A multi-modal large model-based children's clothing personalized intelligent design generation system

Abstract

The application relates to the technical field of artificial intelligence, and discloses a children's clothing personalized intelligent design generation system based on a multi-modal large model, which comprises a multi-modal demand analysis module, a children's clothing portrait construction module, a knowledge constraint reasoning module, a children's clothing design generation module and a feedback output service module. The system standardizes, uniformly encodes and weightedly fuses text description, reference images, fabric data, children information and demand data, outputs a fusion vector and a conflict list, establishes size intervals, growth margins, action margins and constraints in combination with measured information, forms a constraint portrait, retrieves design knowledge and processes conflicts, forms generation conditions and condition sequences, performs reverse denoising based on constraint correction, generates a candidate design through region refinement, consistency verification, safety verification and process analysis, and completes evaluation and modification processing. The application improves the fit, safety and process realizability of design by fusing multi-modal design information and introducing children's constraints.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a personalized intelligent design generation system for children's clothing based on a multimodal large model. Background Technology

[0002] Currently, generative artificial intelligence, multimodal learning, and large-scale model technology have been gradually applied to the design of clothing styles, patterns, and fabric combinations. Related systems typically receive text, images, sketches, or user preferences, and output clothing design drawings after semantic parsing, feature extraction, and generative model processing. Children's clothing design also involves children's age, body type, range of motion, growth margin, and wearing safety, requiring simultaneous handling of personalized expression, size adaptation, fabric selection, and manufacturing processes.

[0003] Regarding the aforementioned intelligent clothing design, Publication No. CN120431199A discloses a clothing design method, system, device, and storage medium based on a large model. This method acquires sample images, descriptive text, and user behavior data, and drives an image generation model through image feature extraction, text semantic understanding, and prompt word generation. Publication No. CN120765778A discloses a clothing design system, method, and storage medium based on multimodal AIGC, mapping text descriptions, clothing style diagrams, fabric images, and user preferences to a unified semantic space, and generating clothing design diagrams by combining style transfer and conditional diffusion models. Authorization Announcement No. CN114491125B uses a multimodal encoding and attention mechanism to uniformly process text, sketches, and local images, then uses an encoder and decoder to generate clothing images; Authorization Announcement No. CN114723843B integrates text and image features to generate virtual clothing.

[0004] In the field of personalized clothing and children's wear design, publication number CN117371173A discloses an intelligent children's wear design method and system based on multimodal perception, interactive emotional intelligence intervention. This method collects individual characteristics, user preferences, and design knowledge, outputs children's wear designs through interactive function design modules and appearance design modules, and makes corrections based on user feedback. Publication number CN120046515A discloses a multimodal personalized clothing design method and system, which uses a multilayer perceptron to form personalized feature vectors, utilizes decision trees to establish design rules, and then employs a nearest neighbor algorithm to output design solutions.

[0005] Furthermore, Publication No. CN117951763A discloses a multimodal data-driven generative fashion-compatible clothing design method and system, which completes design processing through speech recognition, manuscript rendering, clothing attribute interactive design, and aesthetic scoring. Publication No. CN121257282A uses keyword semantic parsing, cultural corpus modeling, style semantic fusion, and a multimodal diffusion model to generate clothing patterns, and optimizes them through structural consistency evaluation. Publication No. CN122087895A discloses a personalized doll clothing customization system based on traditional clothing elements, which parses text, reference images, and hand-drawn sketches, and completes the size adaptation of traditional clothing elements through proportion adjustment and density adjustment.

[0006] As the above solutions demonstrate, existing technologies can already handle multimodal information processing and clothing image generation. However, children's individual information is typically used as general input data in the design process and has not yet been further converted into size ranges, growth margins, movement margins, and hierarchical design constraints. General clothing generation methods often focus on semantic matching, style transfer, or pattern quality, lacking a unified connection between children's body shape, wearing safety, fabric performance, and production processes. Existing children's clothing solutions tend to emphasize voice interaction, temperature-sensitive color changes, or emotional intervention, failing to establish a processing flow that integrates requirements analysis, knowledge reasoning, design generation, and result verification. After the generation results are completed, the joint evaluation of fit, safety, and process feasibility remains insufficient, and local modifications may affect clothing structures and design data that were not adjusted.

[0007] Therefore, this invention provides a personalized intelligent design generation system for children's clothing based on a multimodal large model to address the shortcomings of existing technologies. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a personalized intelligent design generation system for children's clothing based on a multimodal large model, which solves the problem that existing intelligent clothing design cannot simultaneously meet the personalized needs of children's clothing, children's body shape adaptation, wearing safety, and process feasibility.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A personalized intelligent design generation system for children's clothing based on a multimodal large model includes a multimodal demand analysis module, a children's clothing profile construction module, a knowledge-constrained reasoning module, a children's clothing design generation module, and a feedback output service module.

[0010] The multimodal requirement analysis module is used to acquire and standardize design data, perform fusion processing on the design data, and output a fusion vector and a conflict list.

[0011] The children's clothing profile construction module is used to obtain measured information and construct a constraint profile based on the fusion vector and the conflict list, combined with the measured information.

[0012] The knowledge constraint reasoning module is used to retrieve design knowledge based on the constraint profile, handle conflicts in the retrieval process, and form generation conditions and condition sequences.

[0013] The children's clothing design generation module is used to generate and verify designs based on the generation conditions and the condition sequence, and output candidate designs.

[0014] The feedback output service module is used to evaluate the candidate design and output a result package, as well as to receive the modified content and convert the modified content into new design data.

[0015] By adopting the above technical solutions, text descriptions, reference images, fabric data, children's information, and demand data are standardized and integrated to form unified design conditions. Constraint profiles are established by combining children's actual measurement information. Then, design knowledge, individual child constraints, and generation conditions jointly guide the generation, verification, and modification of candidate designs, so that the output results are adapted to children's sizes, wearing activities, safety requirements, fabric properties, and production processes.

[0016] Preferably, the multimodal requirement parsing module receives text descriptions, reference images, fabric data, child information, and requirement data, and uses the text descriptions, reference images, fabric data, child information, and requirement data as input modalities for the design data.

[0017] Preferably, the multimodal demand parsing module performs character encoding unification, clothing terminology standardization, image size processing, children's size unit conversion, and fabric field standardization on the design data, while retaining the original data and data source before standardization processing.

[0018] Preferably, the multimodal requirement analysis module performs quality evaluation on the design data. The quality evaluation includes text completeness evaluation, image clarity evaluation, clothing body completeness evaluation, clothing component recognizability evaluation, fabric texture recognizability evaluation, children's information completeness evaluation, and data source status evaluation. Based on the quality evaluation, missing records and conflict list are established.

[0019] Preferably, the conflict list records the attribute differences between different input modalities according to clothing category, overall silhouette, neckline, sleeve type, waistline, placket, pocket, hem, color, pattern, fabric, and decorative parts.

[0020] Preferably, the multimodal demand parsing module inputs the text description, the reference image, the fabric data, the child information, and the demand data into the corresponding encoding network and outputs feature vectors. The encoding network includes a text encoding network, a visual encoding network, a fabric attribute encoding network, a child feature encoding network, and a user preference encoding network.

[0021] Preferably, the multimodal requirement parsing module calculates the semantic relevance and reliability of each input modality with the design task, determines the fusion weight based on the semantic relevance and the reliability, and performs weighted fusion on the feature vector based on the fusion weight to generate the fusion vector.

[0022] Preferably, the fusion weights are determined according to the clothing attributes, so that the fusion weights of the same input modality on the overall silhouette, color, pattern, fabric and local clothing parts are calculated separately.

[0023] Preferably, the children's clothing profile construction module acquires the measured information including the child's age, height, and weight, and determines the size range, growth margin, and movement margin based on the measured information; body parameters not included in the measured information are used to form inferred data based on a children's body type database.

[0024] Preferably, the size range is determined based on measured values, measurement errors, the child's age stage, gender, height, and weight; the growth margin is determined based on the expected wearing period, clothing category, clothing coverage area, and fabric elasticity; and the movement margin is determined based on the wearing scenario and the displacement of key body points.

[0025] Preferably, the children's clothing portrait construction module establishes safety hard constraints, body fit strong constraints, and aesthetic soft constraints based on the fusion vector and the conflict list, with the processing priority of the safety hard constraints being higher than that of the body fit strong constraints and the aesthetic soft constraints.

[0026] Preferably, the security hard constraint is established based on the currently effective security rules in the security rule base. The security rule includes rule identifier, applicable age group, applicable clothing category, applicable wearing scenario, applicable sales area, restricted object, restriction parameter, restriction value and detection method.

[0027] Preferably, the strong constraints on the fit include size range, growth allowance, movement allowance, connection relationship of garment components, and production process achievable conditions related to production equipment capacity and fabric processing capacity.

[0028] Preferably, the aesthetic soft constraints include color preference, pattern preference, style preference, overall silhouette preference, and decorative component preference.

[0029] Preferably, the children's clothing portrait construction module extracts body regions, clothing components, and fabric attributes from the fusion vector and the conflict list, and establishes associations between the body regions, clothing components, and fabric attributes and the safety hard constraints, the fit strong constraints, and the aesthetic soft constraints to generate the constraint portrait.

[0030] Preferably, the associations include size dependency relationships, spatial coverage relationships, structural connection relationships, prohibition relationships, substitution relationships, and attribute association relationships.

[0031] Preferably, the knowledge constraint reasoning module extracts the child's age stage, size range, clothing category, applicable season, wearing scenario, overall silhouette, color tendency, pattern theme, fabric attributes, safety hard constraints, and production equipment capabilities from the constraint profile as query fields.

[0032] Preferably, the knowledge constraint reasoning module performs structured filtering based on the query fields and performs vector retrieval on the structured filtered design knowledge to obtain candidate knowledge as the design knowledge.

[0033] Preferably, the candidate knowledge includes knowledge of children's clothing styles, colors and patterns, fabrics, children's body shapes, safety rules, pattern making and manufacturing processes, and historical design case studies.

[0034] Preferably, the knowledge constraint reasoning module checks for conflicts between explicit requirements and image attributes in the design data, the constraint profile, and the candidate knowledge, and deletes conflicting attributes or selects alternative structures, alternative fabrics, and alternative production processes in the order of processing safety hard constraints, body fit strong constraints, and aesthetic soft constraints.

[0035] Preferably, the knowledge constraint reasoning module completes the component attributes required for design generation but not included in the design data, and writes the completed component attributes into the instruction diagram to form the generation conditions.

