Text-driven 3D human motion generation method with local generation and global fusion

Abstract

The application discloses a text-driven three-dimensional human body action generation method of local generation and global fusion. First, the input text is decomposed into local descriptions of body parts through a large language model. Then, a parallel diffusion model is used to generate action latent codes for each body part, ensuring the accuracy of local control. Next, a spatial domain fusion network is used to coordinate the actions of each part, forming a natural whole body posture. Finally, the posture is converted into a three-dimensional Gaussian representation, and a two-dimensional diffusion model is used to optimize visual details. Real-time rendering is achieved through Gaussian sputtering. The final trained network model can realize end-to-end generation from text description to three-dimensional human body action. The application solves the problems of inaccurate local control, poor whole body coordination and low rendering efficiency in the prior art, and has significant improvement in action detail performance, semantic alignment and generation efficiency. It can be widely applied in virtual reality, film and game production and other fields.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and computer graphics, specifically to a text-driven 3D human motion generation method that combines local generation and global fusion. Background Technology

[0002] With the rapid development of digital industries such as virtual reality, film and television special effects, game development, and the metaverse, the demand for technology that automatically generates realistic 3D human motion through natural language description is becoming increasingly urgent. This technology can significantly reduce the threshold and cost of 3D content creation and has significant application value.

[0003] Currently, the mainstream methods for text-driven 3D human motion generation are mainly based on generative adversarial networks (GANs) or diffusion models. However, existing technologies have several prominent problems: First, in terms of fine-grained control, the models struggle to accurately align text semantics with specific body parts, often resulting in semantic leakage. For example, when describing a "kicking" motion, the arm may exhibit unreasonable movements. Second, in terms of global coordination, the generated motions often lack natural biomechanical consistency, with incoordination between different body parts. Finally, in terms of visual effects, traditional neural radiation field rendering methods are computationally expensive and difficult to implement in real-time, while mesh-based representations lack rich surface details.

[0004] Therefore, there is an urgent need in this field for a technical solution that can simultaneously address the three key issues of fine local control, whole-body motion coordination, and high-quality real-time rendering. Summary of the Invention

[0005] The purpose of this invention is to provide a text-driven 3D human motion generation method that combines local generation and global fusion. This method provides a 3D human motion generation method that can achieve precise local control, ensure full-body coordination, and efficiently generate high-fidelity visual details.

[0006] To achieve the above functions, this invention designs a text-driven 3D human motion generation method that combines local generation and global fusion, executing the following steps S1-S6 to generate corresponding 3D human motion based on the user-input text:

[0007] Step S1: Receive the user's text description of human body movements in natural language;

[0008] Step S2: Decompose the action text description into multiple preset local action text descriptions of body parts;

[0009] Step S3: For each body part, construct a local motion generation network based on a conditional denoising network, input the local motion text description of each body part, and use a parallel 3D diffusion model to generate the corresponding motion latent code for each body part.

[0010] Step S4: Construct a whole-body motion fusion network, input the motion latent codes corresponding to each body part, perform feature splicing, and introduce a two-stage spatial fusion mechanism to fuse the motion latent codes corresponding to each body part into whole-body posture parameters.

[0011] Step S5: Convert the whole-body pose parameters into a 3D Gaussian point cloud;

[0012] Step S6: Optimize the visual details of the 3D Gaussian point cloud by fractional distillation sampling using a 2D diffusion model, and use a differentiable Gaussian sputtering renderer to render the optimized 3D Gaussian point cloud in real time, outputting the final 3D human motion.

[0013] As a preferred technical solution of the present invention: In step S1, the text description input by the user is simplified by removing role information from the text description to obtain a simplified text description. .

[0014] As a preferred technical solution of the present invention: In step S2, based on the GPT-4 parser, the text description is... The text description of the localized movements of six body parts is broken down as follows: ,in, Text descriptions representing localized movements of body parts. This represents one of the body parts, with a value of , , , , , The numbers represent the head, torso, left arm, right arm, left leg, and right leg, respectively. A collection of body parts; This indicates that the GPT-4 parser will process the text, marking any body parts not mentioned in the user's input text description as "do nothing".

