A multi-dimensional feature-based clothing collocation generation, virtual try-on and transaction system
Patent Information
- Application Number
- CN202611311953.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-27
- Publication Date
- 2026-10-13
AI Technical Summary
[0007]为了解决现有服饰推荐与交互系统存在虚拟试穿效果差且开销高、以及从推荐到交易链路断层难以形成闭环的问题,本申请提出一种基于多维特征的服饰搭配生成、虚拟试穿及交易系统,具体包括:
[0046]本发明通过建立包含品类搭配、风格兼容与颜色搭配规则的审美规则引擎,并将规则约束显式注入推荐与搭配生成流程中,打破了传统仅依赖历史行为或热度的单一推荐模式,生成的整套穿搭方案符合专业审美逻辑且整体协调,有效减少了用户的选购与拼凑时间,降低了决策成本;
Smart Images

Figure CN122905760A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and computer data processing technology, specifically to a system for generating clothing combinations, virtual try-on, and trading based on multi-dimensional features. Background Technology
[0002] With the rapid development of e-commerce and artificial intelligence technologies, online clothing consumption has become the mainstream purchasing method. To improve the user shopping experience and conversion rates, various intelligent recommendation algorithms based on collaborative filtering, content recommendation, or deep learning are widely used on e-commerce platforms. However, existing clothing recommendation and interaction systems still have the following shortcomings in practical applications:
[0003] Recommendations are often limited to a single dimension and lack professional aesthetic rules: Most existing recommendation systems rely solely on users' historical browsing and purchasing behavior data for single-item recommendations (e.g., "What else did users who bought this type of clothing buy?"), lacking a comprehensive consideration of professional aesthetic logic such as clothing category, style compatibility, and color matching. This results in recommendations often being limited to similar or identical products, making it difficult to generate professional and coordinated complete outfit solutions.
[0004] The visual try-on experience is disconnected from the generated experience: Existing virtual try-on technology is often independent of the recommendation system. The try-on model usually requires each garment to be input and re-rendered, which is computationally expensive and slow. At the same time, when trying on multiple garments, there is a lack of efficient visual preprocessing and multimodal generation constraints, which makes the generated dressing images prone to problems such as facial distortion, mismatch of clothing details, or unnatural poses.
[0005] The disconnect between recommendation, experience and transaction: existing clothing recommendations, DIY matching, virtual try-on and e-commerce ordering are often isolated from each other, lacking a complete closed-loop system architecture that can take into account rule engine configuration, high-concurrency request caching, quota management and one-click ordering.
[0006] Therefore, how to build an apparel system that can integrate professional aesthetic rules with multi-algorithm recommendations, provide an efficient and accurate multimodal virtual try-on experience, and establish a closed loop from recommendation to transaction is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0007] To address the problems of poor virtual try-on effects and high costs in existing clothing recommendation and interaction systems, as well as the lack of a closed loop in the recommendation-to-transaction process, this application proposes a clothing matching generation, virtual try-on, and transaction system based on multi-dimensional features, specifically including:
[0008] The data layer is used to store multidimensional feature data of products, user profile data, and user behavior log data; among them, the main product category includes a first-level category, and the first-level category can further include various levels of alternative categories;
[0009] The matching rules engine is used to configure and manage category matching rules, style compatibility matrices, and color matching rules; the matching rules are determined based on actual conditions, such as aesthetic matching rules.
[0010] The multi-algorithm recommendation module includes a rule recommendation unit, a behavior recommendation unit, and a hybrid recommendation unit, which is used to output recommendation results based on the data layer.
[0011] The outfit generation module is used to generate a complete outfit scheme based on the target input, combined with the outfit rule engine and the multi-algorithm recommendation module;
[0012] The virtual try-on module is used to input model images and clothing images into a multimodal generation model to generate images of the try-on results.
[0013] The transaction service module provides a unified API gateway, caching and pagination of recommendation and trial tasks, and user quota and interface rate limiting management to provide an integrated closed-loop service from recommendation to transaction.
[0014] Specifically, the user profile data includes gender, age, height, weight, body shape, chest circumference, waist circumference, hip circumference, shoe size, style preference, color preference, brand preference, and price preference;
[0015] The user behavior log data includes the behavior types such as browsing, adding to cart, favorites and purchasing, behavior timestamps, source pages, session identifiers and device types.
