Intelligent garment pattern generation system based on multi-modal large model

Abstract

The present application relates to the technical field of clothing intelligent pattern making, and discloses an intelligent clothing pattern generation system based on a multi-modal large model, which comprises: an input receiving module that receives clothing input and target human body data, a human body model determination module that determines a target human body model, a multi-modal analysis module that generates clothing semantics and a standardized pattern JSON, a candidate configuration generation module that generates multiple sets of candidate configurations, a paper pattern generation module that generates a candidate two-dimensional paper pattern and reads the paper pattern waistline length, a simulation module that obtains a candidate three-dimensional simulated clothing, a loop parameter reading module that reads the simulated clothing waistline length, the human body same height circumference length and the clothing human body average gap based on the same waist closed transverse loop, a deviation determination module that determines the total deviation accordingly, and a target pattern output module that selects the candidate configuration with the smallest total deviation from the candidate configurations with the determined total deviation as the target pattern and outputs the corresponding result.

Description

Technical Field

[0001] This invention relates to the field of intelligent garment pattern making technology, and more specifically, to an intelligent garment pattern generation system based on a multimodal large model. Background Technology

[0002] In the garment pattern-making process, image recognition, text parsing, parametric CAD, 2D pattern generation, and 3D simulation are typically interconnected by different components. Engineering systems can convert design intentions into pattern data, and then send the cut pieces into a simulation environment to observe their wearing characteristics. In actual implementation, the same design will generate multiple sets of candidate values. 2D patterns can pass checks for closure, seam edges, and units, and 3D simulation can also produce displayable results, but these checks are mostly concentrated in their respective stages.

[0003] The ease allowance of the pattern may be consumed by mesh stretching, local fitting, or suspension after the sewing solution is obtained. If the system selects the result based solely on image appearance, simulation completion status, or experience-based scoring, it is difficult to reliably indicate whether the selected pattern matches the wearing space of the target human body model. For production-oriented data links, it is also necessary to retain verifiable amounts between CAD patterns, garment meshes, and human body models, so that pattern selection can be output together with manufacturable patterns, 3D simulated garments, and result metadata. Summary of the Invention

[0004] This invention provides an intelligent clothing pattern generation system based on a multimodal large model, which solves the technical problems mentioned in the background art.

[0005] To achieve the above-mentioned objectives, the present invention provides the following technical solution.

[0006] The intelligent clothing pattern generation system based on a multimodal large model includes an input receiving module, a human body model determination module, a multimodal analysis module, a candidate configuration generation module, a paper pattern generation module, a simulation module, a loop parameter reading module, a deviation determination module, and a target pattern output module. The input receiving module is used to receive clothing input and target human body data; The human body model determination module is used to determine the target human body model based on the target human body data, which is either the target human body model or standard human body size data. The multimodal parsing module is used to parse clothing input to obtain clothing semantics and generate standardized pattern JSON; The candidate configuration generation module is used to generate multiple sets of candidate configurations based on clothing semantics and the target human body model; The pattern generation module is used to convert candidate configurations into candidate two-dimensional patterns with stitching relationships and to determine the pattern waist length corresponding to the same closed transverse loop at the waist. The simulation module is used to virtually sew candidate two-dimensional paper patterns onto the target human body model to obtain candidate three-dimensional simulated clothing; The loop parameter reading module is used to read the waistline length of the simulated garment, the circumference of the human body at the same height, and the average gap between the garment and the human body based on the closed horizontal loop at the same waist. The deviation determination module is used to determine the total deviation based on the waistline length of the paper pattern, the waistline length of the simulated garment, the circumference of the human body at the same height, and the average gap between the garment and the human body. The total deviation simultaneously represents the deviation in length transfer and the deviation in wearing space. The target pattern output module is used to select the candidate configuration with the smallest total deviation from the candidate configurations with a determined total deviation as the target pattern and output the result.

[0007] Compared with the prior art, the present invention has the following substantial features and significant progress: This invention integrates multimodal analysis, parametric pattern making, 2D pattern generation, 3D simulation, and result output into a single pattern-making process. A single closed horizontal loop at the waist connects candidate 2D patterns, candidate 3D simulated garments, and the target human body model. The system can read the waistline length of the pattern, the waistline length of the simulated garment, the circumference of the human body at the same height, and the average gap between the garment and the human body. It then compares multiple candidate configurations using the total deviation. This reduces the instability caused by relying solely on the simulation completion status or manual scoring for pattern selection, ensuring that the target pattern selection corresponds to the manufacturable patterns, human body models, and 3D garment results, facilitating verification and subsequent editing. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the module architecture and loop verification mechanism of the intelligent clothing pattern generation system based on a multimodal large model according to the present invention. Detailed Implementation

[0009] The following description is provided in conjunction with the accompanying drawings. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the invention; those skilled in the art can make equivalent substitutions or combinations for the specific implementations.

[0010] This embodiment relates to an intelligent garment pattern generation system based on a multimodal large model. This intelligent garment pattern generation system, based on a multimodal large model, is used in intelligent garment pattern making scenarios. It can receive garment input and target human body data, convert the garment input into garment semantics and standardized pattern JSON, and further use the standardized pattern JSON for candidate configuration generation, candidate 2D paper pattern generation, candidate 3D simulation garment generation, reading parameters of the same waist closed horizontal loop, determining the total deviation, and outputting the target pattern.

[0011] In one implementation, the clothing input can be a clothing image, clothing text, or an existing pattern JSON. Clothing image input can come from actual clothing photos, design renderings, or sketches; clothing text input can come from natural language descriptions or natural language modification instructions; existing pattern JSON can come from saved pattern configurations. The target human body data is a target human body model or standard human body size data. When the target human body data is standard human body size data, the human body model determination module determines the target human body model based on the standard human body size data; when the target human body data is already a target human body model, the human body model determination module reads the target human body model and uses it for subsequent candidate 3D simulation clothing generation.

[0012] Reference Figure 1 The intelligent clothing pattern generation system 100 based on a multimodal large model represents the entire system through its outer boundary. The input receiving module 110, located at the starting point of the main link, receives clothing input and target human body data. The human body model determination and multimodal parsing module 120 corresponds to the human body model determination module and the multimodal parsing module. The candidate configuration generation module 130 generates multiple sets of candidate configurations based on clothing semantics and the target human body model. The pattern generation and 2D inspection module 140 corresponds to the related processing of the pattern generation module, configuration inspection and template mapping module, and 2D geometric validity inspection module. The simulation and loop parameter reading module 150 corresponds to the simulation module and the loop parameter reading module. The deviation determination and target pattern output module 160 corresponds to the deviation determination module and the target pattern output module.