[0015] As a preferred technical solution of the present invention, the specific steps of step S3 are as follows:

[0016] Step S3.1: Construct independent conditional denoising networks for each body part. For each independent conditional denoising network, use a local motion encoder pre-trained with large-scale human motion data to learn the pose data of each body part. Text description T of the corresponding part pMapping between; original text description Transformed into pose datasets for various body parts ;

[0017] Step S3.2: Map the pose data of each body part to the corresponding initial motion latent code using a local motion encoder. The diffusion process and the denoising process are performed. The diffusion process is as follows:

[0018] Record the action code corresponding to a body part as The corresponding attitude data is denoted as The diffusion process is modeled as a Markov noise-adding process as follows:

[0019] ;

[0020] in, This represents the diffusion process from adding noise at step t-1 to step t. It is noise sampled from a standard Gaussian distribution. It is the identity matrix. This represents the current diffusion step number. This represents the total number of steps in the diffusion process. Let t be the action code for step t. Let be the action latent code for step t-1. It is the diffusion step size;

[0021] The denoising process is the reverse of the diffusion process:

[0022] ;

[0023] in, This indicates that the denoising process for step t-1 is derived by working backward from step t. Represents variance. This represents the mean;

[0024] loss function for:

[0025] ;

[0026] in, Represents all random variables The joint distribution is averaged; The latent code representing the initial action corresponding to a body part; It is standard Gaussian noise; it follows a standard normal distribution. ; This indicates a trainable neural network.

[0027] As a preferred embodiment of the present invention, the specific steps of step S4 are as follows:

[0028] Step S4.1: Receive the motion latent codes corresponding to each body part, and construct initial global features through feature concatenation operations:

[0029] ;

[0030] in, Represents the initial global features. ; Represents the action codes corresponding to each body part; Indicates a splicing operation;

[0031] Step S4.2: Introduce a two-stage spatial fusion mechanism, wherein the first stage combines the initial global features with the semantic features of the complete action text description. The connection generates transition features through fully connected layers and layer normalization operations. The specific formula is as follows:

[0032] ;

[0033] in, , Indicates a TMR encoder; and These are the learnable weight matrix and bias vector; Presentation layer normalization operation;

[0034] The second stage employs a lightweight FFN network to perform spatial coordination optimization, using transitional input features. The optimized latent code is output as follows:

[0035] ;

[0036] in, This indicates an optimized latent code. , These are the parameters of the first fully connected layer. , For the parameters of the second fully connected layer, This represents the GELU activation function;

[0037] The optimized latent code is decoded into pose parameters of the parameterized human body model, which are used as the whole-body pose parameters, as shown in the following formula:

[0038] ;

[0039] in, Represents the pose parameters of a parameterized human body model; and These are the learnable linear decoder parameters.

[0040] As a preferred technical solution of the present invention, the specific method of step S5 is as follows:

[0041] From the human body mesh of the parametric human body model Among them To add a small amount of volume noise, extract the coordinates of all vertices. Centralized processing Eliminate position offset, The coordinates of the centered vertex. The center of the vertex coordinates is set; the vertex color is also randomly initialized. As the material base, among which Indicates a uniform distribution;

[0042] Constructing a 3D Gaussian point set ,in, Indicates the first A three-dimensional Gaussian point, The position of the i-th 3D Gaussian point is inherited from the vertex coordinates; Let the covariance matrix of the i-th 3D Gaussian point be initialized to the identity matrix. ; The transparency of the i-th 3D Gaussian point is fixed at 0.8; This represents the color of the i-th three-dimensional Gaussian point.

[0043] As a preferred embodiment of the present invention, the specific method of step S6 is as follows:

[0044] Eight camera poses were uniformly sampled in a spherical space with a radius of 3 units. , Representing the k-th camera pose, a multi-view image is generated using a differentiable Gaussian sputtering renderer. , Represents a differentiable Gaussian sputtering renderer. This represents the image from the k-th viewpoint. Represent the Gaussian point parameters; calculate the SDS loss gradient based on the two-dimensional diffusion model:

[0045] ;

[0046] In the formula, Indicates the parameters of the Gaussian point; This is the final SDS loss gradient; The weight function is time-step dependent; This indicates a two-dimensional diffusion model based on the time step of the k-th viewpoint with noise. Action code Text conditions and time step The predicted response Added noise, Standard Gaussian noise, Indicates at time step And the expectation under standard Gaussian noise conditions.

