Personalized image content generation and optimization method and system with fusion generative AI

Abstract

The application provides a personalized image content generation and optimization method and system based on fusion generative AI, relates to the technical field of AI, and comprises the following steps: analyzing the historical interaction track and current intention expression of a target user based on a deep perception network, and constructing a multi-dimensional behavior portrait; identifying a user group with similar generation preferences by using group feature recursive quantization technology, and constructing a group wisdom feature map; performing multi-level mapping analysis on the user features and the group wisdom feature map, anchoring the group affiliation of the target user and visualizing the guide features; constructing a multi-level guide vector according to the current intention expression and the guide features, accurately mapping the hidden space representation of an image generation model, quantifying the deviation in real time, and triggering intelligent calibration; and reconstructing the group distribution structure based on the evaluation data of the target user, and realizing dynamic evolution of the features. The application can accurately grasp the personalized needs of users, and improve the accuracy of image generation and user satisfaction.

Description

Technical Field

[0001] This invention relates to AI technology, and more particularly to a method and system for generating and optimizing personalized image content by incorporating generative AI. Background Technology

[0002] With the rapid development of artificial intelligence technology, generative AI has become an important tool in the field of image content creation. Deep learning architectures such as diffusion models and generative adversarial networks enable machines to generate high-quality image content based on text descriptions or other inputs. However, existing image generation technologies are mainly geared towards general scenarios and lack the ability to deeply understand and adapt to the characteristics and preferences of individual users. When using image generation services, users often need to go through multiple rounds of adjustments to obtain the expected results, which not only reduces creation efficiency but also affects user experience.

[0003] To enhance the personalization of image generation, researchers have begun exploring technical approaches that combine user behavior analysis with generative models. By collecting users' historical interaction data, including input prompts, selected style parameters, and feedback on the generated results, it's possible to infer users' creative preferences to some extent. Simultaneously, recommendation algorithms such as collaborative filtering have been introduced into generative AI systems, attempting to leverage the experience of user groups with similar preferences to optimize the generation results for individual users. However, these exploratory works are still in their early stages, facing numerous limitations in terms of technical implementation and application effectiveness.

[0004] Existing methods provide a rather coarse characterization of user features, typically based on simple tag classification or single-dimensional preference records. This fails to fully capture the complex intentions and diverse needs of users in different creative scenarios, resulting in insufficient accuracy and completeness in personalized modeling and making it difficult to provide effective support for subsequent generation optimization. Summary of the Invention

[0005] This invention provides a method and system for generating and optimizing personalized image content by incorporating generative AI, which can solve the problems in the prior art.

[0006] A first aspect of the present invention provides a method for generating and optimizing personalized image content by incorporating generative AI, comprising:

[0007] Based on deep perception network analysis of target users’ historical interaction trajectory and current intent expression, a multi-dimensional behavioral profile is constructed simultaneously. Based on the multi-dimensional behavioral profile, a group feature recursive quantization technique is used to identify user groups with similar generation preferences, and the distribution structure of generation guidance features of each group in high-dimensional space is depicted to construct a group intelligence feature map.

[0008] The user features contained in the historical interaction trajectory are mapped and analyzed in multiple levels with the wisdom of the crowd feature map, the group affiliation of the target user is anchored, and the guiding features corresponding to the affiliation group are visualized based on the deep feature perspective technology;

[0009] According to the current intention expression and the guiding features, a multi-level guiding vector is constructed and the hidden space representation of the image generation model is accurately mapped, the deviation degree of the generated result and the guiding vector is quantified in real time, and when the deviation exceeds the preset deviation threshold, the vector intelligent calibration is triggered, and the generated result meeting the double constraints is output;

[0010] The generated result is associated and projected with the wisdom of the crowd feature map, the distribution structure of the corresponding group is reconstructed based on the evaluation data of the target user, and the dynamic evolution of the features is realized.

[0011] Based on the multi-dimensional behavior portrait, the user groups with similar generation preferences are identified using group feature recursive quantization technology, and the distribution structure of the generation guiding features of each group in high-dimensional space is described, and the wisdom of the crowd feature map is constructed, including:

[0012] Based on the multi-dimensional behavior portrait, the user groups with similar generation preferences are identified using group feature recursive quantization technology, the evolution chain of behavior trajectory is described, the resonance relationship and entropy distribution between trajectories are captured, the periodic commonality and turning point of user behavior mode in space-time dimension are dynamically analyzed, and the boundary profile and internal cohesion of user groups are determined according to the natural community formed by the resonance relationship in feature space;

[0013] Based on the boundary profile and the internal cohesion, the feature gravity and behavior spectrum of each user group in the multi-dimensional behavior portrait space are described, the dominant tendency and energy distribution proportion of each group in the generation guiding dimension are deconstructed, the high-dimensional distribution structure of the user group generation guiding features is constructed, and the high-dimensional distribution structure is mapped as the wisdom of the crowd feature map.

[0014] Based on the multi-dimensional behavior portrait, the user groups with similar generation preferences are identified using group feature recursive quantization technology, the evolution chain of behavior trajectory is described, the resonance relationship and entropy distribution between trajectories are captured, including:

[0015] Based on the multi-dimensional behavior portrait, the user groups with similar generation preferences are identified using group feature recursive quantization technology, the recursive graph matrix of high-dimensional phase orbit is constructed, the recurrence mode and self-similar structure of the orbit in phase space are quantified through the recursive graph matrix, and the evolution chain and bifurcation features of user behavior trajectory are described according to the recurrence mode and the self-similar structure;

[0016] Based on the evolution chain and the bifurcation feature, the diagonal structure alignment and the texture similarity between different user recursive graph matrices are calculated, and the resonance relationship between trajectories is captured through the combination analysis of the diagonal structure alignment and the texture similarity;

[0017] For the user group with significant resonance relationship, the Shannon entropy and the permutation entropy of the recursive graph matrix are calculated respectively, and the entropy value distribution of the group is obtained by mapping the Shannon entropy and the permutation entropy in the feature space.

[0018] The user features contained in the historical interaction trajectory are mapped and analyzed at multiple levels with the group wisdom feature map, the group affiliation of the target user is anchored, and the guiding features corresponding to the affiliation group are visualized based on the deep feature perspective technology, including:

[0019] The user features contained in the historical interaction trajectory are mapped and analyzed at multiple levels with the group wisdom feature map, the group affiliation of the target user is anchored, and the guiding features corresponding to the affiliation group are visualized based on the deep feature perspective technology, including:

[0020] According to the group affiliation, the corresponding node in the group wisdom feature map is located, the distribution parameters, evolution chain features and entropy labels contained in the corresponding node are deconstructed, and the abstract guiding feature representation of the affiliation group is established by generating the deep perspective of the guiding feature space.

[0021] Based on the abstract guiding feature representation, the group common generating preference is visualized as guiding parameters in the operation dimension, the dominant path in the evolution chain feature is deconstructed to form the visualized expression of the guiding feature.

[0022] By constructing the multi-order projection link of the feature space, the semantic space and the topological space, the affinity relationship between the user feature vector and each group node is quantified, and the affiliation probability distribution is constructed based on the affinity relationship.

[0023] By constructing the multi-order projection link of the feature space, the semantic space and the topological space, the affinity resonance degree of the user feature vector and the distribution parameters of each group node in the feature space is described to obtain the feature affinity, the semantic representation of the user feature vector and the cognitive synergy strength between the dominant guiding features of each group node in the semantic space are described to obtain the semantic affinity, and the coupling depth of the manifold distance of the user feature vector in the group wisdom feature map structure and the topological centrality of each group node in the topological space is constructed to obtain the topological affinity.

[0024] The entropy value label based on each group node reveals the expression tendency of the user feature vector at different cognitive levels, and the synergy mechanism of the three affinities of features, semantics and topology is constructed through the internal consistency of the entropy value label. The contribution proportion of each spatial affinity on different cognitive dimensions is guided according to the synergy mechanism, and a complete affinity relationship is obtained.

[0025] The affinity relationship is condensed into a belonging probability distribution in a probability space, so that adjacent groups form a potential energy difference in the belonging determination process, eliminate the fuzzy area of the group boundary, and obtain a belonging probability distribution with clear dominant direction.

[0026] A multi-level guide vector is constructed, and the hidden space representation of the image generation model is accurately mapped. The deviation degree of the generation result and the guide vector is quantified in real time, and the vector intelligent calibration is triggered when the deviation breaks through the preset deviation threshold, and the generation result meeting the double constraints is output.

[0027] A multi-level guide vector is constructed, and a deep resonance relationship is constructed based on the multi-level guide vector and the hidden space representation of the image generation model, and the image generation model is driven to form an intermediate generation result.

[0028] The hidden layer feature representation is condensed from the intermediate generation result, and the hidden layer feature representation is projected to the multi-dimensional creative space constructed by the multi-level guide vector to depict the image projection of each level.

[0029] According to the creative deviation degree of the intermediate generation result and the multi-level guide vector, the hierarchical resonance strength is constructed based on the dependency association between levels, and the overall deviation degree is obtained by tuning the hierarchical resonance strength. When the overall deviation degree breaks through the preset deviation threshold, the vector calibration mechanism is activated.

[0030] The vector calibration mechanism is executed, the dominant deviation level is located, and the corresponding guide vector is harmonized. The stability index is used to constrain the harmonization amplitude, the multi-level guide vector is reconstructed, a new resonance relationship between the reconstructed multi-level guide vector and the hidden space representation is established, and the deviation perception and calibration process are repeatedly executed until the overall deviation degree reaches the expected harmonious state. The generation result meeting the double constraints is output.

[0031] The hierarchical resonance strength is constructed based on the dependency association between levels, and the overall deviation degree is obtained by tuning the hierarchical resonance strength. When the overall deviation degree breaks through the preset deviation threshold, the vector calibration mechanism is activated, including:

[0032] A hierarchical resonance strength is constructed based on the dependent association between the hierarchical feature vectors, an active degree of information transmission is revealed through an information flow entropy ratio between the hierarchical features based on the hierarchical resonance strength, the active degree reflects a transmission depth and a diffusion law of the features between the hierarchies, a guiding influence between the hierarchies is depicted based on the transmission depth and the diffusion law, and a dynamic evolution chain of the hierarchical features is formed according to the guiding influence;

[0033] The hierarchical resonance strength is harmonically deconstructed based on the dynamic evolution chain, a stability of feature transmission is described by capturing a dominant frequency of resonance fluctuation, a deviation point in the creation process is identified based on the dominant frequency, and a vector calibration mechanism is activated when an overall deviation degree caused by the deviation point breaks through a preset deviation threshold.

