Method and device for automatically generating human body model based on GPT large model

Abstract

The invention discloses a method and device for automatically generating a human body model based on a GPT large model, and the method comprises the steps: analyzing the natural language input of a user through the GPT large model, and extracting human body form parameters, posture parameters and expression parameters; carrying out human body proportion correction, motion correction and expression optimization processing on the extracted parameters to generate three-dimensional grid vertex coordinates, skeleton joint data and facial expression weights; a three-dimensional human body grid model is constructed based on the data, skeleton joints are bound, expressions are driven to change, and finally an interactive human body model is generated through material rendering. According to the method, end-to-end automatic generation from natural language description to a high-precision three-dimensional human body model is realized, the problems of complex process and cross-modal conversion distortion of a traditional modeling process are solved, and the modeling efficiency and precision in the fields of virtual reality, biomechanical simulation and the like are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of human body model generation, and in particular to a method and device for automatically generating a human body model based on a GPT large model. Background Art

[0002] In recent years, large language models such as Generative Pre-trained Transformers (GPT) have made significant progress in the field of natural language processing. However, applying GPT models to human body modeling, especially the automatic generation of 3D human body models, is still in the exploratory stage.

[0003] Currently, human body modeling primarily relies on deep learning techniques such as variational autoencoders (VAEs), generative adversarial networks (GANs), and diffusion models. These methods have achieved success in image and video generation, but remain challenging in generating human motion and posture. For example, a 2023 review of human motion generation by researchers from Peking University and Huawei noted that despite progress, further research is needed to generate natural, realistic, and diverse human motion.

[0004] In addition, researchers from Shanghai Jiao Tong University and other institutions proposed the T2M-GPT model, which combines a vector quantized variational autoencoder (VQ-VAE) with a generative pre-trained Transformer (GPT) to generate high-quality human actions from text descriptions. This approach has made significant progress in human action generation tasks, demonstrating the potential of the GPT model in this field.

[0005] In summary, current research primarily focuses on the application of large GPT models to human pose and motion, while research directly utilizing GPT models to generate complete 3D human models is still immature. Currently, some exploratory work has attempted to apply GPT models to the field of 3D modeling. For example, 3D-GPT is a 3D modeling framework that leverages large language models for guidance and aims to improve the efficiency of automatic content generation and procedural generation. 3D-GPT positions large language models as skilled problem solvers, decomposing procedural 3D modeling tasks into accessible fragments and assigning appropriate agents to each task. 3D-GPT integrates three core agents: a task scheduling agent, a conceptualization agent, and a modeling agent, which together achieve two primary goals: enhancing the initial scene description and seamlessly integrating procedural generation. However, the application of these methods to generating human models requires further research and validation.

[0006] Therefore, how to invent a method to automatically generate a human body model based on the GPT large model, which can automatically extract key information and generate a corresponding three-dimensional human body model by parsing the natural language text input by the user, has become an urgent problem to be solved. Summary of the Invention

[0007] To this end, the present invention provides a method and device for automatically generating a human body model based on a GPT large model. The GPT large model accurately parses natural language descriptions, automatically extracts human body morphology, posture, and expression parameters, and combines inverse kinematics with expression-based optimization technology to efficiently generate ergonomic three-dimensional models, significantly reducing the modeling threshold and improving generation efficiency.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for automatically generating a human body model based on a GPT large model, comprising:

[0009] Parse the natural language input by the user through the GPT large model interface to extract human body shape parameters, posture parameters and expression parameters;

[0010] The extracted human body morphological parameters, the posture parameters and the expression parameters are respectively subjected to human body proportion correction, kinematic correction and expression parameter optimization processing to obtain three-dimensional mesh vertex coordinates, bone joint data, facial key point data and expression weights;

[0011] A three-dimensional human body model is constructed according to the three-dimensional mesh vertex coordinates, topological structure and the skeletal joint data; the joints of the three-dimensional human body model are bound according to the skeletal joint data to obtain a bound three-dimensional human body model; the bound three-dimensional human body model is expression-driven based on the facial key point data and the expression weight, and the bound three-dimensional human body model is visually rendered by setting the material to generate a final human body model.

[0012] As a preferred solution for a method of automatically generating a human body model based on a GPT large model, the human body morphological parameters include: height, weight, shoulder width, waist circumference, hip circumference, leg length and arm length parameters; the posture parameters include: trunk inclination angle, limb rotation angle, joint position and spinal curvature angle parameters; the expression parameters include: mouth corner upward range, eyebrow lifting degree, eyelid closure rate and mouth opening degree parameters.

[0013] As a preferred solution of a method for automatically generating a human body model based on a GPT large model, in the process of performing human body proportion correction processing on the extracted human body morphological parameters and obtaining the three-dimensional mesh vertex coordinates:

[0014] The expression for converting the human body morphological parameters into the three-dimensional mesh vertex coordinates is:

[0015] V'=f shape (S)=V base +B shape S

[0016] Where V' is the coordinate of the three-dimensional grid vertex; Vbase is the reference vertex coordinate of the standard human body model; B shape is the morphological deformation basis matrix, which represents the influence of different morphological parameters on vertex coordinates; S is the morphological parameter;

[0017] Through the strategy of global scaling and local adjustment, the body scaling transformation is performed:

[0018] v' i =λv i +δ i

[0019] Where, v' i is the coordinate of the i-th 3D mesh vertex after the global scaling factor and the local adjustment factor; v i is the original coordinate of the i-th mesh vertex before morphological deformation in the standard human body model; λ is the global scaling factor; δ i is the local adjustment factor.

[0020] As a preferred solution of a method for automatically generating a human body model based on a GPT large model, in the process of kinematically correcting the posture parameters to obtain the skeletal joint data:

[0021] Through inverse kinematics and motion database matching strategies, joint rotation data that conforms to the laws of human motion is calculated;

[0022] The inverse kinematics equation is:

[0023] P end =T joint P base

[0024] Where, P end is the target position of the hand; T joint is the transformation matrix composed of joint rotation matrices; P base is the initial position;

[0025] The optimization objective function of the inverse kinematics equation is:

[0026] J IK =||P end -P target || 2

[0027] Where, J IK is the optimization objective function of inverse kinematics, which represents the current end effector position P end and the expected target position P target The square error between target The three-dimensional position of the hand target specified by the user or obtained after natural language parsing;

[0028] The skeletal joint data is obtained through Jacobian inverse optimization:

[0029] Δθ=J + (P target -P end )

[0030] Where Δθ is the incremental vector of each joint angle in this iteration; J + is the pseudo-inverse of the Jacobian matrix.

