A Multimodal Sewing Pattern Generation Method Based on a Large Garment Model

Abstract

This invention discloses a multimodal sewing pattern generation method based on a large-scale garment model, belonging to the field of intelligent garment manufacturing technology. By receiving multimodal inputs such as text, images, or sketches, a design semantic graph representing garment components and their sewing relationships is constructed using a large-scale garment model, generating an intent vector. This vector is then used to drive the generation of a two-dimensional pattern piece containing parameterized boundary curves and sewing interface declarations. Subsequently, the arc length, torsion angle, and curvature of the sewing boundary pairs of the pattern piece are geometrically verified. If the tolerance threshold dynamically set based on garment category and fabric attributes is not met, the deviation is marked and feedback is used for reconstruction, forming a self-correcting closed loop until a pattern piece that meets sewing requirements is output. This invention does not rely on a pre-built pattern library, supports end-to-end original design, and significantly improves the engineering feasibility and first-time success rate of the generated results through pre-implementation of manufacturing constraints and automatic iterative optimization, achieving efficient transformation from creative input to a manufacturable pattern.

Description

Technical Field

[0001] This invention relates to the field of intelligent garment manufacturing technology, and in particular to a method for generating multimodal sewing patterns based on a large garment model. Background Technology

[0002] In the current field of intelligent apparel design, artificial intelligence technology has been gradually applied to processes such as pattern generation, virtual try-on, and process optimization. However, mainstream solutions still have significant limitations: on the one hand, image- or text-based generative models mostly focus on visual appearance synthesis, outputting only two-dimensional renderings or three-dimensional display models, lacking structured pattern data that can be directly used for production, resulting in "designs that look good but cannot be sewn"; on the other hand, while some systems can output two-dimensional patterns, they heavily rely on pre-built pattern template libraries, achieving customization through retrieval matching and local parameter adjustments, making it difficult to support truly innovative structural designs, and lacking sufficient support for multimodal input. In addition, existing methods generally place manufacturability verification at the end of the process, relying on human experience or post-hoc simulation for rework, failing to embed manufacturing constraints in the generation stage, creating a disconnect between design and production, resulting in low efficiency and high error-tolerance costs.

[0003] To address the aforementioned issues, patent publication number CN121147408A proposes a method for intelligent 3D simulation generation of clothing styles based on a pattern library. This method constructs a cross-modal knowledge graph containing structural parameters, process rules, and visual elements, intelligently optimizes basic pattern templates using a Transformer-PINN hybrid model, and utilizes digital twins for physical simulation verification, thereby improving the accuracy and physical rationality of pattern generation. This solution, to a certain extent, achieves the integration of pattern rules and physical constraints, alleviating the rigidity of traditional template library methods. However, it still relies on a limited pre-stored pattern library and cannot achieve end-to-end original pattern generation without library dependence. Furthermore, its output is primarily 3D simulation, failing to explicitly generate editable and manufacturable programmatic 2D cut pieces, and it lacks a sewnability hard constraint and self-correction closed-loop mechanism based on seam edge geometry. Therefore, it fails to solve the core technical problem addressed by this solution: how to automatically generate sewnable clothing patterns that satisfy both aesthetic creativity and engineering feasibility from arbitrary multimodal inputs, while ensuring a high success rate in one generation. Summary of the Invention

[0004] In view of this, the present invention proposes a multimodal sewing pattern generation method based on a large garment model. Through multimodal-driven end-to-end generation and a self-correcting closed loop based on arc length, torsion angle, and curvature for sewing feasibility, a highly feasible garment pattern output method is achieved, which does not require a pattern library and can be used for manufacturing after a single generation.

[0005] This invention provides a method for generating multimodal sewing patterns based on a large garment model, comprising the following steps: S1. Receive one or more multimodal inputs, the multimodal inputs including text, images, or sketches; S2. Based on the multimodal input, a design semantic graph representing clothing components and their stitching relationships is constructed through a large clothing model, and an intent vector is generated. The design semantic graph is used to represent the shape of clothing components and their stitching relationships. S3. Based on the intent vector, generate a procedural two-dimensional pattern piece containing parameterized boundary curves and seam interface declarations, wherein the parameterized boundary curves and seam interface declarations are used to define the edge characteristics and connection methods of the garment components; S4. For the boundary pairs to be sewn in the programmed two-dimensional cut piece, calculate the geometric properties of the boundary pairs, including arc length, twist angle and curvature; S5. Determine the dynamic geometric tolerance threshold based on the clothing category and fabric attributes, and determine whether the geometric attributes of the boundary pair meet the sewing criterion; if not, mark the deviation and reconstruct the design semantic graph and two-dimensional pattern, repeat the process until the boundary pair meets the conditions or the maximum number of iterations is reached. S6. Output two-dimensional cut pieces that meet the sewing requirements.