[0036] Preferably, the instruction diagram includes clothing category nodes, overall silhouette nodes, front garment piece nodes, back garment piece nodes, neckline nodes, sleeve type nodes, waist node, hem node, placket node, pocket node, color node, pattern node, fabric node, decoration node, and production process node. Each node records attribute values, spatial location, size range, connection relationship, prohibition conditions, and knowledge source.

[0037] Preferably, the knowledge constraint reasoning module calculates the semantic similarity, category matching degree, comfort matching degree, and conflict degree between the candidate knowledge and the constraint profile, and performs a weighted calculation on the semantic similarity, category matching degree, comfort matching degree, and conflict degree to obtain the matching degree.

[0038] Preferably, the knowledge constraint reasoning module selects selected knowledge from the candidate knowledge based on the matching degree, and converts the selected knowledge and the instruction diagram into positive design conditions, negative prohibition conditions, clothing component relationship conditions, spatial location conditions and knowledge evidence conditions, forming the condition sequence.

[0039] Preferably, the children's clothing design generation module generates a control diagram and a prohibited area diagram according to the generation conditions. The control diagram includes a front silhouette diagram, a back silhouette diagram, a clothing component partition diagram, a seam position diagram, a pleat distribution diagram, and a pattern distribution diagram.

[0040] Preferably, the prohibited area map is generated based on the safety hard constraints and is used to record the locations where ropes, small parts, hard decorative parts and sharp parts are not allowed to be installed, and uses the same coordinate system as the control map.

[0041] Preferably, the children's clothing design generation module inputs the control chart, the condition sequence, the constraint profile, and the forbidden region map into the diffusion network, performs reverse denoising based on constraint correction, and generates an intermediate design.

[0042] Preferably, the children's clothing design generation module calculates the denoising direction, fit deviation, safety deviation, and process deviation of the potential data during the denoising iteration process, and adjusts the denoising direction according to the fit deviation, the safety deviation, and the process deviation.

[0043] Preferably, the children's clothing design generation module processes the silhouette, fabric texture, color distribution, and local components during the noise reduction iteration process, and outputs the intermediate design.

[0044] Preferably, the diffusion network processes the silhouette and garment component positions in the early stage of iteration, processes the fabric texture and color distribution in the middle stage of iteration, and processes the seams, decorative components and other local components in the later stage of iteration.

[0045] Preferably, the children's clothing design generation module performs regional refinement, consistency verification, security verification, and process analysis on the intermediate design to generate the candidate design.

[0046] Preferably, the area refinement establishes local areas according to the collar, cuffs, placket, waist, pockets, hem, and decorative parts, and when regenerating the local areas, the garment structure, color, and fabric data outside the local areas are locked.

[0047] Preferably, the consistency verification includes verification of the shoulder line height, side seam length, waist height, hem length, fabric texture direction, and pattern cross-seam position of the front and back designs.

[0048] Preferably, the safety verification includes neckline size verification, cord and strap position verification, small component position verification, sharp component position verification, decorative component fixing area verification, and skin contact surface verification. When a violation of the safety hard constraint is detected, the corresponding intermediate design is deleted.

[0049] Preferably, the process analysis includes identifying the front garment piece, back garment piece, sleeve piece, collar piece, waist section component, placket component, pocket component, hem component, and decorative component, and generating a garment component list, pattern size range, sewing sequence, seam allowance data, binding method, and equipment requirements based on component boundaries, connection relationships, seam positions, pleat density, and fabric properties.

[0050] Preferably, the children's clothing design generation module generates multiple candidate designs by changing the initial random potential data, color combinations, pattern layouts, and local component combinations, and performs repeatability detection based on structural and visual features.

[0051] Preferably, the feedback output service module calculates the consistency score, integration score, safety score, coordination score, and process score of each candidate design, and sorts them according to preset evaluation rules.

[0052] Preferably, the preset evaluation rules include safety hard constraint processing rules, minimum adoption conditions for each score, comprehensive evaluation calculation rules, candidate design ranking rules, and candidate design duplication exclusion rules.

[0053] Preferably, the feedback output service module selects the candidate designs that meet the preset evaluation rules based on the sorting results, and arranges the effect diagrams, detail diagrams, color matching data, pattern and process suggestions to form the result package.

[0054] Preferably, the result package also includes safety hard constraint verification status, assembly verification status, process feasibility verification status, front and back consistency verification status, and items pending manual confirmation.

[0055] Preferably, the feedback output service module receives the modified content and locates the modified content to the corresponding modification node; when a modified content involves multiple clothing parts, the main modification node and associated modification nodes are determined simultaneously.

[0056] Preferably, the feedback output service module locks the structure and data of unaffected nodes and only regenerates the affected areas.

[0057] Preferably, the feedback output service module distinguishes the modified content into current task data and long-term preference data; when the user explicitly chooses to save the corresponding preference, the modified content is recorded as long-term preference data, and the modified content that does not meet the conditions for updating long-term preference data is recorded as current task data.

[0058] Preferably, the feedback output service module converts the regenerated result into the new design data, which includes the original design task identifier, modified content, modified nodes, affected areas, locked nodes, re-executed steps, regenerated results, and constraint verification results.

[0059] Preferably, the system stores the design data, the fusion vector, the conflict list, the constraint profile, the candidate knowledge, the selected knowledge, the instruction graph, the generation conditions, the condition sequence, the intermediate design, the candidate design, the result package, and the new design data in association according to the task identifier.

[0060] Preferably, the system updates the candidate ranking model and lightweight adaptation parameters using candidate designs and modifications that have been manually reviewed, while retaining the basic parameters of the multimodal large model and the model version before the update.

[0061] This invention provides a personalized intelligent design generation system for children's clothing based on a multimodal large model, which has the following beneficial effects: This invention standardizes, evaluates the quality of, uniformly encodes, and weights the text descriptions, reference images, fabric data, children's information, and demand data, and determines the corresponding fusion weights for different clothing attributes. This can reduce the impact of data format differences, missing inputs, and modal conflicts on the design results, and make the generation conditions more accurately reflect user needs and reference image features.

[0062] This invention combines actual measurement information of children to establish size ranges, growth margins, and movement margins, and establishes a correlation between body areas, clothing components, fabric properties, and hard safety constraints, strong fit constraints, and soft aesthetic constraints. This allows children's clothing design to be simultaneously constrained by children's body shape, wearing activities, safety rules, fabric performance, and production process conditions, thereby improving the fit, safety, and process feasibility of candidate designs.

[0063] In the reverse denoising process of the diffusion network, this invention adjusts the denoising direction according to fit deviation, safety deviation, and process deviation, and refines, verifies consistency, verifies safety, and analyzes the intermediate design execution area. After receiving the modification content, it locates the corresponding modified node and locks the unaffected node, which can reduce candidate designs that do not meet the constraints and maintain the stability of other garment structures and data when making local modifications. Attached Figure Description

[0064] Figure 1 This is a framework diagram of the children's clothing personalized intelligent design generation system based on a multimodal large model according to the present invention; Figure 2 This is a schematic diagram illustrating the system deployment and application scenarios of the present invention; Figure 3This is a flowchart illustrating the overall process of the personalized intelligent design generation method for children's clothing according to the present invention. Figure 4 This is a flowchart illustrating the multimodal children's clothing design requirements collection and fusion process of the present invention; Figure 5 Flowchart for constructing a personalized constraint profile for children according to the present invention; Figure 6 Design a reasoning flowchart for the knowledge constraints of this invention; Figure 7 This is a flowchart illustrating the children's clothing design generation and optimization process of the present invention. Figure 8 This is a flowchart of the evaluation output and feedback processing of candidate children's clothing solutions according to the present invention. Detailed Implementation

[0065] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] The data processing device in this embodiment includes a central processing unit (CPU), a graphics processing unit (GPU), random access memory (RAM), non-volatile memory, a network communication interface, and a display output interface.

[0067] Non-volatile memory stores computer programs, neural network parameters, children's clothing design knowledge base, children's body size data, production process data, design data, design task data, candidate designs, user feedback data, and new design data. The central processing unit (CPU) performs design data reception, field standardization, rule judgment, task scheduling, flow control, and result packet arrangement. The graphics processing unit (GPU) performs multimodal coding, cross-modal feature calculation, neural network inference, diffusion network generation, and candidate design evaluation.

[0068] The multimodal large model referred to in this embodiment is a neural network model capable of receiving text descriptions, reference images, fabric data, children's information, and demand data, and performing feature encoding, unified representation, associative reasoning, knowledge retrieval, and conditional generation on the design data. The multimodal large model includes a text encoding network, a visual encoding network, a fabric attribute encoding network, a children's feature encoding network, a user preference encoding network, a cross-modal attention network, a knowledge retrieval network, and a conditional diffusion generation network.

[0069] The text encoding network is used to identify clothing categories, applicable seasons, wearing occasions, design styles, colors, patterns, overall silhouettes, garment components, fabrics, wearing and undressing methods, and prohibited conditions. The visual encoding network is used to identify the main body of the garment, overall silhouette, neckline, sleeve type, waistline, placket, pockets, hem, pattern position, pattern density, decorative components, and fabric surface features in reference images. The fabric attribute encoding network is used to process fabric name, fiber composition, fabric weave, fabric weight, fabric thickness, elongation, breathability, moisture absorption, surface roughness, drape, and available manufacturing processes.

[0070] The child feature encoding network processes data on children's age, gender, height, weight, chest circumference, waist circumference, hip circumference, shoulder width, arm length, back length, torso length, leg length, expected wearing period, and activity scenarios. The user preference encoding network processes requirement data, including user choices regarding color, pattern, style, looseness, wearing and undressing methods, fabric feel, ease of care, and prohibited elements. The cross-modal attention network establishes attribute correspondences between each input modality and the current design task, and calculates the semantic relevance of each input modality.

[0071] The fusion vector, as referred to in this embodiment, is a data representation formed by uniformly encoding and weighted fusion of text descriptions, reference images, fabric data, children's information, and demand data. The conflict list, as referred to in this embodiment, is a data record formed when different input modalities have different attribute values ​​in terms of clothing category, overall silhouette, neckline, sleeve type, waistline, placket, pockets, hem, color, pattern, fabric, or decorative components.