[0047] The present invention also designs an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the text-driven three-dimensional human motion generation method of local generation and global fusion.

[0048] The present invention also designs a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the text-driven three-dimensional human motion generation method of local generation and global fusion.

[0049] Beneficial effects: Compared with the prior art, the advantages of the present invention include:

[0050] This invention achieves precise semantic control of specific body parts through a collaborative design of local diffusion generation and global spatial fusion, while ensuring the natural coordination of the whole body posture. Furthermore, by combining the structural generation capabilities of a 3D diffusion model with the detail optimization capabilities of a 2D diffusion model, it significantly improves visual quality while maintaining motion accuracy, providing strong technical support for the field of 3D human motion generation. Attached Figure Description

[0051] Figure 1 This is a flowchart of a text-driven 3D human motion generation method based on local generation and global fusion according to an embodiment of the present invention;

[0052] Figure 2 This is a schematic diagram of the overall structure of the network model provided in an embodiment of the present invention;

[0053] Figure 3 This is a schematic diagram of a local action generation network provided according to an embodiment of the present invention;

[0054] Figure 4 This is a schematic diagram of a whole-body motion fusion network provided according to an embodiment of the present invention. Detailed Implementation

[0055] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0056] The text-driven 3D human motion generation method based on local generation and global fusion provided in this embodiment of the invention refers to... Figure 1 , Figure 2 Perform the following steps S1-S6 to generate corresponding 3D human body movements based on the text input by the user:

[0057] Step S1: Receive the user's text description of human body movements in natural language;

[0058] To ensure the generality of the generated actions and avoid the model being biased towards specific character features, the text description input by the user needs to be simplified. This involves removing character information from the input text, generalizing a specific character description such as "a character kicks with their left leg" to the standard action expression "a person kicks with their left leg." This yields the simplified text description. .

[0059] Step S2: Decompose the action text description into multiple preset local action text descriptions of body parts;

[0060] Reference Figure 3 Based on the GPT-4 parser, the text description is... The text description of the localized movements of six body parts is broken down as follows: ,in, Text descriptions representing localized movements of body parts. This represents one of the body parts, with a value of , , , , , The numbers represent the head, torso, left arm, right arm, left leg, and right leg, respectively. A collection of body parts; This indicates that the GPT-4 parser is processing the data. Generate precise action descriptions for each body part (such as "left leg quickly extends forward"), and label body parts not mentioned in the user's input text description as "do nothing".

[0061] Step S3: For each body part, construct a local motion generation network based on a conditional denoising network, input the local motion text description of each body part, and use a parallel 3D diffusion model to generate the corresponding motion latent code for each body part.

[0062] The specific steps of step S3 are as follows:

[0063] Step S3.1: For each body part, a parallel encoding architecture is used to construct an independent conditional denoising network for each of the six body parts. For each independent conditional denoising network, a local motion encoder pre-trained on large-scale human motion data is used to learn the pose data of each body part. Text description T of the corresponding part p Mapping between them; local motion encoders optimized for motion verbs, significantly enhancing the accuracy of local semantic representation; original text description Transformed into pose datasets for various body parts ;

[0064] Step S3.2: Map the pose data of each body part to the corresponding initial motion latent code using a local motion encoder. The process involves diffusion and denoising, where the diffusion process removes randomly sampled Gaussian noise. It begins by learning a denoising process, which gradually optimizes the initial action latent code into a meaningful one. The specific diffusion process is as follows:

[0065] Record the action code corresponding to a body part as The corresponding attitude data is denoted as The diffusion process is modeled as a Markov noise-adding process as follows:

[0066] ;

[0067] in, This represents the diffusion process from adding noise at step t-1 to step t. It is noise sampled from a standard Gaussian distribution. It is the identity matrix. This represents the current diffusion step number. This represents the total number of steps in the diffusion process. Let t be the action code for step t. Let be the action latent code for step t-1. It is the diffusion step size; It is the diffusion step size; the initial action latent code for this process. Noise will be added gradually until... It follows a standard normal distribution.

