Behavior perception-based image generation model dynamic fine-tuning method

Abstract

The application relates to the technical field of artificial intelligence image generation, and provides a behavior perception-based image generation model dynamic fine-tuning method. The method constructs a user-specific feature portrait by collecting user operation behavior data, dynamically adjusts the sampling step number, guide coefficient and denoising intensity core parameters of an image generation model according to the user-specific feature portrait, filters high-quality samples based on user effect feedback to perform lightweight incremental fine-tuning to generate a personalized adaptive model, and records the generation effect and feedback data to update the parameter adaptation strategy in a closed loop. The method has the beneficial effects of automatically adapting to the needs of different user groups, not requiring manual parameter adjustment, low personalized adaptation cost, and continuously optimized parameter adaptation capability.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence image generation technology, and in particular to a method for dynamic fine-tuning of an image generation model based on behavior perception. Background Technology

[0002] Existing AI image generation models typically use fixed model parameter configurations, which cannot adapt to the personalized needs of different user groups. Different user groups, such as game concept artists, commercial artists, and portrait retouchers, have significant differences in editing scenarios, operating habits, and generation preferences. Fixed parameter configurations cannot simultaneously meet the generation effect requirements of multiple user groups.

[0003] To improve the generation effect, some solutions allow users to manually adjust core parameters such as sampling steps, guiding coefficients and denoising intensity. However, manual adjustment relies on the user's professional experience, and ordinary users find it difficult to obtain the optimal parameter configuration. Moreover, the adjustment needs to be repeated for each generation task, resulting in low operation efficiency. Summary of the Invention

[0004] This application provides a behavior-aware image generation model dynamic fine-tuning method, which aims to solve the problem that the personalized adaptation of existing image generation models mainly relies on large-scale full-scale fine-tuning, which has high training costs and long cycles, and cannot quickly generate a custom style model based on the historical generation data and effect feedback of a single user.

[0005] In a first aspect, embodiments of this application provide a method for dynamically fine-tuning an image generation model based on behavior awareness, the method comprising: Collect user action data during the image generation process, extract user group type, editing scenario, operation habits and generation preference features based on the action data, and construct a user-specific feature profile; based on the user-specific feature profile, dynamically adjust the sampling step number, guidance coefficient and denoising intensity core parameters of the image generation model, and execute the image generation task based on the adjusted core parameters; Collect user feedback data on the effect of the generated image, select high-quality samples from the user's historical generated images based on the feedback data, input the high-quality samples into the image generation model to perform lightweight incremental fine-tuning, and generate a personalized adaptation model that fits the user's exclusive style. Record the generated effect data and corresponding effect feedback data under different core parameter configurations, update the parameter adaptation strategy based on the generated effect data and effect feedback data, and use the updated parameter adaptation strategy for the next round of dynamic adjustment of core parameters.

[0006] In some embodiments, the collection of user operation behavior data during the image generation process includes: collecting user operation behavior data in real time through the operation log interface of the image generation platform. The operation behavior data includes user-selected model type data, input prompt word data, parameter adjustment record data, editing operation record data, generation count data, and generation duration data; and performing structured storage of the operation behavior data according to user identifier and timestamp to generate a user operation behavior sequence.

[0007] In some embodiments, the step of extracting user group type, editing scenario, operation habits, and generation preference features based on operation behavior data includes: calculating user professional skill indicators based on parameter adjustment record data and editing operation record data in the operation behavior data, and determining user group type based on professional skill indicators; determining editing scenario based on prompt word data in the operation behavior data and the theme and purpose of the generated image; calculating parameter adjustment frequency and parameter adjustment range based on parameter adjustment record data in the operation behavior data, and determining operation habits based on parameter adjustment frequency and parameter adjustment range; and extracting keyword distribution features based on prompt word data in the operation behavior data, and determining generation preferences based on keyword distribution features and user-selected model type data.

[0008] In some embodiments, constructing a user-specific feature profile includes: performing structured encoding on user group type, editing scenario, operation habits, and generation preference features to generate user group type labels, editing scenario labels, operation habit vectors, and generation preference vectors; combining user group type labels, editing scenario labels, operation habit vectors, and generation preference vectors into a user-specific feature profile; storing the user-specific feature profile in a user profile database, and establishing a corresponding mapping relationship between user identifiers and user-specific feature profiles.

[0009] In some embodiments, dynamically adjusting the sampling steps, guidance coefficient, and denoising intensity core parameters of the image generation model based on the user-specific feature profile includes: retrieving the user-specific feature profile of the corresponding user from the user profile database; determining the baseline value range of the sampling steps, guidance coefficient, and denoising intensity core parameters based on the user group type label and editing scene label in the user-specific feature profile; performing fine-grained offset adjustment within the baseline value range based on the operation habit vector and generation preference vector in the user-specific feature profile; and outputting the adjusted sampling steps, guidance coefficient, and denoising intensity core parameters.

[0010] In some embodiments, performing the image generation task based on the adjusted core parameters includes: loading the adjusted sampling steps, guiding coefficients, and denoising intensity core parameters into the inference configuration of the image generation model; receiving prompt words input by the user, performing image generation inference based on the prompt words and the loaded core parameters; outputting the generated image and recording the sampling steps, guiding coefficients, and denoising intensity core parameter configuration used in this generation task.

[0011] In some embodiments, collecting user feedback data on the generated image includes: setting an effect feedback collection entry point on the generated image display interface; collecting user rating data, collection operation data, regeneration operation data, and editing operation data on the generated image through the effect feedback collection entry point; and integrating the rating data, collection operation data, regeneration operation data, and editing operation data into effect feedback data.

[0012] In some embodiments, the step of selecting high-quality samples from the user's historical generated images based on performance feedback data, inputting the high-quality samples into the image generation model to perform lightweight incremental fine-tuning, and generating a personalized adaptation model adapted to the user's unique style includes: calculating the quality score of each image in the user's historical generated images based on the rating data and collection operation data in the performance feedback data; identifying user's historical generated images with quality scores exceeding a preset quality threshold as high-quality samples; forming fine-tuning training sample pairs with the prompt words corresponding to the high-quality samples; inputting the fine-tuning training sample pairs into the image generation model to perform lightweight incremental fine-tuning, updating the low-rank adaptation weight parameters of the image generation model; and encapsulating the updated low-rank adaptation weight parameters into a personalized adaptation model adapted to the user's unique style.

[0013] In some embodiments, recording the generated effect data and corresponding effect feedback data under different core parameter configurations, and updating the parameter adaptation strategy based on the generated effect data and effect feedback data, includes: establishing a corresponding mapping relationship between the core parameter configurations of sampling steps, guiding coefficients, and denoising intensity and the generated effect data; establishing a corresponding mapping relationship between the generated effect data and the effect feedback data; evaluating the adaptation effect of different core parameter configurations based on the effect feedback data, and generating an adaptation effect evaluation result; and updating the parameter adjustment rules in the parameter adaptation strategy based on the adaptation effect evaluation result.

[0014] In some embodiments, the step of using the updated parameter adaptation strategy for the next round of core parameter dynamic adjustment includes: loading the updated parameter adaptation strategy into the parameter dynamic adaptation engine; when the next round of image generation task starts, the parameter dynamic adaptation engine performs dynamic adjustment of the sampling steps, guiding coefficients, and denoising intensity core parameters according to the user-specific feature profile and the updated parameter adaptation strategy; and outputs the sampling steps, guiding coefficients, and denoising intensity core parameters used in the next round of image generation task.

[0015] The method provided in this application has the following beneficial effects: 1. By collecting user operation behavior data and constructing user-specific feature profiles, the system can automatically identify the editing scenarios, operating habits, and generation preferences of different user groups, providing accurate user feature basis for dynamic adaptation of model parameters.

[0016] 2. Based on the user's unique feature profile, the sampling steps, guidance coefficients, and denoising intensity of the image generation model are dynamically adjusted. The optimal parameter configuration that suits the user scenario can be obtained without the user having to make manual adjustments, thereby improving generation efficiency and generation effect.

[0017] 3. Based on user feedback, high-quality samples are selected for lightweight incremental fine-tuning to generate a personalized model that matches the user's unique style. Compared with full fine-tuning, this significantly reduces training costs and time, enabling rapid personalized adaptation at the single-user level.

[0018] 4. By recording the generation effects and user feedback under different core parameter configurations and updating the parameter adaptation strategy in a closed loop, the parameter adaptation capability is continuously optimized as users use the content, thereby continuously improving user satisfaction with the generated content.

[0019] 5. Integrate user behavior perception, dynamic adaptation of model parameters, personalized incremental fine-tuning, and closed-loop optimization mechanism into a complete technical solution to realize the transformation of image generation model from fixed parameter configuration to dynamic adaptation to each individual.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic flowchart illustrating the steps of a behavior-aware image generation model dynamic fine-tuning method provided in an embodiment of this application; Figure 2 This is a schematic diagram illustrating the principle of a dynamic fine-tuning method for an image generation model based on behavior awareness, provided in an embodiment of this application. Figure 3 This is a schematic block diagram of a dynamic fine-tuning system for an image generation model based on behavior awareness, provided in one embodiment of this application. Figure 4This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application.

[0023] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the order described. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0026] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.

[0027] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0028] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0029] Existing AI image generation models typically use fixed model parameter configurations, which cannot adapt to the personalized needs of different user groups. Different user groups, such as game concept artists, commercial artists, and portrait retouchers, have significant differences in editing scenarios, operating habits, and generation preferences. Fixed parameter configurations cannot simultaneously meet the generation effect requirements of multiple user groups.

[0030] To improve the generation effect, some solutions allow users to manually adjust core parameters such as sampling steps, guiding coefficients and denoising intensity. However, manual adjustment relies on the user's professional experience, and ordinary users find it difficult to obtain the optimal parameter configuration. Moreover, the adjustment needs to be repeated for each generation task, resulting in low operation efficiency.

[0031] Furthermore, the personalized adaptation of existing image generation models mainly relies on large-scale full-scale fine-tuning, which is costly and time-consuming, and cannot quickly generate a custom style model based on a single user's historical generation data and performance feedback. Existing solutions also lack a closed-loop optimization mechanism for parameter adaptation effects, and cannot continuously improve parameter adaptation strategies based on user feedback.

[0032] Please refer to Figure 1 This application provides a method for dynamically fine-tuning an image generation model based on behavior awareness, applied to computer devices. The computer devices can be deployed on a single server or a server cluster. They can also be deployed on handheld terminals, laptops, wearable devices, or robots, etc. It should be noted that all information involved in the method provided in this application is extracted with the authorization of the relevant user and in accordance with relevant regulations, and will not infringe on user privacy.

[0033] The provided dynamic fine-tuning method for behavior-aware image generation models includes steps S101 to S103. Details are as follows: Step S101. Collect user operation behavior data during the image generation process, extract user group type, editing scenario, operation habits and generation preference features based on the operation behavior data, and construct a user-specific feature profile; based on the user-specific feature profile, dynamically adjust the sampling step number, guidance coefficient and denoising intensity core parameters of the image generation model, and execute the image generation task based on the adjusted core parameters.

[0034] Specifically, this step is the initial data acquisition and parameter adaptation stage of the method. It is responsible for completing the perception of user operation behavior, the construction of user feature profiles, and the dynamic adjustment of the core parameters of the image generation model. It is the foundational step for realizing personalized image generation.

[0035] First, the system uses a behavior acquisition module deployed at the front end of the image generation platform to collect real-time data on all user actions during the image generation process. The behavior acquisition module non-intrusively monitors all interaction events between the user and the image generation platform's interface, including but not limited to actions such as clicking buttons, sliding sliders, inputting text, uploading images, selecting models, and switching tabs. Each interaction event is recorded as a structured behavior log, containing fields such as a unique user identifier, operation timestamp, operation type identifier, operation object identifier, pre-operation state value, post-operation state value, and operation duration. All behavior logs are arranged in timestamp order, forming a sequence of user actions, and stored in a distributed log storage system for subsequent feature extraction and profile construction.

[0036] While collecting operational behavior data, the system simultaneously collects metadata for each image generation task, including the user's input prompt text, negative prompt text, selected base model type and version number, sampler type used, set random seed value, resolution and aspect ratio of the generated image, and computational resources and time consumed in this generation task. This metadata is associated with the corresponding operational behavior data through a unique identifier for the generation task, ensuring that the context of each generation task can be fully reconstructed during subsequent analysis.

[0037] After collecting operational behavior data, the system performs data cleaning and preprocessing on the raw behavior logs. Data cleaning includes removing duplicate logs, correcting timestamp anomalies, and filtering invalid behavior data generated by test accounts and bot accounts. Preprocessing includes normalizing continuous numerical fields, performing one-hot encoding or label encoding on categorical fields, and performing word segmentation and vectorization on text fields. The cleaned and preprocessed behavior data has a unified data format and numerical range, and can be directly input into the feature extraction module for analysis.

[0038] The feature extraction module performs multi-dimensional feature extraction on the preprocessed operational behavior data, extracting user group type features, editing scenario features, operation habit features, and generation preference features. User group type features are determined by analyzing the complexity of users' parameter adjustment behavior and the professionalism of their editing operations. The system calculates indicators such as the number of times users actively adjust core parameters within a preset statistical period, the magnitude of each adjustment, and the frequency of using advanced editing functions. A comprehensive calculation of the user's professional skill score is then performed, classifying users into three groups based on the score range: novice users, intermediate users, and professional users. Editing scenario features are determined by analyzing the theme content of user prompts, the purpose tags of generated images, and the user's generation activity at different time periods. The system uses a pre-trained text classification model to perform theme classification on user prompts, identifying typical editing scenarios such as game concept art, commercial posters, portrait photography, concept design, and illustration creation. Operation habit features are extracted by statistically analyzing user parameter adjustment time distribution, adjustment order preferences, and commonly used parameter combinations. The system constructs a Markov chain model of user parameter adjustment behavior to capture the state transition patterns during parameter adjustment, forming a feature vector representing user operation habits. The generated preference features are extracted by analyzing the keyword distribution of user prompts, the frequency of use of style words, model selection preferences, and the visual style statistics of generated images. The system uses a topic model to perform topic mining on the user's historical prompts, identify the user's preferred art style, color tone, composition method, and image elements, and form a feature vector representing the user's generated preferences.