[0006] Furthermore, after receiving the multimodal input, a comprehensive quality score is calculated for the k-th modal input. The comprehensive quality score Calculate using the following formula: ; in, The integrity score is calculated using the following formula: ; This represents the number of valid terms in the k-th mode. This indicates the number of responses for the k-th mode; The validity score is calculated using the following formula: ; This represents the number of out-of-bounds terms in the k-th mode. This represents the number of answers received in the k-th modality; The consistency score is calculated using the following formula: ; The semantic representation vector representing the k-th modality. Represents the fused reference semantic vector. This is the numerical stability constant, with a value range of 10. -8 ~10 -6 ; , , The preset weighting coefficients for quality sub-items, and satisfying the following conditions: .

[0007] Furthermore, the parameterized boundary curve adopts an nth-order parameter. The curve represents the parametric equation as follows: ,in, Let i be the nth Bernstein basis function. Let i be the corresponding i-th control point, where i is an integer index and its value range is 0≤i≤n.

[0008] Furthermore, the geometric tolerance threshold includes arc length matching tolerance, torsion angle tolerance, and curvature difference tolerance. The arc length matching tolerance is an absolute length deviation or a relative percentage deviation. The torsion angle tolerance ranges from ±5° to ±15°. The curvature difference tolerance is calculated based on the L2 norm difference of the boundary curvature distribution. The geometric tolerance threshold is dynamically determined according to the current garment category, fabric elastic modulus, and pattern complexity, and is calibrated in three stages: prior initial value, historical samples, and online calibration.

[0009] Furthermore, the sewingability criterion is defined as follows: ; in, , , These represent arc length, angle of twist, and curvature, respectively. , , These represent the arc length matching tolerance, torsion angle tolerance, and curvature difference tolerance, respectively. For indicator functions, , These represent the arc lengths of the first and second edges in the boundary pair to be stitched, respectively. , These represent the torsion angles of the two boundaries at the seam endpoints; , These represent the curvature functions of the two boundary curves as parameters, respectively. The difference in its L2 norm over the entire parameter range.

[0010] Furthermore, the stitch consistency loss is calculated based on the geometric properties of the boundary pairs. And determine whether to trigger reconstruction based on the stitch consistency loss; the stitch consistency loss Defined as: ;in, This indicates the loss of stitch consistency in the current cut pieces. Let represent the set of boundary pairs to be stitched together, where each element (i, j) represents a pair of boundaries to be stitched together. , Let these represent the arc lengths of the i-th and j-th boundaries, respectively. The normalized reference length of this boundary pair is taken as . or Used to eliminate scale effects. This represents the weighting coefficient, used to balance the relative importance of differences in arc length and curvature, and its value ranges from 0.1 to 10. , Let the curvature functions of the i-th and j-th boundaries be represented respectively. The arc length represents the L2 norm. Calculated by integrating the first derivative of the parametric curve C(t): ; among which, L i Let c represent the arc length of the i-th edge. i (t) represents the parametric curve of the i-th edge. This represents the first derivative of the curve with respect to the parameter t. Let represent the vector norm, and t represent the curve parameter.

[0011] Furthermore, the clothing big model adopts a multimodal architecture with Transformer as the backbone, in which text input is encoded by a language big model and image or sketch input is encoded by a visual big model. The two are fused in a cross-modal attention layer to generate the intent vector. The model is trained on a dataset containing aligned samples of two-dimensional cut pieces and three-dimensional wearing states. The samples are automatically labeled with structured semantic tags by a big language model.

[0012] Furthermore, before outputting the final two-dimensional cut piece, a three-dimensional physical simulation of the cut piece is also included. If fabric penetration or local stress exceeds the fabric yield strength occurs during the simulation, the simulation failure area is fed back as an additional deviation to the design semantic graph reconstruction step.

[0013] Furthermore, the method employs a joint loss function for optimization during the training phase. The joint loss function includes: cross-entropy loss for the text generation task and masked L1 loss for the patch geometry parameter regression task, wherein the geometry parameters include boundary control point coordinates and stitching interface identifiers.

[0014] Furthermore, the stitching interface declaration includes the identity of the boundary pair to be matched, the allowed geometric tolerance range, and the seam allowance width parameter, guiding the automated cutting and sewing equipment to perform manufacturing operations.

