A generative AI-driven intelligent interactive system for personalized design of clothing

Abstract

This invention discloses a generative AI-driven intelligent interactive system for personalized clothing design, relating to the field of clothing design technology. The system includes: a fabric physical property database construction module, which digitizes weight, tensile modulus, flexural stiffness, and drape coefficient into feature vectors based on the AATCC standard testing method; an AI design engine module, which embeds the feature vectors as hard constraints into a generative adversarial network to control the generation of a three-dimensional basic pattern; an interactive gesture acquisition module, which converts touchscreen pressure and distribution into dynamic weights for desired stiffness and drape; and a finite element analysis and simulation module, which performs dynamic motion simulations such as walking, sitting, and raising an arm based on a shell element finite element model, calculating stress distribution, pleat density, and temporal changes in drape contour; and converts fabric physical properties into control parameters for the generative model, using dynamic drape performance as interactive feedback to achieve real-time personalized design through the synergy of clothing structure and material.

Description

Technical Field

[0001] This invention relates to the field of clothing design technology, specifically to a generative AI-driven intelligent interactive system for personalized clothing design. Background Technology

[0002] Existing generative AI-driven intelligent interactive systems for personalized clothing design still have the following drawbacks in practical use: Existing AI-powered clothing design systems are mainly divided into two categories: pattern design systems based on 2D image generation and virtual try-on systems based on 3D geometric modeling. The former focuses on the generation of visual elements, using convolutional neural networks or diffusion models to achieve style transfer and image synthesis, but the output is a planar image that cannot reflect the physical nature of clothing as a three-dimensional soft object. The latter constructs geometric models of the human body and clothing through 3D reconstruction technology, supporting static virtual try-on, but it often uses rigid or simplified elastic models to simulate fabrics, ignoring key physical parameters such as fabric anisotropy, bending stiffness, and drape characteristics.

[0003] In the field of fabric physics simulation, traditional methods are mainly based on mass-spring models or positional dynamics, which are computationally efficient but have limited accuracy, making it difficult to accurately simulate the formation and evolution of complex wrinkles. While the finite element method can accurately solve continuum mechanics problems, its computational cost is high, making it difficult to meet the needs of real-time interaction. In recent years, physical information neural networks have attempted to accelerate simulations, but their generalization ability is limited and they lack deep integration with generative design.

[0004] In existing technologies applying generative adversarial networks (GANs) to apparel design, the conditional inputs are mostly style tags or image references, and the conditional control lacks physical interpretability, leading to a disconnect between the generated results and the mechanical behavior of the actual fabric. At the user interaction level, existing systems rely on precise parameter inputs or preset template selections, failing to translate the user's intuitive "tactile" preferences into controllable design variables. This "visual-physical" disconnect results in significant differences between the design drawings and the actual garment's fold patterns and drape when worn dynamically, creating the industry pain point of "beautiful designs that are unwearable." Summary of the Invention

[0005] The purpose of this invention is to provide a generative AI-driven intelligent interactive system for personalized clothing design to solve the above-mentioned problems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a generative AI-driven intelligent interactive system for personalized clothing design, comprising: The fabric physical property database construction module is used to collect the weight, warp and weft tensile modulus, bending stiffness and drape coefficient of different fabrics, and digitize the physical quantities into feature vectors according to the AATCC standard test method. The AI ​​design engine module is used to receive sketches or style keywords input by the user, and embed the feature vectors as hard constraints into the latent space of the generative adversarial network to control the generation of the three-dimensional basic pattern of the garment. An interactive gesture acquisition module is used to acquire the user's desired softness / hardness and desired drape sensitivity input through a touch screen or pressure sensor, and convert the desired softness / hardness and desired drape sensitivity into dynamic weights. The finite element analysis simulation module is used to call a lightweight real-time simulation engine based on finite element analysis to perform virtual dynamic wearing simulation on the three-dimensional basic pattern, and calculate the stress distribution diagram, wrinkle density and temporal changes of the overall drape profile under different combinations of physical property parameters. The structure-material collaborative judgment module is used to compare the temporal changes of the stress distribution map, wrinkle density and overall overhang profile with the dynamic weight to determine whether they match. The dynamic design draft output module is used to output a three-dimensional dynamic design draft with structure-material linkage when judging the match. The three-dimensional dynamic design draft supports 360-degree rotation display and real-time adjustment of the fabric folding and rebound process under standard dynamic motion.

[0007] Furthermore, the fabric physical property database construction module includes: The fabric sample collection unit is used to collect fabric samples of different materials and weaving processes. The AATCC standard test unit is used to determine the drape coefficient of the fabric sample according to the AATCC 202 test method, the bending stiffness of the fabric sample according to the AATCC 139 test method, and the warp and weft tensile modulus of the fabric sample according to the ASTM D5035 standard. The feature vector generation unit is used to combine the drape coefficient, bending stiffness, warp and weft tensile modulus and weight into a multi-dimensional feature vector, and to establish a mapping relationship between the fabric identifier and the multi-dimensional feature vector.

[0008] Furthermore, the AI ​​design engine module includes: The input receiving unit is used to receive two-dimensional sketches or style keywords input by the user; A latent space mapping unit is used to map the feature vectors to constraint nodes in the latent space of a generative adversarial network. Hard constraint embedding unit, used to embed the constraint node as a hard constraint condition into the generator network of the generative adversarial network, wherein the hard constraint condition is used to restrict the geometric topology of the generated three-dimensional basic pattern of clothing. The pattern generation unit is used to output a three-dimensional basic pattern that conforms to the hard constraints based on the generator network.

[0009] Furthermore, the interactive gesture acquisition module includes: The touchscreen pressure acquisition unit is used to acquire the pressure value and pressure distribution area applied by the user on the touchscreen. A softness-hardness mapping unit is used to map the pressure value to a desired softness-hardness parameter; A droop sensitivity mapping unit is used to map the area and duration of the pressure distribution region into a desired droop sensitivity parameter. The dynamic weight generation unit is used to combine the desired stiffness parameter and the desired droop sensitivity parameter into a dynamic weight vector.

[0010] Furthermore, the finite element analysis simulation module includes: Mesh generation unit, used to perform finite element mesh generation on the three-dimensional basic template to generate shell element mesh; The material property assignment unit is used to assign the bending stiffness, shear modulus and areal density in the feature vector to the material property parameters of the shell element mesh. A boundary condition setting unit is used to set the kinematic boundary conditions of the human body model, wherein the kinematic boundary conditions include the motion trajectories of walking, sitting down and raising arms. The dynamic simulation calculation unit is used to calculate the stress distribution, deformation displacement and wrinkle morphology of the shell element mesh under the kinematic boundary conditions based on the explicit finite element algorithm. The time-series data extraction unit is used to extract the changes in stress distribution, deformation displacement and fold morphology over time, and generate the time-series changes in stress distribution map, fold density and overall overhang profile.