[0039] After extracting features across four dimensions, the system performs structured encoding and combination of these features to construct a user-specific feature profile. This profile uses a unified structured data format, including fields such as user group type tags, editing scenario tags, operation habit vectors, generation preference vectors, and profile update timestamps. The system stores the completed user-specific feature profiles in a user profile database, establishing a one-to-one mapping between the user's unique identifier and their profile. The user profile database employs a read-write separation architecture, supporting high-concurrency profile query requests and periodic profile update operations. When a user generates new operational behavior data, the system performs incremental updates to the user-specific feature profile according to a preset update cycle, ensuring that the profile always reflects the user's latest behavioral characteristics and preference changes.

[0040] After constructing a user-specific feature profile, the system dynamically adjusts the core parameters of the image generation model based on the profile content. These core parameters include three key parameters: sampling steps, guidance coefficient, and denoising intensity. These three parameters directly determine the inference process and output quality of image generation. The dynamic parameter adjustment process consists of two stages: The first stage involves determining the baseline range. Based on the user group type label and editing scenario label in the user-specific feature profile, the system queries a pre-built parameter baseline configuration table for the corresponding baseline value ranges of sampling steps, guidance coefficient, and denoising intensity. The parameter baseline configuration table is constructed through statistical analysis of a large amount of generated data from professional users, providing verified optimal parameter value ranges for different user groups and different editing scenarios. For example, for professional users editing game concept art, the baseline range for sampling steps is set to 30 to 50 steps, the baseline range for the guidance coefficient is set to 7.0 to 9.0, and the baseline range for denoising intensity is set to 0.6 to 0.8. For novice users editing portrait photos, the baseline range for sampling steps is set to 20 to 30 steps, the baseline range for the guidance coefficient is set to 5.0 to 7.0, and the baseline range for denoising intensity is set to 0.4 to 0.6. The second stage is fine-tuning the offset. Based on the user's unique feature profile, the system performs fine-tuning offset calculations within the baseline value range. The system inputs the operation habit vector and the generated preference vector into a pre-trained parametric regression model. The model outputs the specific offsets of the three core parameters within the baseline range. The final parameter values ​​are obtained by adding the median of the baseline range to the offsets. The parametric regression model uses a gradient boosting tree or deep neural network structure and is trained under supervision using historical generated data and user feedback data to learn the mapping relationship between user features and optimal parameter configurations.

[0041] After dynamically adjusting the core parameters, the system loads the adjusted sampling steps, guidance coefficients, and denoising intensity into the inference configuration of the image generation model. It receives prompts from the user, performs image generation inference based on the prompts and the loaded core parameters, outputs the generated image, and records all parameter configurations used in this generation task. The image generation model adopts a diffusion model architecture, generating images from random noise through iterative denoising. The sampling steps determine the total number of denoising iterations; more steps result in richer details but take longer. The guidance coefficient determines the strength of the prompts' guidance on the generation process; a higher coefficient results in a more closely matched output but may reduce diversity. The denoising intensity determines the degree of modification to the input image in image-to-image mode; a higher intensity results in a greater difference between the generated output and the input. The system performs inference based on the dynamically adjusted parameter configuration, ensuring that the parameter configuration for each generation task matches the current user's feature profile, thereby achieving personalized image generation effects.

[0042] Step S102. Collect user feedback data on the generated image effect, select high-quality samples from the user's historical generated images based on the feedback data, input the high-quality samples into the image generation model to perform lightweight incremental fine-tuning, and generate a personalized adaptation model that matches the user's exclusive style.

[0043] Specifically, this step is the core model adaptation stage of the method. It is responsible for collecting user feedback, selecting high-quality training samples, and performing lightweight incremental fine-tuning of the personalized adaptation model. It is a key step in realizing user-specific style modeling.

[0044] First, the system sets up multi-dimensional feedback collection entry points on the generated image display interface, collecting user feedback data on the generated images in a combination of explicit and implicit methods. Explicit feedback collection entry points include a star rating control (1 to 5 stars) below the image, a "like" button, a "favorite" button, and a "dislike" button. Users can directly express their satisfaction with the generated image by clicking these controls. Implicit feedback is inferred by analyzing users' subsequent actions after viewing the generated image, including whether the user regenerated the image, downloaded the image to their local computer, sent the image to the editing panel for secondary editing, the duration of time spent viewing the image, and whether they zoomed in to view image details. The system integrates explicit and implicit feedback data to construct a comprehensive feedback record for each generated image.

[0045] During the feedback data collection process, the system performs credibility verification and weight allocation on the feedback data. For explicit rating data, the system analyzes the historical rating distribution of users, identifies users with excessively extreme ratings or abnormal rating patterns, and appropriately reduces the weight of their rating data. For implicit feedback data, the system assigns different weights based on the confidence level of different user actions; for example, the weight of a favorite action is higher than that of a like action, and the weight of a download action is higher than the duration of engagement. The feedback data, after credibility verification and weight allocation, more accurately reflects users' true satisfaction with the generated images.

[0046] After collecting feedback data, the system filters high-quality samples from the user's historical generated images. The system calculates a comprehensive quality score for each image in the user's historical image library. This score considers multiple feedback dimensions, including explicit ratings, collection status, download behavior, editing behavior, regeneration behavior, and dwell time. Each dimension is weighted and summed according to preset weights to obtain the final quality score. Images with quality scores exceeding a preset quality threshold are identified as high-quality samples. High-quality samples represent images that are recognized by the user and meet their aesthetic preferences and generation needs. The preset quality threshold is dynamically adjusted based on the distribution of quality scores in the user's historical generated images to ensure that high-quality samples occupy a reasonable proportion (typically the top 10% to 20%) of all the user's historical images, guaranteeing both sample quality and quantity to meet fine-tuning needs.

[0047] After selecting high-quality samples, the system combines the corresponding prompt words and high-quality images into fine-tuning training sample pairs. Each training sample pair contains prompt word text (including positive and negative prompt words) and a corresponding high-quality image. The prompt words serve as model input, and the image serves as the training target. The system performs data augmentation processing on the training sample pairs, including random cropping, color jittering, and horizontal flipping of the images, and synonym replacement and word order adjustment of the prompt words, to expand the diversity of the training samples and prevent overfitting during fine-tuning. The data-augmented training sample pairs are divided into training and validation sets according to a preset ratio. The training set is used for updating model parameters, and the validation set is used to monitor changes in model performance during fine-tuning.

[0048] After preparing the training samples, the system performs lightweight incremental fine-tuning on the input image generation model using these training samples. Lightweight incremental fine-tuning employs low-rank adaptation technology. Without modifying the original weight parameters of the image generation model, low-rank adaptation matrices are inserted into the model's attention layer and feedforward neural network layer. Only the parameters of these low-rank adaptation matrices are trained, while the original model weights remain frozen. Since the rank of the low-rank adaptation matrix is ​​much smaller than the dimension of the original weight matrix, the number of parameters that need to be trained is only a very small fraction (usually less than 1%) of the total number of parameters in the original model, thus significantly reducing the computational resources and training time required for fine-tuning. The fine-tuning process uses a mini-batch gradient descent optimization algorithm with a small learning rate to avoid compromising the original model's generative ability. Training epochs are stopped early based on validation set performance. During fine-tuning, the system monitors the training loss and validation loss in real time. Training stops when the validation loss no longer decreases for several consecutive epochs, and the optimal low-rank adaptation weight parameters are saved.

[0049] After fine-tuning, the system encapsulates the updated low-rank adaptation weights into a personalized adaptation model tailored to the user's specific style. This personalized model is stored as an independent model file, containing low-rank adaptation weights, model configuration information, training sample statistics, and the model version number. The system establishes a mapping between the user's unique identifier and the personalized adaptation model, storing this mapping in the model management database. When the user subsequently initiates an image generation task, the system automatically loads the corresponding personalized adaptation model, merges the low-rank adaptation weights with the original model weights, and performs inference to ensure the generated result possesses the user's unique style characteristics. The personalized adaptation model supports version management; as high-quality user samples accumulate, the system periodically performs incremental fine-tuning, updating the personalized adaptation model version to ensure the model continuously adapts to the evolution of the user's style preferences.

[0050] Step S103. Record the generated effect data and corresponding effect feedback data under different core parameter configurations. Update the parameter adaptation strategy based on the generated effect data and effect feedback data, and use the updated parameter adaptation strategy for the next round of core parameter dynamic adjustment.

[0051] Specifically, this step is the closed-loop optimization stage of the method, responsible for recording the association between parameter configuration and generated results, evaluating and updating parameter adaptation strategies, and distributing and applying the updated strategies. It is a key step in realizing the continuous evolution of parameter adaptation capabilities.

[0052] First, the system establishes a complete parameter-effect correlation record for each image generation task. The record includes the number of sampling steps, guidance coefficients, and denoising intensity core parameter configurations used in this generation task; the base model type and sampler type used; the input prompt text and random seed value; objective quality evaluation indicators of the generated image (including image sharpness, color richness, composition rationality, prompt relevance, etc., automatically calculated by a pre-trained image quality evaluation model); and user feedback data on the generated image (including explicit ratings, favorites, downloads, edits, regenerations, etc.). The system associates parameter configurations, objective effect indicators, and user feedback data through a unique identifier for the generation task, forming a complete parameter-effect mapping record, which is stored in the parameter-effect database. The parameter-effect database adopts a columnar storage architecture, supporting efficient querying and statistical analysis by user identifier, parameter configuration range, effect rating range, and other dimensions.

[0053] As the generation tasks continue to execute, a large number of parameter-effect mapping records accumulate in the parameter effect database. The system periodically performs statistical analysis on the records in the database, establishing a mapping relationship between the core parameter configurations of sampling steps, guidance coefficients, and denoising intensity and the generated effect data, as well as a mapping relationship between the generated effect data and the effect feedback data. Specifically, the system uses the combination of the three core parameters as the grouping dimension, and statistically analyzes indicators such as the average objective quality score, average user satisfaction score, average generation task time, and user regeneration rate of the generated images under each parameter combination, forming a parameter configuration-effect evaluation comparison table. By analyzing this comparison table, the system can identify which parameter combinations produce the best generation effect for specific user groups and specific editing scenarios, and which parameter combinations have problems with poor effect or low efficiency.

[0054] Based on statistical analysis, the system evaluates the adaptation effect of different core parameter configurations according to feedback data and generates adaptation effect evaluation results. The evaluation process employs a multi-dimensional comprehensive evaluation method, comprehensively scoring each parameter configuration range from three dimensions: generation quality, user satisfaction, and computational efficiency. Generation quality is measured by a weighted average of objective image quality indicators; user satisfaction is measured by a comprehensive indicator of user ratings, collection rate, download rate, and regeneration rate; and computational efficiency is measured by generation time and computational resource consumption. The three dimensions are weighted and summed according to preset weights to obtain the comprehensive adaptation effect score for the parameter configuration. The system compares the comprehensive adaptation effect score with a preset pass / fail threshold to identify parameter configuration ranges with excellent adaptation effects and those with poor adaptation effects. For parameter configurations with excellent adaptation effects, the system increases their recommendation priority in the parameter adaptation strategy; for parameter configurations with poor adaptation effects, the system decreases their recommendation priority or removes them from the strategy.

[0055] Based on the adaptation effect evaluation results, the system updates the parameter adjustment rules in the parameter adaptation strategy. The parameter adaptation strategy, in the form of a decision tree or rule engine, defines recommended values ​​or ranges for sampling steps, guidance coefficients, and denoising intensity under different user group types, editing scenarios, operating habits, and generation preferences. The strategy update process includes: adjusting the branch conditions of each node in the decision tree and the recommended parameter values ​​of the leaf nodes based on the latest adaptation effect evaluation results; adding efficient parameter configuration rules newly discovered in statistical analysis; removing old rules that were confirmed to be ineffective; and adjusting the weight coefficients of different feature dimensions in parameter calculation. The strategy update adopts a combination of offline batch updates and online incremental updates. Offline batch updates perform full strategy retraining during daily off-peak hours, while online incremental updates trigger local rule adjustments in real time when significant changes in effect are detected. The updated parameter adaptation strategy is officially released after gray-scale validation. During the gray-scale validation phase, the new strategy is applied to a portion of the user group to compare the generation effect differences between the new and old strategies. Once it is confirmed that the new strategy is no less effective than the old strategy, it is released to the full user base.

[0056] After the strategy update is complete, the system loads the updated parameter adaptation strategy into the dynamic parameter adaptation engine. The dynamic parameter adaptation engine is a runtime module that performs dynamic adjustments to core parameters. It integrates the latest parameter adaptation strategy, a user-specific feature profile query interface, and a parameter regression inference model. When the next image generation task starts, the dynamic parameter adaptation engine first retrieves the corresponding user's user-specific feature profile from the user profile database based on the user identifier. Then, it inputs the profile features into the updated parameter adaptation strategy, performs dynamic adjustment calculations for the core parameters—sampling steps, guidance coefficients, and denoising intensity—and outputs the core parameter configuration for the next image generation task. The dynamic parameter adaptation engine supports millisecond-level parameter inference response, ensuring that the startup speed of the image generation task is not affected. Simultaneously, the engine records the decision basis and output results for each parameter adjustment for subsequent strategy evaluation and optimization.

[0057] Through the closed-loop execution of the above three steps, the system realizes a complete closed loop from user behavior perception to dynamic parameter adaptation, from effect feedback collection to personalized model fine-tuning, and from effect statistical analysis to strategy continuous optimization. This enables the image generation model to continuously evolve based on each user's unique behavioral characteristics and style preferences, providing a precisely adapted personalized image generation experience for different user groups.