[0015] The present invention has the following advantages over the prior art: By constructing an end-to-end multimodal pattern generation and sewnability verification closed loop, the core problem of existing AI clothing design methods—namely, the unsewnability of generated results or the weak generalization ability due to reliance on template libraries—is effectively solved.

[0016] First, the system uses a large garment model to generate a design semantic graph representing the structure of the parts and the stitching relationship based on multimodal inputs such as text, images or sketches. It then drives the programmable output of two-dimensional cut pieces containing parameterized boundary curves and stitching interface declarations, thereby eliminating the dependence on pre-stored pattern libraries and supporting original structural designs.

[0017] Secondly, to address the question of whether the generated cut pieces are truly sewn, the solution introduces a quantitative verification mechanism for the geometric properties of the seam edges and sets dynamic tolerance thresholds associated with clothing categories and fabric properties to ensure that the cut pieces have geometric matching during physical sewing.

[0018] Furthermore, when the verification fails, the system can automatically label the type and location of the deviation and feed it back to the upstream to reconstruct the semantic graph and pattern pieces, forming a self-correcting iterative loop until the sewing criterion is met or the iteration limit is reached, significantly improving the success rate of one-time generation.

[0019] In addition, the seam interface declaration clearly includes boundary identification, tolerance range and seam allowance parameters, so that the output results can be directly connected to automatic cutting and sewing equipment, connecting the final link from design to manufacturing. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a functional architecture diagram of an embodiment of the present invention; Figure 2 A semantic graph is designed for an example embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figure 1 As shown, this invention provides a multimodal sewing pattern generation method based on a large garment model, comprising the following steps: S1. Receive one or more multimodal inputs, including text, images, or sketches; after receiving the multimodal inputs, calculate a comprehensive quality score for the k-th modal input. The comprehensive quality score Calculate using the following formula: ; in, The integrity score is calculated using the following formula: ; This represents the number of valid terms in the k-th mode. This indicates the number of responses for the k-th mode; The validity score is calculated using the following formula: ; This represents the number of out-of-bounds terms in the k-th mode. This represents the number of answers received in the k-th modality; The consistency score is calculated using the following formula: ; The semantic representation vector representing the k-th modality. Represents the fused reference semantic vector. This is the numerical stability constant, with a value range of 10. -8 ~10 -6 ; , , The preset weighting coefficients for quality sub-items, and satisfying the following conditions: .

[0024] In this implementation, the system receives one or more multimodal inputs from the user, such as text, images, or sketches. First, a modality discrimination module identifies the type of each input and dynamically allocates weights to subsequent processing pathways based on input quality metrics. Then, semantic cleaning and word segmentation are performed on the text, and scale normalization and edge enhancement are applied to the images and sketches to ensure consistency across different modalities. Based on this, a unified semantic embedding representation is constructed, and a semantic indexing mechanism is established for subsequent semantic graph parsing and cross-modal matching optimization. This process employs neural network-based feature extraction and standardization methods to ensure the comparability and stability of input features within a unified semantic space.

[0025] Specifically, the system first receives input data in at least one modality provided by the user, including natural language text, design reference images, or hand-drawn sketches. The system performs preliminary parsing of the input through a front-end interface and then activates the modality discrimination module.

[0026] The modality discrimination module identifies the modality type of each input based on the metadata of the input data and a lightweight classification network, and generates corresponding modality indication signals. Subsequently, the system performs targeted preprocessing for each type of input: for text input, word segmentation and semantic cleaning strategies are used to remove irrelevant words and grammatical noise, and the text is encoded into a context-aware semantic vector sequence through a pre-trained language model LLM; for image input, scale normalization, background removal, and edge enhancement are performed to extract clear clothing outlines and key component regions; for sketch input, topological connectivity restoration and line smoothing are performed to convert discrete strokes into continuous geometric path representations.

[0027] After completing single-modal processing, the system performs quality assessment calculations on all inputs according to the formula described above. Comprehensive consideration of completeness , effectiveness With consistency And accordingly, allocate weights to subsequent fusion pathways. If If the modal input is deemed to have serious quality issues, such as missing content, semantic contradictions, or exceeding boundaries, the system will display "Input Invalid" and refuse to use it for subsequent fusion processing; if If the modality is deemed to have low credibility, the system reduces its path weight during the cross-modal fusion stage to prevent it from having an excessive impact on the final intent vector; if The modality is deemed reliable and assigned normal weights for fusion. Ultimately, the features of each modality are mapped to a unified semantic subspace, forming a structured joint embedding representation, and a semantic index is established to support subsequent semantic graph construction and retrieval matching optimization.

