A text-driven component-based three-dimensional digital human generation method and system

Abstract

This invention relates to a text-driven, component-based 3D digital human generation method and system. The method includes: acquiring accessory components related to the digital human and an initial human body model, and constructing a 3D component asset library; adjusting the human body shape parameters of the initial human body model, selecting target accessory components, and obtaining a 3D digital human base model; constructing a 3D Gaussian-mesh mixture representation of each component based on the 3D digital human base model; introducing a CFD loss function to iteratively optimize the 3D Gaussian-mesh mixture representation of each component, generating a corresponding initial visual appearance; obtaining an overall representation of the digital human based on the initial visual appearance, and optimizing the 3D Gaussian-mesh mixture representation of the overall digital human representation to obtain a component-based 3D digital human. Compared with existing technologies, this invention effectively solves the problems of unreasonable geometric adhesion, chaotic structure, and weak controllability and animation of clothing and hairstyles in existing 3D digital human generation methods.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a text-driven, component-based method and system for generating 3D digital humans. Background Technology

[0002] Text-driven 3D digital human generation technology aims to generate 3D digital humans that match text descriptions, which can then be applied to subsequent production processes such as editing and animation. Traditional 3D character creation requires professional artists to complete multiple stages, including concept design, high-poly model sculpting, topology optimization, UV unwrapping, texture painting, and material addition. This process is cumbersome, time-consuming, costly, and requires a high level of expertise. The emergence of text-driven 3D digital human generation technology significantly reduces modeling time and economic costs, lowers the professional barrier, and supports highly personalized and diverse generation effects. It demonstrates advantages in modeling efficiency and creative richness, and provides the possibility for rapid construction and efficient iteration of 3D digital humans.

[0003] As technological research and application deepen, users are placing higher demands on text-driven 3D digital human generation technology. On one hand, the generated quality needs to be more refined; users expect the generated 3D digital humans to have accurate body structure, distinct clothing layers, and realistic material representation. On the other hand, the need for controllability and editability is increasingly prominent. Building upon automatic generation, users expect to be able to adjust local features such as body proportions, clothing styles, and hairstyles. Furthermore, applicability has become an important consideration. The generated 3D digital humans need to be compatible with animation production pipelines to facilitate subsequent rigging, animation production, and other production application processes, achieving a seamless transition from generation to application. Therefore, high-quality, controllable, easily editable, and highly applicable text-driven 3D digital human generation technology is becoming an important direction for industry innovation and development. Most existing 3D digital human generation technologies utilize AvatarCLIP, HumanNorm, SO-SMPL, and 3D Gaussian splashing techniques to generate 3D digital humans. However, these technologies suffer from the following problems: AvatarCLIP cannot synthesize new clothing styles; HumanNorm's texture generation relies on normal map guidance but lacks accurate text mapping of clothing details; SO-SMPL, while constructing layered clothing through vertex offset, has controllability limited to whether or not clothing is generated, and cannot respond to specific text descriptions about clothing; 3D Gaussian splashing techniques, such as HumanGaussian, do not clearly distinguish between the semantic constraints of clothing and the human body through their bi-branch SDS loss, leading to deviations between clothing styles and text prompts. Alternatively, Chinese patent application CN118229860A provides a text-based 3D human generation method that generates animated virtual human avatars and supports real-time rendering through differentiable rendering under the guidance of a diffusion model, but it still suffers from problems such as unreasonable geometric adhesion, structural chaos, and weak controllability and animation of clothing and hairstyles.

[0004] Therefore, providing a method that can effectively solve the problems of unreasonable geometric adhesion, structural chaos, and weak controllability and animation of clothing and hairstyle in existing 3D digital human generation is a technical problem that needs to be solved. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art by providing a modular 3D digital human generation method that can reuse accessory components. By utilizing the good topological structure of the accessory components themselves, 3D digital human generation is achieved through component assembly, avoiding geometric adhesion in the process of generating 3D digital human, and ensuring that each component is independent yet compatible with each other.