[0016] Specifically, the matching rule engine includes: category matching rules, style compatibility matrix, and color matching rules;
[0017] The multidimensional feature data includes the main product category, brand, color, pattern, style tag, applicable season, available sizes, fabric, price, sales volume, click volume, and rating; the category matching rules define the set of main product categories that can be matched with the main product category, as well as the recommended matching combination slots under each style tag; wherein, the matching combination slots can be fixed categories at each level, or sequentially descending parent category alternatives;
[0018] The style compatibility matrix contains compatibility mapping flags or thresholds between different style tag pairs, which are used to determine whether a product meets the style matching conditions based on the mapping results of the product's style tag and the target style tag in the matrix.
[0019] The color matching rules include the mutual matching mapping relationship between different color label pairs, which is used to determine whether the products meet the color matching conditions based on the mutual matching mapping result between the color labels of the first product and the color labels of the second product.
[0020] The rules are centrally managed in the form of structured data such as dictionaries.
[0021] Specifically, in the multi-algorithm recommendation module:
[0022] The rule recommendation unit assigns a weighted score to five dimensions—price matching, style matching, body size matching, season matching, and brand matching—according to preset weights.
[0023] The behavior recommendation unit assigns corresponding weights to different behavior types and constructs a "user-item" interaction matrix by combining time decay. Based on the matrix, it performs item-based collaborative filtering and then weightedly fuses the collaborative filtering results with TF-IDF content similarity constructed based on product text features.
[0024] The hybrid recommendation unit dynamically allocates the weights of rule-based recommendations and behavioral recommendations based on the user's cold start level. The recommendation results from the rule-based recommendation unit and the behavioral recommendation unit are weighted, merged, and output in sorted order of scores.
[0025] Specifically, the multi-algorithm recommendation module further includes:
[0026] The cold start rollback unit is used to perform multi-level rollback when both user profile and behavioral data are insufficient, sequentially using user profile recommendations, category-based popular recommendations, and global popular recommendations.
[0027] The category diversity filtering unit is used to merge accessories into one slot and allow a maximum of 1 item to be recommended, while each other first-level category can recommend a maximum of 1 item.
[0028] Specifically, the execution process of the matching generation module is as follows:
[0029] Determine the input and anchor points: Use the user-selected or system-recommended products, user-provided products, the seed products with the highest user behavior scores, or the specified styles as inputs, and use categories such as tops, bottoms, skirts, dresses, and suits as anchor points for whole-pool sampling;
[0030] Determine the target style: prioritize the product's own style, followed by the user's preferred style; the default target style is casual, and the style compatibility matrix is used to verify whether the product style and the target style meet the style matching conditions.
[0031] Slot filling: Select products one by one according to the recommended combination slots corresponding to the target style. If the slot is a candidate category, then go back to the first category with stock.
[0032] Vectorized scoring: Vectorize the scores of candidate products for each slot, and randomly sample from the high-scoring candidates based on their scores.
[0033] Variation expansion and deduplication: Generate multiple variation combinations for each style, and deduplicate the variation combinations using the product set as the key;
[0034] Combination scoring and ranking: Price reasonableness is defined by whether the overall price falls within the user's preferred range, and combination completeness is defined. The combination completeness is judged as follows: the combination must contain at least the target product and one combination product; when the number of combination products reaches a preset threshold, a completeness score is obtained; if there is no suitable product in the combination slot, the slot is skipped; each combination is scored based on price reasonableness and combination completeness and sorted in descending order.
[0035] SKU validity verification: Verify whether the inventory of each product in the outfit is valid and available for sale, remove incomplete outfits with products that have been discontinued, and finally output a complete outfit scheme including style, applicable occasion, total price and product details.
[0036] Specifically, when the matching generation module performs vectorized scoring for each slot, the scoring change rules are as follows:
[0037] Price matching adds 0.3 points, style matching adds 0.4 points, and color compatibility adds 0.2 points.
[0038] Recommended products are weighted at a power of 0.6 based on their usage frequency, and then randomly sampled from high-scoring candidates based on their scores.