[0072] The constraint profile referred to in this embodiment is a data set consisting of child size ranges, growth margins, movement margins, safety hard constraints, fit-up strong constraints, aesthetic soft constraints, and their interrelationships, formed based on fusion vectors, conflict lists, and measured information. The generation conditions referred to in this embodiment are design generation data formed by handling conflicts between explicit requirements, image attributes, constraint profiles, and candidate knowledge, and then writing the completed component attributes into the instruction diagram.

[0073] The condition sequence referred to in this embodiment refers to the sequential data of positive design conditions, negative prohibition conditions, clothing component relationship conditions, spatial location conditions, and knowledge evidence conditions formed by converting selected knowledge and instruction graphs. The intermediate design referred to in this embodiment refers to the children's clothing design data output after the diffusion network has completed inverse denoising, but before regional refinement, consistency verification, security verification, and process analysis have been completed.

[0074] The "result package" referred to in this embodiment refers to the data set including renderings, detail images, color matching data, pattern and process suggestions, and verification results generated after the candidate designs have been evaluated and ranked. The "new design data" referred to in this embodiment refers to the design task data generated after locating the modification nodes based on the modifications, locking unaffected nodes, and regenerating the affected areas.

[0075] See Figure 1 The present invention provides a children's clothing personalized intelligent design generation system based on a multimodal large model, including a multimodal demand analysis module, a children's clothing profile construction module, a knowledge constraint reasoning module, a children's clothing design generation module, and a feedback output service module.

[0076] The multimodal requirements analysis module is connected to the children's clothing profile building module, which in turn is connected to the knowledge constraint reasoning module. The knowledge constraint reasoning module is then connected to the children's clothing design generation module, which in turn is connected to the feedback output service module. The feedback output service module is also connected to the multimodal requirements analysis module, used to convert modified content into new design data and use this new design data as input for new design tasks.

[0077] Each module transmits design task data packets via an internal data bus or an Application Programming Interface (API). The design task data packet includes a task identifier, user identifier or anonymous session identifier, data version, input time, clothing category, wearing scenario, design data, data source marker, data credibility, authorization status, processing status, and output address.

[0078] Design data includes text descriptions, reference images, fabric data, child information, and requirement data. Color information is stored as an attribute field in the text descriptions, reference images, fabric data, or requirement data. Design data, fusion vectors, missing records, conflict lists, constraint profiles, candidate knowledge, selected knowledge, instruction diagrams, generation conditions, condition sequences, intermediate designs, candidate designs, result packages, and new design data for the same design task are linked through task identifiers.

[0079] The system establishes a children's clothing design knowledge base and cloud-based data and model services. The knowledge base is connected to the children's clothing profile building module, the knowledge-constrained reasoning module, and the children's clothing design generation module. The cloud-based data and model services are connected to the multimodal requirements analysis module, the children's clothing profile building module, the knowledge-constrained reasoning module, the children's clothing design generation module, and the feedback output service module, used to store model parameters, design knowledge, design task versions, candidate designs, result packages, user feedback records, and model update records.

[0080] The children's wear design knowledge base includes a children's wear style library, a color and pattern library, a fabric database, a children's body shape database, a safety rules library, a pattern and process library, and a historical design case library. Design knowledge in the knowledge base is stored in the form of knowledge entries. Each knowledge entry includes a knowledge entry identifier, knowledge type, applicable age group, applicable clothing category, applicable season, applicable fabric, applicable production equipment, knowledge content, data source, effective version, expiration status, and reliability level.

[0081] The safety rule entries in the safety rule base include rule identifier, applicable age group, applicable clothing category, applicable wearing scenario, applicable sales area, restricted object, restricted parameter, restricted value, detection method, data source, effective date, expiration date, and enforcement level. The system calls the restricted values ​​of the currently effective version according to the sales area, clothing category, and child age group corresponding to the current design task. The restricted values ​​for cord length, detachable small part size, decorative part position, and neckline size are stored in the safety rule base and are not written into the fixed parameters of the multimodal large model.

[0082] The multimodal requirements analysis module includes a data receiving unit, a data standardization unit, an input quality evaluation unit, an attribute identification unit, a multimodal coding unit, and a trusted fusion unit.

[0083] In some embodiments, the data receiving unit receives text descriptions, reference images, fabric data, children's information, and requirement data through a mobile terminal, web terminal, designer workstation, or application programming interface. The text description includes clothing category, applicable season, wearing scenario, design style, color requirements, pattern requirements, overall silhouette requirements, clothing component requirements, fabric requirements, dressing and undressing requirements, and prohibited conditions. The text description is either directly input by the user or formed from voice data after speech recognition processing. The text description formed by speech recognition processing retains the corresponding voice data identifier and speech recognition confidence record. Reference images include overall images of children's clothing, partial images of children's clothing, clothing style diagrams, hand-drawn sketches, pattern images, fabric images, and style reference images.

[0084] The multimodal requirements parsing module stores the source, acquisition time, image size, image type, image purpose, authorization status, and user-specified priority of each reference image. Overall images of children's clothing are used to identify the overall silhouette and proportions of garment components. Detail images of children's clothing are used to identify necklines, cuffs, plackets, waistlines, pockets, hems, or decorative elements. Style reference images are used to identify color organization, pattern density, and surface style.

[0085] The child's information includes age, gender, height, weight, chest circumference, waist circumference, hip circumference, shoulder width, arm length, back length, trunk length, leg length, expected wearing period, and activity scenario. Age is stored in months, height in cm, weight in kg, and chest circumference, waist circumference, hip circumference, shoulder width, arm length, back length, trunk length, and leg length in mm. The child's information also includes the data acquisition method, measurement time, measurement tool, and data status. Data acquisition methods include manual measurement, 3D scanning, historical order data, and inference from a children's body shape database. The measured information in the child's information includes the child's age, height, weight, and other body measurement data obtained through manual measurement, 3D scanning, or user confirmation. Data status includes measured data, user-confirmed data, and inferred data. Inferred data does not overwrite measured data.

[0086] Fabric data includes fabric images, fabric name, fiber composition, fabric weave, fabric weight, fabric thickness, elongation at break, air permeability, moisture absorption, coefficient of friction, drape rating, surface roughness rating, and available manufacturing processes. Fabric weight is measured in g / m². 2 Fabric thickness is measured in mm, breathability in mm / s, and elongation and moisture absorption are recorded as percentages. Requirements data include color preferences, pattern preferences, style preferences, overall silhouette preferences, comfort preferences, ease of wear, fabric feel requirements, ease of care requirements, and prohibited elements.

[0087] In some embodiments, the data standardization unit converts text descriptions into a unified character encoding and performs synonym normalization on garment names, fabric names, garment component names, color names, and manufacturing process names. The data standardization unit converts children's sizes from different sources into unified units, converts fabric attributes into unified fields, and establishes a correspondence between each standardized field and the field identifiers in the children's clothing design knowledge base. The data standardization unit determines the model input size of the reference image according to the input interface of the multimodal large model. The reference image is scaled and edge-filled while maintaining its original aspect ratio; the original reference image, the scaled reference image, and the image transformation parameters are saved separately.

[0088] In some embodiments, the input quality evaluation unit evaluates text completeness, text ambiguity, image clarity, clothing body completeness, clothing component recognizability, fabric texture recognizability, child information completeness, data source status, and the degree of conflict between different input modalities. The input quality evaluation results are normalized to 0–1. The quality judgment value for each target attribute is determined using an annotated data quality validation set. For each target attribute, samples that correctly identify the target attribute are marked as usable samples, and samples that cannot correctly identify the target attribute are marked as unusable samples. Samples are sorted according to the input quality evaluation results, and a cutoff value that meets the system's recognition error rate requirement is selected as the quality judgment value for that target attribute. The quality judgment value is stored together with the multimodal large model version and is re-determined when the multimodal large model version is updated.

[0089] In some embodiments, the attribute recognition unit establishes an attribute table according to clothing category, overall silhouette, neckline, sleeve type, waistline, placket, pockets, hem, color, pattern, fabric, and decorative parts. When text descriptions, reference images, and requirement data have different contents on the same attribute, the attribute recognition unit generates a conflict list. Each conflict record in the conflict list includes the conflicting attribute name, the attribute value of each input source, the quality evaluation result of each input source, the user-specified priority of each input source, the data acquisition time, and the processing status. Missing records include the missing field name, the corresponding input modality, whether the missing field is a necessary field for generation, the method of supplementation, and the processing status.

[0090] In some embodiments, the multimodal coding unit outputs feature vectors through coding networks. These coding networks include a text coding network, a visual coding network, a fabric attribute coding network, a child feature coding network, and a user preference coding network. Text descriptions are input to the text coding network, reference images are input to the visual coding network, fabric data is input to the fabric attribute coding network, child information is input to the child feature coding network, and demand data is input to the user preference coding network. Each coding network outputs feature vectors of the same dimension according to the input interface of the multimodal large model, and each feature vector is normalized to the same numerical range.

[0091] In some embodiments, the trusted fusion unit calculates the semantic relevance of each input modality to the design task using a cross-modal attention network. The input quality evaluation unit determines the reliability of each input modality based on completeness, clarity, source status, and conflict level. Semantic relevance and reliability are used to determine the fusion weights. Different fusion weights are applied to the same input modality across different clothing attributes; the overall fusion weight of a single input modality is not used to replace the attribute-level fusion weights. When the overall silhouette in the reference image is clear but the color is affected by illumination, the reliability of the reference image on the overall silhouette attribute is determined based on the overall silhouette recognition result, and the reliability on the color attribute is determined based on the color quality evaluation result.

[0092] The children's clothing profile construction module includes an age stage determination unit, a body size range calculation unit, a growth margin calculation unit, a movement margin calculation unit, a safety constraint generation unit, a comfort constraint generation unit, a production process constraint generation unit, and a children's clothing constraint relationship diagram construction unit.

[0093] The age group determination unit determines a child's age group based on the measured information, including age, height, and weight. If the child's age is within the corresponding data validity period, that age is used as the basis for age group determination. If the child's age is unavailable or exceeds the data validity period, the age group is determined based on statistical data from the height, weight, and child body type database, and the result is marked as inferred data. The validity period of the child's information is set by the system according to the data type, child's age group, and data collection specifications, and is recorded in the child information field configuration.