[0006] The objective of this invention can be achieved through the following technical solutions: According to one aspect of the present invention, a text-driven, component-based 3D digital human generation method is provided, the method comprising: Obtain accessory components and an initial human body model related to the digital human; preprocess the accessory components based on the initial human body model to construct a three-dimensional component asset library. The initial human body model is adjusted for human body shape parameters. Based on the requirements of digital human generation, a target accessory component is selected from the three-dimensional component asset library. The target accessory component is then adapted to the initial human body model after the body shape parameters are adjusted to obtain a component-based three-dimensional digital human base model. For each component in the three-dimensional digital human model, its vertices are offset to construct a set of triangular prism cells. Based on the set of triangular prism cells, a three-dimensional Gaussian-mesh hybrid representation of the corresponding component is generated. The component includes the target accessory component and the human body model. A CFD loss function is introduced to iteratively optimize the 3D Gaussian-mesh mixture representation of each component to generate the corresponding initial visual appearance. The initial visual appearances are merged to obtain the overall representation of the digital human. The three-dimensional Gaussian-mesh hybrid representation of the overall digital human representation is optimized based on the CFD loss function to obtain a component-based three-dimensional digital human.

[0007] As a preferred technical solution, the method for constructing the aforementioned three-dimensional component asset library is as follows: The accessory components are standardized. The standardization process includes performing basic operations on each accessory component, including rotation, scaling, and translation, so that the accessory components are properly worn on the initial human body model, and the initial coordinates of each accessory component are obtained at this time. The accessory components are categorized, and each category is assigned a number to the accessory components it includes, along with a text description and applicable gender. For each accessory component, an accessory component vector is generated based on its text description, and a three-dimensional component asset library is constructed by combining its number, text description, applicable gender, and initial coordinates.

[0008] As a preferred technical solution, the initial human body model is either a first model or a second model. The first model is constructed based on the SMPL human body template, and the second model is a human body model with all body shape parameters set to preset values.

[0009] As a preferred technical solution, when the initial human body model is the first model, the height and weight of the digital human to be generated are obtained, and the human body shape parameters are adjusted based on the height and weight.

[0010] As a preferred technical solution, when the initial human body model is the second model, the method for adjusting the human body shape parameters is as follows: For the second model, a three-dimensional shape is constructed for each body shape parameter, with values ​​of -1, 0, and 1. A three-dimensional shape of -1 indicates that the actual body shape parameter is at its lower limit; a three-dimensional shape of 0 indicates that the actual body shape parameter is at a preset value; and a three-dimensional shape of 1 indicates that the actual body shape parameter is at its upper limit. The body shape parameters include height, head size, neck length, shoulder width, chest circumference, arm length, waist circumference, leg length, and Adam's apple. The target body shape parameters of the digital human to be generated are collected, the adjacent three-dimensional shapes of the target body shape parameters are determined, the weighted difference of the adjacent three-dimensional shapes is calculated, and the target shape value is obtained, thereby realizing the adjustment of human body shape parameters.

[0011] As a preferred technical solution, the method for constructing the set of triangular prism cells is as follows: for each component, obtain the corresponding vertex and its vertex normal vector, offset each vertex along the positive and negative directions of its corresponding vertex normal vector respectively, generate a triangular prism cell for each vertex, and integrate the triangular prism cells of all vertices to form the set of triangular prism cells for the corresponding component.

[0012] As a preferred technical solution, the method for generating the aforementioned three-dimensional Gaussian-mesh hybrid representation is as follows: Random sampling is performed in each of the triangular prism cells in the aforementioned set of triangular prism cells to obtain multiple centroid coordinates; Using the aforementioned centroid coordinates as the center, three-dimensional Gaussian points are initialized. These three-dimensional Gaussian points and the set of triangular prism cells are then merged to obtain a three-dimensional Gaussian-mesh hybrid representation, as follows: , in, Let represent a set of three-dimensional Gaussian points, and in this set, we have: , Represents a three-dimensional Gaussian point Corresponding to the position of the center of gravity, This represents a learnable scaling vector. Represents a learnable rotation quaternion. Indicates opacity. Represents the spherical harmonic coefficients; Let represent a set of triangular prism cells, and Limited to middle.