[0034] In a second aspect, the embodiment of the present application provides a personalized image content generation and optimization system fused with generative AI, comprising:

[0035] A first unit is configured to analyze a historical interaction trajectory and a current intention expression of a target user based on a deep perception network, and synchronously construct a multi-dimensional behavior portrait; identify a user group with similar generation preferences based on the multi-dimensional behavior portrait using a group feature recursive quantization technology, and describe a distribution structure of generation guiding features of each group in a high-dimensional space to construct a group wisdom feature map;

[0036] A second unit is configured to perform multi-level mapping analysis on user features contained in the historical interaction trajectory and the group wisdom feature map, anchor a group belonging relationship of the target user, and visualize guiding features corresponding to the belonging group based on a deep feature perspective technology;

[0037] A third unit is configured to construct a multi-level guiding vector based on the current intention expression and the guiding features, and accurately map a hidden space representation of an image generation model, real-time quantify a deviation degree of a generation result and the guiding vector, trigger vector intelligent calibration when the deviation breaks through a preset deviation threshold, and output a generation result meeting double constraints;

[0038] A fourth unit is configured to associate and project the generation result with the group wisdom feature map, reconstruct a distribution structure of a corresponding group based on evaluation data of the target user, and realize dynamic evolution of features.

[0039] In a third aspect, the embodiment of the present application provides an electronic device, comprising:

[0040] A processor;

[0041] A memory for storing processor-executable instructions;

[0042] The processor is configured to invoke instructions stored in the memory to perform the method described above.

[0043] In a fourth aspect, the application provides a computer readable storage medium having stored thereon computer program instructions which, when executed by a processor, implement the method described above.

[0044] The beneficial effects of the present application are as follows:

[0045] The present application can comprehensively and accurately understand the personalized needs and preference characteristics of users by analyzing the historical interaction trajectory and current intention expression of the target user through the deep perception network and synchronously constructing a multi-dimensional behavior portrait, and can provide accurate user feature basis for subsequent image content generation, thereby effectively improving the pertinence and accuracy of personalized image generation.

[0046] The present application can fully utilize the collective experience and knowledge of user groups with similar preferences by constructing a group wisdom feature map and performing multi-level mapping analysis, combining individual user features with group wisdom, and simultaneously performing precise mapping and real-time deviation quantification on the hidden space representation of the image generation model through multi-level guidance vectors, triggering an intelligent calibration mechanism when the deviation exceeds a preset threshold, ensuring that the generated results meet both the current intention of the user and the group preference constraints, and significantly improving the quality of image generation and user satisfaction.

[0047] The present application establishes an association projection mechanism between the generated results and the group wisdom feature map, can dynamically reconstruct the distribution structure of the corresponding group based on the evaluation data of the target user, realize the continuous evolution and optimization of features, and make the system have self-learning and iterative improvement capability, which can continuously optimize the generation effect with user feedback, and maintain the long-term effectiveness and adaptability of the technical solution. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 The figure is a flowchart of the personalized image content generation and optimization method of the present application. DETAILED DESCRIPTION

[0049] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0050] The technical solutions of the present application will be described in detail below with specific examples. The following specific examples can be combined with each other, and the same or similar concepts or processes may not be described in detail in some examples.

[0051] Figure 1 The flowchart of the personalized image content generation and optimization method of the fusion generative AI embodiment of the present application is shown in Figure 1 The method comprises:

[0052] Based on the historical interaction trajectory and the current intention expression of the target user, a multi-dimensional behavior portrait is constructed in synchronization. Based on the multi-dimensional behavior portrait, a user group with similar generation preferences is identified using group feature recursive quantization technology, and the distribution structure of the generation guide features of each group in the high-dimensional space is depicted to construct a group wisdom feature map.

[0053] The user features contained in the historical interaction trajectory are mapped and analyzed at multiple levels with the group wisdom feature map to anchor the group affiliation of the target user, and the guide features corresponding to the attribution group are concretized based on deep feature perspective technology.

[0054] According to the current intention expression and the guide features, a multi-level guide vector is constructed and the hidden space representation of the image generation model is precisely mapped, the deviation degree of the generation result and the guide vector is quantified in real time, and when the deviation exceeds the preset deviation threshold, the vector intelligent calibration is triggered, and the generation result that meets the double constraints is output.

[0055] The generation result is associated and projected with the group wisdom feature map, the distribution structure of the corresponding group is reconstructed based on the evaluation data of the target user, and the dynamic evolution of the features is realized.

[0056] In an optional implementation, based on the multi-dimensional behavior portrait, a user group with similar generation preferences is identified using group feature recursive quantization technology, and the distribution structure of the generation guide features of each group in the high-dimensional space is depicted to construct a group wisdom feature map, which comprises:

[0057] Based on the multi-dimensional behavior portrait, a user group with similar generation preferences is identified using group feature recursive quantization technology, the evolution chain of the behavior trajectory is depicted, the resonance relationship and entropy distribution between the trajectories are captured, the periodic commonality and turning point of the user behavior pattern in the time and space dimensions are dynamically analyzed, and the boundary profile and internal cohesion of the user group are determined according to the natural community formed by the resonance relationship in the feature space.

[0058] Based on the boundary profile and the internal cohesion, the feature barycenter and behavior spectrum of each user group in the multi-dimensional behavior portrait space are depicted, the dominant tendency and energy distribution proportion of each group in the generated guide dimension are deconstructed, a high-dimensional distribution structure of the user group generated guide features is constructed, and the high-dimensional distribution structure is mapped as a group wisdom feature map.

[0059] The process of identifying similar generation preference user groups based on the multi-dimensional behavior portrait group feature recursive quantization technology includes multiple detailed technical steps. The system first structures the multi-dimensional behavior portrait data of the user into a time sequence feature matrix, with the rows representing time steps and the columns representing behavior feature dimensions. The time step is set as 1 hour as the basic unit, and each user retains the behavior data of the last 30 days to form a 720x128 feature matrix. The behavior feature dimensions include image generation request frequency, modification operation times, style switching frequency, color adjustment amplitude, composition preference index, completion degree score, interaction stay time, and other key indicators.

[0060] The implementation of recursive quantization analysis requires phase space reconstruction of the behavior trajectory of each user. The time delay embedding method is adopted, with the embedding dimension m being 5 and the delay time τ being 3 time units. During phase space reconstruction, the original one-dimensional time series x(t) is converted into an m-dimensional vector sequence, and the i-th reconstructed vector is [x(i), x(i+τ), x(i+2τ), x(i+3τ), x(i+4τ)]. The reconstructed trajectory forms a complex geometric structure in the 5-dimensional phase space, reflecting the intrinsic dynamic characteristics of user behavior.

[0061] The construction of the recursive matrix requires the calculation of the distance relationship between any two points in the phase space. The Euclidean distance is used for measurement. For the i-th and j-th vector points in the reconstructed trajectory, the distance value d(i,j) is calculated. The determination of the distance threshold ε uses the fixed recursion rate method, with the threshold set to maintain the overall recursion rate at 5%. When d(i,j) is less than the threshold ε, the recursive matrix R(i,j) takes the value of 1, otherwise it takes the value of 0. The dimension of the recursive matrix is equal to the length of the reconstructed trajectory, usually a 715x715 square matrix.

[0062] The depiction of the behavior trajectory evolution chain is achieved through multi-scale time window analysis, with three different scale time windows set: a short-term window of 6 hours, a medium-term window of 3 days, and a long-term window of 10 days. The statistical quantities of user behavior features are calculated within each window, including mean, standard deviation, maximum, minimum, and quartile descriptive statistics. The feature vector of the evolution chain node is composed of these statistical quantities, forming a 15-dimensional node feature representation. The nodes of adjacent time windows are connected by calculating the Mahalanobis distance between the feature vectors, and the nodes with a distance less than the adaptive threshold have an evolution connection between them.

[0063] The capture of the resonance relationship between trajectories adopts the cross-correlation function analysis method, and calculates the correlation coefficient of any two user behavior sequences under different time lags, with the lag range set to -24 to +24 hours. The peak position of the cross-correlation function represents the time synchronization of the two user behavior modes, and the peak size represents the resonance strength. When the maximum value of the cross-correlation function exceeds 0.6 and the corresponding lag time is less than 6 hours, it is determined that there is a strong resonance relationship between the two users. The resonance network is constructed by constructing the resonance relationship between all user pairs as a weighted undirected graph, with the nodes being users and the edge weights being the resonance strength values.

[0064] The calculation of the entropy value distribution adopts the multi-scale sample entropy method, and the user behavior sequence is coarsely granulated, with the scale factor increasing from 1 to 20 step by step. For a scale factor s, the original sequence is divided into non-overlapping segments of length s, and the average value of the values in each segment is taken to form a coarse-grained sequence. The calculation of the sample entropy needs to set the pattern length m to 2 and the tolerance r to 0.15 times the standard deviation of the sequence. The system counts the number of patterns of lengths m and m+1 in the sequence, calculates the negative logarithm of the conditional probability to obtain the sample entropy value. The multi-scale sample entropy curve reflects the complexity characteristics of the user behavior in different time scales.

[0065] The dynamic analysis of the spatiotemporal dimension periodicity commonality of user behavior modes is realized by wavelet packet decomposition technology, and db4 wavelet in the Daubechies wavelet family is selected as the analysis wavelet, and the decomposition layer is set to 6 layers. Wavelet packet decomposition can decompose the behavior signal into different frequency components, and each frequency component corresponds to a specific periodic pattern. The system calculates the energy proportion of each frequency component, and the component with an energy proportion exceeding 8% is identified as a significant periodic pattern. The period length is calculated by the center frequency of the frequency component, and the period intensity is quantified by the energy proportion of the component.

[0066] The turning point recognition adopts the change point detection algorithm, and uses the cumulative sum test method to detect the statistical characteristic mutation point in the behavior sequence. The algorithm calculates the cumulative deviation statistic of the sequence, and marks it as a turning point when the statistic exceeds the critical value corresponding to a confidence level of 95%. The type of the turning point is classified by analyzing the changes of the mean and variance of the sequence in the time period before and after the turning point. The turning point with a significantly increased mean and a decreased variance is marked as an upward stable type, and the turning point with a significantly decreased mean and an increased variance is marked as a downward fluctuation type.