[0011] Furthermore, the finite element analysis simulation module also includes: The sag coefficient dynamic correction unit is used to adjust the bending stiffness parameters of the shell element mesh in real time according to the sag coefficient during the dynamic simulation calculation, so as to simulate the change of sag characteristics of the fabric under dynamic conditions.

[0012] Furthermore, the structure-material collaborative judgment module includes: The feature extraction unit is used to extract structure-material collaborative features from the temporal changes of the stress distribution map, wrinkle density, and overall overhang profile. The weight comparison unit is used to calculate the similarity value between the structure-material collaborative features and the dynamic weights; A threshold determination unit is used to determine whether the similarity value is greater than a preset threshold. A matching tagging unit is used to mark a matching state when the similarity value is greater than a preset threshold.

[0013] Furthermore, the dynamic design draft output module includes: The 3D rendering unit is used to render the structure-material linkage 3D dynamic design draft in real time. The view control unit is used to receive the user's 360-degree rotation command and adjust the viewing angle of the virtual camera; The dynamic playback unit is used to play the animation of the fabric wrinkle formation and rebound process under the standard dynamic motion. The parameter adjustment interface unit is used to receive parameter adjustment instructions input by the user and feed the parameter adjustment instructions back to the AI ​​design engine module and the finite element analysis simulation module to update the three-dimensional dynamic design draft in real time.

[0014] Furthermore, the standard dynamic actions include: Walking motion, used to simulate the periodic stress generated on the hem of clothing by the swinging of the lower limbs when the human body walks; The sitting motion is used to simulate the compression deformation of the fabric in the buttock area and the stretching deformation of the fabric in the waist area when a human body sits down. The arm-raising motion is used to simulate the shear deformation of the fabric in the armpit area and the draping changes of the fabric in the shoulder area when a person raises their arm.

[0015] Furthermore, the system also includes: A generative adversarial network training module is used to train the generator and discriminator of the generative adversarial network based on the feature vectors and corresponding 3D clothing pattern samples in the fabric physical property database. The loss function of the generator includes a fabric physical property reconstruction loss term, and the input of the discriminator includes the feature vectors and the 3D clothing pattern.

[0016] Compared with existing technologies, the generative AI-driven intelligent interactive system for personalized clothing design provided by this invention has the following beneficial effects: This generative AI-driven intelligent interactive system for personalized clothing design transforms abstract mechanical quantities such as weight, tensile modulus, bending stiffness, and drape coefficient into digital feature vectors by constructing a database of fabric physical properties. This enables a scientific quantitative representation of the fabric's "feel," providing precise physical constraints for AI design and eliminating the subjectivity and uncertainty of traditional reliance on experience-based judgment.

[0017] By embedding feature vectors as hard constraints into the latent space of a generative adversarial network, the physical properties of the fabric are directly controlled to generate a three-dimensional basic pattern. This ensures that the generated result conforms to the mechanical feasibility of the fabric, fundamentally avoiding design defects caused by mismatch between the fabric and the pattern, and significantly improving the success rate of the design on the first attempt.

[0018] The lightweight real-time simulation engine based on finite element analysis uses a shell element model and explicit integration algorithm to achieve second-level dynamic simulation while ensuring calculation accuracy. It supports real-time simulation of standard actions such as walking, sitting, and raising hands, allowing users to intuitively "foresee" the dynamic performance of the garment during the design stage, which greatly shortens the traditional "design-sampling-modification" iteration cycle.

[0019] The structure-material collaborative judgment mechanism establishes a closed-loop verification system for design intent by comparing the stress distribution, wrinkle density, and droop profile output from simulation with the dynamic weight of user gesture input. This ensures that the output results meet the user's aesthetic expectations and achieves a precise match between subjective preferences and objective physical laws.

[0020] The interactive gesture acquisition module maps touchscreen pressure and distribution to parameters of softness / hardness and drape sensitivity, enabling users to express their expectations for the dynamic performance of clothing in an intuitive way. This significantly lowers the barrier to entry for the system and expands the design participation capabilities of non-professional users, without requiring professional physics knowledge.

[0021] The final output structure-material linked 3D dynamic design draft supports 360-degree rotation observation and real-time parameter adjustment, providing a more flexible viewing perspective and faster modification response than physical sample garments, realizing true real-time personalized design, and providing key technical support for the digital transformation of the apparel industry. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0023] Figure 1 This is a schematic diagram of the overall system of the present invention. Detailed Implementation

[0024] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0025] Please see Figure 1 A generative AI-driven intelligent interactive system for personalized clothing design includes: The fabric physical property database construction module is used to collect the weight, warp and weft tensile modulus, bending stiffness and drape coefficient of different fabrics, and digitize the physical quantities into feature vectors according to the AATCC standard test method. The core function of this module is to establish a digital characterization system for the mechanical properties of fabrics, transforming the originally abstract "feel" into machine-readable mathematical language. Weight per unit area reflects the thickness of the fabric; warp and weft tensile modulus describes the fabric's tensile strength under warp and weft stress; a higher value indicates that the fabric is less likely to be stretched; bending stiffness characterizes the fabric's resistance to bending deformation, directly related to the crispness or softness of the garment; the drape coefficient is measured using the AATCC 202 standard test method, and its calculation formula is: In the formula, The overhang coefficient; The projected area of ​​the fabric when it hangs naturally; This represents the circular area of ​​the fabric when fully unfolded. A drape coefficient closer to 100% indicates better fabric drape; for example, silk can reach over 85%, while denim is typically below 40%. The AATCC standard testing method ensures the comparability of data measured by different laboratories through standardized clamping devices and image acquisition systems. These physical quantities are combined into a feature vector. , in For weight, For the warp tensile modulus, It is the tensile modulus in the weft direction. For bending stiffness, This represents the "transpose" in mathematics, thus providing precise physical constraints for subsequent AI generation. The technical effect of this digitization process is that it transforms traditional experience-based fabric selection into scientific decision-making based on physical data, eliminating the ambiguity of subjective judgment.

[0026] The AI ​​design engine module receives sketches or style keywords input by the user and embeds the feature vectors as hard constraints into the latent space of the generative adversarial network to control the generation of the three-dimensional basic pattern of the garment. This module is the core of the system's intelligent generation, deeply integrating the user's creative intent with the physical reality of the fabric. The Generative Adversarial Network (GAN) consists of a generator and a discriminator. The generator is responsible for generating data from random noise, while the discriminator is responsible for judging the authenticity of the data. Both improve together through adversarial training. The traditional GAN's latent space is a continuous vector space. Represents a standard Gaussian distributed random latent vector. Representing a Gaussian distribution, usually using This is represented, but this space lacks physical interpretability. This system uses fabric feature vectors... It is explicitly embedded in the latent space as a hard constraint. This represents the conditional latent vector that integrates fabric features. Based on the random latent vectors, conditional latent vectors are formed. This embedding method differs from simple conditional input. Instead, it uses a constrained node architecture design to ensure that the generator meets the physical feasibility of the fabric when generating a basic 3D pattern. For example, when a user selects a fabric with a high drape coefficient, the generator automatically adjusts the hem's wave shape to conform to the natural drape of silk; if a fabric with high bending stiffness is selected, the generated pattern will present a more structured A-line shape. AI-generated patterns are no longer "conceptual diagrams" detached from physical laws, but rather digital patterns with a synergistic structure-material approach that can be directly used in production.