[0013] This implementation retains the main chain of multimodal understanding, pattern parameters, 2D patterns, and 3D simulation. Building upon the system's ability to generate candidate 2D patterns and complete candidate 3D simulated garments from candidate configurations, it further uses the same closed horizontal loop at the waist as a unified geometric object. This ensures that the waistline length of the pattern, the waistline length of the simulated garment, the circumference of the human body at the same height, and the average gap between the garment and the human body correspond to the same wearing position. This allows the total deviation to simultaneously characterize both length transfer deviation and wearing space deviation, enabling the target pattern output module to select the candidate configuration with the smallest total deviation as the target pattern from the candidate configurations with the determined total deviation.

[0014] It should be noted that the system simultaneously determines the baseline wearing height when generating candidate configurations. The baseline wearing height is jointly determined by the waist structure position in the garment semantics and the waist cross-section position in the target human body model; when the garment semantics does not provide a separate waist structure offset, the baseline wearing height adopts the existing waist cross-section position corresponding to the waist structure in the target human body model. The baseline wearing height remains unchanged across all candidate configurations in the same pattern generation task and is transferred along with the target human body model to the pattern generation module, simulation module, and loop parameter reading module.

[0015] The same closed horizontal loop at the waist is represented in the candidate 2D pattern as a closed line formed by the corresponding line segments on each 2D vector piece at the reference wearing height, after being closed by darts, pleats, and side seams; in the candidate 3D simulated garment, it is represented as a closed curve mapped onto the garment mesh by the corresponding line segments of the candidate 2D pattern after virtual stitching and fabric solving; and in the target human body model, it is represented as a cross-section at the same height as the reference wearing height. All three are recorded using the same pattern generation task, the same target human body model, and the same candidate configuration as their indexes.

[0016] If a candidate 2D paper pattern cannot form a closed line at the reference wearing height, or a candidate 3D simulation garment cannot form a closed curve through mapping, or a target human body model cannot form a closed human body cross section at the reference wearing height, then the candidate configuration will not be included in the total deviation comparison.

[0017] The input preprocessing, field integrity check, data type check, value range check, parameterized pattern template mapping, two-dimensional geometric validity check, virtual stitching, fabric solving, and file output in this embodiment can all be implemented using conventional engineering methods in the fields of clothing CAD, three-dimensional clothing simulation, and multimodal analysis.

[0018] Furthermore, the implementation methods of the input receiving module and the input standardization module are as follows: The input receiving module receives clothing input and target human body data. This module can be configured with a file reading interface, a business interface, or a memory object interface to obtain clothing image input, clothing text input, or existing pattern JSON. When receiving clothing image input, the module can read the image file's pixel data, image size, and file identifier; when receiving clothing text input, it can read the text string, text source, and task type; when receiving existing pattern JSON, it can read the field structure and field values. The module also receives target human body data and binds it to the clothing input as part of the same pattern generation task.

[0019] The input receiving module sends the clothing input into the same multimodal parsing process. This "same multimodal parsing process" means that clothing image input, clothing text input, or existing pattern JSON are all processed into input objects that the multimodal parsing module can read, and the clothing semantics are generated by the multimodal parsing module based on the clothing input. For clothing image input, the multimodal parsing module can read the clothing outline, clothing components, and local style information from the image; for clothing text input, the multimodal parsing module can read the clothing category, component composition, discrete style options, and continuous pattern values ​​from the natural language description; for existing pattern JSON, the multimodal parsing module can read existing fields and determine the fields that need to be updated based on natural language modification instructions.

[0020] The input standardization module converts clothing input and target human body data into standard requests that can be received by the multimodal parsing module. Standard requests may include a task identifier, clothing input reference, target human body data reference, unit information, and a set of fields to be parsed. The input standardization module can perform size standardization, color normalization, and effective region cropping on clothing image input; it can perform word segmentation and length checks on clothing text input; and it can perform unit standardization and missing value checks on target human body data.

[0021] The task identifier in the standard request is used to associate the same pattern generation task. The clothing input reference points to clothing image input, clothing text input, or existing pattern JSON. The target human body data reference points to the target human body model or standard human body size data. The unit information is used to limit length-related data to use the same length unit. The set of fields to be parsed limits the clothing categories, component compositions, discrete style options, and continuous pattern values ​​that the multimodal parsing module needs to output. After the input standardization module completes the standard request, it passes the standard request as a whole to the human body model determination module and the multimodal parsing module.

[0022] When the input standardization module performs data validation, it first checks whether the clothing input is readable, then checks whether the target human body data is readable, and then checks whether the unit information can be converted to the same length unit. If the clothing image input cannot parse pixel data, the clothing text input is empty, the existing pattern JSON cannot be parsed into a hierarchical object, or the standard human body size data lacks a waist size source, the input standardization module returns the exception to the input receiving module and stops the clothing input from entering the multimodal parsing module.

[0023] When at least one of the following—the clothing image input, the clothing text input, and the existing pattern JSON—exhibits a formatting error, the input standardization module can return the reason for the error to the input receiving module. Formatting errors include images that cannot be read, empty text, the existing pattern JSON missing a root node, the target human body data missing a necessary size field, or the target human body model failing to load.

[0024] Furthermore, the implementation methods of the human body model determination module and the multimodal analysis module are as follows: The human body model determination module is used to determine the target human body model based on the target human body data. The target human body data can be either a target human body model or standard human body size data. When the target human body data is a target human body model, the module reads the mesh data, pose data, and size identifiers of the target human body model and passes the target human body model to the candidate configuration generation module, simulation module, and loop parameter reading module. When the target human body data is standard human body size data, the module selects a target human body model from an existing target human body model library based on the standard human body size data, or determines the target human body model by performing a standard scaling adjustment on an existing target human body model based on the standard human body size data.

[0025] When determining the target human body model, the human body model determination module must maintain consistency among candidate configurations. For multiple candidate configurations generated under the same clothing semantics, the target human body model must not be replaced with different body types among the candidates; otherwise, the human body circumference at the same height and the average gap between the clothing and the human body cannot be used as a common basis for comparison. After the target human body model is determined, its posture, scale unit, and the coordinate system for reading the human body section at the same height are all shared conditions for subsequent simulation modules and loop parameter reading modules.

[0026] A multimodal parsing module is used to parse clothing input to obtain clothing semantics and generate standardized pattern JSON. The multimodal parsing module may include a pre-trained visual encoder, a language model, a conventional projection layer, or a cross-attention mechanism. The multimodal parsing module extracts clothing outline, clothing components, and local style information from the clothing input and outputs clothing category, component composition, discrete style options, and continuous pattern values. In one specific implementation, the above model components are configured according to the following model structure, dataset, and training parameters to align image and text information into the same inference flow.