[0067] The process of resonance relationships forming natural communities in the feature space is achieved through a community detection algorithm. The Louvain algorithm is used to partition the resonance network into communities, and the algorithm finds the optimal community structure by maximizing the modularity function. The modularity calculation considers the difference between the internal connection density of the network and the expected connection density of the random network; the modularity value ranges from -1 to 1, with a larger value indicating a more pronounced community structure. During the algorithm iteration, adjacent communities are gradually merged until the modularity no longer increases. Each detected community corresponds to a user group, and users within the community share similar generation preference characteristics.

[0068] The user group boundary contour is determined using a concave hull algorithm. This algorithm calculates the set of coordinate points for each user in the feature space, and then applies the α-shape algorithm to construct the group boundary. The α parameter is selected through cross-validation to ensure the boundary encompasses the main distribution area of ​​the group without overfitting outliers. The boundary generated by the concave hull algorithm is a polygonal contour, and the vertex coordinates are stored as an ordered sequence of points. The complexity of the boundary contour is quantified by calculating the ratio of the contour perimeter to the circumference of a circle with an equal area.

[0069] Internal cohesion is calculated based on the spatial density of user feature points within the group. The group centroid is used as a reference point, and its coordinates are the arithmetic mean of the feature vectors of all users within the group. The Euclidean distance of each user to the group centroid reflects their deviation from the group center, and the standard deviation of the distances of all users to the centroid represents the internal dispersion of the group. The cohesion index is defined as the ratio of the maximum distance to the standard deviation of the actual distance, and the maximum distance is calculated from the maximum diameter of the group boundary profile.

[0070] Characterizing the distribution structure of the guiding features generated by each group in high-dimensional space requires extracting the dominant feature patterns of the groups and performing independent component analysis on the behavioral characteristics of users within each group to separate mutually independent feature components. The number of independent components is determined by the information criterion, typically retaining the top few components with a cumulative variance contribution rate of 85%. Each independent component corresponds to a dominant behavioral pattern of the group, and the weight coefficient of the component reflects the importance of that pattern in the group.

[0071] The construction of the swarm intelligence feature map maps the high-dimensional distribution structure to a two-dimensional visualization space. A t-SNE algorithm is used for nonlinear dimensionality reduction mapping, with the perplexity parameter set to one-third of the swarm size and the learning rate set to the square root of the swarm size. The mapping process optimizes the objective function through gradient descent, which measures the consistency between swarm similarity in the high-dimensional space and distance in the low-dimensional space. During iterative optimization, the system monitors the convergence of the objective function value, stopping optimization when the function value changes by less than 0.001 over 50 consecutive iterations.

[0072] In an alternative embodiment, a user group with similar generation preferences is identified based on the multi-dimensional behavior portrait using a group feature recursive quantization technique, a resonance relationship and an entropy value distribution between trajectories are captured by depicting an evolution chain of behavior trajectories, including:

[0073] A user group with similar generation preferences is identified based on the multi-dimensional behavior portrait using a group feature recursive quantization technique, a recursive graph matrix of high-dimensional phase orbits is constructed, a recurrence pattern and a self-similar structure of orbits in phase space are quantified by the recursive graph matrix, and an evolution chain and a bifurcation feature of user behavior trajectories are jointly depicted according to the recurrence pattern and the self-similar structure;

[0074] Based on the evolution chain and the bifurcation feature, the diagonal line structure alignment and the texture similarity between recursive graph matrices of different users are calculated, and the resonance relationship between trajectories is captured by combined analysis of the diagonal line structure alignment and the texture similarity;

[0075] For the user group with significant resonance relationship, the Shannon entropy and the permutation entropy of the recursive graph matrix thereof are calculated respectively, and the entropy value distribution of the group is obtained by mapping the Shannon entropy and the permutation entropy in the feature space.

[0076] The implementation process of identifying a user group with similar generation preferences based on multi-dimensional behavior portraits using a group feature recursive quantization technique requires the establishment of a complete data processing pipeline. The original user behavior data is stored in a distributed database, and each user record contains user identifier, timestamp, behavior type, parameter value, etc. The behavior type field enumerates values including image generation request, style switching operation, color adjustment action, composition modification behavior, completion confirmation operation, etc. 15 basic behavior types. The parameter value field stores specific behavior parameters in JSON format, such as RGB three-channel numerical value of color adjustment amplitude, nine-grid position coordinates of composition preference, millisecond-level time value of stay duration, etc.

[0077] The construction module of multi-dimensional behavior portrait first extracts the behavior sequence within the time window from the original data, and the time window adopts a sliding window mechanism with a window length of 24 hours and a sliding step of 1 hour. The behavior frequency distribution of users in each time window is counted, and the occurrence times, duration mean, parameter variation amplitude, etc. of each behavior type are calculated. The dimension of the behavior portrait vector is determined to be 128, the first 15 dimensions correspond to the frequency characteristics of each behavior type, the 16th to 30th dimensions correspond to the duration characteristics, the 31st to 45th dimensions correspond to the parameter variation amplitude characteristics, and the subsequent dimensions include Markov transition probability, time interval distribution parameter, periodicity intensity coefficient, etc. High-order statistical characteristics of behavior sequence.

[0078] The high-dimensional phase orbit reconstruction module receives a sequence of 128-dimensional behavior portrait vectors as input, each user maintaining a vector sequence of 720 time points. The phase space reconstruction adopts a specific implementation of the Takens embedding theorem, and the embedding dimension is automatically determined by the false nearest neighbor method. The calculation process starts from dimension 1 and gradually increases. When the proportion of false nearest neighbors falls below 5%, the optimal embedding dimension is determined. The delay time is determined by the first local minimum of the mutual information function. The mutual information between the behavior sequence and its delayed version is calculated under different delays, and the delay corresponding to the first local minimum of the mutual information is selected as the optimal delay time. The reconstructed phase orbit is represented as a multi-dimensional vector sequence, each vector containing the values of the original sequence at different delay times.

[0079] The recurrence plot matrix construction module performs distance calculation and thresholding processing on the reconstructed phase orbit. The distance calculation uses weighted Euclidean distance, and the weight of each dimension is determined according to the reciprocal of the variance of the dimension value. Dimensions with large variance obtain smaller weights to ensure that each dimension contributes relatively evenly to distance calculation. The distance threshold is determined using a fixed recurrence rate strategy, with a preset recurrence rate of 5%. The threshold parameter is found by a binary search algorithm to make the actual recurrence rate closest to the preset value. The recurrence plot matrix is stored in a sparse matrix format, only recording the position of elements with a value of 1, which greatly reduces memory usage. The texture feature extraction of the matrix uses a sliding window statistical method, with a window size of 20x20. The distribution density of 1-value elements, the number of connected regions, and the maximum connected region area within the window are calculated.

[0080] The orbit regression mode quantification module analyzes the structural features of the recurrence plot matrix to identify dynamic behavior patterns. Deterministic structure detection is achieved through diagonal line segment analysis. Starting from the main diagonal, scan 50 diagonal lines upward and downward, and calculate the length distribution of continuous 1-value line segments on each diagonal line. The determinism index is calculated as the ratio of the total length of all line segments with a length greater than or equal to 2 to the total number of non-zero elements in the matrix. This index reflects the predictability of the orbit. Laminar structure detection is achieved through vertical line analysis. Scan each column of the matrix and calculate the length distribution of continuous 1-value line segments in the vertical direction. Calculate the average vertical line length as the laminar index. Self-similarity structure detection divides the recurrence plot matrix into 50x50 sub-blocks. Calculate the structural similarity of each sub-block and its surrounding 8 adjacent sub-blocks. Use the normalized cross-correlation coefficient as a measure of similarity.

[0081] The user behavior trajectory evolution chain depicting module divides the recurrence plot matrix into 36 sub-matrices according to time dimension based on time segmentation analysis, each sub-matrix corresponding to a 20-hour behavior pattern time period. The sub-matrix feature vector contains five quantitative indicators, including recurrence rate, certainty index, maximum diagonal length, average vertical line length, and laminar index, within the time period. The transition characteristics between adjacent sub-matrices are measured by calculating the Euclidean distance of the feature vectors, with a distance less than 0.3 indicating smooth evolution, a distance between 0.3 and 0.8 indicating gradual change, and a distance greater than 0.8 indicating sudden evolution. The nodes of the evolution chain include time period identification, feature vector, transition type, and transition strength, while the edges include transition direction, transition probability, and duration.

[0082] The bifurcation feature recognition module detects nonlinear dynamics phenomena in the evolution chain, with bifurcation points identified by calculating the Lyapunov exponent of the evolution trajectory. A region with a Lyapunov exponent greater than 0 indicates chaotic behavior, a region close to 0 indicates quasi-periodic behavior, and a region less than 0 indicates stable behavior. The bifurcation type is determined by analyzing the topological structure changes of the trajectories before and after the bifurcation point, including period-doubling bifurcation, Hopf bifurcation, and saddle-node bifurcation. The bifurcation strength is quantified by calculating the change amplitude of the attractor dimension before and after the bifurcation, with a larger change indicating a stronger bifurcation.

[0083] The diagonal structure alignment degree calculation module of different user recurrence plot matrices uses the dynamic time warping algorithm to extract the main diagonal line and the 5 diagonal lines above and below it from each recurrence plot matrix, forming a feature sequence representing the user behavior time evolution pattern. The dynamic time warping algorithm calculates the optimal alignment path of two user feature sequences, allowing nonlinear stretching matching in the time dimension. The alignment index is defined as the normalized ratio of the cumulative distance of the optimal path to the sequence length, with a smaller value indicating a higher alignment degree. The time complexity of the algorithm is controlled by setting the bandwidth constraint, with a bandwidth of 10% of the sequence length, which significantly reduces the computational overhead while ensuring alignment accuracy.

[0084] The texture similarity analysis module calculates the two-dimensional texture features of the recurrence plot matrix based on the recurrence plot matrix, treating it as a grayscale image and extracting texture features using the gray level co-occurrence matrix method. Four gray level co-occurrence matrices are calculated in 0, 45, 90, and 135 degree directions with a distance of 1 and a distance of 2, resulting in a total of 8 co-occurrence matrices. Four Haralick texture features, including contrast, correlation, energy, and uniformity, are extracted from each co-occurrence matrix, forming a 32-dimensional texture feature vector. Texture similarity is obtained by calculating the cosine similarity of the texture feature vectors of two users, with a similarity value ranging from -1 to 1, with a value closer to 1 indicating more similar textures.