[0068] The denoising process is the reverse of the diffusion process:

[0069] ;

[0070] in, This indicates that the denoising process for step t-1 is derived by working backward from step t. Representing variance, it is usually set as the identity matrix. This represents the mean, which is predicted by a neural network.

[0071] loss function for:

[0072]

[0073] in, Represents all random variables The joint distribution is averaged; The latent code representing the initial action corresponding to a body part; It is standard Gaussian noise; it follows a standard normal distribution. ; This indicates a trainable neural network.

[0074] Step S4: Construct a whole-body motion fusion network, input the motion latent codes corresponding to each body part, perform feature splicing, and introduce a two-stage spatial fusion mechanism to fuse the motion latent codes corresponding to each body part into whole-body posture parameters.

[0075] Reference Figure 4 The specific steps of step S4 are as follows:

[0076] Step S4.1: Receive the six motion latent codes corresponding to each body part, and construct the initial global features through feature concatenation operations:

[0077] ;

[0078] in, Represents the initial global features. ; Represents the action codes corresponding to each body part; Indicates a splicing operation;

[0079] Step S4.2: Introduce a two-stage spatial fusion mechanism, wherein the first stage combines the initial global features with the semantic features of the complete action text description. The connection generates transition features through fully connected layers and layer normalization operations. The specific formula is as follows:

[0080] ;

[0081] in, , Indicates TMR encoder, It carries the overall semantic meaning of the action; and These are the learnable weight matrix and bias vector; Presentation layer normalization operation;

[0082] The second stage employs a lightweight FFN network to perform spatial coordination optimization, using transitional input features. The optimized latent code is output as follows:

[0083] ;

[0084] in, This indicates an optimized latent code. , This represents the GELU activation function, which simulates the nonlinear elastic properties of human tendons. , These are the parameters of the first fully connected layer. , The parameters of the second fully connected layer are used to map the features back to 512 dimensions, which serve as the final optimized whole-body latent code.

[0085] In the spatial domain, cross-part mechanical coupling is established, and the optimization latent code is decoded into parameters of the parametric human body model (SMPL), which are used as whole-body posture parameters, as shown in the following formula:

[0086] ;

[0087] in, Represents the pose parameters of a parameterized human body model; and These are the learnable linear decoder parameters.

[0088] Step S5: Convert the whole-body pose parameters into a 3D Gaussian point cloud;

[0089] The specific method is as follows:

[0090] Initialize the Gaussian points from the human body mesh of the parametric human body model. Among them To add a small amount of volume noise, extract the coordinates of all vertices. Centralized processing Eliminate position offset, The coordinates of the centered vertex. The center of the vertex coordinates is set; the vertex color is also randomly initialized. As the material base, among which Indicates a uniform distribution;

[0091] Constructing a 3D Gaussian point set ,in, Indicates the first A three-dimensional Gaussian point, The position of the i-th 3D Gaussian point is inherited from the vertex coordinates; Let the covariance matrix of the i-th 3D Gaussian point be initialized to the identity matrix. ; The transparency of the i-th 3D Gaussian point is fixed at 0.8 to ensure initial visibility and form a point cloud skeleton with basic anatomical structure; This represents the color of the i-th three-dimensional Gaussian point.

[0092] Step S6: Optimize the visual details of the 3D Gaussian point cloud by fractional distillation sampling using a 2D diffusion model, and use a differentiable Gaussian sputtering renderer to render the optimized 3D Gaussian point cloud in real time, outputting the final 3D human motion.

[0093] The specific method is as follows:

[0094] Appearance details were optimized using a two-dimensional (2D) diffusion model, with eight camera poses uniformly sampled in a spherical space with a radius of 3 units. , Representing the k-th camera pose, a multi-view image is generated using a differentiable Gaussian sputtering renderer. , Represents a differentiable Gaussian sputtering renderer. This represents the image from the k-th viewpoint. Represents the Gaussian point parameters; based on a two-dimensional diffusion model, specifically using the Flux diffusion model as the 2D visual prior, the SDS loss gradient is calculated:

[0095] ;

[0096] In the formula, Indicates the parameters of the Gaussian point; The final SDS loss gradient can be obtained by calculating the expectation and multiplying it by the rendering gradient, resulting in a gradient that can be directly used to update the three-dimensional (3D) Gaussian parameters θ. The weight function is time-step dependent; This indicates a two-dimensional diffusion model based on the time step of the k-th viewpoint with noise. Action code Text conditions and time step The predicted response Added noise, Standard Gaussian noise, Indicates at time step And the expectation under standard Gaussian noise conditions.