[0039] Specifically, the virtual try-on module includes: task submission and preprocessing: receiving model images, single or multiple clothing images uploaded by users or passed through a URL, verifying the format, size, and file size of the images, and performing preprocessing; the preprocessing may include EXIF orientation correction, converting the transparency channel to white background, or proportional scaling and centering;
[0040] Batch collage: When an image involves multiple garments, multiple garment images are stitched together into a single collage according to a preset size grid, enabling batch try-on of multiple garments in a single generation.
[0041] Call the multimodal generation module: Input the clothing image, puzzle image and model image into the multimodal generation model, and inject dressing prompt words. The prompt words constrain the generation model to output the dressing result. The dressing result is a high-definition result image in which only the clothing is replaced, while the model's face, expression, hairstyle, body posture and background remain unchanged, and the style, color, texture and pattern of the clothing are consistent with the input product, and the light and shadow and folds fit the human body posture.
[0042] Specifically, the virtual try-on module also performs result management and retries when calling the multimodal generation model, as follows:
[0043] The multimodal generation process adopts an asynchronous task mechanism. After the client submits the task, it immediately obtains the task identifier. The server background thread executes the generation and uses exponential backoff retries for momentary failures. The client polls the task status to obtain the progress and downloads the high-resolution result image.
[0044] Specifically, the transaction service module works as follows: a unified API gateway is used to encapsulate and forward the recommendation interface and the try-on interface; the recommendation results and try-on results are stored in the cache with task identifiers and a time-to-live (TTL) is set; the first request generates a full set of results and caches them, and subsequent pagination requests directly hit the cache; database transactions are used to ensure the consistency between the user's exchange quota deduction and usage records, and the request frequency is controlled through interface rate limiting.
[0045] Compared with the prior art, the present invention has the following beneficial effects:
[0046] This invention establishes an aesthetic rule engine that includes rules for category matching, style compatibility, and color matching, and explicitly injects rule constraints into the recommendation and outfit generation process. This breaks away from the traditional single recommendation model that relies solely on historical behavior or popularity. The generated complete outfit scheme conforms to professional aesthetic logic and is harmonious as a whole, effectively reducing the time users spend selecting and piecing together outfits and lowering decision-making costs.
[0047] This invention, based on a multimodal generative model and incorporating explicit clothing change prompts, achieves highly realistic clothing replacement while strictly preserving the model's face, body posture, and background. Furthermore, for trying on multiple garments, a grid-based stitching method is used. Figure 1 The system generates outfit changes in one go, eliminating the need for item-by-item model reasoning, which significantly reduces computational resource consumption and improves the efficiency of try-on processing.
[0048] This invention integrates the entire outfit generation, multimodal try-on preview, and transaction order placement into a unified architecture. Through a unified API gateway and task caching mechanism, it ensures consistency of response and results throughout the entire process, realizing an integrated closed loop from "intelligent recommendation - virtual try-on - one-click transaction", which significantly shortens the user's shopping decision and transaction path.
[0049] In addition, this invention also improves the user experience during cold starts, enhances system stability under high concurrency, and provides good business scalability.
[0050] This invention ensures that new users or users with sparse data can still receive reasonable recommendation results by setting up a multi-level cold start rollback mechanism of "user profile - category popularity - global popularity"; it effectively copes with high concurrency requests and instantaneous failures by adopting asynchronous task management, interface rate limiting and exponential backoff retry strategies, ensuring service stability in high-frequency interaction scenarios of e-commerce; aesthetic rules are centrally configured in the rule engine, and business personnel can dynamically update the rules without modifying the underlying algorithm code, reducing operation and maintenance and upgrade costs. Attached Figure Description
[0051] Figure 1 This is a diagram of the overall system architecture.