[0027] The interactive gesture acquisition module is used to acquire the user's desired firmness and desired drape sensitivity input through the touch screen or pressure sensor, and convert the desired firmness and desired drape sensitivity into dynamic weights. This module bridges the gap between the user's subjective aesthetic judgment and the system's objective calculations, achieving intuitive human-computer interaction. Traditional clothing design software requires users to input precise physical parameters, which is extremely unfriendly to non-professional users. This system converts the user's tactile intentions into digital signals through the touchscreen's pressure sensing or a dedicated pressure sensor. Specifically, it measures the pressure applied by the user when pressing the screen. With desired softness and hardness The mapping relationship is as follows: In the formula, This is the hardness conversion factor. The reference offset in the desired hardness / softness; the area of ​​the pressing region. and duration With expected downsizing The mapping relationship is as follows: In the formula, This is the droop sensitivity conversion factor. This represents the baseline offset in the desired verticality. Dynamic weight vector. It was then generated, in which For hardness and softness weights, This is weighted by drape sensitivity. For example, if a user lightly touches a large area of ​​the screen, the system recognizes this as a "soft and draping" intent and generates a highly draped gesture. Value; however, if the user presses harder, it is recognized as an intent to "firm and shape," generating a high value. Value. Users do not need to understand the physical meaning of bending stiffness or sag coefficient; they can express their expectations for the dynamic performance of clothing simply through intuitive gestures, greatly reducing the professional threshold.

[0028] The finite element analysis simulation module is used to call a lightweight real-time simulation engine based on finite element analysis to perform virtual dynamic wearing simulation on a three-dimensional basic pattern, and calculate the stress distribution, wrinkle density and temporal changes of the overall drape profile under different combinations of physical property parameters. This module is the core of the system's physical verification, its function being to predict the dynamic behavior of clothing during actual wear using computational mechanics methods. Finite element analysis is a method that discretizes a continuum into a finite number of elements and approximates the overall deformation by solving for the displacements of the element nodes. Traditional clothing simulation uses a mass-spring model, which has limited accuracy and struggles to handle complex wrinkles. The shell element finite element model based on continuum mechanics considers the bending, shearing, and tensile deformation of the fabric in its constitutive relations. Specifically, for each shell element, its strain energy... Represented as: In the formula, The membrane strain vector; This is the bending strain vector; This is the membrane stiffness matrix, which is related to the tensile modulus in the warp and weft directions; This is the bending stiffness matrix, which is related to bending stiffness. The curvature vector; This is a unit domain. The lightweight real-time simulation engine, through an adaptive time-step algorithm and GPU parallel computing, reduces the time for a single dynamic simulation from hours to seconds using traditional methods. The stress distribution map shows the stress concentration in different areas of the fabric, and the wrinkle density is calculated... The quantification of the surface curvature change rate and the overall drape profile are obtained by extracting the displacement time-series data of the garment's edge nodes. For example, for the skirt hem generated from the aforementioned silk fabric, simulation shows that it produces high-frequency, low-amplitude wavy folds during walking, while denim produces low-frequency, high-amplitude creases. The advantage of this technology is that users can "foresee" the dynamic performance of the finished garment during the design phase, avoiding the problem of fabric and pattern mismatch discovered after sample garment production in the traditional process.

[0029] The structure-material collaborative judgment module is used to compare the temporal changes of stress distribution map, fold density and overall overhang profile with dynamic weights to determine whether they match. This module implements closed-loop verification of design intent, ensuring that the physical simulation results meet the user's aesthetic expectations. The comparison process first requires converting the physical quantities output by the simulation into feature vectors of the same dimension as the dynamic weights. Specifically, it extracts instantaneous stress from the stress distribution map. and peak stress Extracting average fold density from fold density time series data and wrinkle dynamic index Extracting the similarity index of the overhanging contour from the overhanging contour Structure-material co-feature vector Build as: Match Through calculation With dynamic weight vector The weighted Euclidean distance is obtained as follows: In the formula, Structure-material co-feature vector The The components of each dimension; Dynamic weight vector The Components of each dimension These are the weight coefficients for each feature. For feature dimension. When A similarity index >0.8 is considered a match. For example, if a user inputs the intention of "high drape," and the simulation displays the similarity index between the drape profile of a silk skirt and the standard drape shape, then a match is considered successful. =0.9, and the wrinkle dynamic index If it shows gentle fluctuations, then A value close to 1 is considered a match; if the system mistakenly uses the denim parameter, the simulation shows... =0.3 and Intense, then A low value triggers regeneration. This technology establishes a complete closed loop from user intent to physical implementation and intent verification, ensuring the predictability of the design results.

[0030] The dynamic design draft output module is used to output a 3D dynamic design draft with structure-material linkage when judging the match. The 3D dynamic design draft supports 360-degree rotation display and real-time adjustment of the fabric folding and rebound process under standard dynamic motion.

[0031] This module serves as the system's final presentation interface, transforming complex physical simulation results into a digital product that users can intuitively perceive and interact with. Structure-material linkage means that the geometry of the 3D model and the physical properties of the fabric are deeply bound at the data level; any adjustment in one will trigger a real-time response in the other. 360-degree rotation display is achieved through a virtual camera orbiting the model, allowing users to observe garment details from any angle. Standard dynamic movements include walking, sitting, and raising an arm, covering the main mechanical scenarios of everyday wear. Real-time adjustment of the fold formation and rebound process relies on the rapid response capability of the aforementioned finite element simulation engine. When a user adjusts the fabric stiffness parameters via a slider, the system can recalculate and render a new fold shape within 500ms. For example, if a user observes that the folds of a silk dress are too complex when walking, they can input the intention of "more crisp" via gesture, and the system will immediately increase the bending stiffness parameters, regenerating and displaying a new version with fewer folds and a simpler shape. This reduces the traditional "design-sampling-modification" cycle from weeks to minutes, achieving truly real-time personalized design.

[0032] The fabric physical property database construction module includes: The fabric sample collection unit is used to collect fabric samples of different materials and weaving processes. This unit forms the foundation of the database, ensuring the representativeness and diversity of the samples and covering the fabric family commonly used in the apparel industry. Different materials include natural fibers, chemical fibers, and their blends; different weaving techniques include plain weave, twill weave, satin weave, knitted weave, and nonwoven weave. Sample collection requires recording the fabric specifications; this metadata is linked to subsequent physical testing data for traceability and classification. For example, when collecting a "19mm silk crepe satin" sample, it's necessary to record that the warp yarns are 2 / 20 / 22D mulberry silk, the weft yarns are 2 / 20 / 22D mulberry silk, the warp density is 46 threads / cm, the weft density is 36 threads / cm, and it uses a five-end satin weave. A structured and scalable fabric knowledge base has been established, providing a rich physical material library for AI generation.