[0058] In some embodiments, the collection of user operation behavior data during the image generation process includes: collecting user operation behavior data in real time through the operation log interface of the image generation platform. The operation behavior data includes user-selected model type data, input prompt word data, parameter adjustment record data, editing operation record data, generation count data, and generation duration data; and performing structured storage of the operation behavior data according to user identifier and timestamp to generate a user operation behavior sequence.

[0059] This embodiment further details the specific implementation of the operational behavior data collection process, as follows: The front-end interaction layer of the image generation platform integrates a behavior acquisition software development kit (SDK). This kit listens to all user interaction events on the platform interface in a non-intrusive manner. Internally, the kit maintains an event listener registry, automatically binding event listeners to all interactive elements on the page (including buttons, sliders, input boxes, dropdown menus, tabs, image cards, etc.) without explicit calls from business logic. When a user performs any interaction, the corresponding event listener captures the event object, extracts information such as event type, target element identifier, event trigger time, mouse coordinates, and keyboard key presses, encapsulates it into a standardized behavior event object, and sends it to the local event queue.

[0060] The local event queue employs a first-in, first-out (FIFO) buffering mechanism to perform batch aggregation and throttling on frequently occurring behavioral events. For consecutively triggered slider drag events, the toolkit does not send them one by one, but instead aggregates them into a single parameter adjustment record according to a preset time window (e.g., 100 milliseconds), recording the slider's initial value, ending value, adjustment direction, and adjustment magnitude within that time window. For click events, the toolkit sends them immediately after the event is triggered, without aggregation. The local event queue also has offline caching capabilities. When the network connection is interrupted, behavioral events are temporarily stored in the browser's local storage and re-uploaded in batches after the network is restored, ensuring that behavioral data is not lost.

[0061] The collected operational behavior data specifically includes the following six categories: The first category is user-selected model type data, recording the basic model name, version number, fine-tuning model identifier, and sampler type chosen by the user in each generation task, reflecting the user's preference for different model architectures and generation styles; the second category is input prompt word data, recording the positive and negative prompt word texts entered by the user, the prompt word editing process, prompt word template usage, and prompt word length and language type, reflecting the user's expression habits and topic preferences; the third category is parameter adjustment record data, recording the user's adjustments to parameters such as sampling steps, guidance coefficient, denoising intensity, random seed, image resolution, and aspect ratio, including adjustment time, pre-adjustment value, post-adjustment value, and adjustment range, reflecting the user's parameter tuning habits and... The data is categorized into six categories: professional level, editing operation records, and generation frequency data. The fourth category records the secondary editing operations performed by the user after image generation, including partial redrawing, image magnification, style transfer, color adjustment, cropping and rotation, mask editing, etc., reflecting the user's post-processing needs and editing capabilities. The fifth category records the total number of image generation tasks initiated by the user per unit time, the number of generation times in a single session, the number of consecutive regenerations, and the generation frequency distribution under different editing scenarios, reflecting the user's activity level and the intensity of their generation needs. The sixth category records the generation time data, including the total time from submission to completion of each generation task, parameter configuration time, model inference time, result rendering time, and the user's dwell time on the generation result page, reflecting the user's sensitivity to generation efficiency and attention to the results.

[0062] After behavioral event objects are sent from the front end to the back end log receiving service, the receiving service performs format validation and deduplication on the events before writing them to a distributed message queue. The distributed message queue uses a high-throughput, low-latency message middleware as a buffer layer between behavioral data collection and subsequent processing, handling event surges caused by a large number of users operating simultaneously during peak periods. Consumer groups in the message queue consume behavioral events in real time and perform structured storage according to user identifiers and timestamps. Specifically, consumers sort behavioral events of the same user in ascending order of timestamps, assign a globally unique behavioral record number to each event, map the various fields of the event to corresponding columns in the structured storage table, and write the data to the user behavior time-series database. The user behavior time-series database uses a storage engine optimized for time-series data, supporting efficient range queries by user identifier and time range, and supporting automatic downsampling and expiration cleanup of data.

[0063] After structured storage, all user operation behavior data forms a time-ordered sequence of user operation behavior. Each element in the sequence represents a complete operation behavior, including behavior type, behavior object, behavior parameters, behavior timestamp, and behavior context identifier. The user operation behavior sequence is the foundational data for subsequent feature extraction and profile construction. The system supports various sequence operations, such as extracting subsequences by time window, filtering subsequences by behavior type, and grouping and aggregating by generated tasks, to meet the data needs of different analysis scenarios.

[0064] In some embodiments, the step of extracting user group type, editing scenario, operation habits, and generation preference features based on operation behavior data includes: calculating user professional skill indicators based on parameter adjustment record data and editing operation record data in the operation behavior data, and determining user group type based on professional skill indicators; determining editing scenario based on prompt word data in the operation behavior data and the theme and purpose of the generated image; calculating parameter adjustment frequency and parameter adjustment range based on parameter adjustment record data in the operation behavior data, and determining operation habits based on parameter adjustment frequency and parameter adjustment range; and extracting keyword distribution features based on prompt word data in the operation behavior data, and determining generation preferences based on keyword distribution features and user-selected model type data.

[0065] This embodiment further refines the specific implementation method of the multi-dimensional feature extraction step, as detailed below: User group type extraction is achieved through the calculation of professional skill indicators. The system first extracts parameter adjustment and editing operation records from user operation behavior sequences within a preset statistical period (e.g., the last 30 days), calculating the following sub-indicators: Parameter adjustment depth, which is the proportion of times the user actively adjusts the three core parameters (sampling steps, guiding coefficient, and denoising intensity) out of the total number of generation attempts; a higher proportion indicates a greater inclination towards fine-tuning parameters. Advanced function usage rate, which is the proportion of times the user uses advanced functions such as local redrawing, mask editing, control nets, and model fusion out of the total number of operations; a higher proportion indicates a greater mastery of professional skills. Editing operation complexity, which is the number of editing steps, the number of editing tools used, and the total editing time involved in each secondary editing operation; higher complexity indicates stronger post-processing capabilities. Average parameter adjustment amplitude, which is the average change amplitude of each parameter adjustment by the user; moderate and purposeful adjustments indicate that the user understands the meaning of the parameters, while random large adjustments indicate that the user is in the exploratory stage. The system weights and sums the above four sub-indicators according to preset weights to obtain a comprehensive user professional skill score ranging from 0 to 100. Based on the overall score range, users are divided into three groups: those with a score below 40 are novice users, characterized by rarely adjusting parameters, mainly using default configurations, and not using advanced functions; those with a score between 40 and 70 are intermediate users, characterized by occasionally adjusting some parameters, using some advanced functions, and having some editing skills; and those with a score above 70 are professional users, characterized by frequently and meticulously adjusting parameters, being proficient in using all advanced functions, and performing complex and purposeful editing operations.

[0066] The editing scenario extraction is achieved through prompt word theme classification and generation purpose analysis. The system extracts all prompt word data from the user's operation behavior sequence within a recent period and uses a pre-trained text classification model to perform theme classification on each prompt word. The text classification model is trained on a large-scale image-generated prompt word corpus and supports the recognition of dozens of typical editing scenario categories, such as game concept art, character design, scene concept, commercial posters, product advertisements, portrait photography, wedding photography, landscape photography, illustration creation, comic storyboards, icon design, interface materials, architectural visualization, and industrial design. The classification model outputs the probability distribution of each prompt word belonging to each category, and the system takes the category with the highest probability as the editing scenario label for that prompt word. At the same time, the system analyzes the metadata of the generated image and the user's subsequent operations to infer the specific purpose of the generated image. For example, if the user downloads the image and uses it for a presentation, it is determined to be for business purposes; if it is used for social media posting, it is determined to be for social purposes; and if it is used for further editing and creation, it is determined to be for creative purposes. The system combines the prompt word topic classification results with the usage analysis results to statistically analyze the frequency distribution of user generation in different editing scenarios. The editing scenario with the highest frequency is taken as the user's primary editing scenario tag, while secondary editing scenario tags are also recorded to form the user's editing scenario characteristics.

[0067] Operational habits are extracted through parameter adjustment behavior pattern analysis. The system extracts all parameter adjustment records from user operation behavior sequences and calculates the following behavioral pattern indicators: parameter adjustment frequency, i.e., the average number of times a user performs parameter adjustment operations per unit time, reflecting the frequency of parameter adjustment; parameter adjustment range, i.e., the difference between the maximum and minimum historical adjustment values ​​of each core parameter, reflecting the breadth of the user's parameter space exploration; parameter adjustment timing preference, i.e., whether the user prefers to preset parameters before generation or adjust parameters based on the results after generation, reflecting the user's workflow habits; commonly used parameter combinations, i.e., high-frequency co-occurring parameter combinations mined from the user's historical parameter configurations through frequent itemset mining algorithms, reflecting the user's fixed parameter adjustment patterns; and parameter adjustment order preference, i.e., the typical order in which the user adjusts multiple parameters (e.g., adjusting the sampling step count first and then the guiding coefficient), reflecting the user's operational process habits. The system encodes the above behavioral pattern indicators into a fixed-dimensional operational habit vector, where each dimension corresponds to a normalized value of a behavioral pattern indicator. The operational habit vector can quantitatively characterize the user's unique behavioral patterns in parameter adjustment, providing a basis for subsequent refined parameter adaptation.

[0068] Generative preference extraction is achieved through keyword distribution analysis of prompt words and model selection preference analysis. The system performs word segmentation and keyword extraction on the user's historical prompt word data. It uses a term frequency-inverse document frequency (IF-IVF) algorithm to calculate the importance weight of each keyword in the user's prompt word corpus, selecting the keywords with the highest weights to constitute the user's keyword distribution features. These features reflect the user's preferred image themes (e.g., "cyberpunk," "traditional Chinese style," "watercolor"), visual elements (e.g., "light and shadow," "reflection," "depth of field"), style (e.g., "realistic," "cartoon," "oil painting"), and quality modifiers (e.g., "high-definition," "rich in detail," "cinematic"). Simultaneously, the system statistically analyzes the frequency distribution of different model types selected by the user, calculating the user's preference for different base models. For example, whether the user prefers a realistic style model or an anime style model, or a general model or a domain-specific fine-tuned model. The system fuses and encodes the keyword distribution features and model selection preferences to generate the user's generative preference vector. This vector quantitatively represents the user's unique preferences in image generation content and style, providing a style-level basis for personalized model fine-tuning and parameter adaptation.

[0069] In some embodiments, constructing a user-specific feature profile includes: performing structured encoding on user group type, editing scenario, operation habits, and generation preference features to generate user group type labels, editing scenario labels, operation habit vectors, and generation preference vectors; combining user group type labels, editing scenario labels, operation habit vectors, and generation preference vectors into a user-specific feature profile; storing the user-specific feature profile in a user profile database, and establishing a corresponding mapping relationship between user identifiers and user-specific feature profiles.

[0070] This embodiment further details the specific implementation of the user-specific feature profile construction process, as follows: The structured coding process employs appropriate coding methods for each of the four feature dimensions. For user group type features, a category label coding method is used, mapping novice users, intermediate users, and professional users to integer labels 0, 1, and 2 respectively. Corresponding one-hot encoded vectors are generated for model inference scenarios requiring numerical input. For editing scenario features, a multi-label coding method is used. Since users may be involved in multiple editing scenarios simultaneously, the system assigns a binary bit to each typical editing scenario category. The bit corresponding to a user's involved scenario is set to 1, and the bit for an uninvolved scenario is set to 0, forming a fixed-length multi-label binary vector. The category label of the dominant editing scenario is retained for quick querying and rule matching. For operation habit features and generation preference features, a continuous vector coding method is used. Each behavioral pattern indicator and preference indicator is normalized to the 0-1 range and then concatenated in a fixed order into a fixed-length floating-point vector. The vector dimension is determined based on the number of indicators (usually 32 to 128 dimensions). The meaning and value range of each dimension in the vector are clearly defined in the user profile specification, ensuring the comparability of profile vectors from different users.

[0071] After encoding, the system combines the encoding results of the four dimensions into a unified user-specific feature profile structure. The user-specific feature profile adopts an extensible structured object format, containing the following fields: User Unique Identifier field, storing the user's globally unique identifier as the profile's primary key; Profile Version Number field, recording the current profile's data structure version, supporting backward compatibility upgrades; User Group Type Label field, storing the integer label and category name of the user group category; Editing Scene Label field, storing the primary editing scene label, a list of secondary editing scene labels, and a multi-label binary vector; Operation Habit Vector field, storing a fixed-length floating-point vector and metadata descriptions for each dimension of the vector; Generation Preference Vector field, storing a fixed-length floating-point vector and metadata descriptions for each dimension of the vector; Profile Statistics field, storing statistical information such as the time range of behavioral data used to build the profile, the total number of behavioral data entries, and the total number of generated tasks; Profile Quality Score field, storing the profile quality score calculated based on the richness and coverage of behavioral data; profiles with scores below a threshold are marked as low-quality profiles, and their confidence is reduced during parameter adaptation; Profile Creation Time field and Last Update Time field, recording the time of the first profile construction and the time of the most recent incremental update, respectively. All fields are encoded in binary or JSON format according to a unified serialization standard, which facilitates network transmission and persistent storage.

[0072] User-specific feature profiles are stored in a user profile database. The database employs a distributed key-value storage architecture, using the user's unique identifier as the key and the serialized user-specific feature profile as the value. Internally, the database implements read-write separation. Read nodes handle high-frequency profile query requests from the parameter adaptation engine, using memory caching to accelerate the reading of hot user profiles. Write nodes handle profile creation and incremental update operations, employing batch write strategies to reduce storage overhead. The database supports precise queries by user identifier, conditional queries by group type and editing scenario tags, and range queries by update time, meeting the data access needs of different business scenarios. The database also features profile lifecycle management capabilities, automatically archiving profiles of long-term inactive users to cold storage and automatically anonymizing or deleting profiles of deactivated users, ensuring both the economic efficiency of data storage and compliance with user privacy regulations.