[0027] Specifically, the multimodal parsing module consists of a visual encoder, a visual-language projection layer, and an autoregressive language model. The visual encoder employs a Transformer-based image encoding structure. The clothing image input is uniformly scaled to 336×336 pixels and encoded as 14×14 pixel image blocks. The visual encoder includes 24 Transformer layers with a hidden dimension of 1024 and 16 attention heads. The visual-language projection layer uses a two-layer fully connected network. The first layer maps the 1024-dimensional visual features to 4096 dimensions and uses the GELU activation function. The second layer has both 4096 input and output dimensions. The projected visual features serve as a visual tag sequence, which, along with the clothing text input and the text tag sequence corresponding to the target human body data, are fed into the autoregressive language model. The autoregressive language model includes 32 Transformer decoding layers with a hidden dimension of 4096 and 32 attention heads. For editing tasks where both existing pattern JSON and clothing text input coexist, the field names, field values, and natural language modification instructions of the existing pattern JSON are serialized into text tags and then fed into the autoregressive language model.

[0028] In this specific implementation, a total of 30,000 sets of data are used for the supervised fine-tuning of the multimodal parsing module. Each set of data includes clothing input, target human body data, and a standardized pattern JSON tag corresponding to the input and validated by parametric clothing CAD data. Of these, 24,000 sets are used as training data, 3,000 sets as validation data, and 3,000 sets as test data. Data corresponding to the same clothing style is only assigned to one set among the training, validation, and test sets to avoid duplication of the same style across sets. Of the 24,000 sets of training data, 12,000 sets are inputs combining clothing images and clothing text, 4,000 sets are inputs with clothing images alone, 4,000 sets are inputs with clothing text alone, and 4,000 sets are inputs combining existing pattern JSON and natural language modification instructions. Validation and test data are set with the above four types of inputs in corresponding quantities of 1,500, 500, 500, and 500 sets, respectively. The target human body data is bound to the clothing input in each set of data using the size identifier of the target human body model or a standard human body size field.

[0029] During training, the pre-trained parameters of the visual encoder were frozen. The visual-language projection layer was trained, and the query projection and value projection layers of the autoregressive language model were fine-tuned using low-rank adaptation. The rank of the low-rank adaptation was set to 16, the scaling parameter to 32, and the dropout rate to 0.05. The training objective was the autoregressive cross-entropy loss of a normalized JSON sequence. The optimizer was AdamW. The learning rate of the visual-language projection layer was set to 0.0001, the learning rate of the low-rank adaptation parameters of the autoregressive language model was set to 0.00002, the weight decay was set to 0.01, the effective batch size was set to 32, and the number of training epochs was set to 8. The first 5% of the training steps were used for learning rate warm-up, and the maximum input-output sequence length of a single training sample was set to 1536 tags. After each training epoch, the validation loss was calculated using validation data, and the model parameters with the lowest validation loss were retained as test and inference parameters.

[0030] After model training, the parameters of the visual encoder, visual-language projection layer, and autoregressive language model were fixed, and the model was tested using 3000 sets of test data that were not used in training and validation. During testing, each set of model output was checked sequentially to see if the JSON could be parsed, if the clothing category field existed, if the component composition and discrete style options belonged to the allowed enumeration set of the corresponding clothing category, and if the continuous pattern values ​​were numeric and had a uniform length unit. Outputs that failed the above checks were recorded as test failures and were not used as input to the candidate configuration generation module. During actual inference, the decoding temperature was set to 0.1, the top-p was set to 0.9, the maximum output length was set to 1024 tags, and the output format of the root node, field name, enumeration field, and "numerical + unit" data object was restricted according to the preset JSON field structure, so that the output of the multimodal parsing module could be directly read by the candidate configuration generation module.

[0031] Clothing semantics are used to express clothing categories, component composition, discrete style options, and continuous pattern values ​​related to the target human body model. Clothing categories can represent skirts, tops, trousers, or other clothing types; component composition can represent the bodice, sleeves, collar, waist structure, slits, pleats, or other clothing parts; discrete style options can indicate the existence, type, or silhouette of a component; continuous pattern values ​​can represent circumference, length, width, pleat amount, and other values ​​that can be read from the parameterized pattern template. When the multimodal parsing module outputs standardized pattern JSON, the clothing semantics should be organized into a structure with clearly defined fields, a clear hierarchy, and readable by the candidate configuration generation module.

[0032] As a concrete example of a standardized pattern JSON, when the clothing image input represents an A-line midi skirt, the clothing text input represents "straight waist, darts at the front and back, no pleats," and the target body data shows a waist circumference of 720 mm and a hip circumference of 960 mm, the multimodal parsing module can output the following standardized pattern JSON. The clothing category, component composition, and discrete style options in this JSON are used to define the style target, and the continuous CAD field is used to provide the candidate configuration generation module with continuous pattern values ​​that the parametric pattern template can read. { "garment_category": "skirt", "components": { "waistband": "straight", "dart": "front_and_back", "pleat": "none" }, "discrete_style": { "silhouette": "A_line", "length_type": "midi" }, "continuous_cad": { "waist_circumference": {"value": 760, "unit": "mm"}, "skirt_length": {"value": 650, "unit": "mm"}, "waistband_width": {"value": 35, "unit": "mm"}, "dart_intake_total": {"value": 48, "unit": "mm"} } } The standardized pattern JSON mentioned above is the structured parsing result of the multimodal parsing module, not the target pattern determined without comparison. After reading this JSON, the candidate configuration generation module keeps the garment category, component composition, and discrete style options corresponding to garment_category, components, and discrete_style unchanged, and combines them with the size data corresponding to the target human body model to generate multiple sets of continuous CAD values ​​that meet the value range of the parametric pattern template in the circumference, length, width, or pleat amount fields corresponding to continuous_cad. The target pattern is then determined by subsequent pattern generation, 3D simulation, loop parameter reading, and total deviation comparison.

[0033] For editing tasks where both existing pattern JSON and garment text input coexist, the multimodal parsing module generates garment semantics based on these two inputs. The multimodal parsing module only updates fields explicitly involved in the garment text input; fields not involved remain unchanged. This process allows the subsequent candidate configuration generation module to generate candidate configurations based on the existing pattern JSON, and ensures that the modified candidate configurations are still processed by the pattern generation module, simulation module, loop parameter reading module, deviation determination module, and target pattern output module.

[0034] Furthermore, the implementation method of the candidate configuration generation module is as follows: The candidate configuration generation module generates multiple sets of candidate configurations based on clothing semantics and the target human body model. Each candidate configuration is a hierarchical JSON data object representing a set of pattern configurations that can be included in the pattern generation module. Upon receiving the clothing semantics, the module generates multiple sets of consecutive CAD values ​​that satisfy the category field and value range under the same clothing semantics and target human body model. Each set of consecutive CAD values ​​is then merged with the same discrete style options into a hierarchical JSON-formatted candidate configuration.