[0085] The inter-trajectory resonance relationship capturing module fuses the diagonal line structure alignment degree and the texture similarity, designs an adaptive weight distribution mechanism, and dynamically adjusts the weights of the two similarity indexes according to the complexity of the user behavior. The complexity of the behavior is measured by the fractal dimension of the recurrence plot matrix. The texture similarity weight increases for the user behavior mode with high fractal dimension, and the diagonal line structure alignment degree weight increases for the user behavior mode with low fractal dimension. The weight distribution formula is that the texture weight is equal to the ratio of the fractal dimension to the maximum fractal dimension, and the diagonal line weight is equal to 1 minus the texture weight. The comprehensive similarity is calculated by weighted average, and is used as a quantitative index of the inter-user resonance strength.

[0086] The resonance relationship significant user group identification module adopts a density-based clustering algorithm, constructs a user resonance network graph, and connects edges between user pairs with a resonance strength greater than 0.6, with the edge weight being the resonance strength value. The DBSCAN clustering algorithm is used to identify high-density user groups, with the minimum number of included points set to 8 users, and the neighborhood radius automatically determined by the elbow method of the K-distance graph. The algorithm first identifies core points, i.e., users whose number of neighbors reaches the minimum number of included points, and then groups density-reachable users into the same group. The attribution of boundary users is determined by calculating their distances to the core regions of each group, and the group with the closest distance receives the boundary user.

[0087] The recurrence plot matrix Shannon entropy calculation module quantifies the randomness of the user behavior mode in the group. The recurrence plot matrix is divided into 25x25 grid regions according to the row and column coordinates, and the number of elements with a value of 1 in each grid region is counted. The number of elements in each grid region is normalized to a probability distribution, and the probability value is equal to the number of elements in the region divided by the total number of matrix elements. The Shannon entropy is calculated as the negative sum of the product of the probability values of each grid region and their logarithm base 2, and the entropy value ranges between 0 and the logarithm base 2 of 625. A high entropy value indicates that the recurrence points are uniformly and randomly distributed in the matrix, and a low entropy value indicates that the recurrence points are concentrated in a specific region.

[0088] The permutation entropy calculation module analyzes the permutation pattern of the row vectors of the recurrence plot matrix, selects a permutation length of 3, and extracts a permutation pattern composed of 3 consecutive elements from each row of the recurrence plot matrix. There are 8 types of permutation patterns, including 000, 001, 010, 011, 100, 101, 110, and 111, and the frequency of each permutation pattern in all rows is counted. The permutation entropy is calculated as the negative sum of the product of the relative frequencies of each permutation pattern and their logarithm base 2, and the entropy value ranges between 0 and 3. The permutation entropy reflects the complexity of the row vector sequence of the recurrence plot matrix, and a high permutation entropy indicates that the row vector pattern changes complexly, and a low permutation entropy indicates that the row vector pattern is relatively simple and regular.

[0089] The group entropy value distribution characteristic space mapping module constructs a group distribution graph by taking Shannon entropy and permutation entropy as two-dimensional features, eliminates the dimensional difference of the two entropy values by using standardization processing, and the standardization formula is that the entropy value is divided by the standard deviation after subtracting the mean. A two-dimensional coordinate system is established, the horizontal axis is the standardized Shannon entropy, and the vertical axis is the standardized permutation entropy. Each user group corresponds to a coordinate point in the characteristic space, and the coordinate value is the arithmetic mean of the entropy values of all users in the group. The visualization of group distribution adopts the form of scatter plot, the size of the point is proportional to the number of users included in the group, and the color depth reflects the similarity within the group, and the deeper the color, the more consistent the user behavior patterns within the group.

[0090] In an optional embodiment, the user features contained in the historical interaction trajectory are mapped and analyzed at multiple levels with the collective wisdom feature map, the group affiliation of the target user is anchored, and the guiding features corresponding to the affiliation group are visualized based on deep feature perspective technology, including:

[0091] The user features contained in the historical interaction trajectory are mapped and analyzed at multiple levels with the collective wisdom feature map, the affinity relationship between the user feature vector and each group node is quantified by constructing a multi-stage projection link of feature space, semantic space and topological space, the affiliation probability distribution is constructed based on the affinity relationship, and the group affiliation of the target user is anchored according to the dominant peak value of the affiliation probability distribution;

[0092] According to the group affiliation, the corresponding node in the collective wisdom feature map is located, the distribution parameters, evolution chain features and entropy labels contained in the corresponding node are deconstructed, the abstract guiding feature representation of the affiliation group is established by generating a deep perspective of the guiding feature space;

[0093] Based on the abstract guiding feature representation, the generative preferences of the group commonality are visualized as guiding parameters in the operation dimension, the dominant path in the evolution chain feature is fused to deconstruct the control node and parameter strategy, and the visualized expression of the guiding feature is formed.

[0094] In the process of realizing the multi-level mapping analysis of historical interaction trajectory and collective wisdom feature map, the basic feature vector is extracted from the user's historical interaction data. The historical interaction trajectory contains multi-dimensional data such as content generation record, editing operation sequence, tool calling frequency, parameter adjustment amplitude, session duration, etc. The system converts these raw data into standardized numerical feature vectors, and the dimension of each feature vector is set to 512, of which the first 128 dimensions represent content type preference, the middle 192 dimensions represent operation behavior pattern, and the last 192 dimensions represent timing evolution characteristics. For a target user, its historical interaction trajectory shows that it has completed 87 content generation tasks in the past 30 days, of which image tasks account for 63%, text tasks account for 25%, and mixed tasks account for 12%. The average single session duration of this user is 18 minutes, and the parameter adjustment frequency is 7.3 times per session.

[0095] The projection mapping mechanism of the feature space is constructed to compare and calculate the user feature vector with each group node in the collective wisdom feature map. The collective wisdom feature map contains 2048 group nodes, each node representing a user cluster with similar behavior characteristics. The projection of the feature space is realized by calculating the distance measure between the user feature vector and the center vector of each group node. The distance measure adopts the weighted Euclidean distance method, and the weight coefficient is determined according to the inverse variance of the feature dimension. The system calculates that the distance value between the target user and group node A1 is 0.23, the distance value between the target user and group node B7 is 0.31, and the distance value between the target user and group node C2 is 0.19. The smaller the distance value, the higher the affinity.

[0096] At the semantic space level, the user's interaction behavior is semantically analyzed, and the operation sequence is converted into a semantic label sequence. The operation sequence of the target user shows that it frequently uses style conversion functions, detail enhancement tools, and color adjustment modules, which are mapped to a set of semantic labels, including "visual optimization orientation", "fine processing preference", "color sensitive type", etc. The system matches these semantic labels with the group semantic features stored in the group nodes and uses label overlap degree as a measure of semantic affinity. The calculation result shows that the semantic label of the target user has an overlap degree of 78% with the label of group node C2, an overlap degree of 65% with the label of group node A1, and an overlap degree of 52% with the label of group node B7.

[0097] The mapping of the topological space focuses on the consistency of the time-evolution path of user behavior and the group-evolution trajectory, extracts the operation transition pattern of the target user in the time dimension, and constructs the behavior transition topological graph of the individual user. This topological graph records the various intermediate states and their transition probabilities from the initial operation to the final result, for example, when the target user performs an image generation task, there is a 72% probability that he will first perform basic framework construction, then there is an 85% probability that he will enter the detail refinement stage, and finally there is a 91% probability that he will perform global optimization operations. The system calculates the structural similarity between the user topological graph and the standard topological pattern of each group node, and the results show that the topological similarity with group node C2 is 0.83, and the topological similarity with group node A1 is 0.71.

[0098] The affinity calculation results of the three levels of feature space, semantic space, and topological space are integrated to assign a membership probability value to each group node. The membership probability value is obtained by weighted fusion of the affinity calculation results of the three spaces, where the feature space weight is set to 0.3, the semantic space weight is set to 0.35, and the topological space weight is set to 0.35. For the target user, the comprehensive membership probability of group node C2 is 0.76, the membership probability of group node A1 is 0.58, and the membership probability of group node B7 is 0.42. The membership probabilities of the remaining nodes are all less than 0.35. The system identifies the dominant peak in the membership probability distribution as group node C2, and determines the group membership relationship of the target user as C2 group.

[0099] After positioning to group node C2, the multi-dimensional group feature data stored in this node is deconstructed. The distribution parameter part records the statistical distribution characteristics of C2 group members in each feature dimension, including median, interquartile range, skewness coefficient, kurtosis coefficient, etc. The specific data shows that the median of C2 group in the color saturation adjustment amplitude dimension is 32 units, the interquartile range is 17 units, and it presents a right-skewed distribution characteristic with a skewness coefficient of 1.24. In the detail enhancement intensity dimension, the median of this group is 58 units, and the distribution is relatively concentrated with an interquartile range of 9 units. These distribution parameters reveal the concentration trend and dispersion degree of C2 group in operation parameter selection.

[0100] The evolution chain feature records the evolution law of C2 group members in operation strategy over time. The data shows that the average number of parameter adjustments per session of C2 group members at the initial stage of joining the platform is 4.2 times, which increases to 7.8 times after 30 days of use, and stabilizes at about 9.3 times after 90 days of use. The evolution chain also shows that the group members gradually migrate from basic tools to advanced tools, with the proportion of advanced tool usage increasing from 18% at the initial stage to 62% at the mature stage. The dominant path in the evolution chain identifies a typical three-stage evolution mode in this group: the exploration stage focuses on mastering basic parameters, the growth stage focuses on tool combination application, and the mature stage focuses on style customization.

[0101] Entropy value labels quantify the degree of uncertainty in the behavior selection of C2 group, and the calculation results show that the entropy value of the group in the content type selection dimension is 1.86, indicating that the selection has certain concentration but still maintains diversity. The entropy value in the tool call sequence dimension is 2.34, showing that the diversity of operation path is higher. The entropy value in the parameter value distribution dimension is 1.52, indicating that there is obvious preference mode in parameter selection. These entropy value labels provide quantitative basis for understanding the randomness and certainty of group behavior.

[0102] Based on the distribution parameters, evolution chain features, and entropy value labels, a deep perspective model is constructed to establish the abstract representation of C2 group in the guided feature space. The abstract representation adopts a deep neural network architecture, with the input layer receiving the concatenated vector of distribution parameters, evolution chain features, and entropy value labels. After three layers of full connection transformation and nonlinear activation, the output layer generates a 256-dimensional abstract guide feature vector. This vector encodes the core behavior patterns and generation preferences of C2 group in high-dimensional space, specifically represented by the activation patterns of specific dimension clusters and the corresponding relationship with group features.