[0073] The system establishes a one-to-one mapping between user identifiers and user-specific feature profiles. This mapping is stored in a user profile index table, with the user identifier as the primary key. The index table stores the corresponding profile's storage location in the database, its version number, last update time, and cache status. When the parameter matching engine needs to query a user's profile, it first queries the user profile index table to obtain the profile's storage location, and then reads the profile data from the database or cache. When a user profile is updated, the system synchronously updates the version number and update time in the index table and invalidates the corresponding cache, ensuring that subsequent queries retrieve the latest version of the profile. The establishment and maintenance of this mapping are centrally managed by the user profile management module. This module also provides full lifecycle management interfaces for profile creation, updates, queries, deletions, and quality assessments, ensuring the accuracy, timeliness, and availability of user-specific feature profiles.

[0074] In some embodiments, dynamically adjusting the sampling steps, guidance coefficient, and denoising intensity core parameters of the image generation model based on the user-specific feature profile includes: retrieving the user-specific feature profile of the corresponding user from the user profile database; determining the baseline value range of the sampling steps, guidance coefficient, and denoising intensity core parameters based on the user group type label and editing scene label in the user-specific feature profile; performing fine-grained offset adjustment within the baseline value range based on the operation habit vector and generation preference vector in the user-specific feature profile; and outputting the adjusted sampling steps, guidance coefficient, and denoising intensity core parameters.

[0075] This embodiment further details the specific implementation of the dynamic adjustment of core parameters, as follows: When the parameter dynamic adaptation engine receives an image generation task request, it first extracts the user's unique identifier from the request and retrieves the corresponding user-specific feature profile from the user profile database based on the user identifier. If the user is using the service for the first time and the corresponding profile does not exist in the profile database, the engine uses the system's default general profile as a substitute, and simultaneously triggers the profile building process to asynchronously create a unique profile for the user. After retrieving the profile, the engine first checks the profile quality score and the last update time. If the profile quality score is lower than a preset threshold or the update time exceeds the validity period (e.g., 90 days), the engine reduces the confidence weight of the profile in parameter calculation and triggers the profile refresh process within its lifecycle to ensure that parameter adaptation is based on high-quality profile data.

[0076] In the baseline value range determination phase, the engine extracts user group type tags and editing scenario tags from the user-specific feature profile and performs a joint query in a pre-built parameter baseline configuration table. The parameter baseline configuration table is a two-dimensional lookup table with user group type and editing scenario as the joint primary key. Each cell in the table stores the empirical baseline value range for three core parameters under the corresponding group and scenario combination, including the sampling step range, the guidance coefficient range, and the denoising intensity range. The parameter baseline configuration table is constructed through statistical analysis of massive historical generation data from the platform. Specifically, historical generation tasks are grouped by user group type and editing scenario; the parameter distribution used by generation tasks that obtained high user scores in each group is statistically analyzed; and the 25th to 75th percentiles of the distribution are used as the lower and upper limits of the baseline value range. The configuration table supports online updates. As platform data continues to accumulate, the system periodically re-statistically updates the baseline range of each cell to ensure that the baseline values ​​always reflect the optimal parameter practices for the current user group. For example, for a combination of professional users and game concept art editing scenarios, the baseline configuration table provides a sampling step range of 35 to 50 steps, a guiding coefficient range of 7.5 to 9.5, and a noise reduction intensity range of 0.65 to 0.85; for a combination of novice users and portrait photography editing scenarios, the baseline configuration table provides a sampling step range of 20 to 30 steps, a guiding coefficient range of 5.0 to 7.0, and a noise reduction intensity range of 0.4 to 0.6. After obtaining the baseline value range, the engine calculates the median of each parameter range as the baseline point for subsequent fine-tuning offsets.

[0077] In the fine-tuning phase, the engine extracts the operation habit vector and generation preference vector from the user's unique feature profile. These two vectors are concatenated into a joint feature vector, which is then input into a pre-trained parametric regression model. The model outputs the offsets of three core parameters relative to the baseline median. The parametric regression model employs a gradient boosting tree regressor or a multilayer perceptron neural network structure. The model's training data comes from triplets of user feature profiles, parameter configurations, and user feedback ratings from historical generation tasks on the platform. The training objective is to learn the mapping relationship between user features and the optimal parameter offsets, maximizing the generated user ratings under the predicted parameter configuration. Five-fold cross-validation is used to evaluate generalization performance during model training, and a Bayesian optimization algorithm is used to search for the optimal configuration of the model's hyperparameters. The model output offsets are continuous values, corresponding to the sampling step offset (in steps), the guiding coefficient offset (in coefficient values), and the denoising intensity offset (in intensity values). The engine adds the baseline median to the corresponding offset to obtain the initial adjustment value for each parameter. This initial adjustment value is then limited to the upper and lower limits of the baseline value range to ensure that the final parameters do not exceed the empirically validated safe range. For users with lower image quality, the engine performs scaling on the offset (e.g., multiplying by a confidence coefficient of 0.5) to make the parameters closer to the baseline median, reducing the risk of parameter adaptation bias caused by low-quality images.

[0078] After fine-tuning the offset, the engine outputs the final values ​​of three core parameters: sampling steps, guiding coefficients, and denoising intensity. Simultaneously, it records the decision log for this parameter adjustment, including the image version used, the obtained baseline interval, the offset output by the model, the final parameter values, and the confidence score. The decision log is stored in the parameter adaptation decision database for subsequent strategy evaluation and model optimization. The output core parameters are directly passed to the image generation model inference module as the parameter configuration for this generation task.

[0079] In some embodiments, performing the image generation task based on the adjusted core parameters includes: loading the adjusted sampling steps, guiding coefficients, and denoising intensity core parameters into the inference configuration of the image generation model; receiving prompt words input by the user, performing image generation inference based on the prompt words and the loaded core parameters; outputting the generated image and recording the sampling steps, guiding coefficients, and denoising intensity core parameter configuration used in this generation task.

[0080] This embodiment further details the specific implementation of the image generation task execution stage, as follows: After receiving the core parameters output by the parameter dynamic adaptation engine, the image generation model inference module first performs parameter loading and verification operations. Internally, the inference module maintains an inference configuration object containing all the parameters required for image generation inference. The module writes the received sampling steps, guiding coefficients, and denoising intensity into the corresponding fields of the inference configuration object, while simultaneously performing validity checks on the parameters: the sampling steps must be a positive integer and within the range of the model's minimum and maximum number of steps (e.g., 1 to 150 steps); the guiding coefficients must be non-negative real numbers within the range of 0 to 30; and the denoising intensity must be a real number within the range of 0 to 1. If a parameter exceeds the valid range, the inference module truncates the parameter to the boundary value and records a parameter anomaly log. After parameter loading is complete, the inference module configures the denoising scheduler based on the sampling steps, configures the guiding intensity of the classifier-less guiding module based on the guiding coefficients, and configures the initial noise injection ratio and the starting denoising steps in the image-to-image mode based on the denoising intensity, completing all preparations before inference.

[0081] After the parameters are loaded, the inference module receives the prompts input by the user. These prompts are submitted through an input box on the platform's front end and include both positive and negative prompts. The inference module performs preprocessing operations on the prompts, including text cleaning (removing extra spaces and special characters), length truncation (truncating parts exceeding the model's maximum text length), word segmentation and lexicalization (converting the text into a sequence of words that the model can process), and text encoding (converting the sequence of words into text embedding vectors using a text encoder). The preprocessed text embedding vectors, along with the inference configuration, are input into the image generation model, ready for inference execution.

[0082] Image generation inference employs an iterative denoising process using a diffusion model. The inference process begins with pure Gaussian noise (text-to-image mode) or noise added to the input image (image-to-image mode), and progressively denoises within a predetermined number of iterations according to a pre-defined denoising scheduler. In each denoising iteration, the model predicts noise based on the current noisy image, the current time step, and the text embedding vector, and then calculates the denoised image based on the predicted noise and the denoising scheduler. A classifier-free guidance module simultaneously performs conditional prediction (based on positive prompts) and unconditional prediction (based on empty prompts) at each step, weighting the two predictions according to a guidance coefficient to align the generated result with the direction described by the positive prompts. The denoising intensity determines how much original image information is retained in the image-to-image mode; higher intensity results in greater initial noise and a greater difference between the generated result and the input. During inference, the module monitors the denoising loss and numerical stability at each step in real time. When gradient explosion or numerical overflow is detected, gradient clipping and numerical clamping are performed to ensure stable completion of the inference process.

[0083] After inference is complete, the module outputs the generated image data. The output image undergoes post-processing operations, including color space conversion (converting from the model's internal color space to the standard sRGB color space), dynamic range adjustment (mapping the floating-point values ​​output by the model to an 8-bit integer range of 0 to 255), image format encoding (encoding to PNG or JPEG format), and metadata writing (writing information such as the prompts used in this generation, parameter configurations, model version, and random seed into the image's EXIF ​​metadata). The post-processed image is displayed to the user through the platform front-end and simultaneously stored in the user's personal image library.

[0084] While outputting the generated image, the module records all parameter configurations used in this generation task. The records include the specific values ​​of three core parameters: sampling steps, guidance coefficients, and denoising intensity, as well as all metadata such as the basic model type and version, sampler type, random seed, image resolution, aspect ratio, prompt text, negative prompts, generation time, and computational resource consumption. The module associates the parameter configurations and metadata with a unique generation task identifier and writes them into the generation task record database. The data in the generation task record database serves as the foundation for subsequent effect feedback collection, high-quality sample selection, and parameter adaptation strategy updates, ensuring that the parameter configurations and generation effects of each generation task can be fully traced and analyzed.

[0085] In some embodiments, collecting user feedback data on the generated image includes: setting an effect feedback collection entry point on the generated image display interface; collecting user rating data, collection operation data, regeneration operation data, and editing operation data on the generated image through the effect feedback collection entry point; and integrating the rating data, collection operation data, regeneration operation data, and editing operation data into effect feedback data.

[0086] This embodiment further details the specific implementation of the effect feedback data collection process, as follows: The generated image display interface employs a multi-dimensional, multi-layered feedback collection design to comprehensively gather user satisfaction feedback on the generated images without interfering with the user's normal browsing experience. Feedback collection entry points are divided into two categories: explicit feedback entry points and implicit feedback entry points.

[0087] The explicit feedback entry collects feedback data through user-initiated actions and includes the following components: a star rating control, displayed as five stars below the image, allowing users to express their satisfaction by clicking the corresponding number of stars, ranging from 1 star (very dissatisfied) to 5 stars (very satisfied). After a user submits a rating, the control displays the current rating and provides an option to edit it; a "like" and "favorite" button, where the "like" button expresses the user's approval of the image, and the "favorite" button saves the image to the user's personal favorites. Both buttons are toggle-style, highlighting the image and recording the operation time when clicked; quick feedback tags, providing preset tags such as "rich in detail," "satisfactory colors," "good composition," "deviation from prompts," "blurry," and "incompatible style," allowing users to select multiple tags to quickly express their specific evaluation of the image's strengths and weaknesses, reducing the operational cost of feedback; and a "dislike" button, which records the user's negative evaluation of the image and allows them to select specific reasons for dislike (such as incompatible content, poor quality, or incompatible style). The timing and location of the explicit feedback entry have been optimized for user experience. The star rating and like / favorite buttons are always displayed, and the quick feedback label will gradually appear in animation after the user stays for more than a preset time, avoiding premature display and visual interference.

[0088] Implicit feedback entry points infer user satisfaction by analyzing natural user actions after viewing the generated image, without requiring users to actively perform feedback actions. These include the following behavioral signals: Regeneration: Clicking the "Regenerate" button after viewing the image indicates dissatisfaction with the current result; the more times an image is regenerated, the lower the satisfaction. Download: Downloading the image to the user's local device indicates satisfaction with the image quality and a practical need for it; downloading is a strong positive feedback signal. Sending to Edit: Sending the image to the editing panel for further editing indicates the image is generally usable but requires local adjustments; this is a moderately positive feedback signal. Zooming in: [The user...] Clicking on an image to zoom in for details indicates user interest in the image content; the longer the zoom period, the higher the level of attention. Dwell time refers to how long a user stays on the image display page. A moderate dwell time (e.g., 10 to 60 seconds) indicates the user is carefully reviewing and the result is usable. Too short a dwell time suggests the user is unsatisfied and skips it, while too long a dwell time may indicate the user is hesitant or comparing. Page switching behavior—users immediately switching to other pages or closing tabs after viewing an image; combined with dwell time, this can infer satisfaction. Image comparison behavior—users switching back and forth between multiple generated results for comparison indicates they are selecting the best result; the image ultimately selected for download or collection is relatively optimal.

[0089] The system uses a front-end behavior acquisition toolkit to capture the aforementioned explicit and implicit feedback operations in real time. Each feedback operation record includes a user identifier, a generated task identifier, an image identifier, a feedback type, a feedback value, an operation timestamp, and an operation context (such as the current page state and the user's previous operation sequence). The feedback operation records are sent to the back-end feedback processing service, which performs cleaning, deduplication, and association operations on the records, aggregating multiple feedback records for the same image according to the image identifier.

[0090] After aggregation, the system integrates rating data, collection operation data, regeneration operation data, and editing operation data into a unified effect feedback data structure. The effect feedback data includes the following fields: unique image identifier, corresponding user identifier, corresponding generation task identifier, explicit rating value (1 to 5 stars, empty if no rating), whether liked (Boolean), whether collected (Boolean), quick feedback tag list, whether disliked (Boolean), list of reasons for dislike, number of regenerations, whether downloaded, whether sent to editor, number of zoomed-in views and total duration, browsing duration, and feedback data collection time. The integrated effect feedback data is stored in the user feedback database and associated with the corresponding generated image data and parameter configuration data through the generation task identifier, forming a complete four-tuple record of generation task-parameter configuration-generated image-effect feedback. This provides a data foundation for subsequent high-quality sample selection and parameter adaptation strategy optimization.