[0035] The standardized pattern JSON uses clothing category as the upper-level constraint, component composition and discrete style options as style constraints, and continuous pattern values ​​as the source of values ​​that the parameterized pattern template can read. When the candidate configuration generation module reads the standardized pattern JSON, it first confirms the clothing category, then confirms whether the component composition is compatible with the clothing category, and then reads the discrete style options and continuous pattern values. The candidate configuration retains the association identifier with the target human body model, so that the subsequent pattern generation module, simulation module, and loop parameter reading module can confirm that all candidate configurations use the same target human body model.

[0036] The continuous CAD values ​​in the candidate configuration are written into hierarchical JSON according to the fields required by the parametric pattern template. Each continuous CAD value retains both the numerical value and the unit, with the unit being standardized by the input standardization module. The candidate configuration generation module prohibits writing two different field names for the same continuous CAD value in the same candidate configuration; the same field name is also prohibited from representing two different continuous CAD values ​​within the same clothing category.

[0037] Candidate configurations express the same style target. For the same task, the candidate configuration generation module does not change the clothing category, component composition, and discrete style options among candidates, but preserves the differences on continuous CAD values ​​that allow continuous variation, such as circumference, length, and pleat amount.

[0038] The candidate configuration generation module can read clothing semantics from the standardized pattern JSON output by the multimodal parsing module and generate continuous CAD values ​​by combining them with the size data corresponding to the target human body model. The continuous CAD values ​​must satisfy the category field and value range of the current clothing category. For the same candidate configuration, the continuous CAD values ​​and discrete style options are written to the same hierarchical JSON, enabling the configuration checking and template mapping module to check field integrity, data type, and value range.

[0039] When the candidate configuration generation module generates candidate configurations, the number of candidates can be determined by computing resources, task type, or existing decoding strategies of the system. As long as the candidate configurations can express multiple consecutive CAD numerical results under the same garment semantics, the deviation determination module can determine the total deviation based on the pattern waistline length, the simulated garment waistline length, the human body circumference at the same height, and the average gap between the garment and the human body. The target pattern output module then selects the one with the smallest total deviation as the target pattern.

[0040] When the candidate configuration generation module receives a candidate configuration containing out-of-bounds values ​​returned by the configuration checking and template mapping module, it can regenerate the candidate configuration based on clothing semantics and the target human body model. The regenerated candidate configuration must maintain consistency with the same design intent in terms of clothing category, component composition, and discrete style options. Figure 1 This is consistent with, and adjustments are made only within the normal range on continuous CAD values.

[0041] Furthermore, the implementation methods for the configuration check and template mapping module, the pattern generation module, and the two-dimensional geometric validity check module are as follows: The configuration check and template mapping module is used to check the field integrity, data type, and value range of candidate configurations. The field integrity check confirms that the candidate configuration contains the fields required for the current garment category to enter the parameterized pattern template; the data type check confirms that continuous CAD values ​​are numeric fields and discrete style options are enumerated fields allowed by the current garment category; the value range check confirms that continuous CAD values ​​are within the preset garment parameter range corresponding to the garment category and the parameterized pattern template.

[0042] Field integrity checks are performed based on the necessary field set jointly determined by the garment category and the parameterized pattern template. This necessary field set is provided by the parameterized pattern template before execution, and the configuration check and template mapping module confirm the existence of corresponding fields in the candidate configurations field by field. Data type checks are performed according to field roles; continuous CAD values ​​should be convertible to the same length unit, and discrete style options should be enumerated values ​​allowed by the current garment category. Value range checks are performed based on the single-field value range and field combination constraints provided by the parameterized pattern template.

[0043] When a single consecutive CAD value exceeds the preset garment parameter range without affecting the combination of other fields, the configuration check and template mapping module performs boundary truncation. When the combination of consecutive CAD values ​​does not meet the requirements of the parameterized pattern template, the configuration check and template mapping module returns the candidate configuration containing out-of-bounds values ​​to the candidate configuration generation module for regeneration. After boundary truncation, the configuration check and template mapping module performs field integrity, data type, and value range checks again. Only candidate configurations that pass the second check will be mapped to the parameterized pattern template or pattern making program.

[0044] When the configuration check and template mapping module detects out-of-bounds values, it can either perform boundary truncation on the out-of-bounds values ​​according to the preset garment parameter range corresponding to the garment category and parameterized pattern template, or return the candidate configuration containing the out-of-bounds values ​​to the candidate configuration generation module for regeneration. Boundary truncation refers to adjusting continuous CAD values ​​exceeding the preset garment parameter range to the boundary values ​​of the preset garment parameter range. Returning for regeneration means that the candidate configuration generation module regenerates continuous CAD values ​​under the same garment semantics and target human body model.

[0045] The configuration check and template mapping module maps candidate configurations that pass the check to parameterized pattern templates or pattern-making programs based on the garment category. The parameterized pattern template can be a template object from an existing garment CAD system, and the pattern-making program can be a program capable of reading hierarchical JSON and outputting two-dimensional vector cut pieces. Specifically, the configuration check and template mapping module converts the fields and values ​​in the candidate configurations into template parameters that the pattern generation module can execute.

[0046] The pattern generation module converts candidate configurations into candidate 2D patterns with stitching relationships. It can generate 2D vector pieces with realistic dimensions and sutures for virtual stitching based on parametric pattern templates or pattern-making programs. The 2D vector pieces can contain points, lines, curves, darts, pleats, slits, seam allowances, and alignment marks. The sutures for virtual stitching are used in the simulation module to determine the connectivity of the candidate 2D patterns when they are virtually stitched around the target human body model.

[0047] The pattern generation module is also used to determine the waistline length of the pattern corresponding to the same closed horizontal loop at the waist. The module locates the corresponding line segments of the same closed horizontal loop at the waist on each two-dimensional vector pattern piece in the candidate two-dimensional pattern, and after deducting the portion actually incorporated by dart closures or pleats from the candidate two-dimensional pattern, accumulates the waistline length along the corresponding line segments on each pattern piece. The waistline length is represented by the symbol... The waistline length of the pattern is derived from 2D CAD data and represents the effective closed loop length of the waistline after stitching in the candidate 2D pattern.

[0048] When calculating the waistline length of the pattern generation module, it first reads the candidate line segments corresponding to the baseline wearing height on each two-dimensional vector pattern piece, and determines the connection order of each candidate line segment based on the seam lines and alignment marks. For straight candidate line segments, the pattern generation module reads the endpoints of the line segments and calculates the distance between the endpoints; for curved candidate line segments, the pattern generation module calculates the arc length according to the curve control points or discrete points output by the pattern making program. The lengths of all candidate line segments are first unified to the same length unit before being accumulated into the waistline length.

[0049] The portion actually swallowed by the dart closure is determined by the projected length of the closure edges on both sides of the dart along the same horizontal loop of the waist closure; the portion actually swallowed by the pleat fold is determined by the folded length between the pleat fold lines along the same horizontal loop of the waist closure. The pattern generation module only deducts the length of the dart closure and pleat fold that do not participate in the outer waist closure line after wearing, and does not deduct the edge segments of the fabric pieces that still participate in the outer waist closure line. After deduction, the pattern generation module checks whether the corresponding line segments on each fabric piece can be connected by the side seam to form a closure line.