[0028] S2. Based on the multimodal input, a design semantic graph representing clothing components and their stitching relationships is constructed through a large clothing model, and an intent vector is generated. The design semantic graph is used to represent the shape of clothing components and their stitching relationships. The clothing big model adopts a multimodal architecture with Transformer as the backbone, in which text input is encoded by a language big model and image or sketch input is encoded by a visual big model. The two are fused in a cross-modal attention layer to generate the intent vector. It is trained on a dataset containing two-dimensional cut pieces and three-dimensional wearing state aligned samples, and the samples are automatically labeled with structured semantic labels by a big language model.

[0029] The method employs a joint loss function for optimization during the training phase. The joint loss function includes: cross-entropy loss for the text generation task and masked L1 loss for the patch geometry parameter regression task, wherein the geometry parameters include boundary control point coordinates and stitching interface identifiers.

[0030] Specifically, the system uses the multimodal input representation output by S1 to call a pre-trained large clothing model, constructs a structured design semantic graph, and generates an intent vector to drive the generation of cut pieces.

[0031] The large-scale clothing model employs a multimodal architecture with Transformer as its core. Specifically, the original text input is first context-aware encoded using a large-scale language model (LLM) to generate a semantically rich token sequence; image or hand-drawn sketch input is then processed by a large-scale visual model to extract multi-scale spatial features. Both types of features are then fed into a cross-modal attention layer, where text tokens serve as queries and visual features serve as keys and values. By calculating cross-modal association weights, semantic alignment and information fusion are achieved, ultimately outputting a unified multimodal embedding representation.

[0032] The model was trained on the GarmentCode dataset, which contains approximately 100,000 high-quality samples. Each sample consists of a triple: (1) a natural language description or design sketch; (2) the corresponding procedural two-dimensional cutout, containing (3) The three-dimensional mesh state of the pattern piece after being worn on a standard mannequin. Among them, the structured semantic labels in the pattern piece are automatically labeled by a large language model based on the three-dimensional wearing effect and process knowledge, ensuring that the labels have both geometric accuracy and manufacturing semantics.

[0033] During the training phase, the model employs a joint loss function for end-to-end optimization: on the one hand, for the text generation task, cross-entropy loss is used to supervise the language decoder; on the other hand, for the piece geometry parameter regression task, masked L1 loss is used to supervise only the coordinates of the effective boundary control points and the position of the seam interface markers, ignoring the filling area, thereby accurately constraining the shape and connection relationship of the piece.

[0034] Based on the trained model, the system extracts garment component elements and their stitching topological relationships from the fused multimodal embedding, such as "front side seam - back side seam", and constructs a design semantic graph.

[0035] like Figure 2 The diagram shown is a design semantic graph of the T-shirt pattern in an embodiment of the present invention. Nodes represent garment components, such as the front piece, back piece, left sleeve, right sleeve, and neckline ribbing. Directed edges represent the stitching relationships between components, such as shoulder seams, side seams, armhole-sleeve cap, and neckline-ribbing. This structured representation clearly captures the topological connection logic and technological constraints of the garment. Based on this design semantic graph, the present invention proposes a formal modeling method using a directed weighted graph G=(V,E,C), whose construction rules are as follows: Node set V: Each node Represents a garment component, including but not limited to the front piece, back piece, sleeve piece, collar piece, and pocket piece. Node attributes include component type, fabric grain direction, symmetry identifier, and semantic slots; Edge set E: Each edge This indicates the stitching relationship between two components, i.e., components A certain boundary needs to be related to the component The corresponding boundary stitching. Edge attributes include stitch type, seam allowance width, and allowable geometric tolerance range; Constraint Set C: To ensure the process feasibility of the generated pattern, the system applies the following structural constraints, including the no-isolated-parts constraint, contour closure constraint, and topology consistency constraint. The no-isolated-parts constraint means that there are no nodes with a degree of 0 in the graph, i.e., each part participates in at least one stitching relationship; the contour closure constraint means that the boundary curve of each part must form a closed loop (i.e., the start and end points coincide), and the number of stitching interfaces must be even; the topology consistency constraint prohibits non-standard process structures such as "three-way stitching points" (i.e., three parts intersect at the same vertex).