[0013] As a preferred technical solution, the CFD loss function is: , in, Indicates in v, And t follows the mathematical expectation of the value distribution, where v represents the random perspective; Represents the weighting function; Indicates noise level Predict noise, and have , This represents a noisy image. Indicates text prompt words; This represents the signal scaling factor during the diffusion process; This represents the noise scaling factor during the diffusion process; This represents a multi-view consistent Gaussian noise function. express All learnable parameters in, and have , Indicates a random perspective; This refers to a rendered image. As a preferred technical solution, the method for obtaining the rendered image is as follows: Using a 3D Gaussian splashing differentiable rendering method, at random viewpoints The following is a rendering of the image based on the corresponding 3D Gaussian point set; In the 3D Gaussian splash differentiable rendering, 3D Gaussian points are projected onto the camera's imaging plane, and pixels in the rendered image are... Rendering colors at the location The process is as follows: in, Indicates the affected pixels A three-dimensional set of Gaussian points; Represents the color of the i-th 3D Gaussian point; This represents the opacity of the i-th 3D Gaussian point; This indicates the position of the center corresponding to the i-th three-dimensional Gaussian point; Represent the covariance matrix of the i-th three-dimensional Gaussian point; and Defined as a parameter calculated based on opacity and covariance matrix.

[0014] According to a second aspect of the present invention, a text-driven, component-based 3D digital human generation system is provided for implementing the above-described method.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1) Addressing the problems of geometric adhesion, structural chaos, poor visual quality, and difficulty in individually controlling and animating clothing and hairstyles due to integrated generation in existing technologies, this invention constructs a 3D component asset library and a 3D Gaussian-mesh hybrid representation. This enables the construction of a 3D digital human base model by reusing existing 3D asset models, providing a reliable geometric structural foundation and constraints for the generation process. Furthermore, due to the component-based generation approach, individual text descriptions can be provided for each component. If a component needs to be edited locally, its description can be edited separately to regenerate it, thereby achieving local component-level control.

[0016] 2) A two-stage optimization generation strategy with consistent multi-viewpoints is adopted. In the first stage, individual components are optimized independently and geometric constraints are applied to obtain components with clear boundaries and preliminary visual effects. In the second stage, after being stitched together as a whole, the prior boundaries of the components are preserved, which greatly alleviates the problems of blurring and entanglement at the boundaries of different components. Finally, a component-based 3D digital human with harmonious appearance and high visual quality is generated, which improves the shortcomings of the poor appearance and visual quality of existing methods. Attached Figure Description

[0017] Figure 1 This is a framework diagram of the text-driven, component-based 3D digital human generation technology of the present invention. Figure 2 This is a flowchart of the method of the present invention; Figure 3 This is a schematic diagram illustrating the adjustment of body shape parameters according to the present invention; Figure 4 This is a diagram illustrating the retrieval and assembly process of the Agent-based digital human component of the present invention. Figure 5 This is a schematic diagram illustrating the generation of a 3D Gaussian-mesh hybrid representation according to the present invention; Figure 6 The flowchart illustrates the optimization process for introducing perspective consistency in this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0019] To address issues such as unreasonable geometric adhesion, structural chaos, and weak controllability and animation of clothing and hairstyles in existing text-driven 3D digital human generation technologies, a component-based 3D digital human generation method is proposed. The technical framework of this method is as follows: Figure 1 As shown, the detailed process is as follows:Figure 2 As shown, it includes: S1. Obtain accessory components related to the digital human and the initial human body model. Based on the initial human body model, preprocess the accessory components and build a 3D component asset library.

[0020] The initial human body model is a human body model with a well-proportioned body. In this embodiment, the initial human body model is selected as either the first model or the second model. The first model is constructed based on the SMPL human body template and is a general model obtained through training. Because it is trained based on European and American human body scans, it is more in line with the body proportions of European and American people. The second model is a human body model with all body shape parameters set to preset values, which is modeled with a body proportion bias towards Asian people.

[0021] When loading the modeling software, the initial human body model's feet are centered at the origin, with the -y axis representing the front and the +z axis representing the top.