[0050] The calculation relationship for the waistline length of a pattern can be written as: In the formula, Indicates the length of the waistline of the pattern. Indicates the first The lengths of candidate line segments participating in the same waist-closed transverse loop. This indicates the number of candidate line segments participating in the cumulative calculation. Indicates the first The actual length that was encroached upon by the closure of the provincial highway. This indicates the number of provincial highway closures included in the deduction. Indicates the first The actual length that is folded in by the pleats. This indicates the number of pleats / folds that are deducted.

[0051] If the distance between the endpoints of adjacent candidate line segments after the side seam connection exceeds the geometric tolerance of the 2D vector pattern file, the pattern generation module determines that the candidate 2D pattern cannot form the same closed horizontal loop at the waist and returns the corresponding candidate configuration to the configuration check and template mapping module. If the closure error is within the geometric tolerance of the 2D vector pattern file, the pattern generation module calculates the waist length of the pattern according to the closed line after connection and writes the line segment identifiers and pattern identifiers involved in the calculation into the mapping relationship.

[0052] The geometric tolerance of the 2D vector pattern file is determined by the precision of the 2D vector pattern file to which the candidate 2D pattern belongs. When the precision of the 2D vector pattern file is not recorded separately, it is determined by the coordinate precision output by the pattern-making program. When the coordinate precision is not recorded separately by the pattern-making program, it is determined by the default drafting precision of the parameterized pattern template under the current length unit. Before performing the closed-line judgment, the pattern generation module converts the geometric tolerance of the above-mentioned 2D vector pattern files into the length unit in the standard request, and keeps the geometric tolerance of the 2D vector pattern file unchanged across all candidate configurations in the same pattern generation task. The geometric tolerance of the 2D vector pattern file is only used to determine whether the endpoints of the candidate line segments can form a closed line.

[0053] When the pattern generation module determines the waistline length of the pattern, it needs to establish a mapping relationship between the closed horizontal loop of the same waist and the cut piece mesh in the candidate 2D pattern. This mapping relationship is used by the simulation module to complete the candidate 3D simulated garment, so that the loop parameter reading module can read the position corresponding to the closed horizontal loop of the same waist in the garment mesh.

[0054] The 2D geometric validity check module is used to check the 2D geometric validity of candidate 2D patterns that affect virtual stitching and the mapping of the same waist closure transverse loop. 2D geometric validity can include the closure of the pattern outline, self-intersection of curves, presence of seam edges, consistent direction of paired seams, missing pattern pieces, and uniformity of dimensional units. The 2D geometric validity check module returns the candidate configurations corresponding to candidate 2D patterns that fail the check to the configuration check and template mapping module, enabling the configuration check and template mapping module to re-check the candidate configurations or remap the parameterized pattern template.

[0055] The pattern piece outline closure check is completed by checking whether the beginning and end endpoints coincide within the geometric tolerance of the 2D vector pattern piece file; the curve self-intersection check is completed by checking whether there are non-endpoint intersections between the same pattern piece outline and internal curves; the seam edge presence check is completed by checking whether all the seams used for virtual sewing can find paired pattern piece edges; the paired seam line direction consistency check is completed by checking whether the endpoint order and alignment marks of the paired seams are reversed; the pattern piece missing check is completed by checking whether all the pattern piece identifiers required by the candidate configuration have been generated; and the size unit consistency check is completed by checking whether the length unit in the candidate 2D paper pattern is consistent with the unit information in the standard request.

[0056] The 2D geometry validity check module also checks whether the mapping of the closed horizontal loop at the same waistline is complete. If any candidate line segment involved in the waistline length calculation of the pattern lacks a piece identifier, line segment identifier, or mapping identifier with the garment grid, the 2D geometry validity check module determines that the candidate 2D pattern has failed the check. Candidate 2D patterns that fail the check do not enter the simulation module, but instead return the corresponding candidate configuration to the configuration check and template mapping module.

[0057] The pattern generation module and the 2D geometry validity check module provide executable objects for subsequent virtual stitching and fabric solving. If a candidate 2D pattern has an incomplete cut outline, missing seam edges, or inconsistent size units, it may not be able to form a candidate 3D simulation garment, even if the candidate configuration meets the requirements for field completeness and data type. Therefore, the 2D geometry validity check module is located before the simulation module and returns candidate configurations that fail the check to the configuration check and template mapping module.

[0058] Furthermore, the implementation methods of the simulation module, the loop parameter reading module, and the deviation determination module are as follows: The simulation module is used to virtually sew candidate 2D paper patterns onto a target human body model to obtain candidate 3D simulated clothing. Under consistent simulation settings between the virtual sewing and fabric solving processes of each candidate 2D paper pattern, the simulation module performs fabric solving on candidate 2D paper patterns that have passed the 2D geometry validity check module. Simulation settings can include fabric properties, gravity, collision conditions, solver step size, and convergence conditions.

[0059] When comparing candidate configurations under the same garment semantics, the simulation module maintains consistency in the target human body model, target human body model pose, fabric properties, gravity direction, collision conditions, solver step size, and convergence conditions. The simulation module records a simulation setting identifier for each candidate configuration, and the deviation determination module only compares candidate configurations with consistent simulation setting identifiers. If a candidate configuration generates a candidate 3D simulated garment due to inconsistent simulation settings, the deviation determination module will not include that candidate configuration in the overall deviation comparison.

[0060] The simulation is considered incomplete if the fabric solution fails to meet the convergence criteria, the garment mesh is not generated, the correspondence between the candidate 3D simulated garment and the candidate 2D pattern is lost, or the mapping relationship of the same closed horizontal loop at the waist cannot be read after the fabric solution is completed. The simulation module writes the above states into the candidate result states, and the deviation determination module removes candidate configurations with the simulation incomplete state before determining the total deviation.

[0061] When the simulation module performs virtual sewing, it determines the connection relationship between the two-dimensional vector pieces based on the seam lines generated by the pattern generation module for virtual sewing, and places candidate two-dimensional patterns around the same target human body model. After the fabric solution is completed, the simulation module outputs candidate three-dimensional simulation garments. Each candidate three-dimensional simulation garment should correspond one-to-one with the candidate two-dimensional pattern that generated it, so that the target pattern output module can simultaneously output the corresponding two-dimensional vector pattern file and the three-dimensional simulation garment when selecting the target pattern.