[0091] In some embodiments, the step of selecting high-quality samples from the user's historical generated images based on performance feedback data, inputting the high-quality samples into the image generation model to perform lightweight incremental fine-tuning, and generating a personalized adaptation model adapted to the user's unique style includes: calculating the quality score of each image in the user's historical generated images based on the rating data and collection operation data in the performance feedback data; identifying user's historical generated images with quality scores exceeding a preset quality threshold as high-quality samples; forming fine-tuning training sample pairs with the prompt words corresponding to the high-quality samples; inputting the fine-tuning training sample pairs into the image generation model to perform lightweight incremental fine-tuning, updating the low-rank adaptation weight parameters of the image generation model; and encapsulating the updated low-rank adaptation weight parameters into a personalized adaptation model adapted to the user's unique style.

[0092] This embodiment further details the specific implementation of the high-quality sample screening and lightweight incremental fine-tuning process, as follows: High-quality sample selection begins with the user's historical image library. The system retrieves all historical image feedback data from the user feedback database and calculates a comprehensive quality score for each image. The quality score uses a multi-dimensional weighted calculation method, specifically: Quality score equals the normalized value of the explicit score multiplied by the score weight, plus the collection status value multiplied by the collection weight, plus the download status value multiplied by the download weight, plus the positive editing flag multiplied by the editing weight, minus the normalized value of the number of regenerations multiplied by the regeneration penalty weight, and minus the dislike flag multiplied by the dislike penalty weight. The weights for each dimension are preset based on the confidence level of the feedback signal; for example, the collection weight is set to 0.3, the download weight to 0.25, the explicit score weight to 0.2, the regeneration penalty weight to 0.15, and the dislike penalty weight to 0.1. For images without explicit scores, the scoring dimensions are treated as neutral values, without affecting the calculation of other dimensions. The calculated quality score is normalized to the range of 0 to 1; a higher score indicates higher user satisfaction with the image.

[0093] The system presets a quality threshold (e.g., 0.7), identifying images with quality scores exceeding this threshold as high-quality candidate samples. To ensure sample quality, the system performs secondary filtering on the candidate samples: filtering out images with excessively low resolution or obvious artifacts (automatically detected by an image quality assessment model); filtering out samples with excessively short prompts (e.g., less than 5 words) or abnormal prompt content, ensuring that the prompts in the training samples have sufficient semantic information; for multiple high-quality images generated with the same prompt, the one with the highest quality score is selected as the representative sample to avoid duplicate samples causing bias in fine-tuning. The samples after secondary filtering are finally determined as high-quality samples. The number of high-quality samples typically accounts for 10% to 20% of all the user's historical images, ensuring both the purity of the sample quality and the number of samples required for fine-tuning (usually at least 20 to 50 high-quality samples are needed to obtain stable fine-tuning results).

[0094] After selecting high-quality samples, the system pairs the prompt words and high-quality images corresponding to each high-quality sample to form fine-tuning training sample pairs. Each sample pair contains three parts: positive prompt word text, negative prompt word text (if any), and the target image. The system performs data augmentation on the sample pairs to improve the fine-tuning effect: random cropping (cropping ratio not exceeding 10%, maintaining the integrity of the main content), color jittering (small random adjustments to brightness, contrast, and saturation, with an amplitude not exceeding 5%), and horizontal flipping (performed with a 50% probability for asymmetric sensitive images); and synonym replacement (using a pre-trained word vector model to select semantically similar replacement words), word order adjustment (adjusting the order of phrases in the prompt words while maintaining semantic integrity), and length variation (randomly adding or deleting some modifiers). After data augmentation, each original sample pair can generate multiple augmented sample pairs, effectively expanding the training set size. The system randomly divides the augmented sample pairs into training and validation sets in an 8:2 ratio. The training set is used for parameter updates, and the validation set is used to monitor overfitting and early stopping.

[0095] Lightweight incremental fine-tuning employs a low-rank adaptation technique. The system loads a pre-trained base diffusion model, freezing all original weight parameters, which are not involved in gradient updates. Low-rank adaptation matrices are inserted next to the query projection, key projection, value projection, and output projection matrices of the model's cross-attention layers, as well as the corresponding projection matrices of the self-attention layers. Each low-rank adaptation matrix consists of a dimension-reduced matrix (dimension equal to the number of rows in the original matrix multiplied by the rank *r*) and a dimension-increasing matrix (dimension equal to the rank *r* multiplied by the number of columns in the original matrix). The rank *r* is typically 4 to 64, much smaller than the dimension of the original matrix (usually hundreds to thousands), thus the number of parameters in the low-rank adaptation matrix is ​​only a tiny proportion of the original matrix. During fine-tuning, the model's forward propagation uses both the original weights and the low-rank adaptation weights simultaneously. The output is the original matrix output plus the low-rank adaptation matrix output multiplied by a scaling factor. The loss function for fine-tuning uses the standard noise prediction loss of the diffusion model, i.e., the mean squared error between the model's predicted noise and the actual noise. The optimizer employs an adaptive moment estimation algorithm, with a learning rate set to a small value between 1e-5 and 1e-4, a batch size set to 1 to 8 based on GPU memory, and 5 to 20 training epochs with early stopping enabled. During fine-tuning, the system calculates the validation loss on the validation set after each epoch. Training stops when the validation loss no longer decreases for three consecutive epochs, and the low-rank adaptation weights with the lowest validation loss are saved.

[0096] After fine-tuning, the system encapsulates the trained low-rank adaptation weight parameters and corresponding configuration information into a personalized adaptation model. The personalized adaptation model file contains metadata such as the binary data of the low-rank adaptation weights, the rank and position configuration of the adaptation matrix, the base model version identifier, the number and statistics of training samples, fine-tuning hyperparameters, the adaptation model version number, and creation time. The model file uses a standard low-rank adaptation model format and can be directly loaded by mainstream diffusion model inference frameworks. The system stores the personalized adaptation model in a model repository, establishing a mapping relationship between user identifiers and adaptation model versions. When a user initiates an image generation task, the inference module automatically loads the corresponding personalized adaptation model, dynamically merges the low-rank adaptation weights with the original model weights, and performs inference, incorporating the user's unique style preferences into the generated results. The personalized adaptation model supports version iteration. When the user's high-quality sample accumulation reaches the update threshold, the system triggers a new round of incremental fine-tuning, continuing to train and generate a new version based on the old adaptation model, achieving continuous evolution of the user's style model.

[0097] In some embodiments, recording the generated effect data and corresponding effect feedback data under different core parameter configurations, and updating the parameter adaptation strategy based on the generated effect data and effect feedback data, includes: establishing a corresponding mapping relationship between the core parameter configurations of sampling steps, guiding coefficients, and denoising intensity and the generated effect data; establishing a corresponding mapping relationship between the generated effect data and the effect feedback data; evaluating the adaptation effect of different core parameter configurations based on the effect feedback data, and generating an adaptation effect evaluation result; and updating the parameter adjustment rules in the parameter adaptation strategy based on the adaptation effect evaluation result.

[0098] This embodiment further details the specific implementation of the parameter adaptation strategy update step, as follows: The parameter-effect mapping relationship is established starting with the continuous accumulation of generation task records. The system creates a complete record for each image generation task, including a unique task identifier, user identifier, user-specific feature profile snapshot, sampling steps, guidance coefficient, and denoising intensity (three core parameter configurations), other auxiliary parameter configurations, objective effect indicators of the generated image, user feedback data, and a task timestamp. As generation tasks continue to execute, the system accumulates massive amounts of generation task records in a distributed data warehouse. The data warehouse employs a storage strategy of partitioning by date and bucketing by user group and editing scenario, supporting efficient querying and aggregation analysis of large-scale data.

[0099] The system periodically (e.g., during the early morning hours each day) performs batch ETL processing on newly added generation task records in the data warehouse to build a mapping relationship between core parameter configurations and generation effect data. The specific processing flow is as follows: First, the three core parameters are discretized, dividing the continuous parameter values ​​into several intervals. For example, the sampling steps are divided into intervals of 20-25, 25-30, and 30-35, each interval being 5 steps; the guiding coefficient is divided into intervals of 5.0-5.5 and 5.5-6.0, each interval being 0.5; and the denoising intensity is divided into intervals of 0.40-0.45 and 0.45-0.50, each interval being 0.05. Then, the generated task records are grouped and aggregated using user group type, editing scenario, sampling step interval, guiding coefficient interval, and denoising intensity interval as joint grouping keys. Statistical indicators of the generated effect data are calculated for each group, including the mean and standard deviation of the objective image quality score, the mean and standard deviation of the user satisfaction score, the average time spent on the generation task, the user regeneration rate, and the image download rate and collection rate. The grouping keys and corresponding effect statistical indicators are written into the parameter-effect mapping table to form a corresponding mapping relationship between the core parameter configuration and the generated effect data. The parameter-effect mapping table supports roll-up and drill-down queries based on any combination of grouping keys, allowing you to drill down from macro-level statistics for all users and all scenarios to micro-level analysis for specific groups and specific scenarios.

[0100] Building upon the parameter-effect mapping relationship, the system further establishes a corresponding mapping relationship between generated effect data and effect feedback data. Specifically, the system uses objective effect indicators of generated images as independent variables and user subjective effect feedback data as dependent variables, performing correlation analysis and regression modeling. By calculating the Pearson correlation coefficient between objective indicators (such as image clarity, color richness, and prompt word relevance) and subjective feedback (such as user ratings, collection rate, and download rate), the system identifies which objective indicators are highly correlated with user subjective satisfaction. The system employs multiple linear regression or gradient boosting tree regression models to fit the mapping function from objective effect indicators to user satisfaction ratings, enabling the system to predict potential user satisfaction based on objective indicators even before obtaining user feedback. The mapping relationship between generated effects and effect feedback is stored in an effect prediction model library. This model can be used for real-time effect prediction of newly generated images and for supplementing missing user feedback data in parameter adaptation strategy evaluation.

[0101] Based on the aforementioned two-layer mapping relationship, the system evaluates the adaptation effect of different core parameter configurations according to the effect feedback data and generates adaptation effect evaluation results. The evaluation process is executed separately for each user group and editing scenario combination to ensure the scenario-specificity of the evaluation results. For each core parameter configuration combination under a specific group and scenario, the system retrieves the corresponding generation effect statistics from the parameter-effect mapping table, supplements the missing feedback data with the effect prediction model, and calculates the comprehensive adaptation effect score of that parameter configuration combination. The comprehensive adaptation effect score adopts a weighted summation method and includes three dimensions: user satisfaction dimension (weight 0.5), which is calculated by combining the average user rating, collection rate, download rate, and regeneration rate, reflecting the user's subjective recognition of the generated results under the parameter configuration; generation quality dimension (weight 0.3), which is calculated by the average of objective image quality indicators, reflecting the objective technical quality of the generated image under the parameter configuration; and computational efficiency dimension (weight 0.2), which is calculated by the normalized value of generation time and resource consumption, reflecting the time and resource cost efficiency of the parameter configuration. The three dimensions are weighted and summed to obtain a comprehensive adaptation effect score from 0 to 100. The system sorts the scores of all parameter configuration combinations in descending order, identifies the top-scoring parameter configurations as excellent configurations, the bottom-scoring configurations as poor configurations, and those with intermediate scores as average configurations. Simultaneously, the system analyzes the changing trends of parameter configuration scores, comparing score changes over the two most recent statistical periods to identify parameter configurations with continuously improving or declining performance, and generates an adaptation effect evaluation report that includes configuration score rankings, a list of excellent configurations, a list of poor configurations, and trend change analysis.

[0102] Based on the adaptation effect evaluation results, the system updates the parameter adjustment rules in the parameter adaptation strategy. The parameter adaptation strategy adopts a rule engine based on decision trees. The internal nodes of the decision tree are user characteristic conditions (such as whether the user group type is professional users, whether the editing scenario is game concept art, whether a certain dimension of the operation habit vector is higher than a threshold, etc.), and the leaf nodes are the recommended values ​​or value ranges of the core parameters. The strategy update process includes the following operations: For excellent parameter configurations in the evaluation report, the system searches for the corresponding user feature path in the decision tree. If the current recommended configuration under this path is inconsistent with the excellent configuration, the recommended value of the leaf node is updated to the excellent configuration, and the confidence weight of the rule is increased. For poor parameter configurations, the system reduces the confidence weight of the corresponding rule. If the configuration is rated as poor for multiple consecutive periods, the rule is removed from the strategy or replaced with a better adjacent configuration. For newly discovered efficient parameter combinations in the evaluation, the system adds a corresponding rule branch in the decision tree, determines the optimal feature splitting condition through information gain analysis, and inserts the new rule into the appropriate position in the decision tree. For changes in the importance of user feature dimensions, the system recalculates the information gain of each feature in parameter prediction, adjusts the splitting order and splitting threshold of features in the decision tree, and places high-discrimination features at the upper level of the decision tree. After the strategy update, the system performs rule conflict detection to ensure that there are no contradictory parameter recommendation rules under the same user feature path, and performs rule coverage detection to ensure that there are corresponding parameter recommendation rules for all user groups and editing scenario combinations. After undergoing automated testing (simulating strategy execution on historical data and comparing the estimated effects of the old and new strategies) and canary release (first applying it to a small percentage of users to compare the differences in actual effects), the updated parameter adaptation strategy is officially released to the parameter dynamic adaptation engine for the next round of dynamic adjustment of core parameters.

