Search

CN-121999086-A - Personalized design generation method, system, equipment and medium based on user portrait

CN121999086ACN 121999086 ACN121999086 ACN 121999086ACN-121999086-A

Abstract

The application relates to a personalized design generation method, a system, equipment and a medium based on user portraits. The method comprises the steps of analyzing historical interaction data of a target user to generate a structured user preference vector, constructing a dual-path evaluation model, generating an initial design image through a conditional diffusion model based on the structured user preference vector and design task text description, inputting the initial design image into the dual-path evaluation model to obtain a corresponding subjective matching degree score and objective quality score, calculating a comprehensive reward score by combining a calculated fusion weight coefficient, updating the conditional diffusion model with the maximized comprehensive reward score as a target to obtain a personalized design generation model, and generating a personalized design image through the personalized design generation model and a design request of the target user. The method can effectively solve the core problems that subjective preference and objective quality are difficult to quantitatively align and dynamically balance during personalized generation in the prior art.

Inventors

  • DUAN YANTAO
  • TAN KUN

Assignees

  • 北京科技职业大学

Dates

Publication Date
20260508
Application Date
20260126

Claims (9)

  1. 1. The personalized design generating method based on the user portrait is characterized by comprising the following steps: Analyzing historical interaction data of a target user to generate a structured user preference vector, wherein the historical interaction data comprises explicit selection, scoring, editing instructions and implicit behavior feedback of the user on a design image; Constructing a dual-path evaluation model, wherein the dual-path evaluation model comprises a subjective preference matching degree model and an objective design quality model; based on the structured user preference vector and the design task text description, generating an image through a conditional diffusion model to obtain an initial design image; inputting the initial design image into the dual-path evaluation model to obtain a corresponding subjective matching degree score and objective quality score; According to the structured user preference vector and the design task text description, calculating to obtain a fusion weight coefficient; Based on the fusion weight coefficient, carrying out weighted fusion on the subjective matching degree score and the objective quality score to generate a comprehensive rewarding score; Iterative updating is carried out on the conditional diffusion model with the aim of maximizing the comprehensive rewards points, so that a personalized design generating model is obtained; and generating a personalized design image based on the personalized design generation model and the design request of the target user.
  2. 2. The method of claim 1, wherein parsing the historical interaction data of the target user to generate a structured user preference vector, wherein the historical interaction data includes explicit user selections, scoring, editing instructions, and implicit behavioral feedback of the design image, comprises: extracting design images displayed by each interaction, explicit action data of the target user and implicit feedback data obtained through behavior analysis from historical interaction data of the target user, and summarizing the design images, the explicit action data and the implicit feedback data to form an interaction sequence, wherein the interaction sequence is arranged in time sequence; coding each design image in the interaction sequence and each explicit action text description in the explicit action data to obtain an image feature vector and an action text feature vector, and carrying out normalization processing on the implicit feedback data to obtain an implicit feedback scalar; Combining the image feature vector, the action text feature vector and the implicit feedback scalar to obtain a sequence feature; performing time sequence modeling on the sequence features to generate comprehensive implicit expression of a user; And performing multidimensional preference feature mapping on the comprehensive implicit representation of the user to generate a structured user preference vector, wherein the dimension of the structured user preference vector represents interpretable design attribute preference intensity.
  3. 3. The method of claim 1, wherein the constructing a dual path assessment model comprising a subjective preference matching model and an objective design quality model comprises: the method comprises the steps of constructing a subjective preference matching degree model, wherein the subjective preference matching degree model comprises an image tower, a preference tower and a regression head, the image tower is used for encoding an input design image into an image feature vector, the preference tower is used for encoding the structured user preference vector into a preference feature vector, and the regression is used for processing the similarity of the image feature vector and the preference feature vector to generate a subjective matching degree score; constructing a triplet training data set based on historical interaction data of all users and structured user preference vectors of all the users; Training the subjective preference matching degree model through a pairwise ordering loss function based on the triplet training data set to obtain a trained subjective preference matching degree model; Pre-training the deep convolutional neural network through a large-scale professional design gallery to obtain a trained deep convolutional neural network, wherein the pre-training task comprises contrast learning and multitask prediction of design images; carrying out weighted summation on the multidimensional quality attribute prediction result of the trained deep convolutional neural network to obtain a comprehensive score; training the lightweight regression network based on the comprehensive score to obtain a trained lightweight regression network, wherein the trained lightweight regression network is set as an objective design quality model; And integrating the trained subjective preference matching degree model and the objective design quality model to obtain a dual-path evaluation model.
  4. 4. The method of claim 1, wherein said calculating a fused weight coefficient from said structured user preference vector and said design task text description comprises: Extracting semantic features of the design task text description, and generating fusion features based on the semantic features and the structured user preference vector; and inputting the fusion characteristics into a pre-trained dynamic weight prediction network and outputting fusion weight coefficients, wherein the dynamic weight prediction network is composed of a plurality of layers of perceptrons.
  5. 5. The method of claim 1, wherein the weighting the subjective matching score and the objective quality score based on the fusion weight coefficient to generate a composite bonus score comprises: based on the fusion weight coefficient, weighting and fusing the subjective matching degree score and the objective quality score through the following formula to generate a comprehensive rewarding score; Wherein, the For the composite bonus points described, For the said fusion weight coefficient(s), For the subjective matching degree score, For the objective quality score of the object, For a predetermined divergence penalty factor, Is KL divergence.
  6. 6. The method of claim 1, wherein iteratively updating the conditional diffusion model with the goal of maximizing the composite bonus score results in a personalized design generation model, comprising: Determining the current conditional diffusion model as a strategy network to be optimized, and generating a design image through the strategy network to be optimized for a received design request, wherein the design request comprises the structured user preference vector of the target user and the design task text description; Calculating the comprehensive rewards points of each design image, and calculating a dominance function estimated value in strategy gradient update based on the comprehensive rewards points; Calculating strategy loss according to the dominance function estimated value, and carrying out back propagation update on the parameters of the strategy network to be optimized based on the strategy loss to obtain updated parameters; And repeating the steps from generating the design image to generating the updated parameters based on the updated parameters until the performance of the strategy network to be optimized converges, so as to obtain a personalized design generation model.
  7. 7. A personalized design generation system based on a representation of a user, the system comprising: The system comprises a vector generation module, a target user selection module and a user selection module, wherein the vector generation module is used for analyzing historical interaction data of the target user to generate a structured user preference vector, and the historical interaction data comprises explicit selection, scoring, editing instructions and implicit behavior feedback of the user on a design image; the model construction module is used for constructing a dual-path evaluation model, wherein the dual-path evaluation model comprises a subjective preference matching degree model and an objective design quality model; The image generation module is used for generating an image through a conditional diffusion model based on the structured user preference vector and the design task text description to obtain an initial design image; the score calculation module is used for inputting the initial design image into the dual-path evaluation model to obtain a corresponding subjective matching degree score and objective quality score; the weight calculation module is used for calculating to obtain a fusion weight coefficient according to the structured user preference vector and the design task text description; The score fusion module is used for carrying out weighted fusion on the subjective matching degree score and the objective quality score based on the fusion weight coefficient to generate a comprehensive rewards score; The model updating module is used for carrying out iterative updating on the conditional diffusion model with the aim of maximizing the comprehensive rewards points to obtain a personalized design generating model; And the personalized image generation module is used for generating a personalized design image based on the personalized design generation model and the design request of the target user.
  8. 8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method of any of claims 1 to 6 when the computer program is executed.
  9. 9. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method of any of claims 1 to 6.