[0022] This step involves retrieving and assembling a digital human component based on the agent. The detailed steps are as follows: S11. Standardize the accessory components. Standardization includes performing basic operations on each accessory component, such as rotation, scaling, and translation, to ensure that the accessory components fit properly on the initial human body model and to obtain the initial coordinates of each accessory component. The accessory components are stored in the native file format of the modeling software, including the 3D model information of the accessory component itself and the initial coordinates in the modeling software. That is, when the initial human body model and the accessory component are directly loaded in the modeling software, the two are compatible.

[0023] S12. Categorize the accessory components, number the accessory components included in each category, and label them with text descriptions and applicable genders.

[0024] In detail, all accessory components are categorized into six main categories: tops, bottoms, suits, coats, shoes, and hairstyles. Each category is numbered sequentially and labeled with detailed text descriptions and applicable gender information to support efficient semantic retrieval.

[0025] S13. For each accessory component, generate an accessory component vector based on its text description, and construct a 3D component asset library by combining its number, text description, applicable gender, and initial coordinates.

[0026] S2. Adjust the body shape parameters of the initial human body model, select the target accessory component in the 3D component asset library based on the requirements of digital human generation, and adapt the target accessory component to the initial human body model after the body shape parameters are adjusted to obtain the component-based 3D digital human base model.

[0027] S21. Adjustment of human body shape parameters.

[0028] When the initial human body model is the first model, the height and weight of the digital human to be generated are obtained, and the human body shape parameters are adjusted based on the height and weight.

[0029] When the initial human body model is the second model, the method for adjusting the human body shape parameters is as follows: i) For the second model, construct a three-dimensional shape with values ​​of -1, 0 and 1 for each body shape parameter; where a three-dimensional shape of -1 indicates that the actual body shape parameter is at the lower limit; a three-dimensional shape of 0 indicates that the actual body shape parameter is at the preset value; and a three-dimensional shape of 1 indicates that the actual body shape parameter is at the upper limit.

[0030] In this step, body shape parameters include height, head size, neck length, shoulder width, chest circumference, arm length, waist circumference, leg length, and Adam's apple.

[0031] ii) Collect the target body shape parameters of the digital human to be generated, determine the adjacent three-dimensional shapes of the target body shape parameters, perform weighted difference on the adjacent three-dimensional shapes to obtain the target shape value, and realize the adjustment of human body shape parameters.

[0032] The effects of adjusting body shape parameters can be seen in [reference needed]. Figure 3 As shown, the orange humanoid figure represents the initial human body model, and the gray humanoid figure represents the human body model after parameter adjustment.

[0033] S22, Component-based 3D digital human body base model generation.

[0034] S221. Obtain the vertex coordinates of the target accessory component and the corresponding human model vertex coordinates based on the initial coordinates recorded in the 3D asset library, as well as the coordinates of the corresponding human model vertex after adjusting the human body shape parameters. Calculate the deviation between the initial and adjusted human model vertex coordinates, and update the accessory component vertex coordinates based on this deviation to achieve the adaptation between the accessory component and the human model. Figure 3 Vertex group in and For example, suppose there is an accessory component at vertex A of the human body model. To adapt this accessory component to the updated human body model, it is necessary to calculate the relationship between vertex A and vertex B. The deviation between them is used to update the position of the corresponding accessory.

[0035] The process of fitting accessory components to a human body model after body shape parameter adjustments can be represented as follows: ; ; in, This represents the coordinates of vertex i in the initial human body model after the body shape parameters have been adjusted; This represents the initial coordinates of vertex i in the initial human body model; This represents the coordinate difference of the i-th vertex after adjustment of the body shape parameters; Represents the coordinates of the accessory components adapted to the adjusted initial human body model; This represents the vertex coordinates of the accessory component at vertex i in the initial human model.

[0036] S222. Based on the AutoGen framework, a multi-agent workflow is built to receive structured text descriptions input by users. With the help of the Agent's reasoning ability, the component-based 3D digital human model is automatically generated through preset prompts and examples.

[0037] In detail, the process is as follows: Figure 4 As shown, this multi-agent workflow automates the following processes: parsing descriptive text, determining the initial human model type (e.g., female second model) and a series of body shape parameter values ​​based on body shape descriptive words, retrieving accessory components matching the text description from the 3D component asset library based on descriptions of clothing, hairstyle, shoes, etc., and finally calling the plug-in tool interface to load and assemble the human body and accessory component models in the modeling software to obtain the final digital human base model. Specifically, the retrieval step first converts the user description text into vectors, performs a similarity search in the 3D component asset library, returns the top five candidate components with the highest matching degree, and then calls the large model to select the most matching component based on semantics, returning its number.