[0052] Figure 2 A schematic diagram of the data structure for the aesthetic matching rule base;
[0053] Figure 3 Structure diagram of multi-algorithm recommendation module;
[0054] Figure 4 Flowchart for generating clothing matching methods;
[0055] Figure 5 This is a flowchart of the virtual try-on method;
[0056] Figure 6 This is a schematic diagram of a jigsaw puzzle structure for batch try-on. Detailed Implementation Specific Implementation Example 1:
[0058] This embodiment provides a system for generating clothing combinations, virtual try-on, and trading based on multi-dimensional features, combined with... Figure 1 The system architecture is described below. This implementation uses a mobile application or mini-program as the user client. The server consists of application services, data storage, caching, and backend processing components, and calls the recommendation module, outfit generation module, and virtual try-on module through a unified interface. Multimodal generation capabilities can be deployed on a local computing environment or through controlled interfaces to call external image generation services. The client provides entry points for outfit browsing, self-combination, and virtual try-on functions, allowing users to browse system-generated outfits, obtain outfits combined with user preferences, and select and try on clothing. The system provided in this embodiment specifically includes:
[0059] The data layer stores multidimensional feature data of products, user profile data, and user behavior log data. The multidimensional feature data includes the main product category, brand, color, pattern, style tag, applicable season, available sizes, fabric, price, sales volume, click volume, and rating. The main product category includes a first-level category, which can further include second-level and third-level alternative categories.
[0060] Specifically, the user profile data includes gender, age, height, weight, body shape, chest circumference, waist circumference, hip circumference, shoe size, style preference, color preference, brand preference, and price preference;
[0061] The user behavior log data includes browsing, adding to cart, favorites, and purchasing behavior types, behavior timestamps, source pages, session identifiers, and device types. The system can also gradually refine user preferences based on user behavior log data such as browsing, selection, and saving interactions. When a user requests personalized recommendations, the system dynamically selects rule-based recommendations, behavioral recommendations, or a fusion of both, based on the completeness of the user profile, effective interaction information, and recent activity. When user information is limited, the system generates results primarily based on aesthetic rules and known preferences; as interaction information increases, the system can gradually increase the influence of behavioral characteristics on the recommendation results. The fusion method and degree of influence of various recommendation results can be dynamically adjusted according to the user's status, rather than being limited to fixed parameters.
[0062] Combined with a rules engine, Figure 2 The aesthetic matching rules are configured in a structured manner within the rule engine to describe the combination relationships between different clothing categories, the compatibility relationships between different styles, and the coordination relationships between different colors. These rules can be pre-configured by professionals or continuously adjusted based on actual usage, thus providing constraints for subsequent candidate selection and outfit generation. The system performs category diversity control on the candidate results and generates several personalized solutions based on user preferences, occasion requirements, color coordination, and clothing availability. Personalized solutions are not fixed combinations but are adaptively selected from a set of candidates that meet user needs and aesthetic rules. Users can delete or replace clothing in any slot; when replacing, the system retains other clothing that the user has not requested changes to, only re-searches for candidates in the target slot, and excludes clothing already used in the current combination or explicitly excluded by the user. After each adjustment, the system re-checks the category completeness, style consistency, color coordination, and clothing availability of the entire outfit. Outfits that are not yet complete can continue to be previewed and edited, and can be saved or entered into a virtual try-on process after completion. Recommended results and editing status can be temporarily retained for necessary time so that users can continue to adjust; the specific saving method and retention period can be set according to the deployment environment.
[0063] Specifically, the matching rules engine includes: category matching rules, style compatibility matrix, and color matching rules;
[0064] The category matching rules define the set of main product categories that can be matched with the main product category, as well as the recommended matching combination slots under each style tag; the matching combination slots can be fixed categories at each level, or they can be alternative categories at higher levels that are rolled back in sequence.
[0065] The style compatibility matrix contains compatibility mapping flags or thresholds between different style tag pairs, which are used to determine whether a product meets the style matching conditions based on the mapping results of the product's style tag and the target style tag in the matrix.
[0066] The color matching rules include the mutual matching mapping relationship between different color label pairs, which is used to determine whether the products meet the color matching conditions based on the mutual matching mapping result between the color labels of the first product and the color labels of the second product.
[0067] The rules are centrally managed in the form of structured data such as dictionaries.
[0068] Multi-algorithm recommendation module, combined Figure 3 The multi-algorithm recommendation module includes a rule-based recommendation unit, a behavior-based recommendation unit, and a hybrid recommendation unit, which are used to output recommendation results based on the data from the data layer.
[0069] In the multi-algorithm recommendation module:
[0070] The rule recommendation unit assigns a weighted score to five dimensions—price matching, style matching, body size matching, season matching, and brand matching—based on preset weights of 0.3, 0.25, 0.2, 0.15, and 0.1, respectively. It also calculates a comprehensive score by combining factors such as occasion matching, new products, best-selling items, ratings, timeliness, and inventory availability, generating interpretable recommendation reasons and recommendation tags.