[0097] The SDS loss gradient iteratively optimizes the Gaussian point parameters. The main focus is on updating the colors. With transparency This allows the rendering results to approximate the visual features of the text description within the latent space of the diffusion model, especially enhancing details such as skin texture wrinkles and muscle stretching and deformation.

[0098] This invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the text-driven three-dimensional human motion generation method of local generation and global fusion.

[0099] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the text-driven 3D human motion generation method of local generation and global fusion.

[0100] To verify the effectiveness of the invention and the model framework, a quantitative analysis method was used, and experiments were conducted on a computing platform with an NVIDIA GeForce RTX 4090 graphics card using the ThreeStudio framework in PyTorch. The evaluation metrics used were CLIP-S (the CLIP similarity between generated motion rendering frames and text) and part-level motion matching accuracy to assess semantic alignment performance.

[0101] The CLIP metric calculates the cosine similarity between the multi-view rendered image of the generated action and the input text in the CLIP model's latent space. The formula for this calculation is:

[0102] ;

[0103] in and These are CLIP's image encoder and text encoder, respectively. For the k-th viewpoint image, The input is the original text description, and K is the total number of viewpoints (K=8 in this experiment). The higher the CLIP-S value, the better the semantic consistency between the visual content of the generated action and the text description.

[0104] Table 1 presents the CLIP metric values ​​for different methods. This comparative experiment used four cameras to generate 120 rendered images from different angles, calculated the similarity between each selected test image and the text, and took the average. The CLIP similarity was calculated using OpenAI's VIT-L / 14 and Openclip's ViT-bigG-14 models.

[0105] Table 1. Comparative Experimental Results

[0106]

[0107] The experimental results are shown in Table 1. Compared with GaussianDreamer, the method of the present invention achieves performance improvements of 8% and 5% under ViT-L / 14 and ViT-bigG-14 computation methods, respectively.

[0108] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for text-driven 3D human motion generation with local generation and global fusion, characterized in that, Perform the following steps S1-S6 to generate corresponding 3D human motion based on the text input by the user: Step S1: Receive the user's text description of human body movements in natural language; Step S2: Decompose the action text description into multiple preset local action text descriptions of body parts; Step S3: For each body part, construct a local motion generation network based on a conditional denoising network, input the local motion text description of each body part, and use a parallel 3D diffusion model to generate the corresponding motion latent code for each body part. Step S4: Construct a whole-body motion fusion network, input the motion latent codes corresponding to each body part, perform feature splicing, and introduce a two-stage spatial fusion mechanism to fuse the motion latent codes corresponding to each body part into whole-body posture parameters. The specific steps of step S4 are as follows: Step S4.1: Receive the motion latent codes corresponding to each body part, and construct initial global features through feature concatenation operations: ； wherein, represents an initial global feature, ; represents an action latent code corresponding to each body part; represents a concatenation operation; Step S4.2: Introduce a two-stage spatial fusion mechanism, wherein the first stage combines the initial global features with the semantic features of the complete action text description. The connection generates transition features through fully connected layers and layer normalization operations. The specific formula is as follows: ； in, , Indicates a TMR encoder; and The weight matrix and bias vector are learnable. Presentation layer normalization operation; The second stage employs a lightweight FFN network to perform spatial coordination optimization, using transitional input features. The optimized latent code is output as follows: ； in, This indicates an optimized latent code. , These are the parameters of the first fully connected layer. , For the parameters of the second fully connected layer, This represents the GELU activation function; The optimized latent code is decoded into pose parameters of the parameterized human body model, which are used as the whole-body pose parameters, as shown in the following formula: ； in, Represents the pose parameters of a parameterized human body model; and These are the learnable linear decoder parameters; Step S5: Convert the whole-body pose parameters into a 3D Gaussian point cloud; Step S6: Optimize the visual details of the 3D Gaussian point cloud by fractional distillation sampling using a 2D diffusion model, and use a differentiable Gaussian sputtering renderer to render the optimized 3D Gaussian point cloud in real time, outputting the final 3D human motion.