[0038] After obtaining the human body type, body shape parameters, and component number from the large language model, the Agent calls the corresponding plug-in utility functions through the Remote Procedure Call Protocol (RPC) to obtain a component-based 3D digital human base model that initially meets the requirements of the text description.

[0039] S3. For each component in the 3D digital human base model, offset its vertices to construct a set of triangular prism cells, and generate a 3D Gaussian-mesh hybrid representation of the corresponding component based on the set of triangular prism cells.

[0040] In this step, the component assembly includes the target accessory component and the human body model, and the process is as follows: Figure 5 As shown, the detailed steps include: S31. Construct a set of triangular prism cells.

[0041] For each component, obtain the corresponding vertex and its vertex normal vector. Offset each vertex along the positive and negative directions of its corresponding vertex normal vector to form two new surfaces, thus obtaining the triangular prism cell of each vertex. Integrating the triangular prism cells of all vertices is the triangular prism cell set of the corresponding component.

[0042] S32, 3D Gaussian-mesh hybrid representation generation.

[0043] S321. Randomly sample each triangular prism cell in the triangular prism cell set to obtain multiple centroid coordinates.

[0044] S322. Initialize the three-dimensional Gaussian points using the centroid coordinates as the center, and merge the three-dimensional Gaussian points and the triangular prism cell set to obtain the three-dimensional Gaussian-mesh hybrid representation, as follows: , in, Let represent a set of three-dimensional Gaussian points, and in this set, we have: , Represents a three-dimensional Gaussian point Corresponding to the position of the center of gravity, This represents a learnable scaling vector. Represents a learnable rotation quaternion. Indicates opacity. Represents the spherical harmonic coefficients; Let represent a set of triangular prism cells, and Limited to middle.

[0045] The Gaussian point mentioned above is defined by a series of learnable parameters, including: center position. Covariance matrix Opacity and spherical harmonic coefficients Among them, the covariance matrix It can be determined by the scaling matrix. and rotation matrix The formula is: The scaling matrix and rotation matrix are respectively derived from the scaling vector. and rotation quaternions Therefore, a three-dimensional Gaussian point can be represented as... .

[0046] The triangular prism cell set provides a good topological mesh as a structured geometric constraint, which confines the 3D Gaussian points within the triangular prism cell set. This effectively suppresses the floating-point and structural chaos caused by the discrete distribution characteristics of the points and provides stable topological support for subsequent applications such as mesh-based skeletal animation. At the same time, the efficient differentiable rendering characteristics of the 3D Gaussian points are preserved, improving the efficiency of subsequent optimization and generation.

[0047] S4. Introduce a CFD loss function to iteratively optimize the 3D Gaussian-mesh hybrid representation of each component to generate the corresponding initial visual appearance.

[0048] In the individual component optimization generation stage, for each component under the 3D Gaussian-mesh mixture representation, the 3D Gaussian-mesh mixture representation parameters of each component are iteratively optimized by introducing a consistent flow distillation (CFD) loss with multiple perspectives, thereby generating a preliminary visual appearance and clarifying the boundaries to avoid the entanglement and blurring of the boundaries between the body, clothing and hair during the overall optimization process of the digital human.

[0049] S5. Merge the initial visual appearances to obtain the overall representation of the digital human. Optimize the three-dimensional Gaussian-mesh hybrid representation of the overall digital human representation based on the CFD loss function to obtain a component-based three-dimensional digital human.

[0050] In this step, the initial visual appearance obtained in step S4 is first merged, and its form is as follows: , A 3D Gaussian-mesh hybrid representation of a human body model; A 3D Gaussian-mesh hybrid representation of hair; The three-dimensional Gaussian-mesh hybrid representation of garment 1; A three-dimensional Gaussian-mesh hybrid representation of the shoe; , , as well as Let each represent the set of Gaussian points representing the human body model, hair, clothing 1, and shoes, respectively. , , as well as A set of triangular prism cells representing the human body model, hair, clothing 1, and shoes; This indicates that each component 3D Gaussian point set With triangular prism cell set - The parameters are concatenated in an ordered manner on the parameter tensors.

