Three-dimensional garment generation method based on Gaussian model

Abstract

The invention relates to a Gaussian model-based three-dimensional garment generation method, which comprises the following steps of: analyzing data of human body shape parameters, garment types and human body textures from text cues, and initializing and decoupling a garment three-dimensional Gaussian model by using the data; human body shape parameters, clothing types and human body texture data serve as point cloud input, a single three-dimensional Gaussian model is constructed, and the model comprises a clothing vertex set and a body vertex set; inputting the clothing vertex set and the body vertex set into a differentiable renderer, and respectively outputting color rendering graphs and depth rendering graphs of the clothing and the human body; and inputting the text cue word and the rendering graph into the optimized single three-dimensional Gaussian model to generate three-dimensional clothes. Compared with the prior art, the method has the advantages that decoupling of the clothes and the body is achieved, the quality of the produced three-dimensional clothes is higher, and the visual impression is more fidelity.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and more particularly to a three-dimensional clothing generation method based on a Gaussian model. Background Art

[0002] With the increasing popularity of applications such as fashion design, immersive interaction, and virtual fitting, the demand for three-dimensional clothing generation has become increasingly prominent. Clothing, as a key component in the creation of three-dimensional human bodies, determines the realism and usability of three-dimensional virtual characters. It has been widely used in application scenarios such as digital entertainment, Augmented Reality (AR), Virtual Reality (VR), and Mixed Reality (MR).

[0003] In the existing related technologies, there are mainly two ways to generate three-dimensional clothing: First, the generation method based on mesh representation has generality and convenience in the industrial pipeline. However, without post-processing, it is difficult to model complex cloth materials. Second, the method based on implicit representation significantly improves the realism through variational scoring distillation loss, but the generated clothing is coupled with the human body, restricting additional operations such as style modification, virtual dressing, or simulating the real interaction between the clothing and the human body, especially for loose clothing such as long dresses and gowns, and the training process requires extremely large computing resources.

[0004] Therefore, the current clothing generation methods still cannot meet the actual industrial requirements. How to generate three-dimensional clothing that is fully decoupled and has high visual quality remains the current research focus.

[0005] After retrieval, Chinese Patent Application Publication No. CN118781236A discloses a method for generating loose clothing animations for virtual fashion shows, including the following steps: using SMPL human motion sequences and clothing templates as input data; 2) constructing and training a rough clothing deformation prediction network model, and feeding the motion sequences and clothing mesh data into the trained rough clothing deformation prediction network to obtain rough clothing deformations; 3) introducing clothing fit and deformation attributes to enrich the clothing mesh features, and constructing and training a fine clothing deformation prediction network model to obtain loose clothing skinning weights and fine clothing deformations; 4) using the linear blend skinning algorithm to perform skinning deformation on the predicted clothing template to obtain the clothing deformation in the corresponding pose; solving the problem that local wrinkles are not obvious or disappear during the deformation of loose clothing in virtual fashion shows, and enriching the wrinkles of loose three-dimensional clothing during the deformation according to clothing attributes, motion categories, and body motion sequences. This existing patent application has the problem that the clothing and the human body are not decoupled, resulting in deviations in real interaction and real-time performance.

[0006] How to achieve high-visual-quality 3D clothing generation that is fully decoupled from the human body has become a technical problem to be solved. Summary of the Invention

[0007] The purpose of the present invention is to provide a 3D clothing generation method based on a Gaussian model to overcome the defects of the existing technologies described above.

[0008] The purpose of the present invention can be achieved through the following technical solutions:

[0009] According to one aspect of the present invention, there is provided a 3D clothing generation method based on a Gaussian model, the method comprising:

[0010] Step 1: Parse the human body shape parameters, clothing type, and human body texture data from the text prompt, and use this data to initialize and decouple the 3D Gaussian model of the clothing;

[0011] Step 2: Use the human body shape parameters, clothing type, and human body texture data as point cloud inputs to construct a single 3D Gaussian model, which includes a clothing vertex set and a body vertex set;

[0012] Step 3: Input the clothing vertex set and the body vertex set in Step 2 into a differentiable renderer, and respectively output the color rendering map and depth rendering map of the clothing and the human body;

[0013] Step 4: Input the text prompt and the rendering maps in Step 3 into the optimized single 3D Gaussian model to generate 3D clothing.