[0103] When converting the abstract guide feature vector into executable operation-level guide parameters, the mapping relationship between the activation intensity of each dimension in the vector and the specific operation parameters is analyzed. For the color saturation parameter, the average activation intensity of the 23rd to 35th dimension cluster of the abstract vector is 0.68, which maps to the recommended saturation increase range of 28 to 36 units. For the detail enhancement parameter, the activation intensity of the 78th to 94th dimension cluster is 0.82, which maps to the recommended enhancement intensity range of 54 to 62 units. For the contrast parameter, the activation intensity of the 156th to 171st dimension cluster is 0.54, which maps to the recommended contrast adjustment range of 12 to 18 units. These concrete guide parameters constitute the initial recommended configuration scheme for the target user.

[0104] Combined with the dominant path identified in the evolution chain feature, the key regulation nodes of each evolution stage are deconstructed. The regulation nodes in the exploration stage focus on the threshold setting of basic parameters, and the system sets the recommended range boundary of basic parameters for the target user to avoid parameter values exceeding the novice adaptation interval. The regulation nodes in the growth stage shift to tool combination mode guidance, and the system recommends advanced tools with high compatibility with the current proficient tools of the target user, and the recommendation priority is sorted according to the synergy effect coefficient between tools. The regulation nodes in the mature stage focus on the stylized path branches, and the system recommends personalized parameter combination templates according to the historical style tendency of the target user. Each regulation node is configured with a corresponding parameter strategy, which defines the type, intensity, and timing of guide information that the system should provide at that node, ultimately forming a dynamic guide feature expression system covering the entire use cycle of the user.

[0105] In an alternative embodiment, by constructing a multi-order projection link of the feature space, the semantic space and the topological space, the affinity relationship between the user feature vector and each group node is quantified respectively, and the construction of the belonging probability distribution based on the affinity relationship comprises:

[0106] By constructing a multi-order projection link of the feature space, the semantic space and the topological space, the affinity degree of the user feature vector and the distribution parameters of each group node is depicted in the feature space to obtain a feature affinity, the cognitive synergy strength between the semantic representation of the user feature vector and the dominant guiding features of each group node is depicted in the semantic space to obtain a semantic affinity, and the coupling depth of the manifold distance of the user feature vector in the group wisdom feature map structure and the topological centrality of each group node is constructed in the topological space to obtain a topological affinity.

[0107] Based on the entropy value label of each group node, the expression tendency of the user feature vector in different cognitive levels is revealed, the synergy mechanism of the feature, semantic and topological triple affinities is constructed through the internal consistency of the entropy value label, and the contribution proportion of each space affinity in different cognitive dimensions is guided according to the synergy mechanism to obtain a complete affinity relationship.

[0108] The affinity relationship is condensed into a belonging probability distribution in a probability space, so that adjacent groups form a potential energy difference in the belonging determination process, the fuzzy area of the group boundary is eliminated, and a belonging probability distribution with a clear dominant direction is obtained.

[0109] The multi-order projection link construction module establishes data projection channels of the feature space, the semantic space and the topological space, the feature space projection module receives a 128-dimensional behavior portrait vector of a user as input, and adopts principal component analysis technology to project high-dimensional features into a 32-dimensional feature subspace. The principal component selection strategy retains the first 32 principal components with a cumulative variance contribution rate of 92%, the projection matrix is a 128×32 real matrix, and the element value range is between -1 and 1. The semantic space projection adopts a pre-trained text embedding model to convert the semantic description of the user behavior into a 256-dimensional semantic vector, and the semantic description includes text labels such as artistic style, creation theme and technical complexity preferred by the user. The topological space projection maps the positional relationship of the user in the group wisdom feature map into a 64-dimensional topological coordinate vector through a graph embedding algorithm, trains a graph embedding representation by using a Node2Vec algorithm, and sets the random walk length to 40 steps and generates 20 random walk paths for each node.

[0110] The feature affinity calculation module quantifies the resonance degree of the user feature vector and the group node distribution parameters in the feature space. The group node distribution parameters include the mean, covariance matrix, skewness coefficient, kurtosis coefficient, and other statistical descriptors of the group feature vector. The affinity resonance degree is measured by the reciprocal of Mahalanobis distance, which requires the difference between the user feature vector and the group mean vector, as well as the inverse of the group covariance matrix. The singular value decomposition method is used to handle the approximate singular case, and components with singular values less than 0.001 are set to 0.001 to ensure numerical stability. The feature affinity is normalized to 0 to 1 by the Sigmoid function, with a slope factor of 2.5 and a center offset of 0.5.

[0111] The semantic affinity calculation module depicts the cognitive synergy strength between the semantic representation of the user feature vector and the dominant guiding features of each group node. The dominant guiding features of the group node are obtained by weighted aggregation of the semantic vectors of users within the group, and the weight is determined according to the representativeness of the user in the group, which is quantified by the inverse distance between the user and the group center. The cognitive synergy strength is calculated using cosine similarity, which is the inner product of the user semantic vector and the group dominant semantic vector divided by the product of the two vector lengths. The preprocessing of the semantic vector includes L2 norm normalization and centering operation to ensure the geometric meaning of the similarity calculation. The dynamic adjustment mechanism of semantic affinity considers the confidence of the user semantic representation, and the weight of the user semantic affinity is halved when the confidence is less than 0.6.

[0112] The topology affinity calculation module constructs the coupling depth of the manifold distance of the user feature vector in the group wisdom feature map structure and the topological centrality of each group node. The manifold distance is calculated using geodesic distance, and the shortest path from the user node to each group center node is found on the group wisdom feature map by Dijkstra algorithm. The path weight is set as the weighted combination of the Euclidean distance between nodes and the connectivity strength, with a Euclidean distance weight of 0.7 and a connectivity strength weight of 0.3. The group node topological centrality is measured by the feature vector centrality, and the principal eigenvector of the graph adjacency matrix is calculated by power iteration method, with an iteration convergence threshold of 0.0001. The coupling depth is calculated by the product of the negative exponential function of the manifold distance and the topological centrality, with an attenuation coefficient of the exponential function set to 0.15.

[0113] The entropy value label construction module reveals the expression tendency of the user feature vector in different cognitive levels based on the entropy value distribution characteristics of each group node. The group node entropy value label contains statistical descriptions of various entropy value indexes such as recursive graph matrix Shannon entropy, permutation entropy, approximate entropy, and sample entropy. Expression tendency quantization is achieved by calculating the probability density function value of the user individual entropy value and the group entropy value distribution. The probability density function is constructed using the kernel density estimation method, and the kernel function is selected as the Gaussian kernel. The bandwidth parameter is automatically selected through cross-validation. The cognitive level is divided into three dimensions of innovation, regularity, and complexity. Shannon entropy reflects innovation, permutation entropy reflects regularity, and approximate entropy reflects complexity.

[0114] The triple affinity synergy mechanism construction module guides the synergistic effect of each spatial affinity through the internal consistency of the entropy value label. The internal consistency is measured by the Pearson correlation coefficient between the entropy value label vectors. An absolute value of the correlation coefficient greater than 0.6 indicates high consistency, a value between 0.3 and 0.6 indicates moderate consistency, and a value less than 0.3 indicates low consistency. The synergy mechanism dynamically adjusts the weight distribution of each spatial affinity according to the consistency degree. In the case of high consistency, the feature affinity weight is 0.5, the semantic affinity weight is 0.3, and the topology affinity weight is 0.2. In the case of moderate consistency, the weights are adjusted to 0.4, 0.35, and 0.25. In the case of low consistency, the weights are adjusted to 0.35, 0.4, and 0.25.

[0115] The cognitive dimension contribution proportion calculation module determines the contribution distribution of each spatial affinity in the innovation, regularity, and complexity dimensions according to the synergy mechanism. The innovation dimension mainly depends on the semantic affinity, with a basic weight of 0.6. The matching degree between the user innovation entropy value and the group distribution is adjusted, and the weight is increased by up to 0.15 when the matching degree is high. The regularity dimension mainly depends on the feature affinity, with a basic weight of 0.55. The stability of the user behavior pattern is adjusted, and the stability is quantified by the autocorrelation function peak value of the behavior sequence. The complexity dimension mainly depends on the topology affinity, with a basic weight of 0.5. The local clustering coefficient of the user in the graph is corrected, and a high clustering coefficient indicates a dense connection around the user, enhancing the complexity contribution.

[0116] The affinity relationship landscape integration module combines the weighted combination of the three affinities in each cognitive dimension to form a complete affinity relationship description. The integration process uses tensor product operation to perform outer product operation on the 3 affinity vectors and the 3 cognitive dimension weight vectors to obtain a 3x3 affinity relationship tensor. The tensor elements represent the contribution strength of a specific affinity in a specific cognitive dimension, the trace of the tensor represents the overall affinity strength, and the determinant of the tensor represents the stability of the affinity relationship. The visualization of the affinity relationship landscape uses a radar chart form, with three axes representing feature, semantic, and topology affinity. Different colored areas represent different cognitive dimension contributions.

[0117] The probability space condensing module converts the affinity landscape into a belonging probability distribution. The basic form of the probability distribution adopts a Softmax function to convert the comprehensive affinity values of each group into probability values, ensuring that the sum of the belonging probabilities of all groups is equal to 1. The temperature parameter of the Softmax function is set to 1.5. A higher temperature value makes the probability distribution smoother, avoiding excessive concentration on a single group. The probability adjustment mechanism considers the uncertainty of user characteristics, which is quantified by the mean of the variance of each dimension of the feature vector. Users with high uncertainty have a more uniform probability distribution.

[0118] The potential energy gap mechanism construction module eliminates the boundary fuzzy area of adjacent groups in the belonging determination process. Adjacent groups are defined as pairs of groups directly connected in the group wisdom feature map. The connection strength is calculated by the weighted average of the user interaction frequency and feature similarity between groups. The potential energy gap is realized by introducing a competition inhibition mechanism between adjacent groups. When the difference between the belonging probabilities of two adjacent groups is less than 0.15, the potential energy gap adjustment is activated, increasing the probability value of the higher probability group by 0.1 and decreasing the probability value of the lower probability group by 0.1. After adjustment, the probability is normalized again to ensure the rationality of the probability distribution.

[0119] The clear dominant direction generation module determines the final group belonging of the user based on the probability distribution adjusted by the potential energy gap. The dominant direction is determined by the maximum probability principle, and the group with the highest probability becomes the dominant belonging group of the user. The confidence evaluation is calculated by the ratio of the highest probability to the second highest probability. A ratio greater than 2.0 indicates high confidence belonging, a ratio between 1.5 and 2.0 indicates medium confidence, and a ratio less than 1.5 indicates low confidence. In the case of low confidence, a multi-round iteration mechanism is started to recalculate the affinity and adjust the probability distribution, with a maximum of 3 rounds of iteration.