[0036] During the construction process, if the initially generated semantic graph violates the above constraints, the system will activate a conflict resolution mechanism: for incompleteness, nodes are automatically completed based on prior knowledge of the clothing category; for topological conflicts, boundary pairings are adjusted through a local semantic repair module, or the system reverts to S1 to re-parse the input; all correction operations are recorded as structural consistency loss, used to optimize subsequent generation. Finally, the system performs global pooling on the compliant design semantic graph and generates a low-dimensional intent vector through nonlinear mapping. This vector not only encodes the component composition and connection relationships but also implicitly contains manufacturing constraint information, which can directly drive the programmatic pattern generation in S3.

[0037] S3. Based on the intent vector, generate a procedural 2D pattern piece containing parameterized boundary curves and seam interface declarations. The parameterized boundary curves and seam interface declarations define the edge characteristics and connection methods of the garment components. The parameterized boundary curves employ an n-order parameter. The curve represents the parametric equation as follows: ,in, Let i be the nth Bernstein basis function. Let i be the corresponding i-th control point, where i is an integer index and its value range is 0 ≤ i ≤ n. The stitching interface declaration includes the identity of the boundary pair to be matched, the allowed geometric tolerance range, and the seam allowance parameter, which is used to guide the automatic cutting and sewing equipment to perform manufacturing operations.

[0038] In this embodiment, the procedural pattern generation module adopts a hybrid architecture based on graph neural networks and parametric decoders. This module first parses the number, type, and stitching topology of garment components implicit in the intent vector, assigning an independent geometry generation sub-network to each component. Subsequently, for each component, this sub-network outputs a set of control point sequences. And using nth order The curve is parametrically modeled along its boundary, and its parametric equations are as follows: ,in, Let be the Bernstein basis functions of order i. These are two-dimensional plane control points. Where n is of order 3 to 6, preferably order 5. 8-16 control points per component, 5th order Achieving an optimal balance between expressive power and numerical stability, this approach has been widely adopted by mainstream garment CAD systems. Each garment component's complete closed boundary consists of 8 to 16 control points; simple components, such as rectangular pockets and straight-leg trousers, use 8 to 10 control points; complex components, such as dress fronts and suit sleeves, use 12 to 16 control points to accurately depict high-curvature areas such as necklines, armholes, and darts. The distribution of control points follows the fabric grain direction constraint, ensuring warp / weft alignment and preventing cutting deformation. By adjusting the control point positions, various edge shapes, including straight lines, arcs, and free curvatures, can be accurately expressed, ensuring smooth and continuous boundaries to meet the contour accuracy requirements of garment pieces. After generating all component boundaries, the system automatically generates seam interface declarations based on the seam relationships in the design semantic diagram constructed by S2. Each seam interface declaration includes the following three key types of information: Identification: explicitly specifying the two boundaries to be matched, such as "right armhole of the front piece" and "left sleeve cap of the right sleeve"; Allowable geometric tolerance range: including the arc length matching tolerance, torsion angle tolerance, and curvature difference threshold corresponding to this interface. This tolerance can be inherited from global settings or dynamically adjusted according to local process requirements; Seam allowance parameter: specifying the seam allowance width reserved at this seam, used to guide subsequent automatic cutting equipment to extend the seam by the corresponding distance outside the boundary. The final output two-dimensional pattern is encapsulated in a structured data format, containing all components... This format includes a control point list, fabric grain direction indicator, seam joint declaration, and metadata. It can be directly parsed by mainstream garment CAD software or intelligent cutting systems, enabling layout, cutting, and sewing path planning without manual intervention.

[0039] S4. For the boundary pairs to be sewn in the programmed two-dimensional cut piece, calculate the geometric properties of the boundary pairs, including arc length, twist angle and curvature; In this implementation, the system iterates through all the boundary pairs to be stitched specified in the stitching interface declaration generated by S3. For each boundary pair... Calculate its three key geometric properties respectively: (1) Arc length: Suppose the boundary is represented by the parametric curve C(t), then its arc length L is calculated by numerical integration: ;in, The first derivative of the curve with respect to the parameter t. Let represent the Euclidean norm, preferably with the lower limit of integration a=0 and the upper limit of integration b=1. (2) Curvature distribution: at each sampling point t of the curve, calculate the local curvature. For a planar parametric curve, the curvature formula is: The difference in curvature between two sides is measured by the L2 norm of their curvature functions: (3) Twist angle: refers to the angle between the tangent vectors of the two boundaries at the seam endpoints. Let the boundaries be... The unit tangent vector at the endpoints is , for Then the twist angle The range of values ​​is In practical applications, attention should be paid to the absolute value of its deviation from the ideal docking angle (usually 0° or 180°).