[0103] In some embodiments, the step of using the updated parameter adaptation strategy for the next round of core parameter dynamic adjustment includes: loading the updated parameter adaptation strategy into the parameter dynamic adaptation engine; when the next round of image generation task starts, the parameter dynamic adaptation engine performs dynamic adjustment of the sampling steps, guiding coefficients, and denoising intensity core parameters according to the user-specific feature profile and the updated parameter adaptation strategy; and outputs the sampling steps, guiding coefficients, and denoising intensity core parameters used in the next round of image generation task.

[0104] This embodiment further details the specific implementation of the updated strategy's distribution and application phase, as follows: After the parameter adaptation strategy is updated and verified, the system executes the strategy distribution and loading operation. The parameter dynamic adaptation engine internally maintains a strategy runtime environment, including the currently effective parameter adaptation strategy instance, the strategy execution rule engine, the user profile query client, and the parameter regression inference model. The system serializes the updated parameter adaptation strategy into a standard format (such as JSON or ProtocolBuffers) and distributes it to all parameter dynamic adaptation engine instances through the configuration push mechanism of the configuration center. Upon receiving the new strategy, the engine instance first performs strategy format and integrity checks to confirm that the strategy data is complete and the format is valid; then, it performs hot loading of the strategy, atomically replacing the old strategy instance with the new strategy instance without interrupting the ongoing generation task, ensuring that there is no interruption of parameter adaptation service or inconsistency in state during the strategy switch. The engine records the time point of the strategy switch, the old strategy version number, and the new strategy version number, and writes them to the strategy switch log. For engine instances that fail to load, the system automatically rolls back to the old strategy and triggers an alarm to ensure service availability.

[0105] After the strategy is loaded, the parameter dynamic adaptation engine enters the running state of the new strategy. When the next image generation task starts, the engine performs dynamic adjustment of core parameters according to the following process: First, it receives the image generation task request and extracts the user's unique identifier, the preview of the prompt words entered by the user (used to assist in scene judgment), and task context information (such as the user's current function page, recent operation sequence, etc.). Second, it calls the user profile query client based on the user's unique identifier to retrieve the user-specific feature profile of the corresponding user from the user profile database or local cache. The profile data includes user group type labels, editing scene labels, operation habit vectors, and generation preference vectors. If the profile does not exist or has expired, the engine uses the default general profile and triggers asynchronous profile construction. Third, the engine inputs the user-specific feature profile into the updated parameter adaptation strategy rule engine. The rule engine matches the feature conditions of the decision tree layer by layer, starting from the root node and selecting the branch path according to the user features until reaching the leaf node, to obtain the core parameter benchmark recommended value or benchmark value range corresponding to the user feature. Fourth, the engine performs fine-tuning of the offsets based on the baseline recommended values. The operation habit vector and the generation preference vector are regressed into the model as input parameters. The model outputs fine-tuned offsets for the three parameters. The baseline values ​​and offsets are added together and clamped to a valid range to obtain the final parameter values. Fifth, the engine outputs the number of sampling steps, guidance coefficients, and core parameters for denoising intensity used in the next round of image generation. Simultaneously, it generates a decision record for this parameter adaptation, including the strategy version number, image version number, decision tree matching path, baseline value, offset, final parameter values, and decision confidence, and writes this record to the parameter adaptation decision log.

[0106] After the parameters are output, the engine passes them to the image generation model inference module to execute the image generation task. During the generation task, the engine continuously monitors the generation progress and intermediate results. When it detects that the user actively adjusts the parameters during the generation process, the engine records the user's manual adjustment behavior as implicit feedback for subsequent strategy evaluation and model optimization. After the generation task is completed, the engine associates the parameter configuration, generated image identifier, and user identifier for this task. After waiting for user feedback data, it writes the complete parameter-effect-feedback record into the parameter-effect database and enters the next round of strategy statistical analysis and update cycle.

[0107] Through the above strategy distribution and application process, the updated parameter adaptation strategy can take effect in real time in each round of image generation task, so that the system's parameter adaptation capability can continue to evolve with the accumulation of user behavior data and effect feedback data, providing users with increasingly accurate personalized parameter configuration services.

[0108] In some embodiments, please refer to Figure 2 The following describes the overall architecture and complete workflow of the behavior-aware image generation model dynamic fine-tuning system provided in this application.

[0109] like Figure 2 As shown, the overall system architecture is divided into five core parts from bottom to top and from left to right: behavior acquisition layer, AI model engine layer, sample screening and incremental fine-tuning layer, parameter adaptation strategy update layer, and closed-loop feedback mechanism. Each part is connected through data flow and control signals to form a complete personalized image generation closed-loop system.

[0110] The behavior acquisition layer is located in the upper left corner of the system architecture, corresponding to the user's operation in front of the computer and the "behavior acquisition" module in the diagram. This layer is responsible for capturing all user behavior data on the image generation platform in real time, including all interactive operations such as model selection, inputting prompts, adjusting parameters, editing images, browsing results, and submitting feedback. The behavior acquisition layer achieves non-intrusive data acquisition through a behavior acquisition software development kit embedded in the front end. The acquired data is structured to form a sequence of user operation behaviors, serving as the raw input for subsequent user profile construction and parameter adaptation. The arrow in the diagram pointing to the "behavior acquisition" module visually indicates that user behavior is the data input source for the entire system.

[0111] The AI ​​model engine layer is located in the middle of the system architecture, corresponding to the module with the neural network icon and the "AI Model Engine" label in the diagram, and the display on the right showing the generated image. The AI ​​model engine layer is the core computational module for performing image generation inference, integrating a basic diffusion model, a personalized adaptation model loading mechanism, a parameter configuration interface, and an inference scheduler. The engine's parameter configuration interface receives dynamically adjusted values ​​for three core parameters: sampling steps, guidance coefficients, and denoising intensity. The diagram visually demonstrates the adjustability of these three core parameters using three slider controls. The engine performs image generation inference based on user-input prompts and the dynamically configured core parameters, outputting the generated results to the display interface on the right for user viewing. The AI ​​model engine layer also records the parameter configurations and metadata of the generated results used in each generation task for subsequent performance evaluation and strategy optimization.

[0112] The sample selection and incremental fine-tuning layer is located in the lower left corner of the system architecture, corresponding to the user click feedback scenario, the pyramid-shaped sample selection icon, and the "Sample Selection and Incremental Fine-tuning" labeled module in the diagram. This layer is responsible for collecting user feedback data on the generated images, selecting high-quality samples deemed acceptable by the user from historical generated images, pairing these high-quality samples with corresponding prompts to form training sample pairs, and performing lightweight incremental fine-tuning on the image generation model to generate a personalized adaptation model tailored to the user's unique style. The pyramid-shaped icon in the diagram represents the process of progressively selecting a small number of high-quality samples from a large number of historical generated images, while the user's click gesture on the display represents the user providing feedback through interactive operation. The generated personalized adaptation model is fed back to the AI ​​model engine layer and loaded for use in subsequent generation tasks, giving the generated results user-specific style characteristics.

[0113] The parameter adaptation strategy update layer is located in the lower right corner of the system architecture, corresponding to the cylindrical database icon and the "Update Parameter Adaptation Strategy" labeled module in the diagram. This layer records generation effect data and user feedback data under different core parameter configurations. It evaluates the adaptation effect of different parameter configurations through statistical analysis and machine learning methods, identifies efficient and inefficient parameter configurations, and updates the parameter adjustment rules in the parameter adaptation strategy accordingly. The updated parameter adaptation strategy is then fed back to the parameter dynamic adaptation engine in the AI ​​model engine layer for dynamic adjustment of core parameters in the next generation task. The cylindrical database icon in the diagram represents the persistent storage and cumulative analysis of parameter effect data.

[0114] The closed-loop feedback mechanism is located at the bottom of the system architecture, corresponding to the arc-shaped arrow labeled "Closed-Loop Feedback" in the diagram, connecting the parameter adaptation strategy update layer and the behavior collection layer. This closed-loop feedback mechanism signifies that the entire system forms a complete closed loop, from user behavior collection to dynamic parameter adaptation, from displaying generated results to user feedback, and from sample selection and model fine-tuning to continuous strategy optimization. The behavioral and feedback data generated in each round of the generation task flow back to the system, driving user profile updates, personalized model iterations, and parameter adaptation strategy evolution, thus continuously improving the system's personalized adaptation capabilities with use. The arc-shaped arrow in the diagram circles from the lower right to the upper left, visually representing the direction of the closed-loop flow of data and strategies.

[0115] The complete system workflow is as follows: Users access the image generation platform via computer terminals. The behavior acquisition layer collects user operation behavior data in real time and constructs a user-specific feature profile. The parameter dynamic adaptation engine dynamically adjusts three core parameters—sampling steps, guidance coefficient, and denoising intensity—based on the user-specific feature profile and the current parameter adaptation strategy. The AI ​​model engine performs image generation inference based on user prompts and dynamically configured parameters, outputting the generated image to the user. After viewing the generated image, users provide feedback through explicit rating, collection, download, or implicit actions such as regeneration, editing, and dwelling. The sample selection and incremental fine-tuning layer selects high-quality samples based on feedback, performs lightweight incremental fine-tuning to generate a personalized adaptation model, and sends it back to the AI ​​model engine. The parameter adaptation strategy update layer records the correspondence between parameter configuration, generation effect, and user feedback, periodically evaluates and updates the parameter adaptation strategy, and sends the updated strategy back to the parameter dynamic adaptation engine for the next round of parameter adjustment. Through this closed-loop process, the system achieves end-to-end personalized image generation encompassing behavior perception, parameter adaptation, model fine-tuning, and strategy optimization, providing a precisely tailored image generation experience for different user groups.

[0116] In some embodiments, to address the problem that existing parameter adaptation strategies rely on manual rules and offline statistics and cannot adjust parameters in real time based on intermediate results during the generation process, this embodiment provides a parameter adaptive optimization method based on multi-agent reinforcement learning. The dynamic adjustment of sampling steps, guidance coefficients, and denoising intensity is modeled as a multi-agent sequential decision problem, and parameter configuration is optimized in real time during the denoising iteration process of image generation.

[0117] Specifically, the system treats each denoising iteration of the image generation model as a decision moment, and models sampling step scheduling, guidance coefficient adjustment, and denoising intensity control as three independent reinforcement learning agents. Each agent's state space includes the time step index of the current denoising step, the statistical characteristics of the current predicted noise (mean, variance, frequency domain energy distribution), the objective quality indicators of the current intermediate image (sharpness, color distribution, structural similarity), the embedding vector of the user-specific feature profile, and the proportion of denoising steps already performed out of the total steps. Each agent's action space consists of incremental adjustments to the corresponding parameters. For example, the sampling step scheduling agent's action is to increase or decrease the remaining denoising steps (within a preset range), the guidance coefficient adjustment agent's action is to increase or decrease the guidance coefficient by a certain proportion based on the current value, and the denoising intensity control agent's action is to adjust the noise injection ratio in subsequent denoising steps. The three agents share the same reward function, which comprehensively considers the objective quality score of the final generated image, the predicted value of user satisfaction, and the computational resource consumption (total denoising steps). At the end of the generation task, the delayed reward is calculated based on the final result, and the reward is back-allocated to the decision of each step through a credit allocation algorithm.

[0118] The three agents employ a multi-agent reinforcement learning framework with centralized training and distributed execution. During the training phase, the system utilizes the massive amount of generated task data accumulated on the platform to construct a simulation environment. Within this environment, numerous image generation tasks are executed in parallel. The three agents collect trajectory data through interaction with the environment, and a multi-agent proximal policy optimization algorithm is used to jointly train the policy network. Parameter sharing and experience replay mechanisms are introduced during training. The three agents share the underlying state feature extraction network, accelerating training convergence. The experience replay pool stores decision trajectories from different user groups and editing scenarios, ensuring that the training data covers diverse scenarios. After training, the policy network is deployed to the AI ​​model engine layer and runs in inference mode during actual generation tasks.

[0119] In the actual generation process, after each denoising iteration, the three agents independently decide on the parameter adjustment values ​​for the next step based on the state characteristics of the current intermediate image, achieving online adaptive optimization of parameters. For example, when it is detected that the intermediate image details are rich enough, the sampling step scheduling agent can decide to terminate denoising early to save computing resources; when it is detected that the generated result deviates from the direction of the prompt words, the guidance coefficient adjustment agent can decide to increase the guidance strength of subsequent steps to enhance the constraint of the prompt words; when it is detected that the input image features are not retained sufficiently, the denoising intensity control agent can decide to reduce the denoising intensity of subsequent steps to retain more original information. By dynamically optimizing parameters during the generation process, the system can adjust the strategy in real time according to the specific generation situation of each image, achieving a better balance between generation quality and efficiency compared to fixed parameter configuration.

[0120] This embodiment achieves online adaptive optimization of core parameters during image generation through multi-agent reinforcement learning, breaking through the limitations of traditional fixed parameter configuration and offline strategy updates. This enables parameter adaptation to respond to dynamic changes during the generation of each image, further improving the quality and efficiency of personalized image generation.

[0121] In some embodiments, to address the limitations of personalized fine-tuning effects caused by the limited number of high-quality samples per user and the privacy risks associated with centralized training of user image data, this embodiment provides a cross-user style knowledge transfer and privacy protection method based on federated learning. This method achieves style knowledge transfer between similar user groups without disclosing the user's original image data, thereby improving the training effect and data efficiency of personalized adaptation models.

[0122] Specifically, the system clusters users based on similarity in user group type, editing scenario, and generated preference vector, forming several user groups with similar styles. Each user group deploys a federated learning node, which stores the corresponding user's local high-quality samples and personalized adaptation model. The user's original image data is always retained on the local node and is not uploaded to the central server. The federated learning center server maintains a globally shared style model for each user group. This global model encodes the common style knowledge of the group's users, serving as the pre-training foundation for each user's personalized model.