[0094] The body size interval calculation unit processes measured information and inferred data separately. Measured information includes measurement time, measurement method, and measurement tool markings. The body size interval calculation unit forms size intervals based on the allowable error of the measurement tool, repeated measurement results, and data acquisition time. Body parameters for which measured information is not provided are generated from data records in the children's body size database that match age group, gender, height, and weight. The body size interval calculation unit first filters data records by age group and gender, then sorts them by height and weight differences, and forms a matching data set from the sorting results. When the amount of data in the matching data set is lower than the preset minimum statistical sample size, the body size interval calculation unit expands the height and weight matching ranges according to a preset order. The expanded data range, the amount of matching data, and the inference confidence are recorded separately. The minimum statistical sample size is determined based on the dispersion of the corresponding body parameters in the children's body size database and the stability of the inferred interval. The system performs repeated sampling on historical samples. When increasing the sample size no longer causes the inferred interval to change beyond a preset stable value, the corresponding sample size is used as the minimum statistical sample size.

[0095] The growth allowance calculation unit determines the growth allowance based on the expected wearing period, garment category, garment coverage area, child's age group, and fabric elasticity. The growth allowance and wearing ease are stored separately. Wearing ease is used to meet the fit requirements under current wearing conditions, while the growth allowance is used to accommodate changes in body size during the expected wearing period. The growth allowance calculation unit reads growth allowance rules from the pattern processing library that match the child's age group, garment category, garment coverage area, and fabric elasticity. Growth allowance rules include applicable areas, baseline allowance, wearing period adjustment method, fabric elasticity adjustment method, and allowable upper limit.

[0096] The motion margin calculation unit determines the motion type based on the wearing scenario. Everyday activities include arm raising, bending, sitting, and walking, while sports activities include running, jumping, and squatting. The unit generates motion space requirements based on the shoulder, armpit, chest, waist, hip, crotch, and knee areas, and associates these requirements with the corresponding clothing components. The unit also reads the displacement range of key body points and the range of distance changes on the body surface for the corresponding motion from the motion sample database. After unit conversion, the displacement range and distance change range, along with the size range, together form the motion margin.

[0097] The safety constraint generation unit reads the currently valid safety rules corresponding to the child's age group, clothing category, wearing scenario, and sales area. Safety rules include restrictions on cord and strap structures, detachable small parts, sharp parts, decorative part fixing requirements, neck structure restrictions, skin contact surface requirements, and flame retardant requirements. Structures explicitly prohibited by safety rules constitute hard safety constraints. When hard safety constraints conflict, the conflicting design conditions are deleted.

[0098] The comfort constraint generation unit establishes comfort conditions based on fabric thickness, weight, elasticity, surface roughness, breathability, moisture absorption, and skin contact location. Contact surface attributes are set for garment components that directly contact the neck, armpits, waist, and crotch. When fabric attributes do not meet the requirements of the corresponding contact area, processing conditions are generated to add lining, add binding, adjust seam positions, or replace the fabric.

[0099] The production process constraint generation unit establishes production process constraints based on production equipment, fabric type, garment structure, target production batch size, and allowed processes. Production process constraints include seam type, seam allowance, binding method, overlocking method, pleating method, printing method, embroidery method, bonding method, fastening method, and component assembly sequence. Production processes that cannot be completed by the current production equipment are marked as prohibited production processes. Production process constraints are processed according to the handling level of strong fit constraints. Strong fit constraints include size range, growth allowance, movement allowance, garment component connection relationships, and production process feasibility conditions directly related to the above conditions.

[0100] The children's clothing constraint relationship graph construction unit establishes body regions, clothing components, fabric attributes, safety hard constraints, fit strong constraints, and aesthetic soft constraints as nodes or node attributes, respectively. Body regions include the neck, shoulders, chest, waist, hips, upper arms, elbows, crotch, knees, and calves. Clothing components include the front garment piece, back garment piece, neckline, sleeve pieces, cuffs, waistline, placket, pockets, hem, and decorative parts. The children's clothing constraint relationship graph construction unit establishes size dependencies, spatial coverage relationships, structural connection relationships, prohibition relationships, substitution relationships, and attribute association relationships. Each relationship record includes the starting node, ending node, relationship type, constraint level, data source, applicable conditions, current state, and rule version. The children's clothing constraint relationship graph and its node attributes are converted into constraint profiles. Constraint profiles include continuous vector data and structured fields. The structured fields include the child's age group, size range, measured information, inferred data, growth margin, movement margin, hard safety constraints, strong fit constraints, soft aesthetic constraints, permitted fabrics, prohibited fabrics, permitted manufacturing processes, prohibited manufacturing processes, and items requiring manual confirmation. Soft aesthetic constraints include color preferences, pattern preferences, style preferences, overall silhouette preferences, and preferences for decorative components.

[0101] The knowledge-constrained reasoning module includes a query generation unit, a structured retrieval unit, a vector retrieval unit, a constraint conflict handling unit, a requirement completion unit, an instruction graph generation unit, and an evidence recording unit.

[0102] The query generation unit extracts query fields from the constraint profile. These fields include child age group, size range, clothing category, applicable season, wearing scenario, overall silhouette, color preference, pattern theme, fabric attributes, safety hard constraints, and production equipment capabilities. The query generation unit then converts these query fields into structured filtered data and vector query data.

[0103] The structured search unit filters the children's clothing design knowledge base according to children's age group, clothing category, season, fabric type, safety rule version, and production equipment capabilities. The filtering process first excludes design knowledge that is not applicable to the current children's age group, and then excludes design knowledge that is not applicable to the current clothing category, fabric, and production equipment.

[0104] The vector retrieval unit inputs the constraint profile into the knowledge vector encoding network and performs vector retrieval from the filtered design knowledge. The design knowledge obtained from the vector retrieval serves as candidate knowledge. Candidate knowledge includes children's clothing style knowledge, color and pattern knowledge, fabric knowledge, children's body shape knowledge, safety rule knowledge, pattern making and manufacturing process knowledge, and historical design case knowledge. Each candidate knowledge includes a knowledge entry identifier, knowledge type, scope of application, knowledge content, data source, and version information.

[0105] The constraint conflict handling unit checks for conflicts between explicit requirements, image attributes, constraint profiles, and candidate knowledge. Explicit requirements refer to design requirements directly specified by the user and possessing explicit attribute values ​​in text descriptions or requirement data. Image attributes refer to attributes identified from reference images, such as garment category, overall silhouette, neckline, sleeve type, waistline, placket, pockets, hem, color, pattern, fabric, or decorative components. When a safety hard constraint conflicts with an explicit requirement or image attribute, the conflicting explicit requirement or image attribute is deleted, and the conflict rule and alternative attributes are recorded. When a fit-fitting strong constraint conflicts with an image attribute, the fit-fitting strong constraint is retained, and an alternative structure that retains the original color, pattern, or style attributes is selected from candidate knowledge. When a production process constraint conflicts with a candidate garment structure, alternative production processes are first searched. If no alternative production process that meets the current production equipment capabilities exists, alternative garment structures or alternative fabrics are searched.

[0106] The requirement completion unit completes and generates required component attributes that have not been input by the user. Component attributes include neckline opening method, cuff structure, placket structure, waist connection method, lining range, seam allowance, pattern position, and decoration density. The requirement completion unit reads candidate knowledge from high to low matching degree and performs conflict checks on the candidate knowledge. If candidate knowledge conflicts with safety hard constraints, fit strong constraints, or already defined explicit requirements, the candidate knowledge is not used to complete the component attributes. The completed component attributes are written into the instruction diagram. The instruction diagram includes clothing category nodes, overall silhouette nodes, front garment piece nodes, back garment piece nodes, neckline nodes, sleeve type nodes, waist node, hem node, placket node, pocket node, color node, pattern node, fabric node, decoration node, and production process node. Each node in the instruction diagram records attribute values, spatial positions, size ranges, connection relationships, prohibition conditions, and knowledge sources. Generation conditions include the attribute values, spatial positions, connection relationships, size ranges, prohibition conditions, and knowledge sources of each node in the instruction diagram.

[0107] The evidence recording unit records the corresponding explicit requirements, reference image areas, measured information, safety rules, or production process knowledge for each key component attribute in the instruction diagram. Candidate knowledge is sorted by matching degree to form selected knowledge. Candidate knowledge corresponding to hard safety constraints is directly used as selected knowledge, without being limited by the number of ordinary candidate knowledge. The selected knowledge and instruction diagram are converted into a condition sequence. The condition sequence includes positive design conditions, negative prohibition conditions, clothing component relationship conditions, spatial location conditions, and knowledge evidence conditions.

[0108] The children's clothing design generation module includes a structural control diagram generation unit, a conditional diffusion generation unit, a local area refinement unit, a front and back consistency verification unit, a safe fit verification unit, a process reverse analysis unit, and a candidate scheme generation unit. The control diagrams in this embodiment include a front profile diagram, a back profile diagram, a garment component zoning diagram, a seam position diagram, a pleat distribution diagram, and a pattern distribution diagram.

[0109] The structural control chart generation unit generates control charts based on generation conditions. The front and back designs share structural parameters corresponding to shoulder line position, side seam position, waist height, hem length, and sleeve length. The structural control chart generation unit generates a prohibited area map based on safety constraints. The prohibited area map records the locations where drawstrings, small parts, rigid decorative parts, and sharp parts are not allowed, and uses the same coordinate system as the control chart. The structural control chart generation unit performs connectivity checks on the control charts. If there are no connection areas between the sleeve piece and the garment body, between the collar piece and the garment body, or between the placket and the front garment piece, the control chart will not proceed to the next generation process.

[0110] In this embodiment, the diffusion network is a conditional diffusion generation network. The diffusion network receives a control chart, conditional sequence, constraint profile, and forbidden region map. It generates initial random latent data and performs inverse denoising based on constraint corrections. During inverse denoising, the basic denoising network calculates the denoising direction of the latent data, the fit evaluation network calculates fit deviation, the safety evaluation network calculates safety deviation, and the production process evaluation network calculates process deviation. The correction directions corresponding to fit deviation, safety deviation, and process deviation are combined with the basic denoising direction to form the adjusted denoising direction. The diffusion network processes the overall silhouette, front and back proportions, and garment component positions in the early iteration stage; it processes fabric texture, color distribution, pattern distribution, and pleats in the middle iteration stage; and it processes seams, edging, fastening components, and local components in the later iteration stage. An intermediate design is output after inverse denoising is completed.

[0111] The local area refinement unit performs area refinement on the middle design. Area refinement includes neckline area refinement, cuff area refinement, placket area refinement, waistline area refinement, pocket area refinement, hem area refinement, and decorative component area refinement.

[0112] The front and back consistency verification unit performs consistency verification on the middle design. The consistency verification includes shoulder line height verification, side seam length verification, waist height verification, hem length verification, fabric texture direction verification, pattern continuity position verification, and component quantity verification.