[0031] As a preferred solution of a method for automatically generating a human body model based on a GPT large model, in the process of performing expression parameter optimization processing on the expression parameters to obtain the facial key point data and the expression weight:

[0032] The displacement of facial key points is controlled by the expression parameters:

[0033]

[0034] Where, v face,i is the 3D coordinate of the i-th facial key point after expression deformation is applied; As the base face shape; B expr,j is the jth expression base; w j is the weight of the expression parameter; K is the number of expression bases;

[0035] The optimal expression weight is obtained by least square method:

[0036]

[0037] Where, J expr v is the objective function for optimizing expression parameters, which is used to measure the sum of square errors between the coordinates of the i-th key point generated by the model and the coordinates of the real key point; real,i is the true 3D coordinate of the i-th facial key point obtained by calibration, scanning or camera capture; L is the number of facial key points.

[0038] As a preferred solution for the method of automatically generating a human body model based on a GPT large model, a joint angle constraint is introduced during the calculation of joint rotation data to ensure that the joint rotation angle is within the physiological range; the expression of the joint angle constraint is:

[0039] θ min,i ≤θ i ≤θ max,i

[0040] Where θ i is the joint rotation angle; θ min,i is the minimum rotation angle of the joint; θmax,i is the maximum rotation angle of the joint;

[0041] In the process of calculating the three-dimensional vertex coordinates, all vertices are detected using the grid self-intersection detection strategy:

[0042]

[0043] Where, d(v i ,v j ) is the Euclidean distance between the i-th vertex and the j-th vertex; ∈ is the minimum safe distance.

[0044] The present invention also provides a device for automatically generating a human body model based on a GPT large model, comprising:

[0045] The data parameter extraction module is used to parse the natural language input by the user through the GPT large model interface and extract human body morphological parameters, posture parameters and expression parameters;

[0046] a data parameter processing module for performing body proportion correction, kinematic correction, and expression parameter optimization processing on the extracted human body morphological parameters, posture parameters, and expression parameters, respectively, to obtain three-dimensional mesh vertex coordinates, bone joint data, facial key point data, and expression weights;

[0047] A human body model generation module is used to construct a three-dimensional human body model based on the three-dimensional mesh vertex coordinates, topological structure and the skeletal joint data; bind the joints of the three-dimensional human body model according to the skeletal joint data to obtain a bound three-dimensional human body model; drive the bound three-dimensional human body model with facial expressions based on the facial key point data and the expression weights, and visually render the bound three-dimensional human body model by setting materials to generate a final human body model.

[0048] As a preferred solution for a device for automatically generating a human body model based on a GPT large model, in the data parameter extraction module, the human body morphological parameters include: height, weight, shoulder width, waist circumference, hip circumference, leg length and arm length parameters; the posture parameters include: trunk inclination angle, limb rotation angle, joint position and spinal curvature angle parameters; the expression parameters include: mouth corner upward range, eyebrow lifting degree, eyelid closure rate and mouth opening degree parameters.

[0049] As a preferred solution for an apparatus for automatically generating a human body model based on a GPT large model, in the data parameter processing module, in the process of performing human body proportion correction processing on the extracted human body morphological parameters and obtaining the three-dimensional mesh vertex coordinates:

[0050] The expression for converting the human body morphological parameters into the three-dimensional mesh vertex coordinates is:

[0051] V'=f shape (S)=V base +B shape S

[0052] Where V' is the coordinate of the three-dimensional grid vertex; V base is the reference vertex coordinate of the standard human body model; B shape is the morphological deformation basis matrix, which represents the influence of different morphological parameters on vertex coordinates; S is the morphological parameter;

[0053] Through the strategy of global scaling and local adjustment, the body scaling transformation is performed:

[0054] v' i =λv i +δ i

[0055] Where, v' i is the coordinate of the i-th 3D mesh vertex after the global scaling factor and the local adjustment factor; v i is the original coordinate of the i-th mesh vertex before morphological deformation in the standard human body model; λ is the global scaling factor; δ i is the local adjustment factor.

[0056] As a preferred solution of a device for automatically generating a human body model based on a GPT large model, in the data parameter processing module, in the process of performing kinematic correction processing on the posture parameters to obtain the skeletal joint data:

[0057] Through inverse kinematics and motion database matching strategies, joint rotation data that conforms to the laws of human motion is calculated;

[0058] The inverse kinematics equation is:

[0059] P end =T joint P base

[0060] Where, P end is the target position of the hand; T joint is the transformation matrix composed of joint rotation matrices; P base is the initial position;

[0061] The optimization objective function of the inverse kinematics equation is:

[0062] J IK =||P end -P target || 2

[0063] Where, J IKis the optimization objective function of inverse kinematics, which represents the current end effector position P end and the expected target position P target The square error between target The three-dimensional position of the hand target specified by the user or obtained after natural language parsing;

[0064] The skeletal joint data is obtained through Jacobian inverse optimization:

[0065] Δθ=J + (P target -P end )

[0066] Where Δθ is the incremental vector of each joint angle in this iteration; J + is the pseudo-inverse of the Jacobian matrix.

[0067] As a preferred solution of a device for automatically generating a human body model based on a GPT large model, in the data parameter processing module, in the process of performing expression parameter optimization processing on the expression parameters to obtain the facial key point data and the expression weight:

[0068] The displacement of facial key points is controlled by the expression parameters:

[0069]

[0070] Where, v face,i is the 3D coordinate of the i-th facial key point after expression deformation is applied; As the base face shape; B expr,j is the jth expression base; w j is the weight of the expression parameter; K is the number of expression bases;

[0071] The optimal expression weight is obtained by least square method:

[0072]

[0073] Where, J expr v is the objective function for optimizing expression parameters, which is used to measure the sum of square errors between the coordinates of the i-th key point generated by the model and the coordinates of the real key point; real,i is the true 3D coordinate of the i-th facial key point obtained by calibration, scanning or camera capture; L is the number of facial key points.

