimal displacement vector; Based on the optimal displacement vector, the initial text box position is adjusted adaptively to obtain the repositioned text box position.

9. The method for dynamically generating personalized advertisements based on AIGC according to claim 8, characterized in that, Based on the repositioning text box position, the initial advertising material and the product main image data of the product to be promoted are rendered and synthesized into a final personalized advertisement, which is then pushed to the target user, including: Based on the main image data of the product to be promoted, a product mask is extracted using a semantic segmentation model. Based on the product mask, a Poisson fusion of the product and the background is performed to generate a fused base image. Based on the fused base image and the repositioned text box position, the dominant color of the background area is calculated to generate an adaptive text color; A text layer is generated based on the text content of the initial advertising material, the adaptive text color, and the position of the repositioned text box; Based on the fused base image and the text layer, perform Alpha blending to generate the final personalized advertisement; Based on the target user identifier and the final personalized advertisement, a push message package is constructed and sent to the target user.

10. A personalized advertising dynamic generation system based on AIGC, used to implement the personalized advertising dynamic generation method as described in any one of claims 1-9, characterized in that, The system includes: The data acquisition unit is used to acquire historical interaction data and real-time environmental data of target users, as well as product feature data of the products to be promoted, and to construct multi-dimensional feature vectors to generate comprehensive user feature vectors, environmental context feature vectors, and product feature vectors. The prompt word engineering unit is used to generate structured prompt words based on the comprehensive user feature vector, the environmental context feature vector, and the product feature vector, using a preset prompt word construction strategy. An initial content generation unit is used to input the generated prompt words into a multimodal generation model constructed according to the AIGC algorithm to generate initial advertising materials containing background images and text content; The review unit is used to perform compliance checks and quality assessments on the background images and text content of the initial advertising materials; The intelligent layout unit is used to determine the initial text box position set according to the advertising template for the background image of the initial advertising material when the detection is passed and the quality assessment result meets the preset conditions, and to perform layout adaptive adjustment of the initial text box position according to the visual salience to obtain the repositioned text box position. The synthesis and distribution unit is used to render and synthesize the initial advertising material and the main image data of the product to be promoted into a final personalized advertisement based on the position of the repositioning text box, and push it to the target user.

Images