[0113] The safety fit verification unit performs safety checks on the intermediate design. Safety checks include collar size verification, cord and strap position verification, small component position verification, sharp component position verification, decorative component fixing area verification, and skin contact surface verification.

[0114] The process reverse engineering unit performs process analysis on the intermediate design that has passed security verification. Process analysis includes identifying the front and back garment pieces, sleeve pieces, collar pieces, waist components, placket components, pocket components, hem components, and decorative components. Based on component boundaries, connection relationships, seam positions, pleat density, and fabric properties, it generates a component list, pattern size range, sewing sequence, seam allowance data, binding method, and equipment requirements. The intermediate design, after completing area refinement, consistency verification, security verification, and process analysis, forms the candidate design.

[0115] The candidate design generation unit generates multiple candidate designs by modifying the initial random latent data, color combinations, pattern layouts, and combinations of local components. The unit performs duplication checks on the candidate designs using structural and visual features. When two candidate designs simultaneously satisfy both structural and visual duplication conditions, the candidate design with the higher evaluation result is retained.

[0116] The feedback output service module includes a candidate evaluation unit, an output arrangement unit, an interactive modification unit, a feedback classification unit, a user profile update unit, a model parameter update unit, and a service interface unit.

[0117] The candidate evaluation unit calculates a consistency score, fit score, safety score, coordination score, and process score for each candidate design. The consistency score indicates the degree of consistency between the candidate design and the text description, reference images, and requirement data. The fit score indicates the degree of consistency between the candidate design and the size range, growth margin, and motion margin. The safety score indicates the degree to which the candidate design meets safety hard constraints and limitations. The coordination score indicates the degree of coordination between color combinations, pattern distribution, garment component proportions, and front and back structures. The process score indicates the feasibility of manufacturing the candidate design under current fabric and production equipment conditions. Evaluation rules include safety hard constraint handling rules, minimum adoption conditions for each score, comprehensive evaluation calculation rules, candidate design ranking rules, and candidate design duplication exclusion rules. The minimum adoption conditions for each score are determined by a labeled candidate design verification set. Candidate designs that meet the corresponding design requirements are marked as qualified samples, and those that do not are marked as unqualified samples. The minimum adoption conditions for meeting the error acceptance rate requirement are determined based on the evaluation results. If a candidate design violates any safety hard constraint, it will not be included in the results package arrangement process.

[0118] The output arrangement unit selects candidate designs that meet the evaluation rules based on the sorting results and arranges renderings, detail images, color data, pattern design, and process suggestions into a result package. Renderings include front and back design renderings. Detail images include neckline or placket details, cuff or sleeve details, and waist or hem details. When adding pockets, fastening components, or decorative components, corresponding garment component detail images are also generated. Color data includes the color names, color swatch numbers, hue values, brightness values, and saturation values ​​of the primary, secondary, and accent colors. Pattern design includes a garment component list, body size baseline, garment size range, wearing ease, growth allowance, movement allowance, seam allowance, pleat allowance, dart placement, and opening / closing placement. Process suggestions include the number of cut pieces, cutting direction, garment component assembly sequence, seam type, seam allowance, binding method, overlock method, fastening method, printed area, embroidered area, equipment requirements, and quality inspection items.

[0119] The interactive modification unit receives the modified content. The modified content includes text modification data, attribute selection data, and local area selection data. The interactive modification unit locates the modified content to the corresponding modification node. The modification node is the node in the instruction diagram corresponding to the modified content. Color modifications are located to the color node, pattern modifications to the pattern node, sleeve type modifications to the sleeve type node, fabric modifications to the fabric node, and wearing / undressing method modifications to the neckline and placket nodes. When a modification involves multiple nodes, both the primary modification node and associated modification nodes are determined simultaneously. The primary modification node is used to determine the affected area, and associated modification nodes are used to determine the garment components and constraints that need to be re-validated. The interactive modification unit locks the structure and data of unaffected nodes and only regenerates the affected area.

[0120] The feedback classification unit categorizes modified content into current task data and long-term preference data. When the user explicitly chooses to save the corresponding preference, the modified content is marked as long-term preference data. When the user does not explicitly choose to save, the system determines whether to form long-term preference data based on the repeated records of the same attribute preference in different independent design tasks and the reliability of the feedback. Modifications that do not meet the conditions for updating long-term preference data are saved as current task data. After regenerating the affected area, the system combines the original design task identifier, modified content, modified node, affected area, locked node, re-executed steps, regenerated results, and verification results to form new design data.

[0121] The model parameter update unit updates the candidate ranking models and lightweight adaptation parameters using candidate designs and modifications that have been manually reviewed, without overwriting the basic parameters of large multimodal models. Each model parameter update generates a model version number, training data summary, applicable user scope, test results, and rollback address.

[0122] See Figure 2The system is deployed on cloud servers, enterprise internal servers, or designer workstations. Mobile terminals, web terminals, e-commerce platforms, personalized customization platforms, designer workstations, and brand development systems send design data and receive result packages via application programming interfaces (APIs). The cloud server includes load balancing devices, model inference devices, knowledge retrieval devices, task storage devices, and result storage devices. The model inference devices are equipped with graphics processors (GPUs), and the number of GPUs and storage capacity are determined based on the conditional diffusion generation network size, target image size, and the number of concurrent design tasks. The task storage devices store design data, fusion vectors, conflict lists, constraint profiles, generation conditions, condition sequences, intermediate designs, candidate designs, result packages, and new design data according to task identifiers.

[0123] Children's information and user identity data are stored separately. During model inference, the necessary children's information for children's clothing design is retrieved through task identifiers. The training data for the multimodal large model includes authorized children's clothing renderings, front and back design images, detail images, clothing descriptions, clothing component labels, fabric attributes, pattern data, production process records, children's size ranges, safety rule labels, and user feedback data. Personally identified data irrelevant to the children's clothing design task is removed before training.

[0124] Text encoding network, visual encoding network, fabric attribute encoding network, child feature encoding network, and user preference encoding network each generate corresponding feature vectors. Through cross-modal alignment training, different input modalities of the same design task establish a correspondence in a unified feature space. The vector dimension of the unified feature space is determined based on the input interface of the multimodal large model.

[0125] The conditional diffusion generative network is trained using control charts, conditional sequences, constraint profiles, and a target children's clothing design. During training, denoising error, structural error, garment component error, front-back consistency error, safety constraint error, and manufacturing process error are calculated between the generated results and the target children's clothing design. The training termination condition is determined based on an independent validation dataset. Training stops and the current model parameters are saved when the generation error, structural consistency results, safety rule pass results, and manufacturing process feasibility results no longer meet the minimum improvement requirements in the model version configuration over multiple consecutive validation cycles.

[0126] See Figure 3 The personalized intelligent design generation method for children's clothing provided by the present invention includes steps S100, S200, S300, S400 and S500.

[0127] Step S100: Standardize the design data and merge the design data to output a fusion vector and a conflict list. This is performed through the multimodal requirements analysis module.

[0128] Step S200: Construct a constraint profile based on the fusion vector and conflict list combined with measured information. This is executed through the children's clothing profile construction module.

[0129] Step S300 involves retrieving design knowledge based on the constraint profile and handling conflicts to generate conditions and condition sequences. This is executed through the knowledge constraint reasoning module.

[0130] Step S400: Generate and verify candidate designs based on the generation conditions and condition sequence, and output the results. This is performed through the children's clothing design generation module.

[0131] Step S500 involves evaluating the candidate designs, outputting the result package, processing the modifications, and converting them into new design data. This is executed through the feedback output service module.

[0132] See Figure 4 Step S100 includes sub-steps S101, S102 and S103.

[0133] Sub-step S101: Receive text description, reference image, fabric data, children's information, and demand data, and perform standardization processing.

[0134] Receive the text description, reference images, fabric data, child information, and requirement data of the current design task, and use them as input modalities for the design data. Generate a unique task identifier for the current design task, and record the source, receiving time, file type, data version, and authorization status of each input modality.

[0135] The system performs character encoding standardization, punctuation normalization, clothing terminology segmentation, clothing entity recognition, attribute recognition, and negative condition recognition on the text description. Recognized clothing entities include clothing category, overall silhouette, neckline, sleeve type, waistline, placket, pockets, hem, and decorative parts. Recognized attributes include color, pattern, fabric, length, looseness, wearing / removing method, and applicable scenarios. Attribute values ​​corresponding to negative conditions are written to the prohibition condition field, attribute values ​​corresponding to restriction conditions are written to the restriction condition field, and content with explicitly specified user priority is written to the user priority field.

[0136] The reference image undergoes file integrity checks, size reading, orientation correction, color space conversion, and image scaling. The model input size of the reference image is determined based on the input interface of the visual encoding network. The image scaling operation does not alter the original reference image; the scaled reference image is saved separately as a model input file.

[0137] The system performs unit conversion, field integrity checks, and data correlation checks on children's information. Height is uniformly converted to cm, weight to kg, and chest circumference, waist circumference, hip circumference, shoulder width, arm length, back length, torso length, and leg length to mm. Anomalies in children's information are determined based on the statistical range of the corresponding age group and gender in the children's body shape database. Information outside the statistical range is not directly deleted but marked as data awaiting confirmation.

[0138] Perform field standardization on fabric data. Fiber composition is recorded using standardized material names, and fabric weight is uniformly converted to g / m². 2 Fabric thickness is uniformly converted to mm, and air permeability is uniformly converted to mm / s. Fabric images and fabric text attributes are saved separately, and the fabric image recognition results do not overwrite the fabric text attributes entered by the user.

[0139] Sub-step S102: Perform a quality evaluation on the design data and establish a list of missing records and conflicts.

[0140] The system calculates field completeness, clothing terminology coverage, negation condition clarity, and attribute conflict count for the text description. When clothing category, applicable scenario, or other necessary fields are missing, a missing record is created. This record includes the missing field name, corresponding input modality, supplementary requirements, and processing status. If the text description contains mutually exclusive sleeve types, necklines, garment lengths, or opening / closing methods, the corresponding content is added to the conflict list.

[0141] The reference image is used for garment segmentation, garment component recognition, and image quality evaluation. Image quality evaluation is performed on overall silhouette recognition, color recognition, pattern recognition, fabric recognition, and local garment component recognition. If the reference image does not meet the criteria for overall silhouette recognition, it is not used to extract the overall silhouette. If the reference image does not meet the criteria for local garment component recognition, it is not used to extract seams and local garment components, but image attributes that meet the criteria for color or overall style recognition are retained.