[0071] The behavior recommendation unit assigns corresponding weights to different behavior types and constructs a "user-item" interaction matrix based on time decay. Item-based collaborative filtering is then performed based on this matrix, and the results are weighted and fused with TF-IDF content similarity constructed based on product text features. In this embodiment, user behaviors are weighted by type, with the following weight allocation: browsing = 1, adding to cart = 3, favorites = 4, purchasing = 5. A user-item interaction matrix is constructed using a time decay factor, which decays daily over 30 days. Item-based collaborative filtering is performed based on this matrix. Simultaneously, TF-IDF content similarity is constructed based on the product's title, description, style tags, category, brand, color, and other text features. Behavior similarity and content similarity are weighted and fused to form a hybrid similarity matrix for recommendation, where behavior similarity has a weight of 0.6 and content similarity has a weight of 0.4.
[0072] The hybrid recommendation unit dynamically allocates the weights of rule-based recommendations and behavioral recommendations based on the user's cold start level. The recommendation results from the rule-based recommendation unit and the behavioral recommendation unit are weighted, merged, and output in sorted order of scores.
[0073] Preferably, the multi-algorithm recommendation module further includes:
[0074] The cold start rollback unit is used to perform multi-level rollback when both user profile and behavioral data are insufficient, sequentially using user profile recommendations, category-based popular recommendations, and global popular recommendations.
[0075] The category diversity filtering unit is used to merge accessories into one slot and allow a maximum of 1 item to be recommended, while each other first-level category can recommend a maximum of 1 item.
[0076] Combined with the generation module, Figure 4 The matching generation module is used to generate a complete outfit scheme based on the target input, combined with the matching rule engine and the multi-algorithm recommendation module;
[0077] Specifically, the execution process of the matching generation module is as follows:
[0078] Determining Input and Anchor Points: After a user enters the system page, the client submits the user's current needs and necessary conversation information to the recommendation service. User needs may include target items, desired style, occasion, season, price range, or other preferences. If the user has not explicitly specified a target, the system can select seed clothing items based on user profile, recent interaction information, category popularity, and clothing availability, using their respective main clothing categories as styling anchor points. When user information is insufficient, more general category or popularity results can be used as a fallback basis.
[0079] Determining the Target Style: After identifying the anchor point, the system combines the attributes of the anchor garment, user preferences, and current needs to determine the target style, and reads the corresponding matching slots from the rule engine. Candidate garments are first filtered based on criteria such as category, season, size, availability, style compatibility, and color coordination. Then, a comprehensive evaluation is formed by considering price suitability, style suitability, color coordination, and other available features. The specific influence of each evaluation factor can be adjusted based on actual data and operational needs, and is not limited to a fixed proportion. For garments that have appeared frequently recently, the system can appropriately reduce the probability of them being selected again, and adopt an adaptive sampling method with randomness from the candidate set that meets the constraints, to balance matching accuracy and styling diversity.
[0080] Slot filling: Select products one by one according to the recommended combination slots corresponding to the target style. If the slot is a candidate category, then go back to the first category with stock.
[0081] The specific clothing combinations generated by the system are not pre-set templates, nor do they correspond to a unique result. The system dynamically generates the entire outfit based on the user's current request, user preferences, candidate clothing status, historical display data, and random sampling results. Therefore, even if the anchor clothing or target style is the same, different users, different times, or different batches of generation can obtain different combinations of tops, bottoms, shoes, and accessories.
[0082] The system performs completeness and usability checks on the generated results, excluding schemes with excessive duplication, missing key categories, or abnormal clothing status, and outputs several outfit cards for users to browse. Users can save the entire outfit, initiate a virtual try-on, or replace individual garments. During the same browsing process, the system can temporarily save the generated results to maintain consistency in display when flipping through pages or revisiting; if the results become invalid or conditions change, new outfits can be regenerated.
[0083] Vectorized scoring: Vectorize the scores of candidate products for each slot, and randomly sample from the high-scoring candidates based on their scores.