[0123] The federated training process employs a parameter-efficient federated low-rank adaptation method. In each training round, the central server distributes the low-rank adaptation weights of the current globally shared style model to each user node within the group. Each user node performs incremental fine-tuning on the global model using local high-quality samples, updating only the low-rank adaptation matrix parameters. After training, the updated low-rank adaptation weight increments are uploaded to the central server. The central server performs weighted aggregation on the weight increments uploaded by each user. The aggregated weights are dynamically allocated based on the number of local samples and the model quality score for each user. The aggregation result updates the globally shared style model. To protect user privacy, the uploaded weight increments undergo differential privacy noise injection, injecting an appropriate amount of Gaussian noise into the weight gradient, preventing attackers from deducing any individual user's original training data from the aggregated weights. Simultaneously, the system employs a secure aggregation protocol, ensuring that the central server can only see the aggregated results of all user weight increments and cannot see the specific update values ​​of any individual user.

[0124] Based on the globally shared style model, each user node performs local fine-tuning of the personalized adaptation model. Users initialize the model with the low-rank adaptation weights of the global model and perform a few rounds of fine-tuning on high-quality local samples. This allows the model to further adapt to the user's unique preferences while preserving the group's common style knowledge. During local fine-tuning, elastic weight consolidation regularization is employed. This imposes strong regularization constraints on the parameters encoding the group's common knowledge in the global model, preventing local fine-tuning from destroying common knowledge. Weaker constraints are imposed on the parameters encoding the user's personalized features, allowing for full adaptation to the user's unique style.

[0125] When a new user joins and the number of high-quality samples is insufficient, the system matches the most similar user group based on the new user's unique feature profile. It then directly loads the globally shared style model of that group as the new user's initial personalized adaptation model. This allows the new user to achieve group-level style adaptation before accumulating sufficient local samples, enabling personalized services in cold start scenarios. As new user samples accumulate, the system gradually performs local fine-tuning, smoothly transitioning the model from a group-wide common style to a user-specific style.

[0126] This embodiment achieves cross-user style knowledge transfer and collaborative optimization through federated learning, while protecting user data privacy through differential privacy and secure aggregation mechanisms. This not only solves the problem of limited fine-tuning effects caused by insufficient single-user samples, but also meets the compliance requirements for user privacy protection, thereby improving the training efficiency and applicability of personalized adaptation models.

[0127] In some embodiments, to address the problem of confounding variable bias in parameter adaptation strategy evaluation—that is, user satisfaction is affected not only by parameter configuration but also by confounding factors such as prompt word quality, user mood, and random seeds, leading to bias in parameter effect evaluation based on observation data—this embodiment provides a parameter adaptation effect debiasing and counterfactual optimization method based on causal inference. By eliminating the influence of confounding factors through causal modeling, the true causal effect of parameter configuration is accurately estimated, and the parameter adaptation strategy is optimized through counterfactual inference.

[0128] Specifically, the system first constructs a causal graph model of parameter adaptation effects. The causal graph contains the following nodes: core parameter configuration nodes (sampling steps, guidance coefficient, denoising intensity), generation effect nodes (objective image quality and user satisfaction), and confounding variable nodes (prompt quality score, user group type, editing scenario, inferred value of user's current emotional state, random seed, base model version, computational resource load, etc.). Directed edges between nodes represent causal relationships: confounding variables simultaneously affect parameter configuration (users choose parameters based on prompt quality and their own habits) and generation effects (prompt quality directly affects the generation result), while parameter configuration affects generation effects. The system determines the edge structure of the causal graph using domain knowledge and a data-driven structure learning algorithm, and verifies the existence of edges using conditional independence checks.

[0129] Based on a causal graphical model, the system employs an inverse probability weighting method to estimate the average treatment effect of parameter configurations on the generation outcome. Specifically, the system calculates a propensity score for each generated task record, representing the probability of using a specific parameter configuration for that task given a confounding variable. The propensity score is estimated using logistic regression or a gradient boosting tree model, with the confounding variable as input and the parameter configuration category as output. The system uses the inverse of the propensity score as sample weights, deweighting samples with high propensity scores (common configurations) and increasing the weight of samples with low propensity scores (rare configurations), ensuring that the weighted treatment group and control group are balanced in terms of the confounding variable distribution, thereby eliminating selection bias caused by the confounding variable. The average user satisfaction for different parameter configurations is recalculated on the weighted samples to obtain an estimate of the unbiased causal effect of parameter configurations.

[0130] Based on unbiased causal effect estimation, the system executes counterfactual reasoning to optimize the parameter adaptation strategy. Counterfactual reasoning answers the question, "How will user satisfaction change if the user's parameter configuration is changed from A to B?" The system employs a counterfactual reasoning algorithm based on a structural causal model: First, it estimates individual-specific factors for each user (such as style preference parameters and sensitivity parameters) based on current observation data. Then, while keeping the individual-specific factors and confounding variables constant, it modifies the values ​​of the parameter configuration variables and forward calculates the predicted value of the modified generation effect using a causal model. The system performs counterfactual reasoning for each user across all candidate parameter configurations, predicts the user's satisfaction under each configuration, selects the configuration with the highest predicted satisfaction as the optimal parameter recommendation for that user, and updates the corresponding rules in the parameter adaptation strategy.

[0131] The system also performs parameter interpretability analysis based on the causal effect estimation results, quantifies the causal influence direction and degree of each core parameter on the generated effects of different user groups and different editing scenarios, identifies the key parameters and parameter interaction effects that have the greatest impact on the effect, and provides interpretable decision-making basis for manual debugging and optimization of parameter adaptation strategies.

[0132] This embodiment eliminates confounding variable bias in parameter effect evaluation through causal inference, achieving accurate estimation and counterfactual optimization of the true causal effect of parameter configuration. This ensures that the update of parameter adaptation strategy is based on causal relationships rather than spurious correlations, significantly improving the accuracy and robustness of parameter adaptation strategy.

[0133] In some embodiments, to address the problems that a fixed sampling step configuration cannot be dynamically adjusted according to the actual generation difficulty of each image, that oversampling of simple images wastes computational resources, and that insufficient sampling of complex images leads to quality degradation, this embodiment provides a real-time evaluation and dynamic early stopping method for generation quality based on intermediate features of a diffusion model. This method evaluates the generation quality in real time during the denoising iteration process and terminates sampling early when the quality reaches the convergence criterion, thereby achieving adaptive optimization of the sampling step number.

[0134] Specifically, during the denoising iteration of the diffusion model, the system extracts feature maps from the intermediate layers of the model every preset number of steps (e.g., every 5 steps) to construct real-time evaluation metrics for the generated quality. Intermediate features include the predicted noise features output by the denoiser, the attention map of the attention layer, and the latent space representation of the intermediate image. The system extracts the following real-time quality metrics from these intermediate features: a convergence metric, which calculates the cosine similarity and L2 distance between the latent space representations output from two adjacent denoising iterations; when the similarity consistently exceeds a threshold and the distance consistently falls below a threshold, it indicates that the generation result has converged; a detail richness metric, which performs frequency domain analysis on the intermediate latent space representation and calculates the proportion of high-frequency energy to the total energy; a higher proportion of high-frequency energy indicates richer image details; a cue word fit metric, which calculates the cosine similarity between the intermediate image features and the text embedding vector; a higher similarity indicates that the generated content fits the cue word description better; and an artifacts detection metric, which performs artifact detection on the decoded preview of the current intermediate image using a pre-trained lightweight image quality classifier, outputting probability scores for defects such as blurriness, overexposure, color abnormalities, and structural distortions.

[0135] The system inputs the aforementioned real-time quality metrics into a pre-trained early-stop decision model. The model employs a long short-term memory network or temporal convolutional network structure, receiving a sequence of quality metrics from multiple consecutive steps. It outputs a binary classification decision determining whether the currently generated image has reached the target quality, as well as a prediction of the expected quality gain for continued sampling. The early-stop decision model is trained using historical generation task data. The training samples consist of intermediate quality metric sequences at different sampling steps and user satisfaction labels for the final image. The model learns the mapping relationship between the changing trends of the quality metrics and the prediction of the final satisfaction level.

[0136] In the actual generation process, the system starts denoising iterations from the minimum number of sampling steps (e.g., 10 steps). After each step, real-time quality metrics are calculated, and the early stopping decision model is called every preset number of steps to evaluate whether the stopping condition has been met. When the decision model determines that the generated quality has reached the target standard and the expected quality gain from continued sampling is below the threshold, the system terminates the denoising iteration early and outputs the currently generated image. When the decision model determines that the quality has not yet met the standard and continued sampling would yield a significant gain, the system continues to execute denoising iterations until the maximum number of sampling steps is reached. For different user groups and editing scenarios, the system configures different levels of early stopping strictness: for professional users with high-quality requirements, the early stopping threshold is increased to ensure generated quality; for novice users with quick preview scenarios, the early stopping threshold is lowered to prioritize generation speed.

[0137] The computing resources saved by dynamic early stopping can be used to improve concurrent processing capabilities or allocated to complex generation tasks that require more sampling steps. The system maintains a global computing resource scheduler that dynamically allocates GPU computing resources based on the estimated remaining sampling steps for each generation task, thereby optimizing the overall system throughput.

[0138] This embodiment achieves adaptive optimization of the number of sampling steps through real-time quality assessment and dynamic early stopping of intermediate features in the diffusion model. This ensures that each image receives exactly the number of sampling steps that meet the quality requirements, avoiding the waste of oversampling of simple images while guaranteeing the sampling quality of complex images. This significantly improves the computational efficiency and resource utilization of image generation.

[0139] Please see Figure 3 As shown, Figure 3 This is a schematic diagram of the structure of a behavior-aware image generation model dynamic fine-tuning system 200 provided in this application embodiment. The behavior-aware image generation model dynamic fine-tuning system 200 is used to execute the steps of the behavior-aware image generation model dynamic fine-tuning method shown in the above embodiments. The behavior-aware image generation model dynamic fine-tuning system 200 can be a single server or a server cluster, or it can be a terminal, such as a handheld terminal, laptop computer, wearable device, or robot.

[0140] like Figure 3 As shown, the behavior-aware image generation model dynamic fine-tuning system 200 includes: The image acquisition unit 201 is used to collect user operation behavior data during the image generation process, extract user group type, editing scenario, operation habits and generation preference features based on the operation behavior data, and construct a user-specific feature profile; based on the user-specific feature profile, dynamically adjust the sampling step number, guidance coefficient and denoising intensity core parameters of the image generation model, and execute the image generation task based on the adjusted core parameters; The data acquisition unit 202 is used to collect user feedback data on the generated image effect, select high-quality samples from the user's historical generated images based on the feedback data, input the high-quality samples into the image generation model to perform lightweight incremental fine-tuning, and generate a personalized adaptation model that adapts to the user's exclusive style. The dynamic adjustment unit 203 is used to record the generation effect data and corresponding effect feedback data under different core parameter configurations, update the parameter adaptation strategy based on the generation effect data and effect feedback data, and use the updated parameter adaptation strategy for the next round of core parameter dynamic adjustment.

[0141] In some embodiments, the collection of user operation behavior data during the image generation process includes: collecting user operation behavior data in real time through the operation log interface of the image generation platform. The operation behavior data includes user-selected model type data, input prompt word data, parameter adjustment record data, editing operation record data, generation count data, and generation duration data; and performing structured storage of the operation behavior data according to user identifier and timestamp to generate a user operation behavior sequence.

[0142] In some embodiments, the step of extracting user group type, editing scenario, operation habits, and generation preference features based on operation behavior data includes: calculating user professional skill indicators based on parameter adjustment record data and editing operation record data in the operation behavior data, and determining user group type based on professional skill indicators; determining editing scenario based on prompt word data in the operation behavior data and the theme and purpose of the generated image; calculating parameter adjustment frequency and parameter adjustment range based on parameter adjustment record data in the operation behavior data, and determining operation habits based on parameter adjustment frequency and parameter adjustment range; and extracting keyword distribution features based on prompt word data in the operation behavior data, and determining generation preferences based on keyword distribution features and user-selected model type data.

[0143] In some embodiments, constructing a user-specific feature profile includes: performing structured encoding on user group type, editing scenario, operation habits, and generation preference features to generate user group type labels, editing scenario labels, operation habit vectors, and generation preference vectors; combining user group type labels, editing scenario labels, operation habit vectors, and generation preference vectors into a user-specific feature profile; storing the user-specific feature profile in a user profile database, and establishing a corresponding mapping relationship between user identifiers and user-specific feature profiles.

[0144] In some embodiments, dynamically adjusting the sampling steps, guidance coefficient, and denoising intensity core parameters of the image generation model based on the user-specific feature profile includes: retrieving the user-specific feature profile of the corresponding user from the user profile database; determining the baseline value range of the sampling steps, guidance coefficient, and denoising intensity core parameters based on the user group type label and editing scene label in the user-specific feature profile; performing fine-grained offset adjustment within the baseline value range based on the operation habit vector and generation preference vector in the user-specific feature profile; and outputting the adjusted sampling steps, guidance coefficient, and denoising intensity core parameters.

[0145] In some embodiments, performing the image generation task based on the adjusted core parameters includes: loading the adjusted sampling steps, guiding coefficients, and denoising intensity core parameters into the inference configuration of the image generation model; receiving prompt words input by the user, performing image generation inference based on the prompt words and the loaded core parameters; outputting the generated image and recording the sampling steps, guiding coefficients, and denoising intensity core parameter configuration used in this generation task.

[0146] In some embodiments, collecting user feedback data on the generated image includes: setting an effect feedback collection entry point on the generated image display interface; collecting user rating data, collection operation data, regeneration operation data, and editing operation data on the generated image through the effect feedback collection entry point; and integrating the rating data, collection operation data, regeneration operation data, and editing operation data into effect feedback data.