[0142] The system checks children's information for measurement time, measurement method, field completeness, and numerical correlations. When the correlation between a child's height, weight, and body circumference exceeds the statistical range of the children's body shape database, the relevant children's information is marked as data to be confirmed. Actual measurement information and inferred data are recorded separately. Actual measurement information records the measurement time, measurement tool, and measurement method; inferred data records the query conditions, number of matching samples, inference interval, and inference confidence level.

[0143] Calculate illumination uniformity, texture clarity, and subject coverage for the fabric image. If the fabric image does not meet the color recognition criteria, do not extract color attributes from it. If the fabric image does not meet the fabric structure recognition criteria, do not determine the fabric structure from it. If the fabric name and the fabric image recognition result are inconsistent, add the fabric name, recognition result, data source, and reliability to the conflict list.

[0144] Sub-step S103: Perform unified encoding on each input mode and generate a fusion vector based on the fusion weights.

[0145] The text description is input into the text encoding network, the reference image into the visual encoding network, the fabric data into the fabric attribute encoding network, the children's information into the children's feature encoding network, and the demand data into the user preference encoding network. Each encoding network outputs a feature vector of the same dimension.

[0146] A cross-modal attention network calculates the semantic relevance of each input modality to the design task. An input quality evaluation unit calculates reliability based on the completeness, clarity, source state, and conflict level of each input modality. The product of semantic relevance and reliability is used to determine the fusion weights. At least one input modality has a fusion weight greater than 0; when all input modalities have fusion weights of 0, the fusion process stops, and a data supplementation request is output.

[0147] The multimodal demand fusion formula is as follows: ; in, As a fusion vector, it represents the comprehensive representation of text description, reference image, fabric data, children's information and demand data in a unified feature space. The value of each component ranges from -1 to 1, and the vector dimension is determined by the input interface of the multimodal large model. The input modality index represents the sequence number of the input modality participating in the fusion process; The number of input modes represents the total number of input modes actually participating in the current fusion process, and its value ranges from 1 to 5. For the first The semantic relevance of the input modality represents the semantic relevance of the first input modality. The degree of correlation between each input modality and the current design task, with a value ranging from 0 to 1; For the first The reliability of the input mode, which represents the reliability of the first input mode. The completeness, clarity, source status, and conflict status of each input modality range from 0 to 1. For the first The feature vector of the input mode represents the th input mode. The input modalities are represented in a unified feature space after being processed by the corresponding encoding network. The values ​​of each component range from -1 to 1, and the vector dimension is consistent with the fusion vector.

[0148] The trusted fusion unit performs attribute-level fusion based on the attribute reliability of the same input modality across different clothing attributes. After completing the fusion process, it outputs a fusion vector, an attribute reliability table, missing records, a conflict list, and a list awaiting user confirmation, and writes these into the current design task data package.

[0149] See Figure 5 Step S200 includes sub-steps S201, S202 and S203.

[0150] Sub-step S201: Determine the growth margin and movement margin within the size range based on the child's age, height, weight, and other measured information.

[0151] The system retrieves data from a children's body type database based on measured data of age, height, and weight. Data records are first filtered by age group and gender, then sorted by height and weight differences. Matching data records are selected from the sorted results according to the minimum number of samples configured in the children's body type database. If the number of matching data records is insufficient to meet the minimum number of samples, the height and weight matching ranges are expanded according to the database configuration, and the expanded query range, the number of matching samples, and the inference confidence level are recorded.

[0152] For body parameters with available measured values, size ranges are formed using the measured values ​​and measurement errors. Measurement errors are determined based on the calibration data of the measuring instrument, differences in repeated measurements, and the measurement method. For body parameters without available measured values, size ranges are formed based on the distribution of matched data records. Size ranges include a lower limit, upper limit, center value, data source, and confidence level.

[0153] Growth margins are determined based on the expected wearing period, child's age, clothing type, clothing coverage area, and fabric elasticity. Movement types are determined based on the wearing scenario, and movement margins are formed using the displacement of key body points and changes in body surface distance from the movement sample database. Daily activity scenarios at least handle arm raising, sitting, and bending movements, while sports activity scenarios also handle running, jumping, and squatting movements. Movement margins are applied to the corresponding clothing components at the shoulder, armpit, chest, waist, hip, crotch, and knee.

[0154] Sub-step S202 establishes safety hard constraints, combination strong constraints, and aesthetic soft constraints based on the fusion vector and conflict list.

[0155] The system reads safety rules from the safety rule base that are consistent with the child's age group, clothing category, wearing scenario, and sales area. Clothing structures explicitly prohibited by the safety rules constitute hard safety constraints. These hard safety constraints include restrictions on drawstring structures, detachable small parts, sharp parts, decorative part securing requirements, neck structure restrictions, skin contact surface requirements, and flame retardant requirements. The constraint values ​​are provided by the currently effective version of the safety rule base. During verification, the system reads the rule identifier, restricted object, constraint parameters, constraint values, and detection method.

[0156] Strong fit constraints are formed based on size range, growth allowance, movement allowance, fabric properties, and the connection relationships between garment components. Production process constraints are implemented according to the processing level of the strong fit constraints. Production process constraints include production equipment capacity, seam type, seam allowance, binding method, overlocking method, pleating method, printing method, embroidery method, fastening method, and the assembly sequence of garment components. Soft aesthetic constraints are formed based on color, pattern, style, overall silhouette, and decorative component preferences within the fusion vector.

[0157] Constraint conflicts are handled according to the conflict list. When a safety hard constraint conflicts with other conditions, the conflicting condition is deleted. When a combination strong constraint conflicts with an aesthetic soft constraint, the combination strong constraint is retained and an alternative aesthetic structure is retrieved. When aesthetic soft constraints conflict with each other, the processing order is determined according to the user's explicitness and the reliability of the attributes.

[0158] Sub-step S203 establishes relationships between the fabric properties of clothing components in the body area and constraints at all levels, and converts them into constraint profiles.

[0159] Establish nodes for body regions, garment components, fabric attributes, hard safety constraints, strong fit constraints, and soft aesthetic constraints. Establish size dependencies, spatial coverage relationships, structural connection relationships, prohibition relationships, substitution relationships, and attribute association relationships. Establish spatial coverage relationships between neckline nodes and neck nodes; size dependencies between sleeve nodes and shoulder and upper arm nodes; attribute associations between fabric attribute nodes and pleat nodes; and prohibition relationships between prohibited areas and decorative component nodes. Map the fused vectors to candidate constraints and project these constraints onto a feasible set of constraints that satisfy both hard safety constraints and basic structural connection conditions.

[0160] The formula for calculating constrained profiles is as follows: ; in, The constrained image vector represents the design conditions after corrections for size range, growth margin, motion margin, safety hard constraints, combined strong constraints, and production process conditions. The value range of each component is -1 to 1, and the vector dimension is consistent with the fused vector. The candidate constraint vector represents the temporary design conditions to be evaluated during the projection process. The values ​​of each component range from -1 to 1, and the vector dimension is consistent with that of the constraint profile vector. The set of feasible constraints represents the set of candidate constraint vectors that satisfy both the hard safety constraints and the infrastructure connection conditions. This is the demand mapping matrix, which represents the model parameters that transform the fused vector into constraint features; The constraint bias vector represents the basic constraints determined by the child's age stage and clothing category. Each component takes values ​​from -1 to 1, and the vector dimension is the same as the constraint profile vector. The weights for growth and motion margin represent the degree to which the deviation between growth margin and motion margin affects the projection result, and the value ranges from 0 to 1. The safety and comfort weights represent the degree of influence of deviations in safety and skin contact requirements on the projection results, with a value ranging from 0 to 1. The production process weight represents the degree of influence of production process deviation on the projection result, and its value ranges from 0 to 1. The growth and motion margin deviation function represents the degree to which the candidate constraint vector deviates from the growth margin and motion margin requirements, and its value ranges from 0 to 1. The safety and comfort deviation function represents the degree to which a candidate constraint vector violates safety or skin contact requirements, and its value ranges from 0 to 1. This is the production process deviation function, which represents the degree to which the candidate constraint vector exceeds the current production equipment and fabric processing capabilities, and its value ranges from 0 to 1. This is a 2-norm operation, which represents the arithmetic square root of the sum of the squares of the vector's components; This is the minimum position operation, representing the candidate constraint vector that achieves the minimum of the objective function. The sum of the growth and action margin weights, safety and comfort weights, and production process weights is 1. Safety hard constraints are directly limited by the feasible constraint set and are not canceled by reducing their weights.

[0161] The constraint profile simultaneously stores the child's age stage, size range, measured information, inferred data, growth margin, movement margin, safety hard constraints, fit strong constraints, aesthetic soft constraints, permitted fabrics, prohibited fabrics, permitted production processes, prohibited production processes, and items pending manual confirmation.

[0162] See Figure 6 Step S300 includes sub-steps S301, S302 and S303.

[0163] Sub-step S301 involves extracting query fields from the constraint profile, performing filtering and vector retrieval to obtain candidate knowledge.

[0164] Query fields are extracted from the constraint profile. These fields include child's age group, size range, clothing category, season, occasion, overall silhouette, color preference, pattern theme, fabric properties, safety constraints, and production equipment capabilities. The children's clothing design knowledge base is filtered based on child's age group, clothing category, season, fabric type, safety rule version, and production equipment capabilities. The constraint profile is input into a knowledge vector encoding network to form vector query data, and vector retrieval is performed on the filtered design knowledge. The design knowledge obtained from the vector retrieval is used as candidate knowledge. All safety constraints in the safety rule base that match the current design task are used as candidate knowledge, without being limited by the number of ordinary candidate knowledge.

[0165] Sub-step S302: Check for conflicts between the image attribute constraints of the explicit requirements and the candidate knowledge, and form generation conditions.

[0166] Each garment category, overall silhouette, neckline, sleeve type, waistline, placket, pockets, hem, color, pattern, fabric, and decorative components are examined for conflicts between explicit requirements, image attributes, constraint profiles, and candidate knowledge. When a hard safety constraint conflicts with an explicit requirement or image attribute, the corresponding explicit requirement or image attribute is deleted, and the deleted attribute, conflict rule, and alternative attribute are recorded. When a strong fit constraint conflicts with an image attribute, the strong fit constraint is retained, and an alternative structure that retains the original color, pattern, or style attributes is selected from candidate knowledge. When production process conditions conflict with candidate garment structures, alternative production processes are first searched. If no alternative production process that meets the current production equipment capabilities exists, alternative garment structures or alternative fabrics are searched.