[0084] Variation expansion and deduplication: Generate 3 sets of variation combinations for each style, and deduplicate the variation combinations using the product set as the key;
[0085] Combination scoring and ranking: Price reasonableness is defined by whether the overall price falls within the user's preferred range, and combination completeness is defined. The combination completeness is judged as follows: the combination must contain at least the target product and one combination product; when the number of combination products reaches a preset threshold, a completeness score is obtained; if there is no suitable product in the combination slot, the slot is skipped; each combination is scored based on price reasonableness and combination completeness and sorted in descending order.
[0086] SKU validity verification: Verify whether the inventory of each product in the outfit is valid and available for sale, remove incomplete outfits with products that have been discontinued, and finally output a complete outfit scheme including style, applicable occasion, total price and product details.
[0087] Virtual try-on module, combined with Figure 5 The virtual try-on module is used to input model images and clothing images into a multimodal generation model to generate a try-on result image.
[0088] Preferably, when the matching generation module performs vectorized scoring for each slot, the scoring change rule is as follows:
[0089] Price matching adds 0.3 points, style matching adds 0.4 points, and color compatibility adds 0.2 points.
[0090] Recommended products are weighted at a power of 0.6 based on their usage frequency, and then randomly sampled from high-scoring candidates based on their scores.
[0091] The virtual try-on module is used to input model images and clothing images into a multimodal generation model to generate images of the try-on results.
[0092] Users can choose from the candidate clothing provided by the system, or call up previously selected or saved clothing, and can filter by clothing category. The system will only give a prompt when it finds category conflict, style inconsistency or obvious color incompatibility, and can provide alternative candidates that meet the current needs.
[0093] Specifically, the virtual try-on module includes: task submission and preprocessing: receiving model images, single or multiple clothing images uploaded by users or passed through a URL, verifying the format, size, and file size of the images, and performing preprocessing; the preprocessing may include EXIF orientation correction, converting the transparency channel to white background, or proportional scaling and centering;
[0094] Batch jigsaw puzzle: combined with Figure 6 The system explains that when a user selects multiple garments for try-on, it first preprocesses the images of each garment, adjusting their format, orientation, transparent background, subject size, and canvas position. Then, it adaptively arranges these images into a composite image based on the number of garments, ensuring each garment remains intact and minimizing occlusion. The specific arrangement of the composite image can be adjusted according to the number of garments and the proportion of the input images, and is not limited to a fixed number of rows and columns. This allows for batch try-on of multiple garments in a single generation process.
[0095] The multimodal generation module is invoked: Clothing images, jigsaw puzzle images, and model images are input into the multimodal generation model, and dressing prompts are injected. These prompts constrain the generation model to output a dressing result. The dressing result is a high-definition image where only clothing is replaced, while preserving the model's face, expression, hairstyle, body posture, and background. The style, color, texture, and cut of the clothing are consistent with the input product, and the lighting and folds conform to the human posture. In this embodiment, after the user initiates a dressing change, the client submits a person image and the current clothing combination. The person image can be uploaded by the user or selected from images authorized by the user. After necessary verification and preprocessing of the input images, the server inputs the person image and clothing image into the multimodal generation model. Constraint information is used to ensure the model primarily replaces the clothing area, maintaining the stability of the person's face, expression, hairstyle, body posture, and background as much as possible, while preserving the style, color, pattern, and overall cut of the input clothing. This ensures that the occlusion relationships, lighting, and folds in the generated result are adapted to the person's posture.
[0096] Try-on requests are processed by the server in the background, allowing the client to receive a simplified processing progress update and view the results upon completion. In the event of a temporary exception, the system can reprocess the request according to a preset strategy or prompt the user to resubmit, while retaining the currently selected clothing and character images to reduce repetitive operations. Specific task identifiers, status classifications, retry methods, and processing durations can be determined based on the system deployment environment and are not limited. Intermediate data and generated results during processing can be temporarily saved as needed and cleaned up after becoming invalid.
[0097] The generated results can be previewed, regenerated, or saved by the user. When regenerating, the system can retain the current clothing combination and the image of the person, and obtain different try-on results by adjusting the generation conditions; when saving, the system will associate the current outfit information with the generated image for easy viewing or further adjustments later.
[0098] Preferably, the virtual try-on module also performs result management and retries when calling the multimodal generation model, and the execution method is as follows:
[0099] The multimodal generation process adopts an asynchronous task mechanism. After the client submits the task, it immediately obtains the task identifier. The server background thread executes the generation and uses exponential backoff retries for momentary failures. The client polls the task status to obtain the progress and downloads the high-resolution result image.