[0051] Through the above operations, a complete and manageable 3D digital human hybrid representation is constructed, which maintains the original geometric constraints and topological relationships of each component and provides a unified parameter framework for overall optimization. Then, the overall representation of the digital human is generated by CFD loss iterative optimization, and finally a component-based 3D digital human with harmonious appearance and high detail quality is generated.

[0052] In detail, in steps S4 and S5, the single-round CFD loss iterative optimization process is as follows: Figure 6As shown, a 3D model with a hybrid representation is rendered using a differentiable renderer, and noise is superimposed from a multi-view consistent Gaussian noise function based on the current rendering viewpoint. This noise, along with the corresponding text description and the time step sampled under an annealing strategy, is then input into a pre-trained stable diffusion model to predict the noise. The difference between the predicted noise and the original noise is compared to calculate the CFD loss, which guides the gradient optimization of the 3D model. Specifically, a 3D Gaussian splashing differentiable rendering method is used to achieve this at random viewpoints. The following is a rendering of the image based on the corresponding 3D Gaussian point set; wherein, during 3D Gaussian splash differentiable rendering, the 3D Gaussian points are projected onto the camera's imaging plane, and the pixels in the rendering image are... Rendering colors at the location The process is as follows: in, Indicates the affected pixels A three-dimensional set of Gaussian points; Represents the color of the i-th 3D Gaussian point; This represents the opacity of the i-th 3D Gaussian point; This indicates the position of the center corresponding to the i-th three-dimensional Gaussian point; Represent the covariance matrix of the i-th three-dimensional Gaussian point; and Defined as a parameter calculated based on opacity and covariance matrix.

[0053] The method for calculating the CFD loss function is as follows: , in, Indicates in v, And t follows the mathematical expectation of the value distribution, where v represents the random perspective; Represents the weighting function; Indicates noise level Predict noise, and have , This represents a noisy image. Indicates text prompt words; This represents the signal scaling factor during the diffusion process; This represents the noise scaling factor during the diffusion process; This represents a multi-view consistent Gaussian noise function. express All learnable parameters in, and have , Indicates a random perspective; This refers to a rendered image.

[0054] Furthermore, this embodiment also provides a text-driven, component-based 3D digital human generation system for implementing the above-described method.

[0055] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A text-driven, component-based 3D digital human generation method, characterized in that, The methods include: Obtain accessory components and an initial human body model related to the digital human; preprocess the accessory components based on the initial human body model to construct a three-dimensional component asset library. The initial human body model is adjusted for human body shape parameters. Based on the requirements of digital human generation, a target accessory component is selected from the three-dimensional component asset library. The target accessory component is then adapted to the initial human body model after the body shape parameters are adjusted to obtain a component-based three-dimensional digital human base model. For each component in the three-dimensional digital human model, its vertices are offset to construct a set of triangular prism cells. Based on the set of triangular prism cells, a three-dimensional Gaussian-mesh hybrid representation of the corresponding component is generated. The component includes the target accessory component and the human body model. A CFD loss function is introduced to iteratively optimize the 3D Gaussian-mesh mixture representation of each component to generate the corresponding initial visual appearance. The initial visual appearances are merged to obtain the overall representation of the digital human. The three-dimensional Gaussian-mesh hybrid representation of the overall digital human representation is optimized based on the CFD loss function to obtain a component-based three-dimensional digital human.

2. The text-driven, component-based 3D digital human generation method according to claim 1, characterized in that, The method for constructing the aforementioned 3D component asset library is as follows: The accessory components are standardized. The standardization process includes performing basic operations on each accessory component, including rotation, scaling, and translation, so that the accessory components are properly worn on the initial human body model, and the initial coordinates of each accessory component are obtained at this time. The accessory components are categorized, and each category is assigned a number to the accessory components it includes, along with a text description and applicable gender. For each accessory component, an accessory component vector is generated based on its text description, and a three-dimensional component asset library is constructed by combining its number, text description, applicable gender, and initial coordinates.