[0167] Complete the necessary component attributes that were not entered by the user. Completed component attributes include neckline opening / closing method, cuff structure, placket structure, waistband connection method, lining area, seam allowance, pattern position, and decoration density. Write the completed component attributes into the instruction drawing to form the generation conditions.

[0168] Sub-step S303: Calculate the matching degree of candidate knowledge, select the selected knowledge, and generate a condition sequence.

[0169] Semantic similarity, category matching, comfort matching, and conflict degree are calculated for the candidate knowledge and constraint profiles, respectively. Semantic similarity is calculated based on the directional consistency between the constraint profile vector and the candidate knowledge encoding vector. Category matching is determined based on the applicable clothing categories, applicable age groups, and clothing categories of the current design task in the candidate knowledge. Comfort matching is determined based on fabric properties, clothing structure, skin contact points in the candidate knowledge, and comfort requirements in the constraint profile. Conflict degree is determined based on conflict records between the candidate knowledge and hard safety constraints, strong fit constraints, explicit requirements, and determined component attributes.

[0170] The formula for calculating the candidate knowledge matching degree is as follows: ; in, For the first The matching degree of the candidate knowledge, which represents the matching degree of the first candidate knowledge. The applicability of each candidate knowledge to the current design task, with a value ranging from -1 to 1; This is the candidate knowledge index, which represents the sequence number of the candidate knowledge to be included in the matching calculation; This is the semantic similarity weight, which represents the degree to which semantic similarity affects the matching degree, and its value ranges from 0 to 1; This is the category matching weight, which represents the degree to which the category matching score affects the overall matching score. Its value ranges from 0 to 1. The comfort fit weight represents the degree to which the comfort fit affects the overall fit, and its value ranges from 0 to 1. The conflict weight represents the degree to which the conflict reduces the matching score, and its value ranges from 0 to 1. For the first The knowledge encoding vector of the nth candidate knowledge represents the nth... The semantic and attribute content of each candidate knowledge, with each component taking values ​​ranging from -1 to 1, and the vector dimension being consistent with the constraint profile vector; For the first The category matching degree of the candidate knowledge, which represents the category matching degree of the first candidate knowledge. The degree of consistency between the applicable clothing categories and applicable age groups of each candidate knowledge and the current design task, with a value range of 0 to 1; For the first The comfort fit of the candidate knowledge, which represents the comfort fit of the first candidate knowledge. The degree to which the fabric properties and garment structure of each candidate knowledge meet the current comfort requirements, with a value range of 0 to 1; For the first The degree of conflict of the candidate knowledge, which represents the degree of conflict of the first candidate knowledge. The degree to which candidate knowledge conflicts with hard safety constraints, strong constraints of the whole, explicit requirements, or determined component attributes, with a value range of 0 to 1; This is a vector transpose operation. The sum of the semantic similarity weight, category matching weight, comfort matching weight, and conflict weight is 1.

[0171] Candidate knowledge is selected from high to low matching degree, and conflicting candidate knowledge is deleted to obtain the selected knowledge. The selected knowledge and instruction graph are converted into a condition sequence and transmitted to step S400.

[0172] See Figure 7 Step S400 includes sub-steps S401, S402 and S403.

[0173] Sub-step S401: Generate a control chart and a prohibited area map based on the generation conditions.

[0174] Based on the generation conditions, generate the front silhouette boundary, back silhouette boundary, neckline boundary, armhole boundary, sleeve boundary, waist position, placket position, pocket position, hem boundary, seam position, and pleat distribution. Write these structural data into the front silhouette diagram, back silhouette diagram, garment component zoning diagram, seam position diagram, pleat distribution diagram, and pattern distribution diagram to form control charts. Establish shared structural parameters for the front and back designs. These shared structural parameters include shoulder line height, side seam length, waist height, hem length, and sleeve length.

[0175] A prohibited area map is generated based on hard safety constraints. This map records the locations where drawstrings, small parts, rigid decorative parts, and sharp parts are not permitted. The prohibited area map and the control chart use the same coordinate system. A connectivity check is performed on the control chart. If the design involves interconnected garment parts with no boundary contact areas, the control chart does not undergo the reverse denoising process.

[0176] Sub-step S402 involves inputting the control chart condition sequence constraint profile and the forbidden region map into the diffusion network to generate an intermediate design.

[0177] The control chart, condition sequence, constraint profile, and forbidden region map are input into the diffusion network. The diffusion network generates initial random latent data and performs inverse denoising. In each predetermined denoising iteration, the denoising direction, conformity bias, safety bias, and process bias of the latent data are calculated.

[0178] Fit deviation is calculated based on the correspondence between garment component boundaries and size ranges, growth allowances, and movement allowances in the intermediate design. Safety deviation is calculated based on violation records between garment components, decorative components, prohibited areas, and hard safety constraints in the intermediate design. Manufacturing deviation is calculated based on the number of garment components, their connection relationships, seam positions, pleat distribution, fabric properties, and production equipment capabilities in the intermediate design. Noise reduction is adjusted based on fit deviation, safety deviation, and manufacturing deviation.

[0179] The constraint-corrected inverse denoising formula is as follows: ; in, For the first The latent data of the next denoising iteration represents the intermediate state of the current children's clothing design in the normalized latent space, and the value range of each component is -1 to 1; For the first The potential data of the next denoising iteration represents the intermediate state after the current denoising iteration is completed, and the value of each component is in the range of -1 to 1. For denoising iterative indexes, it represents the reverse generation order from the initial random latent data to the intermediate design, and takes a positive integer value; For the first The update step size of each denoising iteration represents the degree of influence of the adjusted denoising direction on the potential data, and its value ranges from 0 to 1. The output of the basic denoising network represents the basic denoising direction under the action of the conditional sequence. It is the set of parameters for the basic denoising network, which represents the parameters of the neural network trained with the children's clothing design training data; This is a conditional sequence, representing a combination of positive design conditions, negative prohibition conditions, clothing component relationship conditions, spatial location conditions, and knowledge evidence conditions. The combination correction weight represents the degree to which the combination deviation adjusts the denoising direction, and its value ranges from 0 to 1. The safety correction weight represents the degree to which the safety deviation adjusts the denoising direction, and its value ranges from 0 to 1. The process correction weight represents the degree to which process deviations adjust the direction of noise reduction, and its value ranges from 0 to 1. The fit deviation represents the degree of deviation between the current design and the size range, growth allowance, and motion allowance, and its value ranges from 0 to 1. For safety deviation, it represents the degree to which the current design violates hard safety constraints, and its value ranges from 0 to 1; This is the process deviation, which represents the degree to which the current design exceeds the fabric processing capacity and production equipment capacity, and its value ranges from 0 to 1. For the first The intensity of random disturbance in the next denoising iteration represents the control value of the candidate design difference, and its value ranges from 0 to 1. For the first The standard normal random vector of the current denoising iteration represents the randomly changed data added in the current denoising iteration; In order to target the The gradient calculation of the latent data in the next denoising iteration represents the direction of the corresponding bias as the latent data changes; This is a numerical trimming operation, which represents restricting the components of the potential data between a specified lower limit and a specified upper limit. The safety correction weight is no less than the fitability correction weight and the process correction weight.

[0180] The process involves three iterations: the initial stage processes the silhouette and position of garment components; the middle stage processes fabric texture and color distribution; and the final stage processes seams, decorative elements, and other local components. After each denoising iteration, the prohibited area map is compared with the design area corresponding to the current potential data. If a restricted garment component appears within the prohibited area, the corresponding restricted garment component is added as a negative prohibition condition. The intermediate design is output after reverse denoising.

[0181] Sub-step S403 involves refining the intermediate design execution area for consistency verification, security verification, and process analysis to form candidate designs.

[0182] Local areas are established for each garment part, including the neckline, cuffs, placket, waistline, pockets, hem, and decorative components, with further refinement applied to the intermediate design areas. Each local area includes the target garment part and the boundary areas directly connected to it. When regenerating local areas, the garment structure, color, and fabric data outside these areas are locked. A consistency check is performed on the intermediate design. This check includes verifying shoulder line height, side seam length, waistline height, hem length, fabric texture direction, and pattern seam crossing position for both the front and back designs. Consistency criteria are determined using a set of labeled front and back paired samples.

[0183] Perform safety checks on intermediate designs. Inspect cords, small parts, sharp parts, decorative parts, collar structures, and skin contact surfaces. If a violation of a hard safety constraint is detected, delete the corresponding intermediate design and do not perform process analysis. Perform process analysis on intermediate designs that pass safety checks.

[0184] Identify the boundaries of the front and back garment pieces, sleeves, collar, waistband, placket, pockets, hem, and decorative elements. Generate a list of garment components based on boundary shape, seam position, symmetry, and connection relationships. Determine the pleat range based on pleat distribution and fabric properties; determine the seam type and feed method based on fabric thickness and weight; and determine the opening / closing method based on neckline structure and fabric elasticity. Pattern parameters that cannot be uniquely determined from the intermediate design are output as ranges and added to the list of items awaiting manual confirmation. After completing area refinement, consistency verification, safety verification, and process analysis, candidate designs are generated.

[0185] Step S500: Evaluate the candidate design, output the result package, process the modifications, and convert them into new design data. See Figure 8 Step S500 includes sub-steps S501, S502 and S503.

[0186] Sub-step S501: Calculate and sort the consistency score, integration score, safety score, coordination score, and process score of the candidate designs.

[0187] Calculate the consistency score, fit score, safety score, coordination score, and process score for each candidate design. The consistency score is determined based on the degree of consistency between the candidate design and the text description, reference images, and requirement data. The fit score is determined based on the degree of consistency between the candidate design and the size range, growth margin, and motion margin. The safety score is determined based on the degree to which the candidate design meets safety hard constraints and limitations. A candidate design is directly removed from the candidate design set if any safety hard constraint is violated.

[0188] The coordination score is determined based on color combination, pattern distribution, proportion of garment components, and consistency between the front and back. The craftsmanship score is determined based on the number of garment components, sewing sequence, fabric processing capabilities, and production equipment capabilities. The values ​​for consistency, fit, safety, coordination, and craftsmanship scores all range from 0 to 1.