[0040] S5. Determine the dynamic geometric tolerance threshold based on the garment category and fabric attributes, and determine whether the geometric attributes of the boundary pair meet the sewing criterion; if not, mark the deviation and reconstruct the design semantic graph and two-dimensional pattern, repeating the process until the boundary pair meets the conditions or reaches the maximum number of iterations; the geometric tolerance threshold includes arc length matching tolerance, torsion angle tolerance, and curvature difference tolerance, wherein the arc length matching tolerance is an absolute length deviation or a relative percentage deviation, the torsion angle tolerance ranges from ±5° to ±15°, and the curvature difference tolerance is calculated based on the L2 norm difference of the boundary curvature distribution; the geometric tolerance threshold is dynamically determined based on the current garment category, fabric elastic modulus, and pattern complexity, and is calibrated in three stages: prior initial value, historical samples, and online calibration.

[0041] The sewingability criterion is defined as follows: ; in, , , These represent arc length, angle of twist, and curvature, respectively. , , These represent the arc length matching tolerance, torsion angle tolerance, and curvature difference tolerance, respectively. For indicator functions, , These represent the arc lengths of the first and second edges in the boundary pair to be stitched, respectively. , These represent the torsion angles of the two boundaries at the seam endpoints; , These represent the curvature functions of the two boundary curves as parameters, respectively. The difference in its L2 norm over the entire parameter range.

[0042] The maximum number of iterations is 2 to 10, and the iteration is terminated early when the decrease in stitch consistency loss between two consecutive iterations is less than the preset convergence threshold.

[0043] Step S5 calculates the stitch consistency loss based on the geometric properties of the boundary pair. And determine whether to trigger reconstruction based on the stitch consistency loss; the stitch consistency loss Defined as: ;in, This indicates the loss of stitch consistency in the current cut pieces. Let represent the set of boundary pairs to be stitched together, where each element (i, j) represents a pair of boundaries to be stitched together. , Let these represent the arc lengths of the i-th and j-th boundaries, respectively. The normalized reference length of this boundary pair is taken as . or Used to eliminate scale effects. This represents the weighting coefficient, used to balance the relative importance of differences in arc length and curvature, and its value ranges from 0.1 to 10. , Let the curvature functions of the i-th and j-th boundaries be represented respectively. The arc length represents the L2 norm. Calculated by integrating the first derivative of the parametric curve C(t): ; among which, L i Let c represent the arc length of the i-th edge. i (t) represents the parametric curve of the i-th edge. This represents the first derivative of the curve with respect to the parameter t. Let represent the vector norm, and t represent the curve parameter.

[0044] In this implementation, the system performs a sewing feasibility criterion verification on each calculated seam-to-seam boundary and determines whether to trigger the reconstruction of the design semantic map and pattern pieces based on the verification results. First, the system dynamically determines three types of geometric tolerance thresholds based on the current garment category (e.g., T-shirt, suit, dress), fabric elastic modulus (reflecting tensile properties), and pattern complexity (e.g., number of seam lines, number of parts): arc length matching tolerance... : Can be set to absolute deviation (e.g., ±3mm) or relative percentage deviation (e.g., ≤5%) to tolerate minor length differences in cut pieces due to fabric stretchability; Torsion angle tolerance The value ranges from ±5° to ±15°, with smaller values ​​used for rigid fabrics and larger torsion allowed for elastic fabrics; curvature difference tolerance. The threshold is set based on the L2 norm difference of the curvature distribution of the two boundaries and calibrated statistically using historical good product samples. The threshold is determined through a three-stage calibration mechanism: Prior initial value setting: initialization based on process manuals and expert experience; historical good product sample verification: extracting similar style data from the pattern library of successfully produced patterns in the factory to calibrate the threshold distribution; online calibration: dynamically fine-tuning the threshold parameters based on user feedback and the actual sewing success rate of the manufacturing equipment. Subsequently, the system applies the sewingability criterion Φ to perform a Boolean judgment on each pair of boundaries: Where L, θ, and κ represent arc length, torsion angle, and curvature, respectively, and I(·) is an indicator function that outputs 1 when the condition is met and 0 otherwise. Φ=1 only when all sub-conditions are met simultaneously, determining that the boundary pair is sewn. If any boundary pair does not meet the criterion, i.e., Φ=0, the system will label the deviation type, such as "arc length out of tolerance" or "torsion angle conflict," and indicate the corresponding garment category and specific location. For example, if the arc length of the left sleeve cap is longer than the front armhole due to an out-of-tolerance error, the control point position will be adjusted, and the corresponding component will be enlarged or reduced; if the torsion angle is out of tolerance due to a deviation in the tangent direction of the sleeve cap endpoint, the control point position will be adjusted. The direction of the control points at the beginning and end of the curve, i.e., adjusting the tangent vector of the control points at the endpoints, is input as a feedback signal to the S2 module, triggering structured modifications and local reconstruction of the design semantic graph. This includes adjusting the relaxation amount of related components, modifying the direction of dividing lines, or adding or removing darts. After reconstruction, the S3-S4 process is re-executed to generate new cut pieces and perform further verification. To avoid infinite loops, the system sets the maximum number of iterations to 2 to 10, preferably 5. Simultaneously, the stitch consistency loss is calculated in each iteration. To monitor the convergence status: ;in, For the seam edge collection, Let λ be the average side length, λ be the curvature weighting coefficient, and arc length be... From the parameter curve Through points Calculation. If two consecutive iterations If the decrease is less than the preset convergence threshold, preferably set to 0.01, the iteration is terminated early, and the current optimal cut piece is output. Finally, when all boundary pairs satisfy Φ=1, or the iteration upper limit / convergence condition is reached, the system exits the loop and enters the output stage.