[0074] As a preferred solution for an apparatus for automatically generating a human body model based on a GPT large model, in the data parameter processing module, a joint angle constraint is introduced during the calculation of joint rotation data to ensure that the joint rotation angle is within the physiological range; the expression for the joint angle constraint is:

[0075] θ min,i ≤θ i ≤θ max,i

[0076] Where θ i is the joint rotation angle; θ min,i is the minimum rotation angle of the joint; θ max,i is the maximum rotation angle of the joint;

[0077] In the process of calculating the three-dimensional vertex coordinates, all vertices are detected using the grid self-intersection detection strategy:

[0078]

[0079] Where, d(v i ,v j ) is the Euclidean distance between the i-th vertex and the j-th vertex; ∈ is the minimum safe distance.

[0080] The present invention has the following advantages: the present invention parses the natural language input by the user through the GPT large model interface, extracts human body morphological parameters, posture parameters and expression parameters; performs human body proportion correction, kinematic correction and expression parameter optimization processing on the extracted human body morphological parameters, posture parameters and expression parameters, respectively, to obtain three-dimensional mesh vertex coordinates, bone joint data, facial key point data and expression weights; constructs a three-dimensional human body model according to the three-dimensional mesh vertex coordinates, topological structure and the bone joint data; performs joint binding on the three-dimensional human body model according to the bone joint data to obtain a bound three-dimensional human body model; performs expression driving on the bound three-dimensional human body model based on the facial key point data and the expression weights, and performs visual rendering on the bound three-dimensional human body model by setting materials to generate a final human body model. This invention pioneered the seamless integration of language parsing, parameter optimization, and 3D reconstruction. It utilizes the GPT large model to accurately extract multi-dimensional parameters such as morphology, posture, and expression implicit in natural language. By integrating anthropometric constraints, inverse kinematics algorithms, and expression-based deformation technology, it transforms abstract parameters into mesh vertex coordinates, skeletal joint data, and expression weights that conform to biomechanical laws. Ultimately, an interactive 3D human model with complete topology, natural movement, and vivid expressions is generated in real time on the modeling platform. This invention completely overturns the traditional manual modeling modeling model that relies on professional software, compressing the modeling cycle from hours to seconds. It supports the simultaneous parsing and generation of composite instructions, solving the problems of semantic loss and physical distortion in cross-modal conversion, and providing efficient, accurate, and low-threshold digital human body generation solutions for fields such as virtual reality, medical simulation, and game development. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are merely exemplary, and those skilled in the art can, without inventive effort, derive other implementation drawings based on the provided drawings.

[0082] The structures, proportions, sizes, etc. illustrated in this specification are intended solely to complement the contents disclosed herein and to facilitate understanding and reading by persons skilled in the art. They are not intended to limit the conditions under which the present invention may be implemented and therefore have no substantive technical significance. Any structural modifications, changes in proportions, or adjustments in sizes, without affecting the efficacy and objectives of the present invention, shall remain within the scope of the technical contents disclosed herein.

[0083] Figure 1 Schematic diagram of a flow chart of a method for automatically generating a human body model based on a GPT large model provided in Example 1 of the present invention;

[0084] Figure 2 This is a schematic diagram of a specific implementation process of a method for automatically generating a human body model based on a GPT large model provided in Example 1 of the present invention;

[0085] Figure 3 This is a schematic diagram of a user interaction interface in a method for automatically generating a human body model based on a GPT large model provided in Example 1 of the present invention;

[0086] Figure 4 A schematic diagram of key point binding of a three-dimensional digital human in a method for automatically generating a human body model based on a GPT large model provided in Example 1 of the present invention;

[0087] Figure 5 Schematic diagram of human body models with different shapes and expressions in a method for automatically generating a human body model based on a GPT large model provided in Example 1 of the present invention;

[0088] Figure 6 This is a schematic diagram of the architecture of an apparatus for automatically generating a human body model based on a GPT large model provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0089] The following describes the implementation of the present invention using specific embodiments. Those skilled in the art will readily understand the other advantages and benefits of the present invention from the disclosure herein. Obviously, the embodiments described are only a portion of the present invention, not all of it. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are intended to fall within the scope of protection of the present invention.

[0090] Example 1

[0091] See also Figure 1 and Figure 2 Embodiment 1 of the present invention provides a method for automatically generating a human body model based on a GPT large model, comprising the following steps:

[0092] S1. Parse the natural language input by the user through the GPT large model interface to extract human body morphological parameters, posture parameters and expression parameters;

[0093] S2, performing body proportion correction, kinematic correction and expression parameter optimization processing on the extracted human body morphological parameters, the posture parameters and the expression parameters, respectively, to obtain three-dimensional mesh vertex coordinates, bone joint data, facial key point data and expression weights;

[0094] S3. Construct a three-dimensional human body model according to the three-dimensional mesh vertex coordinates, topological structure and the skeletal joint data; perform joint binding on the three-dimensional human body model according to the skeletal joint data to obtain a bound three-dimensional human body model; perform expression driving on the bound three-dimensional human body model based on the facial key point data and the expression weight, and perform visual rendering on the bound three-dimensional human body model by setting the material to generate a final human body model.

[0095] In this embodiment, in step S1, the natural language input by the user is parsed through the GPT large model interface to extract human body morphological parameters, posture parameters and expression parameters;

[0096] Specifically, existing large-scale modeling interfaces (such as Kimi and Deepseek) process user input in natural language. Through carefully designed prompts, the model is guided to understand the text content and extract the morphological, posture, and expression parameters required for human body modeling. By establishing a structured input and output method, the accuracy, stability, and usability of parameter analysis are ensured, providing standardized input data for subsequent 3D modeling.