[0033] The AATCC standard test unit is used to determine the drape factor of fabric samples according to the AATCC 202 test method, the bending stiffness of fabric samples according to the AATCC 139 test method, and the warp and weft tensile modulus of fabric samples according to the ASTM D5035 standard. This unit ensures data accuracy by obtaining comparable and reliable physical quantity data through internationally standardized testing methods. The AATCC 202 drape test uses a circular specimen clamped in the center, measuring the projected area through image analysis; the AATCC 139 bending stiffness test uses the heart-shaped ring method to measure the length change of the fabric when it forms a ring under its own weight; the ASTM D5035 tensile test uses the strip method, stretching the material at a constant rate on a universal testing machine and recording the force-displacement curve and tensile modulus. Calculated by the slope of the linear segment: In the formula, For force increment; The width of the sample; For length increments; This is the initial clamping length. For example, in the above test of silk crepe satin, the drape coefficient D=87%, bending stiffness B=12mN·cm, and warp tensile modulus were obtained. =2.1 GPa, weft tensile modulus =1.8 GPa. These standardized data ensure the consistency of test results across different batches and laboratories, which is the fundamental guarantee of the reliability of the eigenvectors.

[0034] The feature vector generation unit is used to combine the drape coefficient, bending stiffness, warp and weft tensile modulus and weight into a multi-dimensional feature vector, and to establish the mapping relationship between fabric identification and multi-dimensional feature vector. This unit is crucial for data structuring, its role being to integrate disparate physical quantities into a machine learning input format. Multidimensional feature vectors Dimensions =5, each dimension corresponds to a physical attribute. The mapping relationship is stored in key-value pair form, where the key is a unique identifier for the fabric and the value is a feature vector. To eliminate the influence of dimensional differences on AI training, normalization processing is required: In the formula, For the normalized first Dimensional features; The i-th eigenvalue of the multidimensional eigenvector before normalization. This is the mean of all samples in this dimension; This represents the standard deviation. For example, the original vector of silk crepe satin. After normalization, it becomes A standardized mapping from physical entities to mathematical representations was established, enabling AI systems to "understand" the physical nature of fabrics.

[0035] The AI ​​design engine module includes: The input receiving unit is used to receive two-dimensional sketches or style keywords input by the user; This unit serves as the entry point for user intent, accommodating multiple input methods and lowering the barrier to entry. Two-dimensional sketches can be drawn by hand on a touchscreen or uploaded as image files; the system uses edge detection algorithms to extract the garment's outline. Style keywords are parsed into structural semantic tags through a natural language processing module. For example, if a user draws an outline that is narrow at the top and wide at the bottom and labels it "flowing long skirt," the system recognizes it as having the structural features of an "A-line skirt + high hem." This allows non-professional users to express their design intent intuitively, without needing to master professional CAD software.

[0036] The latent space mapping unit is used to map feature vectors to constraint nodes in the latent space of the generative adversarial network. This unit is the core technology for achieving structure-material synergy; its role is to establish the mathematical relationship between physical parameters and the generation space. Constraint nodes are control units inserted into specific layers of the generator network, and their weight matrices... With feature vectors Associated via a fully connected layer: In the formula, To constrain the hidden vector; For activation functions; This is the bias vector. This mapping is not a simple concatenation, but a learned transformation that allows changes in the feature vector to smoothly influence the geometric properties of the generated output. For example, when the sag coefficient... As the value continuously changes from 0.4 to 0.85, the resulting skirt hem circumference correspondingly increases from 0.9. Expanded to 1.3 This creates a natural, gradual transition. It enables continuous and adjustable control of the fabric's physical properties, allowing users to explore the effects of different fabrics in real time during the design process.

[0037] Hard constraint embedding unit is used to embed constraint nodes as hard constraints into the generator network of generative adversarial network. Hard constraints are used to restrict the geometric topology of the three-dimensional basic pattern of the generated clothing. This unit is crucial for ensuring the generated results meet physical feasibility requirements; its role is to transform the mechanical limits of the fabric into boundary conditions for geometric generation. The difference between hard and soft constraints is that the former must be strictly satisfied, while the latter allows for some deviation. This system ensures that the entire generation process, from low to high resolution, is physically constrained by connecting constraint nodes to each upsampling layer of the generator. Specifically, for bending stiffness... Higher fabric density and constraint nodes will limit the generated surface to Gaussian curvature. The generator generates negative value regions because high-stiffness fabrics are difficult to form complex saddle-shaped wrinkles. Geometric topological constraints are implemented through a graph neural network, ensuring that the generated mesh manifold is free of self-intersections and holes. For example, when dealing with denim with high bending stiffness, the generator automatically avoids generating sharp wrinkles at the elbows and knees, instead generating smooth, curved transitions that conform to the actual wearing shape of denim. This fundamentally eliminates generation results that "mismatch between fabric and pattern," such as outputs that violate physical laws, like "generating a flowing hem of a silk skirt from denim."

[0038] The pattern generation unit is used to output a 3D basic pattern that meets hard constraints based on the generator network; This unit serves as the execution terminal for AI design, outputting 3D digital clothing that can be directly used for subsequent simulations. The basic 3D pattern is represented in the form of a triangular or quadrilateral mesh, including vertex coordinates. ( (Number of vertices) and face index ( (The number of facets is used). The generator network employs a progressive growth strategy, gradually increasing the resolution from 4×4 to 256×256, with each resolution stage subject to constraint node control. The output pattern already contains the basic structure of clothing, but physical simulation has not yet been performed; its form is static geometry under a "standard mannequin pose." For example, given the aforementioned intention of a "flowing long dress" and the characteristics of silk fabric, the generator unit outputs a hem circumference of 1.3. A quadrilateral mesh model with a skirt length reaching the ankle and 12 pleats at the waist was created. The topology of this model was optimized for the low bending stiffness of silk, and the number of pleats was matched to the drape of the fabric. This transforms the user's vague intentions into a precise, physically feasible 3D digital prototype, laying the foundation for subsequent dynamic simulations.

[0039] The interactive gesture acquisition module includes: The touchscreen pressure acquisition unit is used to acquire the pressure value and pressure distribution area applied by the user on the touchscreen. This unit is the sensing front end for tactile interaction, its function being to capture the mechanical information when a user's finger contacts the screen. Modern capacitive touchscreens can detect both contact area and pressure, with sampling frequencies typically no lower than 60Hz. Pressure value The pressure distribution area is obtained by converting the capacitance signal through calibration curves and extracted from the touch point cloud data using image processing algorithms. For example, the pressure value is collected when a user lightly touches the screen with their index finger. =0.5N, the contact area is elliptical, with a major axis of 8mm and a minor axis of 6mm, and the contact area is... If the user presses down forcefully, Increased to 2.0N, the contact area increased due to soft tissue deformation. It transforms users' unconscious tactile behaviors into quantifiable digital signals, providing raw data for intent recognition.