[0045] S6. Output a two-dimensional cut piece that meets the sewing requirements. Before outputting the final two-dimensional cut piece, a three-dimensional physical simulation of the cut piece is performed. If fabric penetration or local stress exceeds the fabric yield strength occurs during the simulation, the simulation failure area is fed back as an additional deviation to the design semantic graph reconstruction step.

[0046] Specifically, the system will use the current version of procedural 2D cropping (including...) The boundary, seam interface declaration, and fabric grain direction information are input into a high-precision garment physics simulation engine. This engine constructs a digital fabric model based on user-specified fabric physical parameters, including but not limited to elastic modulus, shear stiffness, Poisson's ratio, yield strength, and thickness. It then automatically and virtually sews the fabric pieces together at the seam interfaces to form a complete three-dimensional garment mesh. Subsequently, the system performs dynamic wearing simulation on a standard human body model, simulating the drape under gravity and typical movements such as raising arms and bending over. During this process, two key failure modes are monitored in real time: fabric penetration: when two non-sewn areas of the fabric pieces illegally overlap in three-dimensional space, it indicates a structural conflict in the two-dimensional unfolding; local stress exceeding limits: when the equivalent stress in a certain area exceeds the yield strength threshold of the fabric, it indicates that tearing, permanent deformation, or discomfort may occur after actual sewing. If the simulation does not detect the above problems, the fabric pieces are deemed to fully meet engineering and wearing requirements. The system outputs the final two-dimensional fabric piece file in standard formats such as DXF and custom JSON for direct use by automatic cutting beds and intelligent sewing equipment. If the simulation detects failure areas such as underarm penetration or shoulder stress concentration, the system will locate the corresponding 2D cut piece position and mark the area as an additional deviation feedback signal, which, along with the failure type, will be transmitted to the S2 module. The S2 module will then reconstruct the design semantic map accordingly, for example, by adding inserts in the penetration area, introducing dart transfer in high-stress areas, or adjusting the relaxation amount distribution. The S3-S5 process will then be retried until the 3D simulation also passes verification.

[0047] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for generating multimodal sewing patterns based on a large garment model, characterized in that, Includes the following steps: S1. Receive one or more multimodal inputs, the multimodal inputs including text, images, or sketches; S2. Based on the multimodal input, a design semantic graph representing clothing components and their stitching relationships is constructed through a large clothing model, and an intent vector is generated. The design semantic graph is used to represent the shape of clothing components and their stitching relationships. S3. Based on the intent vector, generate a procedural two-dimensional pattern piece containing parameterized boundary curves and seam interface declarations, wherein the parameterized boundary curves and seam interface declarations are used to define the edge characteristics and connection methods of the garment components; S4. For the boundary pairs to be sewn in the programmed two-dimensional cut piece, calculate the geometric properties of the boundary pairs, including arc length, twist angle and curvature; S5. Determine the dynamic geometric tolerance threshold based on the clothing category and fabric attributes, and determine whether the geometric attributes of the boundary pair meet the sewing criterion; if not, mark the deviation and reconstruct the design semantic graph and two-dimensional pattern, repeat the process until the boundary pair meets the conditions or the maximum number of iterations is reached. S6. Output two-dimensional cut pieces that meet sewing requirements; The sewingability criterion is defined as follows: ； in, , , These represent arc length, angle of twist, and curvature, respectively. , , These represent the arc length matching tolerance, torsion angle tolerance, and curvature difference tolerance, respectively. For indicator functions, , These represent the arc lengths of the first and second edges in the boundary pair to be stitched, respectively. , These represent the torsion angles of the two boundaries at the seam endpoints; , These represent the curvature functions of the two boundary curves as parameters, respectively. The L2 norm difference is given over the entire parameter range, and the torsion angle is the angle between the tangent vectors of the two boundaries at the stitching endpoints.