[0097] After a user enters a natural language description, the system first uses a large model to perform semantic parsing, identifying human features within the text and classifying them according to predefined rules. For body morphology descriptions, feature recognition and mapping methods are used to automatically extract multi-dimensional morphological parameters such as height, weight, shoulder width, waist circumference, hip circumference, leg length, and arm length. For posture descriptions, grammatical analysis and contextual reasoning are used to analyze posture parameters such as torso tilt angle, limb rotation angle, joint position, and spinal curvature angle. For facial expressions, keyword matching and sentiment analysis are used to extract expression parameters such as mouth corner lift, eyebrow lift, eyelid closure rate, and mouth opening. These morphology, posture, and expression parameters, along with other necessary fine-tuning parameters, are then fed into the subsequent parameter optimization and 3D model generation modules to ensure that the final model aligns with the user's textual intent and meets ergonomic standards. Table 1 shows the mapping relationships between some natural language expressions and multiple parameters.

[0098]

[0099]

[0100]

[0101] Table 1 Mapping relationship between natural language and multiple parameters

[0102] In this embodiment, in order to improve the parameter parsing accuracy and adaptability of the large model, a high-quality human parameter extraction dataset can be constructed through fine-tuning training for specific tasks. The dataset covers a variety of human feature descriptions, different expressions and application scenarios, and annotates the corresponding standardized human parameters. In the data preprocessing stage, the existing human body measurement database (such as ergonomic statistics, medical measurement data, relevant national standards, etc.) is integrated to construct parameter correspondences, and the parameter inference ability is optimized in combination with the deep learning algorithm to ensure that the large model can intelligently complete the key parameters when faced with incomplete or vague descriptions. For example, when the user only enters "height 175 cm, well-proportioned body", the system can automatically calculate a reasonable weight range, limb proportions and other morphological parameters to ensure that the generated human body model conforms to the actual measurement data.

[0103] Furthermore, a rule-based validation mechanism is employed to check the rationality of extracted parameters. Anthropometric data is used to establish proportional constraints, such as arm length should not exceed a specific proportional range relative to height, and leg length should conform to a specific statistical distribution. If parameter anomalies are detected, the system can automatically adjust or prompt the user to correct their input. Furthermore, a reinforcement learning optimization strategy can be incorporated to continuously adjust the parameter parsing model based on user feedback, improving the stability and usability of the parsing results.

[0104] In this embodiment, in step S2, the extracted human body morphological parameters, the posture parameters, and the expression parameters are subjected to human body proportion correction, kinematic correction, and expression parameter optimization processing, respectively, to obtain three-dimensional mesh vertex coordinates, bone joint data, facial key point data, and expression weights;

[0105] Specifically, the extracted multi-dimensional morphological parameters, posture parameters and expression parameters are further corrected and optimized, and converted into the data format required for three-dimensional modeling. The core idea of ​​this step is "mesh structure + bone joint drive": at the mesh level, the surface skin is composed of N vertices and patches, and at the bone level, the deformation of the mesh is controlled by the joint transformation matrix. Its goal is to map the high-level semantic parameters output by the first module, such as height, weight, rotation angles of each joint, facial expression weights, etc., into three-dimensional mesh vertex coordinates, mesh topology information and bone joint parameters to ensure that the generated model meets anthropometric standards and has natural motion characteristics and computational efficiency.

[0106] Among them, morphological parameter optimization and grid point coordinate generation:

[0107] Morphological parameters mainly include height, weight, shoulder width, leg length, arm length, etc. These parameters determine the basic shape of the human body. The present invention uses a morphological parameter mapping model to map these parameters to the grid vertex coordinates of the three-dimensional model, making it consistent with the human physiological structure and adaptable to different body shapes.

[0108] Convert the morphological parameters into three-dimensional vertex coordinates:

[0109] The human body mesh in the 3D human body model consists of N vertices V = {v1,v2,···,v N}, each vertex is determined by the morphological parameters. Assume that the input morphological parameters are S = {s1,s2,···,s m}, the conversion relationship can be expressed as:

[0110] V'=f shape (S)=V base +B shape S

[0111] Where V' is the coordinate of the three-dimensional grid vertex; V base is the reference vertex coordinate of the standard human body model; B shape is the morphological deformation basis matrix, which represents the influence of different morphological parameters on vertex coordinates.

[0112] The optimization goal is to minimize the morphological deformation error:

[0113]

[0114] Where, J shapeis the objective function of morphological optimization, which represents the sum of square errors between the i-th vertex and the real measurement data after the model is mapped by the morphological basis parameters; V real,i are the vertex coordinates of real human body measurement data.

[0115] Through the strategy of global scaling and local adjustment, the body scaling transformation is performed:

[0116] v' i =λv i +δ i

[0117]

[0118] δ i =B region S

[0119] Where, v' i is the coordinate of the i-th 3D mesh vertex after the global scaling factor and the local adjustment factor; v i is the original coordinate of the i-th mesh vertex before morphological deformation in the standard human body model; λ is the global scaling factor; δ i is the local adjustment factor; h user is the actual measured height of the user; h base B is the reference height corresponding to the standard human body model; region Adjust the basis matrix for local morphology

[0120] In this embodiment, the posture parameters determine the rotation angle and displacement of each joint of the human body, affecting the movement of the skeleton and mesh deformation. The inverse kinematics + motion database matching method is used to calculate the joint rotation data that conforms to the laws of human movement.

[0121] Specifically, the conversion of posture parameters to joint rotation data:

[0122] Assume that the human body consists of M joints, and the rotation matrix of each joint is R i , and its calculation formula is:

[0123] R i =f pose (p i )

[0124] Where p i Joint parameters provided by the first module, f pose is the posture mapping function;

[0125] Rotation matrix R i By quaternion q i express:

[0126] q i =[wquat ,x,y,z]

[0127] Where w quat is the scalar component of the quaternion; x, y, z are the vector components of the quaternion, corresponding to the three-dimensional projection of the rotation axis vector multiplied by the rotation angle;

[0128] The optimization goal of the joint rotation matrix is:

[0129]

[0130] Where, J pose is the objective function for posture optimization, which represents the predicted i-th joint rotation matrix R i and the i-th joint rotation matrix R in the real database real,i The sum of squared errors between real,i It is the real joint data in the human body measurement database.