[0062] The loop parameter reading module is used to read the waistline length of the simulated garment, the circumference of the human body at the same height, and the average gap between the garment and the human body based on a closed horizontal loop at the same waist. The module reads the waistline length of the simulated garment according to the mapping relationship between the closed horizontal loop at the same waist and the garment mesh. The waistline length of the simulated garment is represented by the symbol... The value is derived from the clothing mesh in the candidate 3D simulation clothing, and represents the arc length of the corresponding closed curve of the same waist closed horizontal loop in the candidate 3D simulation clothing.

[0063] When the loop parameter reading module reads the waistline length of the simulated garment, it locates the sequence of grid points corresponding to the same closed horizontal loop at the waist in the garment mesh of the candidate 3D simulated garment, based on the pattern identifier, line segment identifier, and mapping identifier written by the pattern generation module. The grid point sequence is arranged according to the connection order in the candidate 2D pattern. When the first and last grid points coincide or are within the garment mesh closure tolerance, the grid point sequence is considered to form a closed curve. The waistline length of the simulated garment is obtained by accumulating the 3D distance between adjacent grid points.

[0064] The garment mesh closure tolerance is determined by the mesh resolution of the candidate 3D simulated garment. When the garment mesh file records mesh positioning accuracy, that accuracy is used. When the garment mesh file does not record mesh positioning accuracy, the mesh closure accuracy supported by the distance between adjacent mesh points in the corresponding mesh point sequence of the same waistline closed transverse loop is used. When the simulation system has provided a closure curve reading tolerance, that tolerance is used. Before reading the waistline length of the simulated garment, the loop parameter reading module converts the garment mesh closure tolerance to a unit consistent with the pattern waistline length and maintains this tolerance unchanged across all candidate configurations in the same pattern generation task. The garment mesh closure tolerance is used to determine whether the mesh point sequence forms a closed curve.

[0065] The calculation relationship for the waistline length of simulated clothing can be written as: In the formula, Indicates the length of the waistline in simulated clothing. This indicates the first closed horizontal loop corresponding to the same waistline in the clothing mesh. Grid points, Indicates and The next adjacent grid point, This indicates the number of grid points involved in calculating the waistline length of the simulated garment. For closed curves, and They are in the same closed position.

[0066] If any mapping point is missing from the grid point sequence, or if the first and last grid points cannot form a closed curve within the clothing grid closure tolerance, the loop parameter reading module will not output the simulated clothing waistline length and will mark the candidate configuration as a failed read of the same waistline closed transverse loop. This candidate configuration will not be included in the total deviation comparison.

[0067] The loop parameter reading module reads the body circumference at the height corresponding to the closed horizontal loop at the same waist. The body circumference at the same height is represented by a symbol. This indicates that the circumference of the target human body model is located at the same height as the closed transverse loop at the same waistline. The circumference at this height provides the baseline circumference of the target human body model and is used for scale normalization of the total deviation.

[0068] When the loop parameter reading module reads the circumference of the human body at the same height, it determines the intercept plane based on the reference wearing height and uses this intercept plane to intersect with the human body mesh of the target human body model. The intersection points obtained from the intersection are sorted according to the connection order of the target human body model surface to form a closed cross section at the same height. The circumference of the human body at the same height is obtained by accumulating the distances between adjacent intersection points on the cross section at the same height.

[0069] The relationship between the height and circumference of the human body can be written as: In the formula, Indicates the circumference of a person at the same height. Denotes the first cross-section at the same height as the target human body model. A point of intersection, Indicates and The next adjacent intersection point, This indicates the number of intersection points involved in the calculation of the circumference of a human body at the same height. For a closed cross-section of a human body at the same height, and They are in the same closed position.

[0070] If the intercepted plane cannot form a closed cross section at the same height as the target human body model, or if the reading result of the human body circumference at the same height is empty, non-numerical, or less than or equal to zero, the loop parameter reading module will not output the human body circumference at the same height, and the deviation determination module will remove the candidate configuration.

[0071] The loop parameter reading module reads the corresponding position of the same waist-closed transverse loop in the candidate 3D simulation garment and the average gap between the garment and the human body at the same height cross-section. The average gap between the garment and the human body is represented by the symbol. The average distance between the candidate 3D simulation garment and the target human body model is derived from the geometric relationship between them. It represents the average normal distance from the corresponding position of the same closed transverse loop at the waist in the candidate 3D simulation garment to the same height section of the human body. The average gap between the garment and the human body is used to characterize the wearing space formed by the candidate 3D simulation garment relative to the target human body model.

[0072] When calculating the average gap between the garment and the human body, the loop parameter reading module uses the grid point sequence used to calculate the waistline length of the simulated garment as the sampling point set. For each sampling point, the loop parameter reading module finds the nearest cross-sectional point corresponding to that sampling point on the same height cross-section of the human body, and uses the distance from the sampling point to the nearest cross-sectional point as the gap distance of that sampling point. If the candidate 3D simulated garment penetrates the target human body model at that sampling point, the gap distance of that sampling point is recorded as a negative value, and the candidate configuration is removed based on the candidate result elimination rules.

[0073] The calculation relationship for the average gap between clothing and the human body can be written as: In the formula, Indicates the average gap between clothing and the human body. This indicates the number of grid points involved in calculating the waistline length of the simulated garment. This indicates the first closed horizontal loop corresponding to the same waistline in the clothing mesh. Grid points, Indicates the cross-section of the human body at the same height as The corresponding nearest cross-section point, Indicates the first The sign factor of each sampling point. When no penetration occurs, Take the positive sign; when penetration occurs, Take the negative sign.

[0074] If the number of grid points in the closed transverse loop at the same waist is less than the minimum number required to form a closed curve, or if any sampling point cannot find the nearest corresponding cross-section point on the same height of the human body, the loop parameter reading module will not output the average gap between the garment and the human body and will mark the candidate configuration as a failure to read the average gap between the garment and the human body. This candidate configuration will not be included in the total deviation comparison.

[0075] The deviation determination module is used to determine the total deviation based on the waistline length of the paper pattern, the waistline length of the simulated garment, the circumference of the human body at the same height, and the average gap between the garment and the human body. The deviation determination module can first determine the paper pattern simulation difference. The paper pattern simulation difference is represented by symbols. The calculation relationship is as follows: In the formula, This indicates poor paper pattern simulation. Indicates the length of the waistline of the pattern. This represents the waistline length of the simulated garment. The pattern simulation difference characterizes the length transfer deviation at the corresponding position of the closed horizontal loop at the same waist after the candidate 2D pattern enters the virtual sewing and fabric solution. The larger the pattern simulation difference, the greater the difference between the corresponding length of the closed horizontal loop at the same waist in the candidate 2D pattern and the corresponding length of the closed horizontal loop at the same waist in the candidate 3D simulated garment.