[0014] Preferably, the Step 1 includes:

[0015] 1-1: Parse the human body shape parameters, clothing type, and human body texture data from the text prompt, use this data to initialize and decouple the 3D Gaussian model of the clothing, select a parameterized human body model according to the human body shape parameters, and obtain the clothing mesh body and the human body mesh body according to the clothing type and the parameterized human body model;

[0016] 1-2: Uniformly sample according to the vertex set of the clothing type according to the ratio of the clothing mesh body and the human body mesh body to obtain the clothing vertex set and the body vertex set;

[0017] 1-3: Uniformly scale all the clothing vertex sets, set the mean color, and do not apply rotation;

[0018] 1-4: Uniformly scale all the body vertex sets, do not apply rotation, sample the color of the input texture, and assign the corresponding color to each body point.

[0019] Preferably, the process of constructing the single 3D Gaussian model includes:

[0020] 2-1: Set two corresponding groups of labels for the clothing vertex set and the body vertex set respectively;

[0021] 2-2: Merge the clothing vertex set and the body vertex set, and perform unified spatial alignment and scaling on the merged vertex set;

[0022] 2-3: Set the single three-dimensional Gaussian model parameters and gradient masks according to the clothing vertex set, the body vertex set, and their respective corresponding labels;

[0023] 2-4: Use the stochastic gradient descent method to estimate the parameters of the three-dimensional Gaussian to fit the scene.

[0024] Preferably, for the depth value d(p) of N ordered Gaussian instances overlapping on a pixel in the single three-dimensional Gaussian model, perform point-based α blending calculation to obtain:

[0025]

[0026] where d i is the depth value of the i-th Gaussian, and α i is the opacity of the i-th Gaussian point, is the value defined by the i-th Gaussian at the query point p.

[0027] Preferably, use a differentiable renderer to output a color rendering map and a depth rendering map for the clothing and the human body respectively, specifically:

[0028]

[0029] where x b and d b are the color rendering map and the depth map of the human body respectively, and x c and d c are the color rendering map and the depth map of the clothing respectively, is the human body part of the single three-dimensional Gaussian model, is the clothing part of the single three-dimensional Gaussian model, is the human body part of the differentiable renderer, is the clothing part of the differentiable renderer.

[0030] Preferably, step 3 further includes: estimating the human body pose from the color rendering map using the DWPose detection network.

[0031] More preferably, the position update of the three-dimensional Gaussian points in the single three-dimensional Gaussian model follows exponential decay, and the mean absolute error loss and the structural similarity loss are used in training;

[0032] The single three-dimensional Gaussian model uses an adaptive clothing cutting strategy for geometric constraints.

[0033] Preferably, the single three-dimensional Gaussian model uses a bidirectional fractional distillation loss to simultaneously supervise the rendered images of the clothing and the human body, specifically:

[0034]

[0035]

[0036] Among them, λ 1 and λ 2 are coefficients for balancing the influence weights between the RGB image and the depth image, x b-t , d b-t and are respectively the color rendering image, the depth image, and the initial noise of the human body at time t, x c-t , d c-t and are respectively the color rendering image, the depth image, and the initial noise of the clothing at time step t; is the loss function for the human body rendering image; is the loss function for the individual clothing rendering image; E[*] is the mathematical expectation of the image at time step t, is the mathematical expectation of the human body rendering image at time step t, is the mathematical expectation of the human body depth image at time step t; is the mathematical expectation of the clothing rendering image at time step t; is the mathematical expectation of the clothing depth image at time step t; w t is to control the noise schedule and is determined by the sampler of the diffusion model; ∈ φ is the noise predictor; P c is the three-dimensional Gaussian point set of the clothing part; y is the text input; θ is the training parameter.