[0147] In some embodiments, the step of selecting high-quality samples from the user's historical generated images based on performance feedback data, inputting the high-quality samples into the image generation model to perform lightweight incremental fine-tuning, and generating a personalized adaptation model adapted to the user's unique style includes: calculating the quality score of each image in the user's historical generated images based on the rating data and collection operation data in the performance feedback data; identifying user's historical generated images with quality scores exceeding a preset quality threshold as high-quality samples; forming fine-tuning training sample pairs with the prompt words corresponding to the high-quality samples; inputting the fine-tuning training sample pairs into the image generation model to perform lightweight incremental fine-tuning, updating the low-rank adaptation weight parameters of the image generation model; and encapsulating the updated low-rank adaptation weight parameters into a personalized adaptation model adapted to the user's unique style.

[0148] In some embodiments, recording the generated effect data and corresponding effect feedback data under different core parameter configurations, and updating the parameter adaptation strategy based on the generated effect data and effect feedback data, includes: establishing a corresponding mapping relationship between the core parameter configurations of sampling steps, guiding coefficients, and denoising intensity and the generated effect data; establishing a corresponding mapping relationship between the generated effect data and the effect feedback data; evaluating the adaptation effect of different core parameter configurations based on the effect feedback data, and generating an adaptation effect evaluation result; and updating the parameter adjustment rules in the parameter adaptation strategy based on the adaptation effect evaluation result.

[0149] In some embodiments, the step of using the updated parameter adaptation strategy for the next round of core parameter dynamic adjustment includes: loading the updated parameter adaptation strategy into the parameter dynamic adaptation engine; when the next round of image generation task starts, the parameter dynamic adaptation engine performs dynamic adjustment of the sampling steps, guiding coefficients, and denoising intensity core parameters according to the user-specific feature profile and the updated parameter adaptation strategy; and outputs the sampling steps, guiding coefficients, and denoising intensity core parameters used in the next round of image generation task.

[0150] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the above-described behavior-aware image generation model dynamic fine-tuning system and its modules can be referred to the corresponding content in the various embodiments of the above-described behavior-aware image generation model dynamic fine-tuning method, and will not be repeated here.

[0151] The aforementioned dynamic fine-tuning method for behavior-aware image generation models can be implemented as a computer program, which can, for example... Figure 3 It runs on the system shown.

[0152] Please see Figure 4 , Figure 4 This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application. The computer device includes a processor, a memory, and a network interface connected via a device bus, wherein the memory may include a storage medium and internal memory.

[0153] The storage medium can store operating devices and computer programs. The computer program includes program instructions that, when executed, cause the processor to perform any method for dynamically fine-tuning a behavior-aware image generation model.

[0154] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0155] Internal memory provides an environment for the execution of computer programs in non-volatile storage media. When these computer programs are executed by the processor, the processor can perform any dynamic fine-tuning method for behavior-aware image generation models.

[0156] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the terminal to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0157] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0158] In one embodiment, the processor is configured to run a computer program stored in memory to perform the following steps: Collect user action data during the image generation process, extract user group type, editing scenario, operation habits and generation preference features based on the action data, and construct a user-specific feature profile; based on the user-specific feature profile, dynamically adjust the sampling step number, guidance coefficient and denoising intensity core parameters of the image generation model, and execute the image generation task based on the adjusted core parameters; Collect user feedback data on the effect of the generated image, select high-quality samples from the user's historical generated images based on the feedback data, input the high-quality samples into the image generation model to perform lightweight incremental fine-tuning, and generate a personalized adaptation model that fits the user's exclusive style. Record the generated effect data and corresponding effect feedback data under different core parameter configurations, update the parameter adaptation strategy based on the generated effect data and effect feedback data, and use the updated parameter adaptation strategy for the next round of dynamic adjustment of core parameters.

[0159] In some embodiments, the collection of user operation behavior data during the image generation process includes: collecting user operation behavior data in real time through the operation log interface of the image generation platform. The operation behavior data includes user-selected model type data, input prompt word data, parameter adjustment record data, editing operation record data, generation count data, and generation duration data; and performing structured storage of the operation behavior data according to user identifier and timestamp to generate a user operation behavior sequence.

[0160] In some embodiments, the step of extracting user group type, editing scenario, operation habits, and generation preference features based on operation behavior data includes: calculating user professional skill indicators based on parameter adjustment record data and editing operation record data in the operation behavior data, and determining user group type based on professional skill indicators; determining editing scenario based on prompt word data in the operation behavior data and the theme and purpose of the generated image; calculating parameter adjustment frequency and parameter adjustment range based on parameter adjustment record data in the operation behavior data, and determining operation habits based on parameter adjustment frequency and parameter adjustment range; and extracting keyword distribution features based on prompt word data in the operation behavior data, and determining generation preferences based on keyword distribution features and user-selected model type data.

[0161] In some embodiments, constructing a user-specific feature profile includes: performing structured encoding on user group type, editing scenario, operation habits, and generation preference features to generate user group type labels, editing scenario labels, operation habit vectors, and generation preference vectors; combining user group type labels, editing scenario labels, operation habit vectors, and generation preference vectors into a user-specific feature profile; storing the user-specific feature profile in a user profile database, and establishing a corresponding mapping relationship between user identifiers and user-specific feature profiles.

[0162] In some embodiments, dynamically adjusting the sampling steps, guidance coefficient, and denoising intensity core parameters of the image generation model based on the user-specific feature profile includes: retrieving the user-specific feature profile of the corresponding user from the user profile database; determining the baseline value range of the sampling steps, guidance coefficient, and denoising intensity core parameters based on the user group type label and editing scene label in the user-specific feature profile; performing fine-grained offset adjustment within the baseline value range based on the operation habit vector and generation preference vector in the user-specific feature profile; and outputting the adjusted sampling steps, guidance coefficient, and denoising intensity core parameters.

[0163] In some embodiments, performing the image generation task based on the adjusted core parameters includes: loading the adjusted sampling steps, guiding coefficients, and denoising intensity core parameters into the inference configuration of the image generation model; receiving prompt words input by the user, performing image generation inference based on the prompt words and the loaded core parameters; outputting the generated image and recording the sampling steps, guiding coefficients, and denoising intensity core parameter configuration used in this generation task.

[0164] In some embodiments, collecting user feedback data on the generated image includes: setting an effect feedback collection entry point on the generated image display interface; collecting user rating data, collection operation data, regeneration operation data, and editing operation data on the generated image through the effect feedback collection entry point; and integrating the rating data, collection operation data, regeneration operation data, and editing operation data into effect feedback data.

[0165] In some embodiments, the step of selecting high-quality samples from the user's historical generated images based on performance feedback data, inputting the high-quality samples into the image generation model to perform lightweight incremental fine-tuning, and generating a personalized adaptation model adapted to the user's unique style includes: calculating the quality score of each image in the user's historical generated images based on the rating data and collection operation data in the performance feedback data; identifying user's historical generated images with quality scores exceeding a preset quality threshold as high-quality samples; forming fine-tuning training sample pairs with the prompt words corresponding to the high-quality samples; inputting the fine-tuning training sample pairs into the image generation model to perform lightweight incremental fine-tuning, updating the low-rank adaptation weight parameters of the image generation model; and encapsulating the updated low-rank adaptation weight parameters into a personalized adaptation model adapted to the user's unique style.

[0166] In some embodiments, recording the generated effect data and corresponding effect feedback data under different core parameter configurations, and updating the parameter adaptation strategy based on the generated effect data and effect feedback data, includes: establishing a corresponding mapping relationship between the core parameter configurations of sampling steps, guiding coefficients, and denoising intensity and the generated effect data; establishing a corresponding mapping relationship between the generated effect data and the effect feedback data; evaluating the adaptation effect of different core parameter configurations based on the effect feedback data, and generating an adaptation effect evaluation result; and updating the parameter adjustment rules in the parameter adaptation strategy based on the adaptation effect evaluation result.

[0167] In some embodiments, the step of using the updated parameter adaptation strategy for the next round of core parameter dynamic adjustment includes: loading the updated parameter adaptation strategy into the parameter dynamic adaptation engine; when the next round of image generation task starts, the parameter dynamic adaptation engine performs dynamic adjustment of the sampling steps, guiding coefficients, and denoising intensity core parameters according to the user-specific feature profile and the updated parameter adaptation strategy; and outputs the sampling steps, guiding coefficients, and denoising intensity core parameters used in the next round of image generation task.

[0168] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to implement the steps of the behavior-aware image generation model dynamic fine-tuning method provided in any embodiment of this application.

[0169] The computer-readable storage medium may be an internal storage unit of the computer device described in the foregoing embodiments, such as the hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the computer device.

[0170] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for dynamic fine-tuning of an image generation model based on behavior awareness, characterized in that, include: Collect user action data during image generation, extract user group type, editing scenario, operation habits and generation preference features based on action data, and build a user-specific feature profile; Based on the user's unique feature profile, the sampling steps, guidance coefficients, and core parameters of the image generation model are dynamically adjusted, and the image generation task is executed based on the adjusted core parameters. Collect user feedback data on the effect of the generated image, select high-quality samples from the user's historical generated images based on the feedback data, input the high-quality samples into the image generation model to perform lightweight incremental fine-tuning, and generate a personalized adaptation model that fits the user's exclusive style. Record the generated effect data and corresponding effect feedback data under different core parameter configurations, update the parameter adaptation strategy based on the generated effect data and effect feedback data, and use the updated parameter adaptation strategy for the next round of dynamic adjustment of core parameters.

2. The method according to claim 1, characterized in that, The data collected on user actions during the image generation process includes: The operation log interface of the image generation platform collects user operation behavior data in real time during the image generation process. The operation behavior data includes user-selected model type data, input prompt word data, parameter adjustment record data, editing operation record data, generation count data, and generation duration data. The operation behavior data is structured and stored according to user identifier and timestamp to generate a sequence of user operation behavior.

3. The method according to claim 1, characterized in that, The extraction of user group types, editing scenarios, operating habits, and generation preference features based on operational behavior data includes: Based on the parameters in the operational behavior data, adjust the recorded data and edit the operational record data to calculate the user's professional skill index, and determine the user group type based on the professional skill index; The editing scenario is determined based on the prompt words in the operation behavior data and the theme and purpose of the generated image. Calculate the parameter adjustment frequency and range based on the parameter adjustment record data in the operation behavior data, and determine the operation habits based on the parameter adjustment frequency and range; Keyword distribution features are extracted from the prompt words in the operational behavior data, and generation preferences are determined based on the keyword distribution features and the model type data selected by the user.

4. The method according to claim 3, characterized in that, The construction of user-specific feature profiles includes: The user group type, editing scenario, operation habit and generation preference features are structured and encoded to generate user group type labels, editing scenario labels, operation habit vectors and generation preference vectors; Combine user group type tags, editing scenario tags, operation habit vectors and generated preference vectors to create a user-specific feature profile; Store user-specific feature profiles in a user profile database to establish a mapping relationship between user identifiers and user-specific feature profiles.

5. The method according to claim 1, characterized in that, The process of dynamically adjusting the sampling steps, guidance coefficients, and denoising intensity of the image generation model based on the user's unique feature profile includes: Retrieve the user-specific feature profile of the corresponding user from the user profile database; The baseline value range of the core parameters, namely sampling steps, guidance coefficient, and denoising intensity, is determined based on the user group type label and editing scenario label in the user-specific feature profile. Based on the user's unique feature profile, the operation habit vector and the generated preference vector are used to perform fine-grained offset adjustments within the baseline value range; Output the adjusted sampling steps, guiding coefficients, and core parameters of denoising intensity.

6. The method according to claim 1, characterized in that, The step of performing the image generation task based on the adjusted core parameters includes: The adjusted sampling steps, guiding coefficients, and denoising intensity core parameters are loaded into the inference configuration of the image generation model; Receive prompts from the user and perform image generation inference based on the prompts and the loaded core parameters; Output the generated image and record the core parameter configurations used in this generation task, including the number of sampling steps, guiding coefficients, and denoising intensity.

7. The method according to claim 1, characterized in that, The data collected includes user feedback on the generated image's effect, including: Set up an entry point for collecting effect feedback in the generated image display interface; The system collects user ratings, collection data, regeneration data, and editing data for the generated images through the feedback collection portal. The rating data, collection data, regeneration data, and editing data are integrated into performance feedback data.

8. The method according to claim 7, characterized in that, The process of selecting high-quality samples from users' historical generated images based on performance feedback data, inputting these high-quality samples into the image generation model for lightweight incremental fine-tuning, and generating a personalized adaptation model tailored to the user's unique style includes: Calculate the quality score for each image in the user's historical generated images based on the rating data and collection operation data in the effect feedback data; Historical generated images of users whose quality scores exceed a preset quality threshold are identified as high-quality samples. Combine the prompt words corresponding to high-quality samples with high-quality samples to form fine-tuning training sample pairs; Lightweight incremental fine-tuning is performed on the input image generation model using fine-tuned training samples, updating the low-rank adaptation weight parameters of the image generation model. The updated low-rank adaptation weight parameters are encapsulated into a personalized adaptation model that adapts to the user's unique style.

9. The method according to claim 1, characterized in that, The process of recording the generated effect data and corresponding effect feedback data under different core parameter configurations, and updating the parameter adaptation strategy based on the generated effect data and effect feedback data, includes: Establish a mapping relationship between the core parameter configurations of sampling steps, guiding coefficients, and denoising intensity, and the generated effect data; Establish a mapping relationship between generated effect data and effect feedback data; The adaptation effect of different core parameter configurations is evaluated based on the feedback data, and the adaptation effect evaluation results are generated. Update the parameter adjustment rules in the parameter adaptation strategy based on the adaptation effect evaluation results.

10. The method according to claim 1, characterized in that, The step of using the updated parameter adaptation strategy for the next round of dynamic adjustment of core parameters includes: Load the updated parameter adaptation strategy into the parameter dynamic adaptation engine; When the next image generation task is launched, the parameter dynamic adaptation engine dynamically adjusts the core parameters of sampling steps, guidance coefficients and denoising intensity based on the user's unique feature profile and the updated parameter adaptation strategy. Output the number of sampling steps, guiding coefficients, and core parameters of denoising intensity used in the next round of image generation.

Images