[0189] The comprehensive evaluation formula for candidate designs is as follows: ; in, For the first The comprehensive evaluation result of the candidate design represents the first The degree to which each candidate design comprehensively meets the requirements of consistency, fit, safety, coordination and process feasibility, with a value ranging from -1 to 1; This is the candidate design index, which represents the sequence number of the candidate design participating in the evaluation; The set of evaluation items represents the set of evaluation items corresponding to consistency score, integration score, safety score, coordination score, and process score; The evaluation item index represents the serial number of the evaluation item in the set of evaluation items; For the first The evaluation weight of the evaluation item represents the evaluation weight of the first item. The degree of influence of each evaluation item on the comprehensive evaluation result ranges from 0 to 1, and the sum of all evaluation weights is 1. For the first The candidate design in the first The evaluation score on the evaluation item represents the score of the first evaluation item. The candidate design satisfies the first The degree required for each evaluation item, with a value range of 0 to 1; To evaluate the stability parameter of the denominator, it represents a preset positive number to avoid the denominator of the comprehensive evaluation calculation being 0, and its value is less than the minimum effective calculation precision of the unstable term in the evaluation denominator. The penalty weight for violating hard constraints represents the degree to which a violation of hard constraints reduces the overall evaluation result, and its value ranges from 0 to 1. For the first The degree of hard constraint violation of the candidate design, which represents the degree of violation of the first hard constraint. The degree to which each candidate design violates hard safety constraints or basic structural connection conditions is evaluated, with a value ranging from 0 to 1. A weighted harmonic calculation is used for the overall evaluation. A decrease in any evaluation score leads to a corresponding decrease in the overall evaluation result.

[0190] The evaluation rules stipulate that candidate designs that violate hard safety constraints are directly deleted. Candidate designs whose combined score does not meet the minimum adoption conditions for combined designs are returned to step S400, and candidate designs whose process score does not meet the minimum adoption conditions for process designs are returned to step S300. Candidate designs that meet the evaluation rules are ranked from highest to lowest according to the comprehensive evaluation results.

[0191] Sub-step S502: Based on the sorting results, arrange the effect diagrams, detail diagrams, color matching data, layout, and process suggestions to form a result package.

[0192] Candidate designs that meet the evaluation rules are selected based on the ranking results. A front and back design rendering are generated for each selected candidate design. The front and back design renderings use the same display scale, background, and lighting conditions. Detailed images are generated based on the garment components. These detailed images must include at least the neckline or placket detail, cuff or sleeve detail, and waist or hem detail. Color scheme data is generated. This data includes the color names, color swatch numbers, hue values, lightness values, and saturation values ​​for the primary, secondary, and accent colors.

[0193] Generate fabric recommendation data. This data includes fabric name, fiber composition, fabric weight range, fabric thickness range, elasticity requirements, breathability requirements, applicable garment components, and alternative fabrics. Generate pattern. The pattern includes a list of garment components, body size baseline, garment size range, ease allowance, growth allowance, movement allowance, seam allowance, pleat allowance, dart placement, and closure placement. Generate process suggestions. These suggestions include the number of cut pieces, cutting direction, garment component assembly sequence, seam type, seam allowance, binding method, overlock method, fastening method, printed areas, embroidered areas, equipment requirements, and quality inspection items.

[0194] The renderings, detailed images, color scheme data, pattern design, and process suggestions are compiled into a results package. The results package also includes the status of safety hard constraints verification, assembly verification, process feasibility verification, front and back consistency verification, and items pending manual confirmation.

[0195] Sub-step S503: Receive the modified content, locate the modified node, and regenerate the affected area to form new design data.

[0196] This system receives user modifications to the overall design, color, pattern, neckline, sleeve type, waistline, skirt length, looseness, fabric, decorative parts, safety features, and production costs. Modifications include the object of the modification, modification requirements, modification time, user operation method, and whether to save as a long-term preference. The modifications are then mapped to the corresponding modification nodes in the instruction diagram. Color modifications are mapped to the color node, pattern modifications to the pattern node, sleeve type modifications to the sleeve type node, fabric modifications to the fabric node, and wearing / removing method modifications to the neckline and placket nodes. When a modification involves multiple garment components, both the primary modification node and related modification nodes are determined simultaneously.

[0197] Lock the structure and data of unaffected nodes. When only color or pattern is modified, lock the overall silhouette, garment component structure, and pattern. When only neckline or sleeve type is modified, lock non-related garment component and fabric data. When fabric replacement is involved, re-perform comfort condition verification and production process condition verification.

[0198] Modifications are categorized into current task data and long-term preference data. When a user explicitly chooses to save a corresponding preference, the modification is recorded as long-term preference data. Modifications that do not meet the conditions for updating long-term preference data are recorded as current task data. The affected area is determined based on the main modification node, and the clothing components and constraints requiring re-verification are determined based on the associated modification nodes.

[0199] Regenerate the affected area, locking out the structures and data outside the area. Combine the original design task identifier, modified content, modified nodes, affected area, locked nodes, re-executed steps, regenerated results, and constraint verification results into new design data. The new design data serves as input for a new design task, and steps S300, S400, or S500 are re-executed as needed.

Claims

1. A personalized intelligent design generation system for children's clothing based on a multimodal large model, characterized in that, include: The multimodal requirements analysis module is used to acquire and standardize design data, perform fusion processing on the design data, and output a fusion vector and a conflict list; The children's clothing profile construction module is used to acquire measured information and construct a constraint profile based on the fusion vector and the conflict list, combined with the measured information. The knowledge-constrained reasoning module is used to retrieve design knowledge based on the constraint profile, handle conflicts during the retrieval process, and generate conditions and condition sequences. The children's clothing design generation module is used to generate and verify designs based on the generation conditions and the condition sequence, and output candidate designs. The feedback output service module is used to evaluate the candidate design and output a result package, as well as to receive the modified content and convert the modified content into new design data.

2. The personalized intelligent design generation system for children's clothing based on a multimodal large model according to claim 1, characterized in that, The multimodal requirements parsing module is used to acquire and standardize design data, perform fusion processing on the design data, and output a fusion vector and a conflict list, including the following processing steps: The system receives text descriptions, reference images, fabric data, children's information, and demand data as input modalities for the design data. The design data is standardized and its quality is evaluated to establish missing records and a conflict list. The design data is uniformly encoded and fused to output the fused vector.

3. The personalized intelligent design generation system for children's clothing based on a multimodal large model according to claim 2, characterized in that, The multimodal requirement parsing module is used to uniformly encode and fuse the design data, including the following processing steps: The text description, the reference image, the fabric data, the child information, and the demand data are respectively input into the corresponding encoding network, and feature vectors are output. Calculate the semantic relevance and reliability of each input modality with the design task, and determine the fusion weights; The feature vectors are weighted and fused according to the fusion weights to generate the fusion vector.

4. The personalized intelligent design generation system for children's clothing based on a multimodal large model according to claim 1, characterized in that, The children's clothing profile construction module is used to construct a constraint profile based on the fusion vector and the conflict list, combined with the measured information, including the following processing steps: Based on the child's age, height, and weight included in the measured information, determine the size range, growth margin, and movement margin; Based on the fusion vector and the conflict list, establish hard safety constraints, strong integration constraints, and soft aesthetic constraints; Extract body regions, clothing components, and fabric attributes from the fusion vector and the conflict list, and associate the body regions, clothing components, and fabric attributes with the safety hard constraints, the fit strong constraints, and the aesthetic soft constraints to generate the constraint profile.

5. The personalized intelligent design generation system for children's clothing based on a multimodal large model according to claim 1, characterized in that, The knowledge-constrained reasoning module is used to retrieve design knowledge based on the constraint profile and handle conflicts to form generation conditions and condition sequences, including the following processing steps: Extract query fields from the constraint profile, perform filtering and vector retrieval, and obtain candidate knowledge as the design knowledge; Check for conflicts between explicit requirements and image attributes in the design data, the constraint profile, and the candidate knowledge; complete the component attributes and write them into the instruction diagram to form the generation conditions. Calculate the matching degree of the candidate knowledge, select the selected knowledge from the candidate knowledge according to the matching degree, and convert the selected knowledge and the instruction graph into the condition sequence.

6. The personalized intelligent design generation system for children's clothing based on a multimodal large model according to claim 5, characterized in that, The knowledge-constrained reasoning module is used to calculate the matching degree of the candidate knowledge, including the following processing steps: Calculate the semantic similarity, category matching degree, comfort matching degree, and conflict degree between the candidate knowledge and the constrained profile, respectively; The matching degree is obtained by weighting the semantic similarity, the category matching degree, the comfort matching degree, and the conflict degree.

7. The personalized intelligent design generation system for children's clothing based on a multimodal large model according to claim 1, characterized in that, The children's clothing design generation module is used to generate and verify designs based on the generation conditions and the condition sequence, and output candidate designs, including the following processing steps: Generate a control map and a prohibited area map based on the aforementioned generation conditions; The control chart, the condition sequence, the constraint profile, and the forbidden region map are input into the diffusion network, and inverse denoising is performed based on constraint correction to generate an intermediate design. The intermediate design is refined into regions, verified for consistency, verified for security, and analyzed for process, to generate the candidate design.

8. The personalized intelligent design generation system for children's clothing based on a multimodal large model according to claim 7, characterized in that, The children's clothing design generation module is used to perform inverse denoising based on constraint correction and generate intermediate designs, including the following processing steps: During the denoising iteration process, the denoising direction, fitability deviation, safety deviation, and process deviation of the potential data are calculated. The noise reduction direction is adjusted based on the fit deviation, the safety deviation, and the process deviation. During the iteration process, the silhouette, fabric texture, color distribution, and local components are processed to output the intermediate design.

9. The personalized intelligent design generation system for children's clothing based on a multimodal large model according to claim 1, characterized in that, The feedback output service module is used to evaluate the candidate design and output a result package, including the following processing steps: Calculate the consistency score, integration score, safety score, coordination score, and process score for each candidate design, and sort them according to the preset evaluation rules; Based on the sorting results, the candidate designs that meet the preset evaluation rules are selected, and the renderings, detail images, color matching data, pattern and process suggestions are arranged to form the result package.

10. The personalized intelligent design generation system for children's clothing based on a multimodal large model according to claim 1, characterized in that, The feedback output service module is used to receive modified content and convert the modified content into new design data, including the following processing steps: Receive the modified content and locate the modified content to the corresponding modification node; Lock the structure and data of unaffected nodes, distinguish the modified content into current task data and long-term preference data, regenerate the affected area, and convert the regenerated result into the new design data.

Images