[0040] The hardness-softness mapping unit is used to map pressure values ​​to desired hardness-softness parameters. ; This unit establishes a mapping relationship between mechanical inputs and physical properties, serving to interpret the physical meaning of user gestures. The mapping function employs a piecewise linear model to distinguish the correspondence between "light touch - soft" and "hard press - firm": In the formula, The slope of the light-touch segment; The slope of the pressed segment; Pressure threshold; This represents the maximum softness value. This is a medium hardness value. For example, When =0.5N, =0.9; When =2.0N, =0.2. Users can intuitively control the force to express their subjective preference for the feel of the clothing without needing to know the technical parameters of the bending stiffness.

[0041] The droop sensitivity mapping unit is used to map the area and duration of the pressure distribution region to the desired droop sensitivity parameter. This unit captures the spatiotemporal characteristics of user gestures, its function being to distinguish the different aesthetic intentions expressed by "instantaneous taps" and "continuous stroking." (Expected vertices) With contact area Positive correlation with duration Negative correlation: In the formula, This is the area weighting coefficient; This is the time decay factor. For example, when a user presses and holds their finger over a large area of ​​the screen, =0.85; while rapid tapping, even over a large area, It also dropped to 0.3. This enriched the semantic dimension of the interaction, enabling users to precisely express subtle preferences for the dynamic performance of clothing through spatiotemporal combinations of gestures.

[0042] The dynamic weight generation unit is used to combine the desired stiffness and drape parameters into a dynamic weight vector.

[0043] This unit integrates multi-dimensional interactive input, its purpose being to form a unified, computable representation of intent. Dynamic weight vector. The generation of this parameter needs to consider the coupling relationship between two parameters, because stiffness and drape sensitivity are not independent in reality. A covariance matrix is ​​used. Describe its relevance: In the formula, For the Cholesky decomposition of the lower triangular matrix, satisfying For example, user input =0.9、 =0.1, the system identified it as a contradictory input, and corrected it using covariance constraints. =0.9、 =0.8, and prompts the user that "extremely soft fabrics are usually accompanied by good drape." Ensure the physical rationality of the user's intention and avoid generating design solutions that do not conform to the laws of materials science.

[0044] The finite element analysis simulation module includes: Mesh generation unit, used to perform finite element mesh generation on the 3D basic template to generate shell element mesh; This element is a preprocessing step in physical simulation, its function being to discretize the continuous geometric model into computable mechanical elements. Shell elements are a specialized element type for simulating thin-walled structures, with each node having 5 degrees of freedom. Meshing employs an adaptive algorithm, automatically refining the mesh in regions of high curvature and sparsening it in flat regions to balance computational accuracy and efficiency. Element size... With fabric thickness Relationship satisfaction / >5. To avoid shear self-locking. For example, for fabric thickness. For silk fabric with a thickness of 0.2 mm, the unit size was set to 1 mm in the pleated area and 5 mm in the flat area of ​​the garment, with the total number of units controlled within 100,000 to ensure real-time calculation. This provides a high-quality discrete model for subsequent mechanical calculations, ensuring both the capture of pleat details and meeting real-time requirements.

[0045] Material property assignment unit, used to assign the bending stiffness, shear modulus and areal density in the feature vector to the material property parameters of the shell element mesh; This unit realizes the physicalization of "digital fabrics," its role being to transform abstract parameters from the database into material constitutive models recognizable by the simulation model. Bending stiffness The bending stiffness matrix is ​​directly assigned to the shell element; shear modulus. Describes a fabric's ability to resist shear deformation, and is related to its warp and weft tensile moduli through anisotropic constitutive relations; areal density. Used to calculate the inertia term. For orthotropic fabrics, the constitutive matrix... for: In the formula, , It is the tensile modulus in the warp and weft directions; , Poisson's ratio; This is the in-plane shear modulus. For example, the shear modulus of silk crepe satin. =2.1 GPa, =1.8GPa, =0.3GPa, =0.15, after assignment, the shell unit possesses the mechanical "identity" of silk. This ensures that the simulation model's physical response to specific fabrics is realistic and credible, as different fabrics will exhibit drastically different mechanical behaviors under the same pattern.

[0046] The boundary condition setting unit is used to set the kinematic boundary conditions of the human body model. The kinematic boundary conditions include the motion trajectories of walking, sitting, and raising arms. This unit defines the simulation's "stage" and "motions," serving to simulate the mechanical environment of clothing during actual wear. The human body model employs a skeleton-skin structure, with the skeleton defining joint movements and the skin acting as the contact body for the clothing. Kinematic boundary conditions are driven by keyframe interpolation or motion capture data, defining the angles of each joint. Changes over time. For example, walking is defined as the angle of hip flexion and extension: In the formula, This refers to the hip flexion-extension angle. The duration is 1.5 Hz, and the step frequency is 1.5 Hz. The knee and ankle joints move in unison, following the cyclical pattern of human gait. The sitting motion is defined as hip flexion from 0° to 90° while the knee flexes to 90°, lasting 2 seconds. The arm-raising motion is defined as shoulder abduction from 0° to 180° while the elbow remains slightly flexed, lasting 1.5 seconds. These boundary conditions are applied to the contact surface between the clothing and the body using the Lagrange multiplier method to simulate the constraint relationship during wear. This ensures the simulation scenario closely resembles real-world wearing conditions and guarantees the practical reference value of the verification results.

[0047] The dynamic simulation calculation unit is used to calculate the stress distribution, deformation displacement and wrinkle morphology of shell element mesh under kinematic boundary conditions based on the explicit finite element algorithm. This unit is the computational core of the physical simulation, and its function is to solve the mechanical response of the fabric under dynamic loads. The explicit finite element algorithm uses the central difference method for time integration, and its recursive formula is as follows: In the formula, It is the displacement vector; It is the velocity vector; This is the quality matrix; This is the vector of external forces; This is the internal force vector; This is the index of the current time step. For a given time step, the CFL condition must be met to ensure stability: In the formula, Minimum unit size; For material wave velocity, ,in For elastic modulus, Density. Stress distribution is calculated using the strain-stress relationship at the element Gaussian point, deformation displacement is the nodal displacement vector, and wrinkle morphology is extracted through curvature analysis in post-processing. For example, in a walking simulation of a silk skirt, the hem node reaches its maximum lateral displacement at t=0.5s. =15cm, corresponding stress =5MPa, curvature analysis shows that a wavelength is formed at the edge of the skirt. =8cm periodic pleats. Through precise physical numerical calculations, the dynamic aesthetic performance and mechanical safety of the garment are predicted.

[0048] The time-series data extraction unit is used to extract the changes in stress distribution, deformation displacement, and fold morphology over time, and generate stress distribution maps, fold density, and time-series changes in the overall overhang profile.