2. The text-driven 3D human motion generation method based on local generation and global fusion according to claim 1, characterized in that, In step S1, the text description input by the user is simplified by removing role information from the text description to obtain a simplified text description. .

3. The text-driven 3D human motion generation method based on local generation and global fusion according to claim 2, characterized in that, In step S2, the text description is parsed using the GPT-4 parser. The text description of the localized movements of six body parts is broken down as follows: ,in, This represents the parsed text description of the local movements of a body part. This represents one of the body parts, with a value of , , , , , The numbers represent the head, torso, left arm, right arm, left leg, and right leg, respectively. A collection of body parts; This indicates that the GPT-4 parser is processing the text; body parts not mentioned in the user's input text description are labeled as "donothing".

4. The text-driven 3D human motion generation method based on local generation and global fusion according to claim 3, characterized in that, The specific steps of step S3 are as follows: Step S3.1: Construct independent conditional denoising networks for each body part. For each independent conditional denoising network, use a local motion encoder pre-trained with large-scale human motion data to learn the pose data of each body part. Original text description T of the corresponding part p Mapping between; original text description Transformed into pose datasets for various body parts ; Step S3.2: Map the pose data of each body part to the corresponding initial motion latent code using a local motion encoder. The diffusion process and the denoising process are performed. The diffusion process is as follows: Record the action code corresponding to a body part as The corresponding attitude data is denoted as The diffusion process is modeled as a Markov noise-adding process as follows: ； in, This represents the diffusion process from adding noise at step t-1 to step t. It is noise sampled from a standard Gaussian distribution. It is the identity matrix. This represents the current diffusion step number. This represents the total number of steps in the diffusion process. Let t be the action code for step t. Let be the action latent code for step t-1. It is the diffusion step size; The denoising process is the reverse of the diffusion process: ； in, This indicates that the denoising process for step t-1 is derived by working backward from step t. Represents variance. This represents the mean; loss function for: ； in, Represents all random variables The joint distribution is averaged; The latent code representing the initial action corresponding to a body part; It is standard Gaussian noise; it follows a standard normal distribution. ; This indicates a trainable neural network.

5. The text-driven 3D human motion generation method based on local generation and global fusion according to claim 4, characterized in that, The specific method for step S5 is as follows: From the human body mesh of the parametric human body model ,in To add a small amount of volume noise, extract the coordinates of all vertices. Centralized processing Eliminate position offset, The coordinates of the centered vertex. The center of the vertex coordinates is set; the vertex color is also randomly initialized. As the material base, among which Indicates a uniform distribution; Constructing a 3D Gaussian point set Where i represents the i-th three-dimensional Gaussian point; The position of the i-th 3D Gaussian point is inherited from the vertex coordinates; Let the covariance matrix of the i-th 3D Gaussian point be initialized to the identity matrix. ; The transparency of the i-th 3D Gaussian point is fixed at 0.8; This represents the color of the i-th three-dimensional Gaussian point.

6. The text-driven 3D human motion generation method based on local generation and global fusion according to claim 5, characterized in that, The specific method for step S6 is as follows: Eight camera poses were uniformly sampled in a spherical space with a radius of 3 units. , Representing the k-th camera pose, a multi-view image is generated using a differentiable Gaussian sputtering renderer. , Represents a differentiable Gaussian sputtering renderer. This represents the image from the k-th viewpoint. Represent the Gaussian point parameters; calculate the SDS loss gradient based on the two-dimensional diffusion model: ； In the formula, Indicates the parameters of the Gaussian point; This is the final SDS loss gradient; The weight function is time-step dependent; This indicates a two-dimensional diffusion model based on the time step of the k-th viewpoint with noise. Action code Text conditions and time step The predicted response Added noise, Standard Gaussian noise, Indicates at time step And the expectation under standard Gaussian noise conditions.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the text-driven 3D human motion generation method of local generation and global fusion as described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the text-driven 3D human motion generation method of local generation and global fusion as described in any one of claims 1 to 6.

Images