[0100] The transaction service module provides a unified API gateway, caching and pagination of recommendation and trial tasks, and user quota and interface rate limiting management to provide an integrated closed-loop service from recommendation to transaction.
[0101] Specifically, the transaction service module is as follows:
[0102] A unified API gateway is used to encapsulate and forward the recommendation and try-on interfaces in a unified manner;
[0103] The recommendation results and fitting results are stored in the cache with task identifiers and a time-to-live (TTL) is set. The first request generates the full results and caches them, and subsequent pagination requests directly hit the cache.
[0104] Database transactions are used to ensure consistency between user replacement quota deductions and usage records, and API rate limiting is used to control request frequency.
[0105] The scope of this invention is not limited to the above-described embodiments; a combination of one or more specific embodiments can also achieve the purpose of the invention.
Claims
1. A system for generating clothing combinations, virtual try-on, and trading based on multi-dimensional features, characterized in that: include: The data layer is used to store product data, user profile data, and user behavior log data; The matching rules engine is used to configure and manage category matching rules, style compatibility matrices, and color matching rules; The multi-algorithm recommendation module includes a rule recommendation unit, a behavior recommendation unit, and a hybrid recommendation unit, which is used to output recommendation results based on the data layer. The outfit generation module is used to generate a complete outfit scheme based on the target input, combined with the outfit rule engine and the multi-algorithm recommendation module; The virtual try-on module is used to input model images and clothing images into a multimodal generation model to generate images of the try-on results. The transaction service module provides a unified API gateway, caching and pagination of recommendation and trial tasks, and user quota and interface rate limiting management to provide an integrated closed-loop service from recommendation to transaction.
2. The system for generating clothing combinations, virtual try-on, and trading based on multi-dimensional features according to claim 1, characterized in that: The product data includes the main product category, brand, color, pattern, style tag, applicable season, available sizes, fabric, price, sales volume, click volume, and rating. The user profile data includes gender, age, height, weight, body shape, chest circumference, waist circumference, hip circumference, shoe size, style preference, color preference, brand preference, and price preference; The user behavior log data includes the behavior types such as browsing, adding to cart, favorites and purchasing, behavior timestamps, source pages, session identifiers and device types.
3. The system for generating clothing combinations, virtual try-on, and trading based on multi-dimensional features according to claim 1, characterized in that, The matching rule engine specifically includes: category matching rules, style compatibility matrix, and color matching rules; The category matching rules define the set of main product categories that can be matched with the main product category, as well as the recommended matching combination slots under each style tag; the matching combination slots can be fixed categories at each level, or they can be alternative categories at higher levels that are rolled back in sequence. The style compatibility matrix contains compatibility mapping flags or thresholds between different style tag pairs, which are used to determine whether a product meets the style matching conditions based on the mapping results of the product's style tag and the target style tag in the matrix. The color matching rules include the mutual matching mapping relationship between different color label pairs, which is used to determine whether the products meet the color matching conditions based on the mutual matching mapping result between the color labels of the first product and the color labels of the second product. The rules are centrally managed in the form of structured data such as dictionaries.
4. The system for generating clothing combinations, virtual try-on, and trading based on multi-dimensional features according to claim 1, characterized in that, In the multi-algorithm recommendation module: The rule recommendation unit assigns a weighted score to five dimensions—price matching, style matching, body size matching, season matching, and brand matching—according to preset weights. The behavior recommendation unit assigns corresponding weights to different behavior types and constructs a "user-item" interaction matrix by combining time decay. Based on the matrix, it performs item-based collaborative filtering and then weightedly fuses the collaborative filtering results with TF-IDF content similarity constructed based on product text features. The hybrid recommendation unit dynamically allocates the weights of rule-based recommendations and behavioral recommendations based on the user's cold start level. The recommendation results from the rule-based recommendation unit and the behavioral recommendation unit are weighted, merged, and output in sorted order of scores.