[0049] This unit structures the simulation results, providing standardized input data for subsequent decision-making modules. Time-series data is extracted at a fixed sampling rate to form a time-series matrix. The stress distribution map represents the scalar stress field. Visualization of the garment surface, where x, y, z are three-dimensional spatial coordinates. For time intervals, color gradations are typically used, with red representing high-stress areas and blue representing low-stress areas; wrinkle density Defined as: In the formula, for The number of peak wrinkles detected at any given time; The surface area of ​​the garment. The overall drape profile is determined by extracting the coordinate set of the nodes at the hem of the garment. It means that, among them, Representing the Each contour node at time... coordinates Represents the total number of contour nodes, calculate its The convex hull area of ​​the overall overhang profile at any given time and the position of the center of mass This is used to quantify changes in the drape and shape of a silk skirt. For example, during a walking cycle, At 0.8 Up to 1.2 Fluctuations between, among which, The area enclosed by the static silhouette of the garment and the lateral swing amplitude of the center of mass within ±5cm form the basis of the structure-material collaborative feature vector. This transforms complex continuous simulation results into discrete, comparable data features, facilitating automated matching and judgment.

[0050] The finite element analysis simulation module also includes: The sag coefficient dynamic correction element is used to adjust the bending stiffness parameters of the shell element mesh in real time according to the sag coefficient during dynamic simulation calculations, so as to simulate the changes in the sag characteristics of the fabric under dynamic conditions. This unit represents a key improvement to the standard finite element model, addressing the discrepancy between the static drape coefficient and dynamic drape behavior. The standard drape coefficient is measured under static conditions, but fabrics exhibit "dynamic drape" during dynamic wear, deviating from the static value. This is resolved by introducing a dynamic correction factor. Adjust bending stiffness in real time : In the formula, Static bending stiffness; This is a dynamic correction factor; For reference overhang coefficient; This refers to the angular velocity of the fabric's movement. For fabrics with a high drape coefficient, the effective bending stiffness decreases during dynamic movement, resulting in a more flowing motion; for fabrics with a low drape coefficient, the effective bending stiffness increases, leading to a stiffer movement. For example, in walking simulations, the angular velocity of a silk skirt's hem... =2rad / s, after correction =0.8 The skirt's sway amplitude increased by 20% compared to static predictions, which is more consistent with actual observation. This significantly improves the realism of dynamic simulation and compensates for the shortcomings of traditional static parameters in dynamic scenes.

[0051] The structure-material collaborative judgment module includes: The feature extraction unit is used to extract structure-material co-features from the temporal changes of stress distribution map, wrinkle density and overall overhang profile; This unit is crucial for data dimensionality reduction and feature engineering, its role being to extract key indicators directly related to user experience from massive simulation data. Structure-material synergistic features are comprehensive indicators describing the coupling relationship between "fabric properties, pattern structure, and dynamic performance," including: 1. Stress uniformity index ,in, For stress standard deviation, The value represents the average stress; the closer the value is to 1, the more uniform the stress distribution and the more comfortable the wear. 2. Wrinkle Dynamic Index ,in, For time period, The density of folds produced in an extremely short time. The rate of change of dynamic fold density with respect to time. This represents an extremely small, instantaneous change, indicating the degree to which the wrinkles change over time; the larger the value, the stronger the dynamic effect. 3. Suspension stability index ,in, This is the instantaneous ratio of the dynamic area to the static area. The closer the value is to 1, the closer the dynamic overhang shape is to the static ideal shape.

[0052] For example, for a silk A-line skirt, the extracted... =0.85, =0.42, =0.95. These features comprehensively reflect the degree of matching between the fabric and the pattern, serving as the basis for subsequent comparisons. The effect of this technology lies in transforming complex physical simulation results into interpretable and quantifiable aesthetic indicators, establishing a bridge between engineering parameters and aesthetic experience.

[0053] The weighted comparison unit is used to calculate the similarity value between the structure-material co-features and the dynamic weights; This unit enables a quantitative comparison between user intent and simulation results, serving to objectively evaluate whether the design meets user expectations. This is due to the structure-material co-feature vector... With dynamic weights Different dimensions require projection matrices. Will Mapped to Space: Similarity value Calculation using cosine similarity: The value range is [-1, 1], which is normalized to [0, 1] in practical applications. For example, if a user expects a silk skirt to be "soft and drapey," then the projected characteristics of the skirt are considered. Calculated =0.98, highly matched; if denim is mistakenly used, , =0.35, mismatch. It provides an objective and repeatable evaluation standard, eliminating the ambiguity of subjective judgment.

[0054] The threshold judgment unit is used to determine whether the similarity value is greater than a preset threshold. This unit executes the decision-making logic, its role being to make binary matching judgments based on similarity values. Preset threshold. It is usually set to 0.8, but can be adaptively adjusted based on user feedback data. The judgment logic is as follows: when When the value is within the fuzzy range of [0.7, 0.8), the system can trigger a "suggest fine-tuning" prompt instead of directly determining failure. For example, a silk dress... =0.98>0.8, therefore a match is found; denim skirt =0.35 < 0.8, which is considered a mismatch, triggering a regeneration process. Clear pass / fail criteria have been established to ensure that only designs that meet user intent can proceed to the final output stage.

[0055] The matching tagging unit is used to mark a matching state when the similarity value is greater than a preset threshold.

[0056] This unit serves as a status indicator for process control, transmitting verification results to subsequent modules. Matching status is stored as a Boolean or enumerated value in the design draft metadata, serving as the basis for version control and quality traceability. When marked as MATCHED, the design draft enters the output module; when marked as UNMATCHED, the system records the reason for the mismatch and feeds it back to the AI ​​design engine module to guide regeneration; when marked as UNCERTAIN, it enters the manual review queue. For example, the metadata for a silk dress design draft includes: Status=MATCHED, Similarity=0.98, Features=[0.85, 0.92, 0.95, 0.42, 0.87], Timestamp=2024-01-15T09:30:00Z. This achieves automated quality control of the design process, ensuring consistency and traceability of the output.

[0057] The dynamic design draft output module includes: The 3D rendering unit is used for real-time rendering of 3D dynamic design drafts that link structure and material; This unit forms the technological foundation for visual presentation, transforming geometric models and physical simulation results into realistic visual images. Real-time rendering employs a physically-based rendering pipeline, considering the fabric's reflective and transmissive properties, as well as surface details. For silk fabrics, low roughness, high transmittance, and anisotropic reflection are set during rendering to simulate its sheen; for denim, high roughness, opacity, and fabric texture mapping are used. The rendering frame rate is no less than 30fps to ensure smooth dynamics. For example, when the viewpoint of a silk dress rotates, the sheen of the hem changes with the viewing angle, creating a "silky" effect, while the folds exhibit alternating light and shadow due to changes in curvature. The effect of this technology is to provide a visual effect close to that of realistic photography, allowing users to accurately predict the appearance and texture of the finished garment.