[0037] Preferably, an adaptive clothing cutting strategy is introduced to perform geometric constraint training on the single three-dimensional Gaussian model.

[0038] Preferably, the method further includes step 5: registering the three-dimensional clothing template mesh sequence to obtain an animatable three-dimensional clothing, specifically:

[0039] 5-1. Use the iterative closest point algorithm to achieve the initial alignment between the three-dimensional clothing and the first clothing mesh;

[0040] 5-2: Find the nearest clothing mesh point of the three-dimensional clothing Gaussian point;

[0041] 5-3: Use the deformation matrix between the clothing mesh vertices to calculate the transformation of the three-dimensional clothing Gaussian point, and finally output the animatable three-dimensional clothing.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1) The present invention parses data on human body parameters, clothing types and human body texture from text prompt words, uses these data to initialize a decoupled clothing three-dimensional Gaussian model, constructs a single three-dimensional Gaussian model, effectively realizes the decoupling of clothing and body, inputs a clothing vertex set and a body vertex set into a differentiable renderer, inputs the output rendering image and the obtained text prompt words into the optimized single three-dimensional Gaussian model based on decoupling, generates three-dimensional clothing, and only updates the clothing vertex set during optimization. The generated three-dimensional clothing has higher quality and more authentic visual perception, while reducing the demand for computing resources.

[0044] 2) The present invention proposes to use a bidirectional fractional distillation loss to optimize a single 3D Gaussian model, optimize the 3D character clothing under the conditions of a color rendering image, a depth image, and a posture image, and improve the visual quality of the 3D clothing.

[0045] 3) The single three-dimensional Gaussian model of the present invention uses an adaptive clothing cutting strategy for geometric constraints, which can improve the geometric integrity of loose clothing during the generation process.

[0046] 4) Compared with the existing 3D clothing generation methods, the generated 3D clothing by the method of the present invention is superior to the current most advanced methods in terms of appearance and geometric quality, and can further realize realistic clothing animation under the prior of the registration grid sequence, and can realize virtual dressing and realistic clothing animation, which will effectively promote the high-quality development of digital interactive entertainment. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of the process of the present invention;

[0048] Figure 2 It is a schematic diagram of the principle of the method of the present invention;

[0049] Figure 3 A schematic diagram of a clothing vertex set corresponding to a parameterized human body model in the present invention;

[0050] Figure 4 This is a high-fidelity clothing rendering of the three-dimensional clothing generation method based on the Gaussian model in the present invention.

[0051] Figure 5 Comparison of clothing renderings generated by the method of the present invention and the existing related technology. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0053] Example 1

[0054] This embodiment relates to a 3D clothing generation method based on a Gaussian model, which has a stronger ability to decouple clothing from the human body, generates 3D clothing with higher geometric quality and more authentic visual perception, and the entire model training process is low-cost and does not require expensive hardware equipment. Figure 1 , the steps are as follows:

[0055] Step 1: Initialization and preprocessing of the 3D character model: parse the acquired text prompt words to obtain data on human body parameters, clothing type, and human body texture, use these data to initialize the decoupled clothing 3D Gaussian model, select a parameterized human body model based on the human body parameters, and perform geometric uniform sampling on the parameterized human body model;

[0056] Step 2: Take the data of human body parameters, clothing type and human body texture as point cloud input, build a single 3D Gaussian model, including the vertex set of clothing and body, and perform parameter optimization and gradient transfer on some vertices in 3D space;

[0057] Step 3: Input the vertex sets of clothing and body into the differentiable renderer, output the color rendering and depth rendering of clothing and body respectively, and use the bidirectional fractional distillation loss to supervise the pre-trained diffusion model to perform fractional distillation optimization on the single 3D Gaussian model constructed in step 2;

[0058] Step 4: Introduce an adaptive clothing cutting strategy (cutting strategy for loose clothing), and geometrically constrain the model trained in step 3 in the network training module; input the text prompt words and renderings into the optimized single 3D Gaussian model to generate 3D clothing.