[0120] The data case shows the application effect of the technology on a data set containing 950 users and 8 groups. The user feature vector dimension is 128, projected to a 32-dimensional feature space through principal component analysis. The semantic space uses a 768-dimensional pre-trained embedding vector, compressed to 256 dimensions. The embedding dimension of the topological space is set to 64. The average value of the calculated feature affinity is 0.67, the average value of the semantic affinity is 0.58, and the average value of the topological affinity is 0.72. The collaborative mechanism analysis shows that 74% of the users have a high consistency entropy label, 21% have a medium consistency, and 5% have a low consistency.

[0121] In an alternative embodiment, a multi-level guide vector is constructed and the hidden space representation of the image generation model is accurately mapped. The deviation degree of the generation result from the guide vector is quantified in real time. When the deviation exceeds the preset deviation threshold, the vector intelligent calibration is triggered, and the generation result that meets the dual constraints is output, including:

[0122] A multi-level guidance vector is constructed, a deep resonance relationship is constructed based on the multi-level guidance vector and a latent space representation of an image generation model, and the image generation model is driven to form an intermediate generation result;

[0123] An implicit layer feature representation is condensed from the intermediate generation result, the implicit layer feature representation is projected to a multi-dimensional creative space constructed by the multi-level guidance vector, and an image projection of each level is depicted;

[0124] According to the creative deviation degree of the intermediate generation result and the multi-level guidance vector, a level resonance strength is constructed based on the dependency association between levels, the overall deviation degree is obtained by harmonizing the level resonance strength, and a vector calibration mechanism is activated when the overall deviation degree exceeds a preset deviation threshold;

[0125] The vector calibration mechanism is executed, the dominant deviation level is located and the corresponding guidance vector is harmonized, the stability index is used to constrain the harmonization amplitude, the multi-level guidance vector is reconstructed, a new resonance relationship between the reconstructed multi-level guidance vector and the latent space representation is established, the deviation perception and calibration process are repeatedly executed until the overall deviation degree reaches the expected harmonious state, and a generation result satisfying the double constraints is output.

[0126] In the operation process of the latent space of the image generation model, the multi-level guidance vector is constructed by using a hierarchical semantic coding architecture. The input text description is converted into an initial vector representation by a word embedding layer, and the dimension of the initial vector is set to 512. Each dimension corresponds to a specific semantic component. In the semantic decomposition stage, the initial vector is decomposed along different abstraction levels to form three levels of high-level concept vectors, middle-level attribute vectors, and low-level detail vectors. The high-level concept vector captures the overall style and theme information, and its dimension is 128. The numerical range of each component is between -1 and 1. The middle-level attribute vector encodes attribute features such as color distribution and spatial layout, and its dimension is 192. The low-level detail vector is responsible for expressing fine-grained features such as texture and edge, and its dimension is 192. The three level vectors are recombined into a complete 512-dimensional guidance vector through splicing operation.

[0127] When establishing the deep resonance relationship between the multi-level guidance vector and the hidden space representation of the image generation model, a feature alignment module is used for space mapping. This module receives the guidance vector as input and converts it into a representation that matches the hidden space of the generation model through a three-layer fully connected network. The first fully connected network maps the 512-dimensional guidance vector to a 1024-dimensional intermediate representation, with a leaky rectified linear unit as the activation function and a negative slope parameter of 0.2. The second layer maintains a dimension of 1024 and introduces a residual connection structure to maintain information flow. The third layer compresses the representation to a 768-dimensional hidden space vector required by the generation model. The hidden space vector is embedded into each decoding layer of the generation model through an injection mechanism, injecting guidance information at the corresponding level at the 3rd, 6th, and 9th layers to form a hierarchical feature modulation effect. The generation model performs forward propagation based on the injected hidden space representation and outputs an intermediate generated result with a resolution of 512 by 512 pixels.

[0128] The process of extracting hidden layer feature representation from the intermediate generated result uses a multi-scale feature acquisition strategy. A feature extraction network is applied to the generated image data, which contains four downsampling stages. Each stage reduces the spatial resolution by half and doubles the number of channels. The initial input is a 512 by 512 by 3 RGB image. After the first stage of convolution operation, a 256 by 256 by 64 feature map is obtained, with a convolution kernel size of 3 by 3, a step size of 2, and a padding of 1. The second stage outputs a 128 by 128 by 128 feature map. The third stage generates a 64 by 64 by 256 feature map. The fourth stage produces a 32 by 32 by 512 deep layer feature representation. Global average pooling operations are performed on the feature maps of the four stages to compress the spatial dimension to a single value, resulting in stage feature vectors with dimensions of 64, 128, 256, and 512 respectively. The four vectors are combined into a 960-dimensional comprehensive hidden layer feature representation through concatenation.

[0129] Projecting the hidden layer feature representation into the multi-dimensional creative space constructed by the multi-level guidance vector requires establishing a feature mapping relationship. The definition of the creative space is based on the three levels of the multi-level guidance vector, corresponding to three subspaces. The high-level subspace has a dimension of 128, the middle-level subspace has a dimension of 192, and the low-level subspace has a dimension of 192. The 960-dimensional hidden layer feature representation is mapped to the three subspaces through three independent projection networks. The high-level projection network contains two fully connected structures, the first layer compresses the 960-dimensional representation to 256-dimensional, and the second layer outputs a 128-dimensional high-level image projection. The middle-level projection network also uses a two-layer structure, with an intermediate layer dimension of 384 and an output layer dimension of 192. The low-level projection network has an intermediate layer dimension of 384 and an output dimension of 192. The three image projection vectors reflect the feature distribution state of the intermediate generated result at different abstraction levels.

[0130] When calculating the degree of deviation of the intermediate generated result from the creative direction of the multi-level guide vector, the difference indicators are measured for each level. The high-level deviation indicator is obtained by calculating the Euclidean distance between the high-level image projection and the high-level concept vector. Specifically, the difference between the corresponding positions of the two 128-dimensional vectors is squared, and the square root of the sum of the squares of all positions is taken to obtain the scalar distance value. When the high-level concept vector has a value of 0.8 in a certain dimension and the corresponding image projection has a value of 0.6 in that dimension, the contribution of the difference in that dimension is 0.04. After accumulating the differences of all 128 dimensions and taking the square root, assuming the accumulated sum is 2.56, the high-level deviation indicator is 1.6. The middle-level deviation indicator and the low-level deviation indicator are calculated in the same way, measuring the distance between the middle-level attribute vector and the middle-level image projection, and the low-level detail vector and the low-level image projection, respectively, in a 192-dimensional space. Assuming the middle-level deviation indicator is 2.3 and the low-level deviation indicator is 1.9.

[0131] To construct the level resonance strength, the dependency correlation weight between levels is introduced, which reflects the influence of different levels on the final generated effect and is pre-set according to the characteristics of the generation task. For generation tasks that focus on overall style consistency, the high-level dependency weight is set to 0.5, the middle-level dependency weight is set to 0.3, and the low-level dependency weight is set to 0.2. Multiply the deviation indicators of each level by the corresponding dependency weight, the weighted deviation of the high level is 1.6 multiplied by 0.5, which equals 0.8, the weighted deviation of the middle level is 2.3 multiplied by 0.3, which equals 0.69, and the weighted deviation of the low level is 1.9 multiplied by 0.2, which equals 0.38. Sum the three weighted deviation values to obtain the overall deviation degree, which is 1.87 in this example. The pre-set deviation threshold is determined according to the generation quality requirement, which is set to 1.5. The current overall deviation degree 1.87 exceeds the threshold, triggering the vector calibration mechanism.

[0132] When executing the vector calibration mechanism, the dominant deviation level is located by comparing the weighted deviation values of each level. In the current example, the high-level weighted deviation 0.8 is the maximum value, and the high level is determined as the dominant deviation level. Harmonization is performed on the high-level concept vector, and the harmonization direction is determined by the difference vector between the high-level image projection and the high-level concept vector. Each dimension of the difference vector represents the direction and relative amplitude of the adjustment needed for that dimension. A positive value indicates that the corresponding dimension of the original vector needs to be increased, and a negative value indicates that the corresponding dimension of the original vector needs to be decreased. A stability constraint coefficient is introduced to control the harmonization amplitude, which is dynamically adjusted according to the current iteration round. The initial round is set to 0.3, and gradually decays to 0.1 as the iteration proceeds. Multiply the difference vector by the stability constraint coefficient to obtain the adjustment vector, and add the adjustment vector to the original high-level concept vector to complete the vector harmonization. Assuming that the difference value of a certain dimension is negative 0.2 and the stability constraint coefficient is 0.3, the adjustment amount of that dimension is negative 0.06, and the original value 0.8 is adjusted to 0.74.

[0133] When reconstructing the multi-level guidance vector, the original high-level vector is replaced by the harmonized high-level concept vector, and the middle-level attribute vector and the low-level detail vector remain unchanged. A new 512-dimensional complete guidance vector is formed through splicing operation. The reconstructed guidance vector is re-input into the feature alignment module to perform the same mapping conversion as the initial process to generate a new 768-dimensional hidden space representation. The new hidden space representation is injected into the image generation model again to trigger a new round of forward propagation process to produce an updated intermediate generation result. The complete process of feature extraction, projection mapping, and deviation calculation is repeatedly performed on the updated result to obtain a new overall deviation value. Assuming that the overall deviation is reduced to 1.62 after one calibration, which is still higher than the threshold of 1.5, the next round of calibration is continued. In the second round of calibration, the middle-level weighted deviation becomes the dominant factor, and the harmonization operation is performed on the middle-level attribute vector. The iterative process continues until the overall deviation is reduced to 1.48, which meets the condition of a harmonious state below the threshold, and the calibration cycle is terminated.

[0134] When outputting the generated result that meets the dual constraints, the dual constraints include deviation constraint and stability constraint. The deviation constraint requires the overall deviation to be lower than the preset threshold to ensure the consistency of the generated content with the guidance intention. The stability constraint prevents the generation collapse caused by excessive correction of the vector by limiting the harmonization amplitude. The final generated result retains the 512 by 512 pixel image data generated in the last iteration, which is highly aligned with the original guidance vector in the semantic level and meets the output standard of the generation model in the visual quality. The execution time of the entire process depends on the initial deviation degree and the calibration convergence speed. In a typical scenario, 3 to 5 iterations are needed to complete the calibration process.