[0058] The view control unit is used to receive the user's 360-degree rotation command and adjust the viewing angle of the virtual camera; This unit enables interactive viewing, allowing users to examine the design from any angle. The virtual camera uses a spherical coordinate system. control, The distance from the camera to the center of the model. It is the azimuth angle. This is the polar angle. User input via mouse dragging or touchscreen gestures is mapped to coordinate changes. In the formula, Sensitivity coefficient for horizontal viewing angle control This is the sensitivity coefficient for vertical viewing angle control. This refers to the amount of horizontal displacement of the user's gesture or mouse cursor on the screen. This represents the vertical displacement of the user's gesture or mouse cursor on the screen. Camera movement uses smooth interpolation to avoid abrupt changes. For example, dragging from left to right rotates the viewpoint horizontally around the model, allowing the user to observe the fit of the side seams; dragging from bottom to top raises the viewpoint, allowing the user to observe the drape of the neckline. This provides a more flexible observation method than physical sample garments, allowing users to easily view back and side details that are difficult to see when wearing the garment.

[0059] The dynamic playback unit is used to play animations of fabric wrinkling and rebound processes under standard dynamic motion. This section showcases the dynamic aesthetics of the design, aiming to present the realistic performance of clothing in motion. The animation playback is based on time-series data calculated using finite element simulation, with the frame rate consistent with the simulation sampling rate and adjustable playback speed. Wrinkle formation and rebound are manifestations of the fabric's viscoelasticity; silk, due to its low elastic modulus, rebounds slowly, while denim, due to its high elastic modulus, rebounds quickly. During playback, users can pause and view frame by frame to observe the wrinkle patterns at specific moments. For example, during a sitting motion, the waist fabric of a silk skirt forms radial wrinkles at t=0.8 s; when standing up at t=2.0 s, the wrinkles gradually flatten, but some memory wrinkles remain—this process is fully recorded and displayed. This allows users to evaluate the aesthetics and practicality of the clothing in real-world wearing scenarios, such as whether the skirt will produce persistent wrinkles when standing up after sitting down.

[0060] The parameter adjustment interface unit is used to receive parameter adjustment commands input by the user and feed them back to the AI ​​design engine module and the finite element analysis simulation module to update the 3D dynamic design draft in real time.

[0061] This unit enables closed-loop optimization of the design, supporting users in iterative improvements based on observations. Parameter adjustment commands include: fabric change, pattern fine-tuning, and gesture re-input. The interface uses intuitive controls such as sliders and knobs, with numerical changes transmitted to the backend in real time. For example, if a user observes that the pleats of a silk dress are too intricate, they can use the slider to reduce the "desired stiffness" from 0.9 to 0.6. This command is fed back to the AI ​​design engine, triggering a regeneration. The new pattern is then verified again through finite element simulation, and the entire cycle is completed within 10 seconds. This reduces the traditional "design-sampling-modification" cycle in clothing design from weeks to seconds, achieving truly real-time interactive design.

[0062] Standard dynamic movements include: Walking motion, used to simulate the periodic stress generated on the hem of clothing by the swinging of the lower limbs when the human body walks; This action serves as a fundamental scenario for evaluating the comfort and aesthetics of everyday wear, examining the dynamic response of clothing under cyclic loads. The biomechanical characteristics of walking include: cadence 1.2-1.8 Hz, stride length 0.5-0.8 m, hip flexion / extension angle ±30°, and knee flexion / extension angle ±40°. For skirts, the cyclic stress on the hem primarily originates from air resistance and inertial forces caused by leg movement; the stress amplitude is related to fabric density and swing speed. In the formula, The amplitude of the periodic stress experienced by the hem of the garment. areal density The speed of leg swing. Silk skirts because... Low stress amplitude, flowing and graceful; heavy woolen skirts due to High, with large stress amplitude, and cumbersome swing. This movement reveals the impact of fabric inertia on dynamic aesthetics, helping users choose fabric-pattern combinations suitable for everyday walking.

[0063] The sitting motion is used to simulate the compression deformation of the fabric in the buttock area and the stretching deformation of the fabric in the waist area when a human body sits down. This action is a key scenario for evaluating the rationality and comfort of clothing structure, its purpose being to examine the stress distribution and wrinkle formation of the garment under significant deformation. The sitting action involves 90° hip flexion, 90° knee flexion, compression from the buttocks contacting the seat, and stretching from the lumbar spine leaning backward. The strain in the compression area... Up to -30% strain in the tensile region Up to +15%. High-elasticity fabrics absorb strain through deformation, while low-elasticity fabrics release strain through wrinkles. For example, when sitting, silk trousers form radial wrinkles at the hips and fine horizontal wrinkles at the waist. After standing up, the hip wrinkles quickly return to their original shape, leaving only slight wrinkles at the waist. The technical advantage of this technique is that it allows for prediction of the garment's appearance and comfort in a sitting position, avoiding the selection of fabrics that are prone to permanent wrinkling at the hips.

[0064] The arm-raising motion is used to simulate the shear deformation of the fabric in the armpit area and the draping changes of the fabric in the shoulder area when a person raises their arm.

[0065] This action is a crucial scenario for evaluating the garment's cut and armhole design, serving to test the fit and freedom of movement during upper limb activity. The arm-raising motion involves 180° shoulder abduction, slight elbow flexion, and shear deformation of the underarm fabric, causing the shoulder fabric to shift from a natural drape to horizontal stretching. Shear strain. With armhole depth Height of the sleeve mountain Related: An armhole that is too shallow or a sleeve cap that is too low will result in Excessive shear modulus can cause a "hanging" effect when the fabric's shear stiffness is insufficient. For example, a silk shirt with a low shear modulus will exhibit noticeable diagonal wrinkles under the armpits when the arms are raised, causing the shoulder fabric to slip and the garment to rise by 3cm. In contrast, a similar shirt with lining, where the shear modulus is increased to G=1.0 GPa, will only experience a 1cm rise. This adjustment optimizes the design parameters of the armhole and sleeve cap, ensuring a proper fit for functional activities.

[0066] The system also includes: The Generative Adversarial Network (GAN) training module is used to train the generator and discriminator of the GAN based on the feature vectors and corresponding 3D clothing pattern samples in the fabric physical property database. The generator's loss function includes a fabric physical property reconstruction loss term, and the discriminator's input includes the feature vectors and 3D clothing patterns.

[0067] This module forms the foundation for the system's intelligent capabilities. Its function is to learn the mapping relationship between fabric properties and pattern structure through a data-driven approach. Training data includes: input as feature vectors. and random noise The output is a 3D clothing pattern mesh. Generator loss function It consists of three parts: In the formula, Weighting coefficients to counteract losses, To combat the loss, the Wasserstein distance is used to measure the difference in distribution between the generated pattern and the real pattern. The weighting coefficients for reconstruction loss, To assess the reconstruction loss, the geometric difference between the generated pattern and the target pattern is measured using the Chamfer distance; Weighting coefficients for the loss in fabric physical property reconstruction. The formula for calculating the fabric physical property reconstruction loss is as follows: In the formula, The feature vector of the actual (or target) physical properties of the fabric. The generated pattern is input into a pre-trained physical property prediction network to obtain a predicted feature vector. This loss term ensures that the generator not only generates patterns that "look like" patterns, but also patterns that are "physically correct." Discriminator The input is Yes, the output is a truth score, and its loss function is... This measures the ability to distinguish between real and generated data pairs. For example, when training with silk feature vectors as input, the generator must output a pattern with a large hem circumference and many pleats; if it outputs a straight-cut pattern, This will significantly increase, forcing the generator to adjust. This imbues the generative adversarial network with physical common sense, ensuring that the generated results conform to the mechanical laws of the fabric from the outset, thus greatly reducing the failure rate of subsequent simulation verification.