[0059] Step 5: Register the template mesh sequence and post-process the 3D garment output from step 4 to obtain a drivable 3D garment with a clear geometric structure.

[0060] Example 2

[0061] This embodiment relates to a 3D clothing generation method based on a Gaussian model, which differs from Embodiment 1 in the following aspects:

[0062] The initialization and preprocessing process of the decoupled clothing 3D Gaussian model in step 1 includes the following operations:

[0063] 1-1: Parse the human body shape parameters, clothing type, and human body texture data from the obtained text prompts, use these data to initialize and decouple the 3D Gaussian model of the clothing, and select a parametric human body model according to the human body shape parameters;

[0064] 1-2: According to the vertex set of the clothing type, sample clothing points according to the ratio of the clothing mesh body and the human body mesh body. By default, the human body is initialized with 100,000 vertices; the clothing vertex set corresponding to the parametric human body model is as Figure 3 shown.

[0065] The calculation expression for the clothing vertex sampling quantity is as follows:

[0066]

[0067] Among them, P c is the sampling quantity of the clothing vertices, P b is the sampling quantity of the human body points, idx is the predefined vertex set, and V b is the total number of vertices of the human body mesh body.

[0068] 1-3: Perform uniform scaling on all clothing vertex sets, set the mean color, and do not apply rotation;

[0069] 1-4: Perform uniform scaling on all body vertex sets, do not apply rotation, sample the color of the input texture, and assign the corresponding color to each body point.

[0070] The process of constructing a single 3D Gaussian model in step 2 includes the following operations:

[0071] 2-1: Set two corresponding sets of labels for the vertex sets of the clothing and the body respectively;

[0072] 2-2: Merge the clothing vertex set and the body vertex set, and perform unified spatial alignment and scaling on the two vertex sets;

[0073] 2-3: Set the single 3D Gaussian model parameters and gradient masks according to the clothing vertex set, body vertex set, and corresponding labels;

[0074] 2-4: Adopt the stochastic gradient descent method to reasonably fit the scene by estimating the parameters of the 3D Gaussian;

[0075] In this embodiment, the body and clothing of the 3D character are modeled as a single 3D Gaussian model. For rendering the multi-view depth maps of the 3D character, this embodiment obtains them by performing point-based α blending calculation on the depth values of N ordered Gaussian instances overlapping on the pixels:

[0076]

[0077] Among them, di is the depth value of the i-th Gaussian, α i is the opacity of the i-th Gaussian point, is the value defined by the i-th Gaussian at the query point p.

[0078] It should be noted that the single-step depth data in the initial stage of training cannot accurately determine the occlusion relationship between two objects. Directly merging the rendered images of the body and clothing through depth will cause the rendered images to be partially transparent and produce unsatisfactory results. In addition, using the transparency rendering as a preprocessing mask to merge the rendered images of the body and clothing may result in uneven black edges due to the ellipsoidal representation characteristics of the three-dimensional Gaussian points. Therefore, if the character's clothing and body are represented as two Gaussian models, it is impossible to fit the clothing to the body during the overall optimization process, resulting in errors during the optimization generation process.

[0079] 2-5: The position update of the three-dimensional Gaussian points follows exponential decay, and the mean absolute error loss and structural similarity loss are used during training;

[0080] The expression of the loss function is as follows:

[0081]

[0082] where, is the mean absolute error loss, is the structural similarity loss, λ is a constant used to balance the two losses, and λ = 0.2 is used in this embodiment.

[0083] During each backpropagation process, the sampled body points are tracked, and their gradients are masked after backpropagation. This embodiment uses the method of masking the gradients to avoid optimizing the selected set of three-dimensional Gaussian of the body. During the adaptive densification process, the gradient accumulation matrix of the three-dimensional Gaussian points with respect to the position parameters is masked, and only partial valid gradients are used for update in this embodiment until rasterization is completed. In addition, only the clothing three-dimensional Gaussian points are selected for densification or pruning judgment during optimization to reduce the computational overhead.