[0076] The deviation determination module can also determine the gap perimeter difference. The gap perimeter difference is represented by the symbol... The calculation relationship is as follows: In the formula, Indicates the difference in circumference of the gap. Indicates the length of the waistline in simulated clothing. Indicates the circumference of a person at the same height. This represents the average gap between the garment and the human body. The difference in the perimeter of the gap is used to characterize the deviation of the candidate 3D simulated garment from the wearing space formed by the target human body model. This calculation relationship is based on the correspondence between the same closed transverse loop at the waist and the cross-section at the same height of the human body, and reads the degree of deviation based on the conventional geometric relationship that the perimeter increment is formed by the average normal distance.

[0077] The deviation determination module comprehensively considers the differences in paper pattern simulation, gap perimeter differences, and the circumference of the human body at the same height to determine the total deviation. The total deviation is represented by the symbol... The calculation relationship is as follows: In the formula, Indicates the total deviation. This indicates poor paper pattern simulation. Indicates the difference in circumference of the gap. This represents the circumference of the human body at the same height. The total deviation simultaneously characterizes the deviation in length transfer and the deviation in wearing space, and is normalized to scale using the circumference of the human body at the same height. The deviation determination module determines the total deviation for each group of candidate configurations that have completed the generation of candidate 3D simulation clothing and have not been eliminated.

[0078] Before determining the total deviation, the deviation determination module can first eliminate candidate results that have not completed the simulation or have obvious penetration, misalignment, or slippage of fabric pieces. This elimination process is used to remove candidate results that cannot form valid candidate 3D simulated garments. For the remaining candidate configurations, the deviation determination module determines the total deviation according to the same calculation relationship.

[0079] Obvious penetration is determined by whether there are multiple consecutive negative gap distances at sampling points on the same closed transverse loop at the waist of the candidate 3D simulation garment. If the negative gap distance causes the average gap between the garment and the human body to fail to represent the wearing space of the candidate 3D simulation garment outside the target human body model, the deviation determination module will eliminate the candidate configuration. Piece misalignment is determined by whether the paired seams used for virtual sewing maintain corresponding alignment marks in the candidate 3D simulation garment. If the alignment marks of the paired seams do not correspond, the deviation determination module will eliminate the candidate configuration. Slippage is determined by whether the position of the same closed transverse loop at the waist in the candidate 3D simulation garment deviates from the area corresponding to the reference wearing height. If the deviation causes the loop parameter reading module to fail to read the waistline length of the simulation garment corresponding to the reference wearing height, the deviation determination module will eliminate the candidate configuration.

[0080] Consecutive negative gap distances refer to the negative gap distances where at least three sampling points appear consecutively in the grid point sequence of the same closed transverse loop at the waist, following the order of adjacent grid points. If the same closed transverse loop at the waist is stored as a higher-density sampling point sequence after resampling by the simulation system, the consecutive negative gap distances are still determined according to the number of consecutive sampling points that can cover at least one local garment grid segment, and the same determination method is maintained across all candidate configurations for the same pattern generation task.

[0081] The deviation determination module eliminates candidate configurations in the following order: first, candidate configurations that have not been completed in simulation are eliminated; then, candidate configurations that have failed to read the closed transverse loop at the same waist are eliminated; then, candidate configurations that have obvious penetration, misalignment or slippage of the cut pieces are eliminated; and finally, the paper pattern simulation difference, gap perimeter difference and total deviation are determined for the remaining candidate configurations.

[0082] Furthermore, the implementation methods for the target template output module and the editing and review process are as follows: The target pattern output module selects the candidate configuration with the smallest total deviation from the candidate configurations with a determined total deviation as the target pattern and outputs the result. The target pattern output module receives the total deviation of each candidate configuration output by the deviation determination module, and compares the remaining candidate configurations after eliminating those that have not completed simulation or have obvious penetration, misalignment, or slippage. The target pattern output module selects the one with the smallest total deviation as the target pattern, ensuring that the target pattern simultaneously satisfies the length transfer relationship from the candidate 2D paper pattern to the candidate 3D simulated garment and the wearing space relationship formed by the candidate 3D simulated garment relative to the target human body model.

[0083] The target pattern output module outputs a standardized pattern JSON file corresponding to the target pattern, a 2D vector pattern file with actual dimensions and seam markings, and a 3D simulated garment. The standardized pattern JSON file records the garment semantics, discrete style options, and continuous CAD values ​​corresponding to the target pattern; the 2D vector pattern file records the 2D vector cut pieces and seams used for virtual sewing corresponding to the target pattern; and the 3D simulated garment records the candidate 3D simulated garment results corresponding to the target pattern.

[0084] The target pattern output module also retains the following in the result metadata: pattern waistline length, simulated garment waistline length, human body circumference at the same height, average gap between garment and human body, and total deviation. The result metadata can be used as a field in the standardized pattern JSON or as a record associated with the output file. The result metadata is used to verify why the target pattern was selected and to re-execute candidate configuration generation, pattern generation, simulation, loop parameter reading, deviation determination, and target pattern output in subsequent editing tasks.

[0085] In one editing and review process, the input receiving module receives natural language modification instructions in the form of clothing text input and receives existing pattern JSON. The multimodal parsing module reads the existing pattern JSON and updates only the fields explicitly involved in the clothing text input; the candidate configuration generation module generates multiple sets of candidate configurations based on the updated clothing semantics and the target human body model; the configuration checking and template mapping module, pattern generation module, two-dimensional geometric validity checking module, simulation module, loop parameter reading module, deviation determination module, and target pattern output module execute their processing sequentially. The modified pattern formed in this way still uses the one with the smallest total deviation as the target pattern.

[0086] When the total deviation of multiple candidate configurations is the same, the target version output module can process them in parallel according to the existing candidate order, generation confidence, or user preference.

[0087] When the total deviation of multiple candidate configurations is the same, the target pattern output module prioritizes selecting the target pattern according to the existing candidate order in the system; if the existing candidate order in the system does not exist, the target pattern is selected according to the generation confidence score already output by the multimodal parsing module or the candidate configuration generation module; if the generation confidence score also does not exist, the target pattern is selected according to the user preference explicitly given by the user in the clothing text input.

[0088] As an optional implementation, the clothing input can be either a clothing image or clothing text, and the target human body data can be standard human body size data. The input standardization module converts the clothing input and target human body data into a standard request; the multimodal parsing module obtains the clothing semantics and generates a standardized pattern JSON; the candidate configuration generation module generates multiple sets of candidate configurations; the pattern generation module generates multiple sets of candidate two-dimensional patterns and reads the waistline length of each candidate two-dimensional pattern; the simulation module obtains multiple sets of candidate three-dimensional simulated clothing; the loop parameter reading module reads the waistline length of the simulated clothing, the human body circumference at the same height, and the average gap between the clothing and the human body for each set of candidate three-dimensional simulated clothing; the deviation determination module determines the total deviation; and the target pattern output module selects the pattern with the smallest total deviation as the target pattern.