2. The method as described in claim 1, characterized in that, Step S1 further includes: after receiving the multimodal input, calculating the comprehensive quality score for the k-th modal input. The calculation formula is as follows: in, The integrity score is calculated using the following formula: This represents the number of valid terms in the k-th mode. This indicates the number of responses for the k-th mode; The validity score is calculated using the following formula: ; This represents the number of out-of-bounds terms in the k-th mode. This represents the number of answers received in the k-th modality; The consistency score is calculated using the following formula: ; The semantic representation vector representing the k-th modality. Represents the fused reference semantic vector. Here is the numerical stability constant, and its range is [range missing]. ; , , The preset weighting coefficients for quality sub-items, and satisfying the following conditions: .

3. The method as described in claim 1, characterized in that, The parameterized boundary curve adopts an nth-order parameterization. The curve represents the parametric equation as follows: ,in, Let i be the nth Bernstein basis function. Let i be the corresponding i-th control point, where i is an integer index and its value range is 0≤i≤n.

4. The method as described in claim 1, characterized in that, The dynamic geometric tolerance threshold includes arc length matching tolerance, torsion angle tolerance, and curvature difference tolerance. Arc length matching tolerance is an absolute length deviation or a relative percentage deviation. The torsion angle tolerance ranges from ±5° to ±15°. The curvature difference tolerance is calculated based on the L2 norm difference of the boundary curvature distribution. The geometric tolerance threshold is dynamically determined according to the current garment category, fabric elastic modulus, and pattern complexity, and is calibrated in three stages: prior initial value, historical samples, and online calibration.

5. The method as described in claim 1, characterized in that, Step S5 calculates the stitch consistency loss based on the geometric properties of the boundary pair. And determine whether to trigger reconstruction based on the stitch consistency loss; the stitch consistency loss Defined as: ;in, This indicates the loss of stitch consistency in the current cut pieces. Let represent the set of boundary pairs to be stitched together, where each element (i,j) represents a pair of boundaries that need to be stitched together. , Let i and j represent the arc lengths of the i-th and j-th boundaries, respectively. The normalized reference length of this boundary pair is taken as . or , This represents the weighting coefficient, with a value ranging from 0.1 to 10. , Let the curvature functions of the i-th and j-th boundaries be represented respectively. The arc length represents the L2 norm. Calculated by integrating the first derivative of the parametric curve C(t): ;in, Indicates the first The arc length of the strip, Indicates the first Parametric curves of the strip, Indicates the curve with respect to parameters The first derivative, Describing the vector norm, This represents the curve parameters.

6. The method as described in claim 1, characterized in that, Before outputting the final two-dimensional cut piece, a three-dimensional physical simulation of the cut piece is also included. If fabric penetration or local stress exceeds the yield strength of the fabric occurs during the simulation, the simulation failure area is fed back as an additional deviation to the design semantic graph reconstruction step.

7. The method as described in claim 1, characterized in that, The clothing big model adopts a multimodal architecture with Transformer as the backbone, in which text input is encoded by a language big model and image or sketch input is encoded by a visual big model. The two are fused in a cross-modal attention layer to generate the intent vector. It is trained on a dataset containing two-dimensional cut pieces and three-dimensional wearing state aligned samples, and the samples are automatically labeled with structured semantic labels by a big language model.

8. The method as described in claim 7, characterized in that, The method employs a joint loss function for optimization during the training phase. The joint loss function includes: cross-entropy loss for the text generation task and masked L1 loss for the patch geometry parameter regression task, wherein the geometry parameters include boundary control point coordinates and stitching interface identifiers.

9. The method as described in claim 1, characterized in that, The seam interface declaration includes the identity of the boundary pair to be matched, the allowed geometric tolerance range, and the seam allowance width parameter, guiding the automated cutting and sewing equipment to perform manufacturing operations.

Images