[0131] In this embodiment, the posture parameters are subjected to kinematic correction processing to obtain the skeletal joint data.

[0132] Specifically, when the user enters a natural language description, such as "raise your right hand", the joint rotation angle that satisfies this action needs to be solved. The inverse kinematics equation is:

[0133] P end =T joint P base

[0134] Where, P end is the target position of the hand; T joint is the transformation matrix composed of joint rotation matrices; P base is the initial position;

[0135] The optimization objective function of the inverse kinematics equation is:

[0136] J IK =||P end -P target || 2

[0137] Where, J IK is the optimization objective function of inverse kinematics, which represents the current end effector position P end and the expected target position P target The square error between target The three-dimensional position of the hand target specified by the user or obtained after natural language parsing;

[0138] The skeletal joint data is obtained through Jacobian inverse optimization:

[0139] Δθ=J + (P target -P end )

[0140] Where Δθ is the incremental vector of each joint angle in this iteration; J + is the pseudo-inverse of the Jacobian matrix.

[0141] In this embodiment, when the skeleton joints rotate, the skeleton binding algorithm is used to update the grid point position. The present invention uses a linear skinning deformation method to calculate the deformation of the vertices controlled by the skeleton:

[0142]

[0143] Where, ω ij is the influence weight of the vertex on the jth joint; R j is the rotation matrix of joint j; t j is the translation vector of the joint.

[0144] In this embodiment, the expression parameters are subjected to expression parameter optimization processing to obtain the facial key point data and the expression weight.

[0145] Specifically, assuming that there are L key points on the face, the displacement of each key point is controlled by the expression parameter:

[0146]

[0147] Where, v face,i is the 3D coordinate of the i-th facial key point after expression deformation is applied; As the base face shape; B expr,j is the jth expression base; w j is the weight of the expression parameter; K is the number of expression bases;

[0148] The optimal expression weight is obtained by least square method:

[0149]

[0150] Where, J expr v is the objective function for optimizing expression parameters, which is used to measure the sum of square errors between the coordinates of the i-th key point generated by the model and the coordinates of the real key point; real,i is the true 3D coordinate of the i-th facial key point obtained by calibration, scanning or camera capture; L is the number of facial key points.

[0151] In this embodiment, during the calculation of joint rotation data, a joint angle constraint is introduced to keep the joint rotation angle within the physiological range; the expression of the joint angle constraint is:

[0152] θmin,i ≤θ i ≤θ max,i

[0153] Where θ i is the joint rotation angle; θ min,i is the minimum rotation angle of the joint; θ max,i is the maximum rotation angle of the joint;

[0154] In the process of calculating the three-dimensional vertex coordinates, all vertices are detected using the grid self-intersection detection strategy:

[0155]

[0156] Where, d(v i ,v j ) is the Euclidean distance between the i-th vertex and the j-th vertex; ∈ is the minimum safe distance.

[0157] The parameter conversion ideas and example formulas in this embodiment can also be directly referenced in actual projects by the parameter conversion strategies and formulas in the SMPL / SMPL-X model, as shown in Table 2:

[0158]

[0159]

[0160] Table 2 Parameter conversion strategy and formula implementation list

[0161] In this embodiment, in step S3, a three-dimensional human body model is constructed based on the three-dimensional mesh vertex coordinates, topological structure and the skeletal joint data; the joints of the three-dimensional human body model are bound according to the skeletal joint data to obtain a bound three-dimensional human body model; the bound three-dimensional human body model is expression-driven based on the facial key point data and the expression weight, and the bound three-dimensional human body model is visually rendered by setting the material to generate a final human body model.

[0162] Specifically, the system first reads the mesh vertex coordinates, topology, and skeletal joint data, and creates a 3D human model within the modeling platform. The mesh is constructed based on optimized morphological parameters, ensuring that the overall proportions, dimensions, and topology of the human model conform to actual measurement standards. After the system imports the data, it optimizes the mesh structure to eliminate possible geometric anomalies such as self-intersections, patch distortion, and mesh damage, ensuring the topological integrity of the model. Furthermore, the mesh normal direction is adjusted to match lighting calculations and rendering requirements, laying the foundation for subsequent model optimization.

[0163] The human body model is then bound to the joints based on the skeletal joint data, the joint hierarchy is associated with the mesh model, and bone influence weights are assigned to each vertex. Based on bone-driven mesh deformation (Skinning), the system calculates the degree to which mesh vertices are affected by bone rotation, ensuring that when the bone joints move, the mesh vertices can deform accordingly, maintaining the coordination between the surface skin and the bone structure. After the skeletal binding is completed, the system adjusts the initial posture based on the posture parameters to put the human body model in a state that meets the user's settings, such as standing, sitting, walking, or other preset actions.

[0164] Furthermore, expression parameters can be used to adjust the deformation of key facial points, ensuring natural expression changes that conform to the laws of facial muscle movement. Expression control is based on preset deformation weights. The system calculates the offset of each key point based on the input expression parameters and maps it to a three-dimensional mesh, enabling the model to simulate facial expressions such as smiles, frowns, and surprise. After expression adjustment is completed, the system ensures that joint deformation is synchronized with mesh updates. That is, when the human body model makes movement adjustments, the deformation of the mesh surface is dynamically updated with the movement of the bones, avoiding abnormal skin stretching or self-intersection problems, and improving the naturalness and stability of the model.

[0165] Finally, the mesh model is selectively assigned a material to complete the visual rendering. The skin color setting is based on the material assignment method. Users can choose a standard skin color scheme or personalize it through mapping to enhance the realism of the model. In addition, the model interaction function can be provided in the modeling platform, allowing users to rotate, scale, and adjust the model perspective in a three-dimensional environment, and further modify the posture, expression, and material to meet application requirements. The generated human models with different shapes and expressions are as follows: Figure 5 As shown; the interactive interface of the human body model is as follows Figure 3 shown.