[0089] As an optional implementation, multiple candidate configurations undergo checks for field integrity, data type, value range, and two-dimensional geometric validity, and the simulation module generates candidate 3D simulated garments for each configuration. However, after virtual stitching and fabric solving, different candidate configurations exhibit varying total deviations in the corresponding pattern waistline length, simulated garment waistline length, human body circumference at the same height, and average gap between the garment and the human body. The deviation determination module determines the total deviation based on the same calculation relationship. The target pattern output module does not select the target pattern based on manual scoring but instead selects the candidate configuration with the smallest total deviation from the candidate configurations with the determined total deviation as the target pattern.

[0090] As an optional implementation, if the waistline lengths of the two candidate configurations are similar, but the waistline length of the simulated garment obtained after fabric calculation for one of the candidate configurations differs significantly from the waistline length of the paper pattern, then the paper pattern simulation error of that candidate configuration increases. Conversely, if the waistline length of the simulated garment for the other candidate configuration is close to the waistline length of the paper pattern, but the difference in the gap perimeter formed by the human body's height circumference and the average gap between the garment and the human body is large, then the gap perimeter difference of that candidate configuration increases. The deviation determination module adds the paper pattern simulation error and the gap perimeter difference and normalizes it using the human body's height circumference. The target pattern output module then selects the one with the smallest total deviation as the target pattern.

[0091] As an optional implementation, the configuration check and template mapping module has already generated a parameterized pattern template input, but the candidate 2D patterns output by the pattern generation module have 2D geometric validity issues that affect the mapping of virtual stitching or the same waist closure transverse loop. The 2D geometric validity check module returns the candidate configurations corresponding to the candidate 2D patterns that fail the check to the configuration check and template mapping module. The configuration check and template mapping module re-checks the candidate configurations, and if necessary, returns candidate configurations containing out-of-bounds values ​​to the candidate configuration generation module for regeneration.

[0092] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and all such modifications and variations fall within the scope defined by the appended claims.

Claims

1. An intelligent garment pattern generation system based on a multimodal large model, comprising an input receiving module, a human body model determination module, a multimodal analysis module, a candidate configuration generation module, a pattern generation module, a simulation module, a loop parameter reading module, a deviation determination module, and a target pattern output module, characterized in that: The input receiving module is used to receive clothing input and target human body data; The human body model determination module is used to determine the target human body model based on the target human body data, which is either the target human body model or standard human body size data. The multimodal parsing module is used to parse clothing input to obtain clothing semantics and generate standardized pattern JSON; The candidate configuration generation module is used to generate multiple sets of candidate configurations based on clothing semantics and the target human body model; The pattern generation module is used to convert candidate configurations into candidate two-dimensional patterns with stitching relationships and to determine the pattern waist length corresponding to the same closed transverse loop at the waist. The simulation module is used to virtually sew candidate two-dimensional paper patterns onto the target human body model to obtain candidate three-dimensional simulated clothing; The loop parameter reading module is used to read the waistline length of the simulated garment, the circumference of the human body at the same height, and the average gap between the garment and the human body based on the closed horizontal loop at the same waist. The deviation determination module is used to determine the total deviation based on the waistline length of the paper pattern, the waistline length of the simulated garment, the circumference of the human body at the same height, and the average gap between the garment and the human body. The total deviation simultaneously represents the deviation in length transfer and the deviation in wearing space. The target pattern output module is used to select the candidate configuration with the smallest total deviation from the candidate configurations with a determined total deviation as the target pattern and output the result.

2. The intelligent clothing pattern generation system based on a multimodal large model according to claim 1, characterized in that, The clothing input includes clothing image input, clothing text input, or existing pattern JSON. The input receiving module is used to send the clothing input into the same multimodal parsing process so that the multimodal parsing module can generate the clothing semantics based on the clothing input.

3. The intelligent clothing pattern generation system based on a multimodal large model according to claim 1, characterized in that, It also includes an input standardization module, which is used to convert the clothing input and the target human body data into standard requests that can be received by the multimodal parsing module.

4. The intelligent clothing pattern generation system based on a multimodal large model according to claim 1, characterized in that, The multimodal parsing module is used to extract clothing outline, clothing components and local style information from the clothing input, and output clothing category, component composition, discrete style options and continuous pattern values.

5. The intelligent clothing pattern generation system based on a multimodal large model according to claim 1, characterized in that, The candidate configuration generation module is used to generate multiple sets of continuous CAD values ​​that satisfy the category field and value range based on the same clothing semantics and the target human body model, and to merge each set of continuous CAD values ​​with the same discrete style options into a hierarchical JSON-formatted candidate configuration.

6. The intelligent clothing pattern generation system based on a multimodal large model according to claim 1, characterized in that, It also includes a configuration check and template mapping module, which is used to check the field integrity, data type and value range of the candidate configuration, perform boundary truncation for out-of-bounds values ​​according to the preset clothing parameter range corresponding to the clothing category and parameterized pattern template, or return the candidate configuration containing out-of-bounds values ​​to the candidate configuration generation module for regeneration, and map the candidate configuration that passes the check to the parameterized pattern template or pattern making program according to the clothing category.

7. The intelligent clothing pattern generation system based on a multimodal large model according to claim 1, characterized in that, The pattern generation module is used to generate two-dimensional vector pieces with real dimensions and seam lines for virtual sewing. After deducting the portion actually swallowed by dart closure or pleat folding from the candidate two-dimensional pattern, the waistline length of the pattern is accumulated along the corresponding line segments on each piece.

8. The intelligent clothing pattern generation system based on a multimodal large model according to claim 6, characterized in that, It also includes a two-dimensional geometric validity check module, which is used to check the two-dimensional geometric validity of the virtual stitching of the candidate two-dimensional paper pattern and the mapping of the same waist closure transverse loop, and return the candidate configuration corresponding to the candidate two-dimensional paper pattern that fails the check to the configuration check and template mapping module.

9. The intelligent clothing pattern generation system based on a multimodal large model according to claim 8, characterized in that, The simulation module is used to perform fabric solving on candidate 2D patterns that have passed the 2D geometric validity check module under a consistent simulation setting between the virtual sewing and fabric solving processes of each candidate 2D pattern; the loop parameter reading module is used to read the waistline length of the simulated garment according to the mapping relationship between the same waist closed horizontal loop and the garment mesh, read the human body circumference at the height corresponding to the same waist closed horizontal loop, and read the average gap between the garment and the human body at the same height section of the corresponding position of the same waist closed horizontal loop in the candidate 3D simulated garment.

10. The intelligent clothing pattern generation system based on a multimodal large model according to claim 1, characterized in that, The target pattern output module is used to output a standardized pattern JSON corresponding to the target pattern, a two-dimensional vector paper pattern file with real size and seam marks, and a three-dimensional simulated garment, and retains the paper pattern waistline length, simulated garment waistline length, human body circumference at the same height, average gap between the garment and the human body, and total deviation in the result metadata.

Images