[0135] In an optional implementation, a hierarchical resonance strength is constructed based on the dependency association between levels, the overall deviation is obtained by harmonizing the hierarchical resonance strength, and a vector calibration mechanism is activated when the overall deviation breaks through a preset deviation threshold, including:

[0136] A hierarchical resonance strength is constructed based on the dependency association between levels, the activity level of information transmission is revealed through the information flow entropy ratio between hierarchical features based on the hierarchical resonance strength, the activity level reflects the transmission depth and diffusion law of features between levels, the guiding influence between levels is depicted based on the transmission depth and the diffusion law, and a dynamic evolution chain of hierarchical features is formed based on the guiding influence;

[0137] The hierarchical resonance strength is deconstructed based on the dynamic evolution chain, the stability of feature transmission is described by capturing the dominant frequency of resonance fluctuation, the deviation point in the creation process is identified based on the dominant frequency, and a vector calibration mechanism is activated when the overall deviation caused by the deviation point breaks through a preset deviation threshold.

[0138] In the implementation process, the dependency relationship between the feature vectors of each level is quantitatively analyzed, and the first-level feature vector A1 = [0.82, 0.35, 0.67, 0.91], the second-level feature vector A2 = [0.74, 0.42, 0.59, 0.88], and the third-level feature vector A3 = [0.69, 0.51, 0.63, 0.85] are extracted. For the feature vectors between adjacent levels, the system calculates the difference degree of the numerical value at each dimension position, and the difference value between the first level and the second level in the first dimension is calculated as 0.08, the difference value in the second dimension is 0.07, the difference value in the third dimension is 0.08, and the difference value in the fourth dimension is 0.03. The system comprehensively evaluates all the difference values, and the square sum of the difference values of each dimension is added and then the square root is taken, and the dependency relationship degree between levels is 0.135. By traversing all adjacent level pairs, the system constructs a complete dependency relationship matrix, which describes the connection tightness of features between different levels.

[0139] Based on the dependency relationship matrix, the hierarchical resonance strength index is constructed. For the creative network structure containing five levels, the dependency relationship degree between the first level and the second level is recorded as 0.135, the dependency relationship degree between the second level and the third level is recorded as 0.142, the dependency relationship degree between the third level and the fourth level is recorded as 0.168, and the dependency relationship degree between the fourth level and the fifth level is recorded as 0.193. The system progressively accumulates the dependency relationship values between adjacent levels, and assigns a position weight coefficient to each relationship degree. The closer the position is to the output layer, the higher the weight of the relationship degree. In the specific calculation, the relationship degree between the first level and the second level is multiplied by the weight 0.8, the relationship degree between the second level and the third level is multiplied by the weight 0.9, the relationship degree between the third level and the fourth level is multiplied by the weight 1.0, and the relationship degree between the fourth level and the fifth level is multiplied by the weight 1.2. The system adds the weighted relationship degree values and divides by the total weight to obtain the comprehensive hierarchical resonance strength of 0.162.

[0140] According to the hierarchical resonance strength calculation information flow entropy ratio, the number of activated feature nodes in each level is counted, the first level activated node number is 128, the second level activated node number is 96, the third level activated node number is 84, the fourth level activated node number is 72, and the fifth level activated node number is 58. The system carries out ratio operation on the activated node number of adjacent levels, the node retention rate from the first level to the second level is 0.75, the node retention rate from the second level to the third level is 0.875, the node retention rate from the third level to the fourth level is 0.857, and the node retention rate from the fourth level to the fifth level is 0.806. The system takes the logarithm value of the retention rate of each level and normalizes it to obtain the information flow entropy values of each level, which are 0.32, 0.15, 0.17 and 0.24 respectively. The system carries out ratio calculation on the information flow entropy values of adjacent levels to obtain the entropy ratio sequence 2.13, 0.88 and 1.41. The entropy ratio sequence reveals the activity degree of information in the hierarchical transmission process, and the larger the value, the more active the information transmission between the levels, and the wider the feature diffusion range.

[0141] Based on the information flow entropy ratio, the transmission depth and diffusion law are determined, the levels with entropy ratio value greater than 1.5 are marked as high active transmission area, the levels with entropy ratio value between 0.8 and 1.5 are marked as medium active transmission area, and the levels with entropy ratio value less than 0.8 are marked as low active transmission area. In this case, the first level to the second level belongs to the high active transmission area, the transmission depth score is 9.2, indicating that the feature information can be deeply transmitted to the subsequent level; the second level to the third level belongs to the low active transmission area, the transmission depth score is 4.6, indicating that the feature transmission exists attenuation; the third level to the fourth level belongs to the medium active transmission area, the transmission depth score is 6.8. The system constructs a diffusion law model, divides the feature activation distribution of each level according to the space region, counts the concentration of feature activation in each space region, and the concentration less than 0.4 indicates that the diffusion law presents divergent type, the concentration between 0.4 and 0.7 indicates that the diffusion law presents balanced type, and the concentration higher than 0.7 indicates that the diffusion law presents convergent type.

[0142] According to the description of the guiding influence between levels according to the transmission depth and the diffusion law, the transmission depth score and the concentration of the diffusion law are cross-analyzed. When the transmission depth score is higher than 7.5 and the diffusion law is divergent, the guiding influence of the level on the subsequent level is evaluated as strong guiding type, and the influence coefficient is set to 1.35; when the transmission depth score is between 5.0 and 7.5 and the diffusion law is balanced, the guiding influence of the level on the subsequent level is evaluated as medium guiding type, and the influence coefficient is set to 1.0; when the transmission depth score is lower than 5.0 and the diffusion law is convergent, the guiding influence of the level on the subsequent level is evaluated as weak guiding type, and the influence coefficient is set to 0.65. In this case, the system determines that the first level is a strong guiding type, the second level is a weak guiding type, the third level is a medium guiding type, and the fourth level is a medium guiding type. The system connects the guiding influence coefficients of each level in time sequence to form a dynamic evolution chain [1.35, 0.65, 1.0, 1.0], which reflects the change trajectory of the transmission strength of the feature between different levels.

[0143] Based on the dynamic evolution chain, the harmonic deconstruction of the level resonance strength is performed, and the influence coefficient sequence in the dynamic evolution chain is taken as a discrete signal. The system performs frequency domain decomposition processing on the signal. The system constructs a sampling window with a length of 8, periodically extends the influence coefficient sequence [1.35, 0.65, 1.0, 1.0] to obtain an extended sequence [1.35, 0.65, 1.0, 1.0, 1.35, 0.65, 1.0, 1.0]. The system performs frequency component extraction on the extended sequence and identifies three main frequency components. The first frequency component has an amplitude of 0.87 and a period of 2 levels, the second frequency component has an amplitude of 0.43 and a period of 4 levels, and the third frequency component has an amplitude of 0.21 and a period of 8 levels. The system determines the frequency component with the largest amplitude as the dominant frequency. In this case, the period of the dominant frequency is 2 levels and the amplitude is 0.87. Based on the periodic characteristics of the dominant frequency, the system performs periodic alignment analysis on the influence coefficients of each level. When the influence coefficient of a level deviates from the theoretical value of the dominant frequency by more than 0.35, the level is marked as a potential deviation point.

[0144] The stability of the feature transfer is characterized by the dominant frequency, and the ratio of the amplitude of the dominant frequency to the sum of the amplitudes of all frequency components is calculated to obtain a dominant frequency proportion of 0.576. This proportion reflects the stability of the dominant mode in the feature transfer process. When the dominant frequency proportion is higher than 0.65, the system determines that the feature transfer has high stability, and the creation process shows strong regularity. When the dominant frequency proportion is between 0.45 and 0.65, the system determines that the feature transfer has moderate stability, and the creation process has some fluctuations. When the dominant frequency proportion is lower than 0.45, the system determines that the feature transfer has low stability, and the creation process shows irregularity. In this case, the stability is evaluated as moderate stability, and the stability score is 58.3. The system further analyzes the distribution characteristics of the non-dominant frequency components. When the number of non-dominant frequency components exceeds 5 and the amplitude distribution dispersion is greater than 0.6, the system determines that there is a multi-mode competition phenomenon, and the stability score is reduced by 15 percentage points.

[0145] Based on the dominant frequency identification of the deviation point in the creation process, the actual influence coefficient of each level is compared with the theoretical influence coefficient predicted by the dominant frequency. The actual influence coefficient of the second level is 0.65, and the theoretical coefficient predicted according to the periodicity of the dominant frequency should be 1.15. The difference between the two is 0.50, which exceeds the deviation threshold of 0.35, so the second level is identified as a deviation point. The system records the position information, deviation degree value, and deviation direction type of the deviation point. In this case, the deviation point is located at the second level, the deviation degree is negative deviation of 0.50, and the deviation direction is inhibitory deviation. The system constructs a deviation point influence range evaluation model to analyze the influence propagation depth of the deviation point on subsequent levels. When the influence of the deviation point decays within two levels to less than 30% of the initial deviation degree, it is determined to be a local deviation. When the influence of the deviation point still maintains more than 30% of the initial deviation degree in three or more levels, it is determined to be a diffusion type deviation.

[0146] The overall deviation degree index is calculated by weighting and accumulating the deviation degree values of all identified deviation points, and higher weight coefficients are given to deviation points located in the front-end level. The weight coefficient of the second level deviation point is set to 1.25. The system multiplies the deviation degree 0.50 of this deviation point by the weight coefficient 1.25 to obtain a weighted deviation degree of 0.625. The system also considers the adjustment effect of the stability score on the overall deviation degree, and subtracts the stability score 58.3 from 100 to obtain an instability index of 41.7. The instability index is converted into an adjustment factor of 1.417. The system multiplies the weighted deviation degree 0.625 by the adjustment factor 1.417 to obtain a final overall deviation degree of 0.886. The system sets a preset deviation threshold of 0.75. When the overall deviation degree value exceeds this threshold, the system determines that the creation process deviates significantly and needs to start the vector calibration mechanism for intervention and adjustment to ensure that the feature transfer returns to the expected trajectory.