[0166] In a possible embodiment, three specific examples are provided as follows:

[0167] Example 1: With just a simple, natural language input like "170cm tall, 42cm shoulder width, 30% smile," two morphological parameters (height and shoulder width) and one expression parameter (smile intensity) are mapped simultaneously. Using a shape deformation formula, height and shoulder width are mapped to spatial offsets of mesh vertices. Then, using an expression superposition formula, the smile base is applied to facial vertices. The result is a 3D human model that conforms to anthropometric standards and exhibits a natural smile.

[0168] Example 2: When the input "right arm raised 45°, arm length 58cm" is given, the system extracts the arm length morphological parameter and the shoulder lift angle gestural parameter. The morphological component first scales the corresponding mesh area according to the arm length. Then, using posture correction, the 45° shoulder rotation is precisely applied to the skeletal joints. Skin weights are then used to achieve joint-driven mesh tracking, ultimately achieving a realistic 3D effect of the dynamic arm lift.

[0169] Example 3: When a user describes "sitting with eye height 90cm, smile 10%, and eyebrow raise 5%," the system simultaneously recognizes the sitting eye height morphological parameter and two expression fine-tuning parameters (smile and eyebrow raise). This demonstrates that the present invention can extract multi-dimensional parameters from a single natural language input and simultaneously reflect posture (slightly sitting down), expression (smile and eyebrow raise), and local morphological changes on a 3D model, meeting the needs of complex interactive scenarios.

[0170] Among them, in the process of generating 3D human body models in the above three examples, the transmission and calculation of parameters are as follows:

[0171] As shown in Table 3:

[0172]

[0173]

[0174] Table 3 Parameter transfer and calculation in examples

[0175] In summary, the present invention parses the natural language input by the user through the GPT large model interface to extract human body morphological parameters, posture parameters and expression parameters; the extracted human body morphological parameters, posture parameters and expression parameters are respectively subjected to human body proportion correction, kinematic correction and expression parameter optimization processing to obtain three-dimensional mesh vertex coordinates, bone joint data, facial key point data and expression weights; a three-dimensional human body model is constructed according to the three-dimensional mesh vertex coordinates, topological structure and the bone joint data; the three-dimensional human body model is joint-bound according to the bone joint data to obtain a bound three-dimensional human body model; the bound three-dimensional human body model is expression-driven based on the facial key point data and the expression weight, and the bound three-dimensional human body model is visually rendered by setting the material to generate a final human body model. This invention pioneered the seamless integration of language parsing, parameter optimization, and 3D reconstruction. It utilizes the GPT large model to accurately extract multi-dimensional parameters such as morphology, posture, and expression implicit in natural language. By integrating anthropometric constraints, inverse kinematics algorithms, and expression-based deformation technology, it transforms abstract parameters into mesh vertex coordinates, skeletal joint data, and expression weights that conform to biomechanical laws. Ultimately, an interactive 3D human model with complete topology, natural movement, and vivid expressions is generated in real time on the modeling platform. This invention completely overturns the traditional manual modeling modeling model that relies on professional software, compressing the modeling cycle from hours to seconds. It supports the simultaneous parsing and generation of composite instructions, solving the problems of semantic loss and physical distortion in cross-modal conversion, and providing efficient, accurate, and low-threshold digital human body generation solutions for fields such as virtual reality, medical simulation, and game development.

[0176] It should be noted that the method of the embodiments of the present disclosure can be performed by a single device, such as a computer or server. The method of the embodiments of the present disclosure can also be applied in a distributed scenario, where multiple devices cooperate to perform the method. In such a distributed scenario, one of the multiple devices may only perform one or more steps of the method of the embodiments of the present disclosure, and the multiple devices will interact with each other to complete the method.

[0177] It should be noted that the above description is limited to some embodiments of the present disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in an order different from that described in the above embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0178] Example 2

[0179] See also Figure 6Embodiment 2 of the present invention further provides a device for automatically generating a human body model based on a GPT large model, comprising:

[0180] The data parameter extraction module 001 is used to parse the natural language input by the user through the GPT large model interface and extract the human body shape parameters, posture parameters and expression parameters;

[0181] The data parameter processing module 002 is used to perform body proportion correction, kinematic correction and expression parameter optimization on the extracted human body morphological parameters, posture parameters and expression parameters, and obtain three-dimensional mesh vertex coordinates, bone joint data, facial key point data and expression weights;

[0182] The human body model generation module 003 is used to construct a three-dimensional human body model based on the three-dimensional mesh vertex coordinates, topological structure and the bone joint data; bind the joints of the three-dimensional human body model according to the bone joint data to obtain a bound three-dimensional human body model; drive the bound three-dimensional human body model based on the facial key point data and the expression weight, and visually render the bound three-dimensional human body model by setting the material to generate a final human body model.

[0183] In this embodiment, in the data parameter extraction module 001, the human body morphological parameters include: height, weight, shoulder width, waist circumference, hip circumference, leg length and arm length parameters; the posture parameters include: trunk inclination angle, limb rotation angle, joint position and spinal curvature angle parameters; the expression parameters include: mouth corner upward amplitude, eyebrow lifting degree, eyelid closure rate and mouth opening degree parameters.

[0184] In this embodiment, in the data parameter processing module 002, in the process of performing human body proportion correction processing on the extracted human body morphological parameters and obtaining the three-dimensional mesh vertex coordinates:

[0185] The expression for converting the human body morphological parameters into the three-dimensional mesh vertex coordinates is:

[0186] V'=f shape (S)=V base +B shape S

[0187] Where V' is the coordinate of the three-dimensional grid vertex; V base is the reference vertex coordinate of the standard human body model; B shape is the morphological deformation basis matrix, which represents the influence of different morphological parameters on vertex coordinates; S is the morphological parameter;

[0188] Through the strategy of global scaling and local adjustment, the body scaling transformation is performed:

[0189] v'i =λv i +δ i

[0190] Where, v' i is the coordinate of the i-th 3D mesh vertex after the global scaling factor and the local adjustment factor; v i is the original coordinate of the i-th mesh vertex before morphological deformation in the standard human body model; λ is the global scaling factor; δ i is the local adjustment factor.