3. The text-driven, component-based 3D digital human generation method according to claim 1, characterized in that, The initial human body model is either a first model or a second model. The first model is constructed based on the SMPL human body template. The second model is a human body model with all body shape parameters set to preset values.

4. The text-driven, component-based 3D digital human generation method according to claim 3, characterized in that, When the initial human body model is the first model, the height and weight of the digital human to be generated are obtained, and the human body shape parameters are adjusted based on the height and weight.

5. The text-driven, component-based 3D digital human generation method according to claim 3, characterized in that, When the initial human body model is the second model, the method for adjusting the human body shape parameters is as follows: For the second model, a three-dimensional shape is constructed for each body shape parameter, with values ​​of -1, 0, and 1. A three-dimensional shape of -1 indicates that the actual body shape parameter is at its lower limit; a three-dimensional shape of 0 indicates that the actual body shape parameter is at a preset value; and a three-dimensional shape of 1 indicates that the actual body shape parameter is at its upper limit. The body shape parameters include height, head size, neck length, shoulder width, chest circumference, arm length, waist circumference, leg length, and Adam's apple. The target body shape parameters of the digital human to be generated are collected, the adjacent three-dimensional shapes of the target body shape parameters are determined, the weighted difference of the adjacent three-dimensional shapes is calculated, and the target shape value is obtained, thereby realizing the adjustment of human body shape parameters.

6. The text-driven, component-based 3D digital human generation method according to claim 1, characterized in that, The method for constructing the set of triangular prism cells is as follows: For each component, obtain the corresponding vertex and its vertex normal vector, offset each vertex along the positive and negative directions of its corresponding vertex normal vector respectively, generate a triangular prism cell for each vertex, and integrate the triangular prism cells of all vertices to form the set of triangular prism cells for the corresponding component.

7. The text-driven, component-based 3D digital human generation method according to claim 1, characterized in that, The method for generating the aforementioned 3D Gaussian-mesh hybrid representation is as follows: Random sampling is performed in each of the triangular prism cells in the aforementioned set of triangular prism cells to obtain multiple centroid coordinates; Using the aforementioned centroid coordinates as the center, three-dimensional Gaussian points are initialized. These three-dimensional Gaussian points and the set of triangular prism cells are then merged to obtain a three-dimensional Gaussian-mesh hybrid representation, as follows: ， in, Let represent a set of three-dimensional Gaussian points, and in this set, we have: , Represents a three-dimensional Gaussian point Corresponding to the position of the center of gravity, Represents a learnable scaling vector. Represents a learnable rotation quaternion. Indicates opacity. Represents the spherical harmonic coefficients; Represents a set of triangular prism cells, and Limited to middle.

8. The text-driven, component-based 3D digital human generation method according to claim 1, characterized in that, The CFD loss function is as follows: ， in, Indicates in v, And t follows the mathematical expectation of the value distribution, where v represents the random perspective; Represents the weighting function; Indicates noise level Predict noise, and have , This represents a noisy image. Indicates text prompt words; This represents the signal scaling factor during the diffusion process; This represents the noise scaling factor during the diffusion process; This represents a multi-view consistent Gaussian noise function. express All learnable parameters in, and have , Indicates a random perspective; This refers to a rendered image.

9. A text-driven, component-based 3D digital human generation method according to claim 8, characterized in that, The method for obtaining the rendered image is as follows: Using a 3D Gaussian splashing differentiable rendering method, at random viewpoints The following is a rendering of the image based on the corresponding 3D Gaussian point set; In the 3D Gaussian splash differentiable rendering, 3D Gaussian points are projected onto the camera's imaging plane, and pixels in the rendered image are... Rendering colors at the location The process is as follows: in, Indicates the affected pixels A three-dimensional set of Gaussian points; Represents the color of the i-th 3D Gaussian point; This represents the opacity of the i-th 3D Gaussian point; This indicates the position of the center corresponding to the i-th three-dimensional Gaussian point; Represent the covariance matrix of the i-th three-dimensional Gaussian point; and Defined as a parameter calculated based on opacity and covariance matrix.

10. A text-driven, component-based 3D digital human generation system, characterized in that, The system is used to implement the method as described in any one of claims 1 to 9.

Images