Description

Personalized design generation method, system, equipment and medium based on user portrait Technical Field The invention belongs to the technical field of artistic design generation, and particularly relates to a personalized design generation method, system, equipment and medium based on user portraits. Background With the rapid development of the technology for generating content by artificial intelligence, particularly the breakthrough progress of a diffusion model in the field of image generation, automatic design generation is an important auxiliary tool in the fields of planar design, marketing material manufacture, user interface conception and the like. The tools can quickly generate a large number of design sketches or schemes according to natural language descriptions input by users, greatly improve creation efficiency and reduce the threshold of professional design. On this basis, in order to further improve the practicability and user satisfaction of the generated result, the industry is focused on how to make the generated design not only meet general aesthetic, but also fit the unique taste and preference of the specific user, namely, realize "personalized design generation". Existing personalization methods either focus on generating works with general aesthetic quality from text cues or attempt to personalize with a small number of user examples or feedback. However, these methods have a fundamental limitation in that the "personalization" and "specialization" of the generation are difficult to combine. First, existing methods rely heavily on lagging, unquantified user feedback (e.g., final selection or scoring), and lack of a quantitative model to evaluate "user preference" in real-time during the generation process, results in generation with blindness. Second, the lack of explicit, quantifiable, specialized criteria modeling of what is a good design by the system can lead to poor quality, out-of-specification designs that cannot be actively circumvented to cater to some potentially immature, non-specialized preferences of the user (e.g., too-dazzling color schemes, chaotic layouts). When the subjective preferences of the user conflict with objective design criteria, the prior art lacks a dynamic, adaptive decision mechanism to intelligently weigh both, often only to deal with each other, either to generate professional but not user favorite designs or to generate user favorite but not professional designs. The problem of alignment of subjective preference and objective quality becomes a core bottleneck which restricts the trend of personalized design generation technology to practicality and specialization. Disclosure of Invention Based on the foregoing, it is necessary to provide a personalized design generating method capable of deeply fusing user personalized preferences and professional design criteria and realizing quantitative evaluation and dynamic trade-off of the user personalized preferences and the professional design criteria in the generating process, so as to produce design works with high professional level and satisfactory users. In a first aspect, the present application provides a personalized design generation method based on a user portrait, including: Analyzing historical interaction data of a target user to generate a structured user preference vector, wherein the historical interaction data comprises explicit selection, scoring, editing instructions and implicit behavioral feedback of the user on a design image; Constructing a dual-path evaluation model, wherein the dual-path evaluation model comprises a subjective preference matching degree model and an objective design quality model; Based on the structured user preference vector and the design task text description, generating an image through a conditional diffusion model to obtain an initial design image; inputting the initial design image into a dual-path evaluation model to obtain a corresponding subjective matching degree score and objective quality score; According to the structured user preference vector and the design task text description, calculating to obtain a fusion weight coefficient; Based on the fusion weight coefficient, weighting and fusing the subjective matching degree score and the objective quality score to generate a comprehensive rewarding score; Iterative updating is carried out on the conditional diffusion model by taking the maximized comprehensive rewards as a target to obtain a personalized design generation model; And generating a personalized design image based on the personalized design generation model and the design request of the target user. Further, historical interaction data of the target user is analyzed to generate a structured user preference vector, wherein the historical interaction data comprises explicit selection, scoring, editing instructions and implicit behavior feedback of the user on the design image, and the method comprises the following steps: extracting desi