5. The system for generating clothing combinations, virtual try-on, and trading based on multi-dimensional features according to claim 1, characterized in that, The multi-algorithm recommendation module also includes: The cold start rollback unit is used to perform multi-level rollback when both user profile and behavioral data are insufficient, sequentially using user profile recommendations, category-based popular recommendations, and global popular recommendations. The category diversity filtering unit is used to merge accessories into one slot and allow a maximum of 1 item to be recommended, while each other first-level category can recommend a maximum of 1 item.
6. The system for generating clothing combinations, virtual try-on, and trading based on multi-dimensional features according to claim 1, characterized in that, The specific execution process of the matching generation module is as follows: Determine the input and anchor points: Use the user-selected or system-recommended products, user-provided products, the seed products with the highest user behavior scores, or the specified styles as inputs, and use categories such as tops, bottoms, skirts, dresses, and suits as anchor points for whole-pool sampling; Determine the target style: prioritize the product's own style, followed by the user's preferred style; the default target style is casual, and the style compatibility matrix is used to verify whether the product style and the target style meet the style matching conditions. Slot filling: Select products one by one according to the recommended combination slots corresponding to the target style. If the slot is a candidate category, then go back to the first category with stock. Vectorized scoring: Vectorize the scores of candidate products for each slot, and randomly sample from the high-scoring candidates based on their scores. Variation expansion and deduplication: Generate multiple variation combinations for each style, and deduplicate the variation combinations using the product set as the key; Combination scoring and ranking: Price reasonableness is defined by whether the overall price falls within the user's preferred range, and combination completeness is defined. The combination completeness is judged as follows: the combination must contain at least the target product and one combination product; when the number of combination products reaches a preset threshold, a completeness score is obtained; if there is no suitable product in the combination slot, the slot is skipped. Each outfit is scored based on its price reasonableness and the completeness of its combination, and then ranked in descending order. SKU validity verification: Verify whether the inventory of each product in the outfit is valid and available for sale, remove incomplete outfits with products that have been discontinued, and finally output a complete outfit scheme including style, applicable occasion, total price and product details.
7. The system for generating clothing combinations, virtual try-on, and trading based on multi-dimensional features according to claim 6, characterized in that, The specific rules for score changes during slot-by-slot vectorized scoring in the matching generation module are as follows: Price matching adds 0.3 points, style matching adds 0.4 points, and color compatibility adds 0.2 points. Recommended products are weighted at a power of 0.6 based on their usage frequency, and then randomly sampled from high-scoring candidates based on their scores.
8. The system for generating clothing combinations, virtual try-on, and trading based on multi-dimensional features according to claim 1, characterized in that, The virtual try-on module specifically includes: task submission and preprocessing: receiving model images, single or multiple clothing images uploaded by users or passed through a URL, verifying the format, size, and file size of the images, and performing preprocessing; the preprocessing may include EXIF orientation correction, converting the transparency channel to white background, or proportional scaling and centering; Batch collage: When an image involves multiple garments, multiple garment images are stitched together into a single collage according to a preset size grid, enabling batch try-on of multiple garments in a single generation. Call the multimodal generation module: Input the clothing image, puzzle image and model image into the multimodal generation model, and inject dressing prompt words. The prompt words constrain the generation model to output the dressing result. The dressing result is a high-definition result image in which only the clothing is replaced, while the model's face, expression, hairstyle, body posture and background remain unchanged, and the style, color, texture and pattern of the clothing are consistent with the input product, and the light and shadow and folds fit the human body posture.
9. A system for generating clothing combinations, virtual try-on, and trading based on multi-dimensional features according to claim 8, characterized in that, The virtual try-on module also performs result management and retries when calling the multimodal generation model, specifically as follows: The multimodal generation process adopts an asynchronous task mechanism. After the client submits the task, it immediately obtains the task identifier. The server background thread executes the generation and uses exponential backoff retries for momentary failures. The client polls the task status to obtain the progress and downloads the high-resolution result image.
10. A system for generating clothing combinations, virtual try-on, and trading based on multi-dimensional features according to claim 1, characterized in that, The transaction service module specifically includes: A unified API gateway is used to encapsulate and forward the recommendation and try-on interfaces in a unified manner; The recommendation results and fitting results are stored in the cache with task identifiers and a time-to-live (TTL) is set. The first request generates the full results and caches them, and subsequent pagination requests directly hit the cache. Database transactions are used to ensure consistency between user replacement quota deductions and usage records, and API rate limiting is used to control request frequency.