[0147] In a second aspect, the embodiment of the present application provides a personalized image content generation and optimization system based on a fusion generative AI, comprising:

[0148] A first unit is configured to analyze a historical interaction trajectory and a current intention expression of a target user based on a deep perception network, and synchronously construct a multi-dimensional behavior portrait; based on the multi-dimensional behavior portrait, a user group with similar generation preferences is identified by using a group feature recursive quantization technology, and a distribution structure of generation guide features of each group in a high-dimensional space is depicted to construct a group wisdom feature map;

[0149] A second unit is configured to perform multi-level mapping analysis on user features contained in the historical interaction trajectory and the group wisdom feature map, anchor a group belonging relationship of the target user, and based on a deep feature perspective technology, the guide features corresponding to the belonging group are visualized;

[0150] A third unit is configured to construct a multi-level guide vector according to the current intention expression and the guide features, and perform accurate mapping on a hidden space representation of an image generation model, to quantize a deviation degree of a generation result and the guide vector in real time, and when the deviation exceeds a preset deviation threshold, trigger vector intelligent calibration, and output a generation result meeting double constraints;

[0151] A fourth unit is configured to establish an association projection between the generation result and the group wisdom feature map, reconstruct a distribution structure of a corresponding group based on evaluation data of the target user, and realize dynamic evolution of features.

[0152] In a third aspect, the embodiment of the present application provides an electronic device, comprising:

[0153] A processor;

[0154] A memory for storing processor-executable instructions;

[0155] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0156] In a fourth aspect, the embodiment of the present application provides a computer-readable storage medium having computer program instructions stored thereon, and the computer program instructions are executed by a processor to implement the method described above.

[0157] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions loaded thereon for executing various aspects of the present application.

[0158] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions recorded in the above embodiments can be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A personalized image content generation and optimization method integrating generative AI, characterized in that, include: Based on deep perception networks, analyze the target user's historical interaction trajectory and current intent expression, and simultaneously construct a multi-dimensional behavioral profile; Based on the multidimensional behavioral profile, the group feature recursive quantization technique is used to identify user groups with similar generation preferences, and the distribution structure of the generation guidance features of each group in high-dimensional space is depicted to construct a group intelligence feature map. The user characteristics contained in the historical interaction trajectory are mapped and analyzed at multiple levels with the collective intelligence feature map to anchor the group affiliation relationship of the target user, and the guiding features corresponding to the affiliation group are visualized based on deep feature perspective technology, including: The user characteristics contained in the historical interaction trajectory are mapped and analyzed in a multi-level manner with the collective intelligence feature map. By constructing multi-order projection links in feature space, semantic space and topological space, the affinity relationship between user feature vectors and each group node is quantified respectively. Based on the affinity relationship, the affiliation probability distribution is constructed. The group affiliation relationship of the target user is anchored according to the dominant peak of the affiliation probability distribution. Based on the group affiliation relationship, locate the corresponding node in the group intelligence feature map, deconstruct the distribution parameters, evolution chain features and entropy labels contained in the corresponding node, and establish an abstract guiding feature representation of the affiliation group by generating a deep perspective of the guiding feature space; Based on the abstract guidance feature representation, the common generation preferences of the group are visualized as guidance parameters in the operational dimension. The dominant path deconstruction control nodes and parameter strategies in the evolutionary chain features are integrated to form a concrete expression of the guidance feature. Based on the current intent expression and the guidance features, a multi-level guidance vector is constructed and accurately mapped to the latent space representation of the image generation model. The deviation between the generated result and the guidance vector is quantified in real time. When the deviation exceeds the preset deviation threshold, vector intelligent calibration is triggered, and the generated result that satisfies the dual constraints is output. The generated results are correlated and projected with the collective intelligence feature map, and the distribution structure of the corresponding group is reconstructed based on the evaluation data of the target users, so as to realize the dynamic evolution of features.

2. The method according to claim 1, characterized in that, Based on the multi-dimensional behavioral profile, the recursive quantization technique of group features is used to identify user groups with similar generation preferences, and the distribution structure of the generation guidance features of each group in high-dimensional space is characterized to construct a collective intelligence feature map, including: Based on the multidimensional behavioral profile, the group feature recursive quantization technology is used to identify user groups with similar generation preferences. By depicting the evolution chain of behavioral trajectories, the resonance relationship and entropy distribution between trajectories are captured. The periodic commonalities and turning points of user behavior patterns in the spatiotemporal dimension are dynamically analyzed. Based on the natural community formed in the feature space by the resonance relationship, the boundary contour and internal cohesion of the user group are determined. Based on the boundary contour and the internal cohesion, the feature centroid and behavior spectrum of each user group in the multi-dimensional behavior profile space are depicted, and the dominant tendency and energy distribution ratio of each group in the generation guidance dimension are deconstructed. A high-dimensional distribution structure of the user group generation guidance features is constructed, and the high-dimensional distribution structure is mapped into a collective intelligence feature map.

3. The method according to claim 2, characterized in that, Based on the aforementioned multidimensional behavioral profile, recursive quantization technology of group features is used to identify user groups with similar generation preferences. By depicting the evolutionary chain of behavioral trajectories, the resonance relationship and entropy distribution between trajectories are captured, including: Based on the multidimensional behavioral profile, the recursive quantization technique of group features is used to identify user groups with similar generation preferences, construct a recursive graph matrix of high-dimensional phase orbits, and obtain the regression pattern and self-similar structure of the orbit in phase space through the quantization of the recursive graph matrix. The evolution chain and bifurcation features of the user behavior trajectory are jointly depicted according to the regression pattern and the self-similar structure. Based on the evolutionary chain and the bifurcation feature, the diagonal structure alignment and texture similarity between the recursive graph matrices of different users are calculated, and the resonance relationship between trajectories is captured by the combination analysis of the diagonal structure alignment and texture similarity. For user groups with significant resonance relationships, the Shannon entropy and permutation entropy of their recursive graph matrix are calculated respectively, and the entropy distribution of the group is obtained by mapping the Shannon entropy and permutation entropy in the feature space.

4. The method according to claim 1, characterized in that, By constructing multi-order projection links in feature space, semantic space, and topological space, the affinity relationships between user feature vectors and nodes of each group are quantified respectively. Based on these affinity relationships, an affiliation probability distribution is constructed, including: By constructing a multi-order projection link of feature space, semantic space and topological space, feature affinity is obtained by depicting the affinity and resonance degree between user feature vectors and the distribution parameters of each group node in the feature space. Semantic affinity is obtained by characterizing the semantic representation of user feature vectors and the cognitive synergy strength between the dominant and guiding features of each group node in the semantic space. Topological affinity is obtained by constructing the manifold distance of user feature vectors in the collective intelligence feature graph structure and the coupling depth of the topological centrality of each group node in the topological space. The entropy labels of each group node reveal the expression tendencies of user feature vectors at different cognitive levels. The inherent consistency of the entropy labels is used to construct a collaborative mechanism of feature, semantic and topological affinity. Based on the collaborative mechanism, the contribution ratio of each spatial affinity in different cognitive dimensions is guided to obtain a complete affinity relationship picture. The affinity relationship picture is condensed into an affiliation probability distribution in the probability space, so that the potential energy difference between adjacent groups is formed in the affiliation determination process, eliminating the fuzzy area of ​​the group boundary and obtaining an affiliation probability distribution with a clear dominant orientation.

5. The method according to claim 1, characterized in that, A multi-level guiding vector is constructed and accurately mapped to the latent space representation of the image generation model. The deviation between the generated result and the guiding vector is quantified in real time. When the deviation exceeds a preset deviation threshold, intelligent vector calibration is triggered. The output of the generated result that satisfies dual constraints includes: A multi-level guiding vector is constructed, and a deep resonance relationship is built based on the multi-level guiding vector and the latent space representation of the image generation model to drive the image generation model to form an intermediate generation result; The hidden layer feature representation is condensed from the intermediate generated results, and the hidden layer feature representation is projected onto the multi-dimensional creative space constructed by the multi-level guiding vector to depict the image projection of each level. Based on the degree of creative deviation between the intermediate generated results and the multi-level guiding vectors, a hierarchical resonance intensity is constructed based on the inter-level dependency relationship. The hierarchical resonance intensity is harmonized to obtain the overall deviation. When the overall deviation exceeds the preset deviation threshold, the vector calibration mechanism is activated. The vector calibration mechanism is executed to locate the dominant deviation level and harmonize the corresponding guiding vector. Based on the stability index constraint, the harmonization amplitude is constrained, and the multi-level guiding vector is reconstructed. The reconstructed multi-level guiding vector is then used to establish a new resonance relationship with the latent space representation. The deviation perception and calibration process is executed cyclically until the overall deviation reaches the expected harmonious state, and the generation result that satisfies the dual constraints is output.

6. The method according to claim 5, characterized in that, A hierarchical resonance intensity is constructed based on the inter-level dependencies. The overall deviation is obtained by harmonizing the hierarchical resonance intensity. When the overall deviation exceeds a preset deviation threshold, a vector calibration mechanism is activated, including: The hierarchical resonance intensity is constructed based on the dependency relationship between feature vectors at each level. The information transmission activity is revealed by the information flow entropy ratio between hierarchical features according to the hierarchical resonance intensity. The activity reflects the transmission depth and diffusion law of features between levels. The guiding influence between levels is described based on the transmission depth and diffusion law. The dynamic evolution chain of hierarchical features is constructed based on the guiding influence. Based on the dynamic evolution chain, the hierarchical resonance intensity is harmonic deconstructed. By capturing the dominant frequency of the resonance fluctuation, the stability of the characteristic transmission is characterized. Based on the dominant frequency, deviation points in the creation process are identified. When the overall deviation caused by the deviation point exceeds the preset deviation threshold, the vector calibration mechanism is activated.

7. A personalized image content generation and optimization system integrating generative AI, used to implement the method of any one of claims 1-6, characterized in that, include: The first unit is used to analyze the target user's historical interaction trajectory and current intent expression based on deep perception network, and simultaneously build a multi-dimensional behavioral profile. Based on the multidimensional behavioral profile, the group feature recursive quantization technique is used to identify user groups with similar generation preferences, and the distribution structure of the generation guidance features of each group in high-dimensional space is depicted to construct a group intelligence feature map. The second unit is used to perform multi-level mapping and analysis of the user characteristics contained in the historical interaction trajectory and the collective intelligence feature map, anchor the group affiliation relationship of the target user, and visualize the guiding features corresponding to the affiliation group based on deep feature perspective technology. The third unit is used to construct a multi-level guidance vector based on the current intention expression and the guidance features, and to accurately map the latent space representation of the image generation model. It quantifies the degree of deviation between the generated result and the guidance vector in real time. When the deviation exceeds the preset deviation threshold, it triggers intelligent vector calibration and outputs the generated result that satisfies the dual constraints. The fourth unit is used to establish a correlation projection between the generated result and the collective intelligence feature map, and to reconstruct the distribution structure of the corresponding group based on the evaluation data of the target user, so as to realize the dynamic evolution of the features.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images