[0191] In this embodiment, in the data parameter processing module 002, in the process of performing kinematic correction processing on the posture parameters to obtain the skeletal joint data:

[0192] Through inverse kinematics and motion database matching strategies, joint rotation data that conforms to the laws of human motion is calculated;

[0193] The inverse kinematics equation is:

[0194] P end =T joint P base

[0195] Where, P end is the target position of the hand; T joint is the transformation matrix composed of joint rotation matrices; P base is the initial position;

[0196] The optimization objective function of the inverse kinematics equation is:

[0197] J IK =||P end -P target || 2

[0198] Where, J IK is the optimization objective function of inverse kinematics, which represents the current end effector position P end and the expected target position P target The square error between target The three-dimensional position of the hand target specified by the user or obtained after natural language parsing;

[0199] The skeletal joint data is obtained through Jacobian inverse optimization:

[0200] Δθ=J + (P target -P end )

[0201] Where Δθ is the incremental vector of each joint angle in this iteration; J+ is the pseudo-inverse of the Jacobian matrix.

[0202] In this embodiment, in the data parameter processing module 002, in the process of performing expression parameter optimization processing on the expression parameters to obtain the facial key point data and the expression weight:

[0203] The displacement of facial key points is controlled by the expression parameters:

[0204]

[0205] Where, v face,i is the 3D coordinate of the i-th facial key point after expression deformation is applied; As the base face shape; B expr,j is the jth expression base; w j is the weight of the expression parameter; K is the number of expression bases;

[0206] The optimal expression weight is obtained by least square method:

[0207]

[0208] Where, J expr v is the objective function for optimizing expression parameters, which is used to measure the sum of square errors between the coordinates of the i-th key point generated by the model and the coordinates of the real key point; real,i is the true 3D coordinate of the i-th facial key point obtained by calibration, scanning or camera capture; L is the number of facial key points.

[0209] In this embodiment, in the data parameter processing module 002, a joint angle constraint is introduced during the calculation of the joint rotation data to ensure that the joint rotation angle is within the physiological range; the expression of the joint angle constraint is:

[0210] θ min,i ≤θ i ≤θ max,i

[0211] Where θ i is the joint rotation angle; θ min,i is the minimum rotation angle of the joint; θ max,i is the maximum rotation angle of the joint;

[0212] In the process of calculating the three-dimensional vertex coordinates, all vertices are detected using the grid self-intersection detection strategy:

[0213]

[0214] Where, d(v i ,v j) is the Euclidean distance between the i-th vertex and the j-th vertex; ∈ is the minimum safe distance.

[0215] It should be noted that the information interaction, execution process, etc. between the modules of the above-mentioned system are based on the same concept as the method embodiment in Example 1 of the present application, and the technical effects they bring are the same as those of the method embodiment of the present application. For specific contents, please refer to the description in the method embodiment shown above in the present application, and no further details will be given here.

[0216] Example 3

[0217] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium, in which a program code for a method for automatically generating a human body model based on a GPT large model is stored. The program code includes instructions for executing embodiment 1 or any possible implementation thereof, a method for automatically generating a human body model based on a GPT large model.

[0218] Computer-readable storage media can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0219] Example 4

[0220] Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor;

[0221] The processor and the memory communicate with each other via a bus; the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute a method for automatically generating a human body model based on a GPT large model according to Example 1 or any possible implementation thereof.

[0222] Specifically, the processor can be implemented by hardware or by software. When implemented by hardware, the processor can be a logic circuit, an integrated circuit, etc.; when implemented by software, the processor can be a general-purpose processor, which is implemented by reading software code stored in a memory. The memory can be integrated into the processor or located outside the processor and exist independently.

[0223] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable systems. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website, computer, server or data center to another website, computer, server or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode.

[0224] Obviously, those skilled in the art will appreciate that the various modules or steps of the present invention described above can be implemented using a general-purpose computing system. They can be centralized on a single computing system or distributed across a network of multiple computing systems. Alternatively, they can be implemented using program code executable by a computing system, and thus, they can be stored in a storage system and executed by the computing system. In some cases, the steps shown or described herein can be performed in a different order than that shown, or they can be fabricated into separate integrated circuit modules, or multiple modules or steps can be fabricated into a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0225] Although the present invention has been described in detail above using general descriptions and specific embodiments, it will be apparent to those skilled in the art that modifications and improvements may be made thereto. Therefore, such modifications and improvements, without departing from the spirit of the present invention, are intended to be within the scope of protection claimed herein.

Claims

1. A method for automatically generating a human body model based on a GPT large model, characterized in that: include: Parse the natural language input by the user through the GPT large model interface to extract human body shape parameters, posture parameters and expression parameters; The extracted human body morphological parameters, the posture parameters and the expression parameters are respectively subjected to human body proportion correction, kinematic correction and expression parameter optimization processing to obtain three-dimensional mesh vertex coordinates, bone joint data, facial key point data and expression weights; A three-dimensional human body model is constructed according to the three-dimensional mesh vertex coordinates, topological structure and the skeletal joint data; the joints of the three-dimensional human body model are bound according to the skeletal joint data to obtain a bound three-dimensional human body model; the bound three-dimensional human body model is expression-driven based on the facial key point data and the expression weight, and the bound three-dimensional human body model is visually rendered by setting the material to generate a final human body model.

2. The method for automatically generating a human body model based on a GPT large model according to claim 1, characterized in that: The human body morphological parameters include: height, weight, shoulder width, waist circumference, hip circumference, leg length and arm length parameters; the posture parameters include: trunk inclination angle, limb rotation angle, joint position and spinal curvature angle parameters; the expression parameters include: mouth corner upward range, eyebrow lifting degree, eyelid closure rate and mouth opening degree parameters.

3. The method for automatically generating a human body model based on a GPT large model according to claim 2, characterized in that: In the process of performing human body proportion correction processing on the extracted human body morphological parameters to obtain the three-dimensional mesh vertex coordinates: The expression for converting the human body morphological parameters into the three-dimensional mesh vertex coordinates is: V′=f shape (S)=V base +B shape S Where V′ is the coordinate of the 3D mesh vertex; V base is the reference vertex coordinate of the standard human body model; B shape is the morphological deformation basis matrix, which represents the influence of different morphological parameters on vertex coordinates; S is the morphological parameter; Through the strategy of global scaling and local adjustment, the body scaling transformation is performed: v′ i =λv i +δ i Where v′ i is the coordinate of the i-th 3D mesh vertex after the global scaling factor and the local adjustment factor; v i is the original coordinate of the i-th mesh vertex before morphological deformation in the standard human body model; λ is the global scaling factor; δ i is the local adjustment factor.