[0068] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A generative AI-driven intelligent interactive system for personalized clothing design, characterized in that, include: The fabric physical property database construction module is used to collect the weight, warp and weft tensile modulus, bending stiffness and drape coefficient of different fabrics as physical quantities, and digitize the physical quantities into feature vectors according to the AATCC standard test method. The AI ​​design engine module is used to receive sketches or style keywords input by the user, and embed the feature vectors as hard constraints into the latent space of the generative adversarial network to control the generation of the three-dimensional basic pattern of the garment. An interactive gesture acquisition module is used to acquire the user's desired softness / hardness and desired drape sensitivity input through a touch screen or pressure sensor, and convert the desired softness / hardness and desired drape sensitivity into dynamic weights. The finite element analysis simulation module is used to call a lightweight real-time simulation engine based on finite element analysis to perform virtual dynamic wearing simulation on the three-dimensional basic pattern, and calculate the stress distribution diagram, wrinkle density and temporal changes of the overall drape profile under different combinations of physical property parameters. The structure-material collaborative judgment module is used to compare the temporal changes of the stress distribution map, wrinkle density and overall overhang profile with the dynamic weight to determine whether they match. The dynamic design draft output module is used to output a three-dimensional dynamic design draft with structure-material linkage when judging the match. The three-dimensional dynamic design draft supports 360-degree rotation display and real-time adjustment of the fabric folding and rebound process under standard dynamic motion.

2. The generative AI-driven intelligent interactive system for personalized clothing design according to claim 1, characterized in that, The fabric physical property database construction module includes: The fabric sample collection unit is used to collect fabric samples of different materials and weaving processes. The AATCC standard test unit is used to determine the drape coefficient of the fabric sample according to the AATCC 202 test method, the bending stiffness of the fabric sample according to the AATCC 139 test method, and the warp and weft tensile modulus of the fabric sample according to the ASTM D5035 standard. The feature vector generation unit is used to combine the drape coefficient, bending stiffness, warp and weft tensile modulus and weight into a multi-dimensional feature vector, and to establish a mapping relationship between the fabric identifier and the multi-dimensional feature vector.

3. The generative AI-driven intelligent interactive system for personalized clothing design according to claim 1, characterized in that, The AI ​​design engine module includes: The input receiving unit is used to receive two-dimensional sketches or style keywords input by the user; A latent space mapping unit is used to map the feature vectors to constraint nodes in the latent space of a generative adversarial network. Hard constraint embedding unit, used to embed the constraint node as a hard constraint condition into the generator network of the generative adversarial network, wherein the hard constraint condition is used to restrict the geometric topology of the generated three-dimensional basic pattern of clothing. The pattern generation unit is used to output a three-dimensional basic pattern that conforms to the hard constraints based on the generator network.

4. The generative AI-driven intelligent interactive system for personalized clothing design according to claim 1, characterized in that, The interactive gesture acquisition module includes: The touchscreen pressure acquisition unit is used to acquire the pressure value and pressure distribution area applied by the user on the touchscreen. A softness-hardness mapping unit is used to map the pressure value to a desired softness-hardness parameter; A droop sensitivity mapping unit is used to map the area and duration of the pressure distribution region into a desired droop sensitivity parameter. The dynamic weight generation unit is used to combine the desired stiffness parameter and the desired droop sensitivity parameter into a dynamic weight vector.

5. The generative AI-driven intelligent interactive system for personalized clothing design according to claim 1, characterized in that, The finite element analysis simulation module includes: Mesh generation unit, used to perform finite element mesh generation on the three-dimensional basic template to generate shell element mesh; The material property assignment unit is used to assign the bending stiffness, shear modulus and areal density in the feature vector to the material property parameters of the shell element mesh. A boundary condition setting unit is used to set the kinematic boundary conditions of the human body model, wherein the kinematic boundary conditions include the motion trajectories of walking, sitting down and raising arms. The dynamic simulation calculation unit is used to calculate the stress distribution, deformation displacement and wrinkle morphology of the shell element mesh under the kinematic boundary conditions based on the explicit finite element algorithm. The time-series data extraction unit is used to extract the changes in stress distribution, deformation displacement and fold morphology over time, and generate the time-series changes in stress distribution map, fold density and overall overhang profile.

6. The generative AI-driven intelligent interactive system for personalized clothing design according to claim 5, characterized in that, The finite element analysis simulation module also includes: The sag coefficient dynamic correction unit is used to adjust the bending stiffness parameters of the shell element mesh in real time according to the sag coefficient during the dynamic simulation calculation, so as to simulate the change of sag characteristics of the fabric under dynamic conditions.

7. The generative AI-driven intelligent interactive system for personalized clothing design according to claim 1, characterized in that, The structure-material collaborative judgment module includes: The feature extraction unit is used to extract structure-material collaborative features from the temporal changes of the stress distribution map, wrinkle density, and overall overhang profile. The weight comparison unit is used to calculate the similarity value between the structure-material collaborative features and the dynamic weights; A threshold determination unit is used to determine whether the similarity value is greater than a preset threshold. A matching tagging unit is used to mark a matching state when the similarity value is greater than a preset threshold.

8. The generative AI-driven intelligent interactive system for personalized clothing design according to claim 1, characterized in that, The dynamic design draft output module includes: The 3D rendering unit is used to render the structure-material linkage 3D dynamic design draft in real time. The view control unit is used to receive the user's 360-degree rotation command and adjust the viewing angle of the virtual camera; The dynamic playback unit is used to play the animation of the fabric wrinkle formation and rebound process under the standard dynamic motion. The parameter adjustment interface unit is used to receive parameter adjustment instructions input by the user and feed the parameter adjustment instructions back to the AI ​​design engine module and the finite element analysis simulation module to update the three-dimensional dynamic design draft in real time.

9. The generative AI-driven intelligent interactive system for personalized clothing design according to claim 1, characterized in that, The standard dynamic actions include: Walking motion, used to simulate the periodic stress generated on the hem of clothing by the swinging of the lower limbs when the human body walks; The sitting motion is used to simulate the compression deformation of the fabric in the buttock area and the stretching deformation of the fabric in the waist area when a human body sits down. The arm-raising motion is used to simulate the shear deformation of the fabric in the armpit area and the draping changes of the fabric in the shoulder area when a person raises their arm.

10. The system according to any one of claims 1 to 9, characterized in that, Also includes: A generative adversarial network training module is used to train the generator and discriminator of the generative adversarial network based on the feature vectors and corresponding 3D clothing pattern samples in the fabric physical property database. The loss function of the generator includes a fabric physical property reconstruction loss term, and the input of the discriminator includes the feature vectors and the 3D clothing pattern.

Images