4. The method for automatically generating a human body model based on a GPT large model according to claim 3, characterized in that: In the process of performing kinematic correction processing on the posture parameters to obtain the skeletal joint data: Through inverse kinematics and motion database matching strategies, joint rotation data that conforms to the laws of human motion is calculated; The inverse kinematics equation is: P end =T joint P base Where, P end is the target position of the hand; T joint is the transformation matrix composed of joint rotation matrices; P base is the initial position; The optimization objective function of the inverse kinematics equation is: J IK =||P end -P target || 2 Where, J IK is the optimization objective function of inverse kinematics, which represents the current end effector position P end and the expected target position P target The square error between target The three-dimensional position of the hand target specified by the user or obtained after natural language parsing; The skeletal joint data is obtained through Jacobian inverse optimization: Δθ=J + (P target -P end ) Where Δθ is the incremental vector of each joint angle in this iteration; J + is the pseudo-inverse of the Jacobian matrix.

5. The method for automatically generating a human body model based on a GPT large model according to claim 4, characterized in that: In the process of performing expression parameter optimization processing on the expression parameters to obtain the facial key point data and the expression weight: The displacement of facial key points is controlled by the expression parameters: Where, v face,i is the 3D coordinate of the i-th facial key point after expression deformation is applied; As the base face shape; B expr,j is the jth expression base; w j is the weight of the expression parameter; K is the number of expression bases; The optimal expression weight is obtained by least square method: Where, J expr The objective function for optimizing expression parameters is used to measure the sum of square errors between the coordinates of the i-th key point generated by the model and the coordinates of the real key point; v real,i is the true 3D coordinate of the i-th facial key point obtained by calibration, scanning or camera capture; L is the number of facial key points.

6. The method for automatically generating a human body model based on a GPT large model according to claim 5, characterized in that: In the process of calculating joint rotation data, joint angle constraints are introduced to keep the joint rotation angle within the physiological range; the expression of the joint angle constraint is: i min,i ≤θ i ≤θ max,i Where θ i is the joint rotation angle; θ min,i is the minimum rotation angle of the joint; θ max,i is the maximum rotation angle of the joint; In the process of calculating the three-dimensional vertex coordinates, all vertices are detected using the grid self-intersection detection strategy: Where, d(v i ,v j ) is the Euclidean distance between the i-th vertex and the j-th vertex; ∈ is the minimum safe distance.

7. A device for automatically generating a human body model based on a GPT large model, characterized in that: include: The data parameter extraction module is used to parse the natural language input by the user through the GPT large model interface and extract human body morphological parameters, posture parameters and expression parameters; a data parameter processing module for performing body proportion correction, kinematic correction, and expression parameter optimization processing on the extracted human body morphological parameters, posture parameters, and expression parameters, respectively, to obtain three-dimensional mesh vertex coordinates, bone joint data, facial key point data, and expression weights; A human body model generation module is used to construct a three-dimensional human body model based on the three-dimensional mesh vertex coordinates, topological structure and the skeletal joint data; bind the joints of the three-dimensional human body model according to the skeletal joint data to obtain a bound three-dimensional human body model; drive the bound three-dimensional human body model with facial expressions based on the facial key point data and the expression weights, and visually render the bound three-dimensional human body model by setting materials to generate a final human body model.

8. The device for automatically generating a human body model based on a GPT large model according to claim 7, characterized in that: In the data parameter extraction module, the human body morphological parameters include: height, weight, shoulder width, waist circumference, hip circumference, leg length and arm length parameters; the posture parameters include: trunk inclination angle, limb rotation angle, joint position and spine bending angle parameters; the expression parameters include: mouth corner upward range, eyebrow lifting degree, eyelid closure rate and mouth opening degree parameters.

9. The device for automatically generating a human body model based on a GPT large model according to claim 8, characterized in that: In the data parameter processing module, in the process of performing human body proportion correction processing on the extracted human body morphological parameters and obtaining the three-dimensional mesh vertex coordinates: The expression for converting the human body morphological parameters into the three-dimensional mesh vertex coordinates is: V′=f shape (S)=V base +B shape S Where V′ is the coordinate of the 3D mesh vertex; V base is the reference vertex coordinate of the standard human body model; B shape is the morphological deformation basis matrix, which represents the influence of different morphological parameters on vertex coordinates; S is the morphological parameter; Through the strategy of global scaling and local adjustment, the body scaling transformation is performed: v′ i =λv i +δ i Where v′ i is the coordinate of the i-th 3D mesh vertex after the global scaling factor and the local adjustment factor; v i is the original coordinate of the i-th mesh vertex before morphological deformation in the standard human body model; λ is the global scaling factor; δ i is the local adjustment factor.

10. The device for automatically generating a human body model based on a GPT large model according to claim 9, characterized in that: In the data parameter processing module, in the process of performing kinematic correction processing on the posture parameters to obtain the skeletal joint data: Through inverse kinematics and motion database matching strategies, joint rotation data that conforms to the laws of human motion is calculated; The inverse kinematics equation is: P end =T joint P base Where, P end is the target position of the hand; T joint is the transformation matrix composed of joint rotation matrices; p base is the initial position; The optimization objective function of the inverse kinematics equation is: J IK =||P end -P target || 2 Where, J IK is the optimization objective function of inverse kinematics, which represents the current end effector position P end and the expected target position P target The square error between target The three-dimensional coordinates of the hand's desired target, specified by the user or obtained after natural language parsing; The skeletal joint data is obtained through Jacobian inverse optimization: Δθ=J + (P target -P end ) Where Δθ is the incremental vector of the rotation angle of each joint in this iteration; J + is the pseudo-inverse of the Jacobian matrix.

Images