[0084] 2-6: The single three-dimensional Gaussian model is trained using the Adam optimization algorithm, with a learning rate of 0.00001, a batch size of 8, and a total of 3000 iterations of training.

[0085] The process of score distillation optimization in step 3, as Figure 2 , includes the following operations:

[0086] 3-1: Use a differentiable renderer to render the color map and depth map of the model in step 2 respectively, with a resolution of 1024x1024;

[0087] Benefiting from the single three-dimensional Gaussian model created in step 2, the present invention can achieve the separated rendering of the clothing of a three-dimensional character. The rendering process expression of the differentiable renderer is as follows:

[0088]

[0089] where x b and d b are respectively the color rendering map and the depth map of the human body, and x c and d c are respectively the color rendering map and the depth map of the clothing, is the human body part of the single three-dimensional Gaussian model, is the clothing part of the single three-dimensional Gaussian model, is the human body part of the differentiable renderer, is the clothing part of the differentiable renderer.

[0090] 3-2: Estimate the human body pose from the color map using the DWPose detection network;

[0091] 3-3: Inject the pose map and the text prompt into the pre-trained bidirectional diffusion model at the same time to generate the target image;

[0092] Specifically, in order to inject the dual-channel information of RGB and depth, the present invention uses the UNet network of the diffusion model as the basis and introduces additional pose conditions by means of channel-level splicing.

[0093] 3-4: Use the denoising loss to simultaneously supervise the rendering maps of the individual clothing and the character (i.e., the human body);

[0094] The expression of the denoising loss function is as follows:

[0095]

[0096] where λ 1 and λ 2 are the coefficients for balancing the influence weights between the RGB map and the depth map, x b-t , d b-t and are respectively the color rendering map, the depth map and the initial noise of the human body at time t, and x c-t , d c-t and are respectively the color rendering map, the depth map and the initial noise of the clothing at time step t; is the loss function for the human body rendering map; is the loss function for the individual clothing rendering map; E[*] is the mathematical expectation of the image at time step t, is the mathematical expectation of the human body rendering map at time step t, is the mathematical expectation of the human body depth map at time step t; is the mathematical expectation of the clothing rendering at time step t; is the mathematical expectation of the clothing depth map at time step t; w t To control the noise scheduling, it is determined by the sampler of the diffusion model; ∈ φ is the noise predictor; P c is the three-dimensional Gaussian point set of the clothing part; y is the text input; θ is the training parameter.

[0097] According to the differentiable renderer The obtained rendering and the character rendering are calculated at the same time and clothing renderings Overseeing characters and costumes.

[0098] 3-5: The camera distance range is [1.5, 2.0], the field of view range is [40°, 70°], and the elevation range is:

[0099] [-30°, 30°], and the azimuth range is [-180°, 180°].

[0100] The adaptive garment cutting strategy in step 4 includes the following operations:

[0101] 4-1: For the model optimization process in step 3, adaptive 3D Gaussian point encryption and pruning are performed every 100 steps between 300 and 800 iterations;

[0102] 4-2: A total of 2500 iterations are performed, with a pruning at 1800 steps and a scaling factor of 0.5.

[0103] The registration template grid sequence in step 5 includes the following operations:

[0104] 5-1: Extract the 3D garment generated in step 4 and use ICP registration (iterative closest point algorithm) to achieve preliminary alignment between the garment and the first simulated mesh;

[0105] 5-2: Accelerate and optimize the nearest neighbor search and build a KD tree (K-Dimensional Tree) to find the nearest grid point of each clothing Gaussian point;

[0106] 5-3: Use the deformation matrix between mesh vertices to calculate the transformation of clothing Gaussian points, and finally output an animated three-dimensional clothing.

[0107] Example 3

[0108] This embodiment relates to the verification of a three-dimensional clothing generation method based on a Gaussian model.

[0109] The performance of the method of the present invention is verified by comparing the generation of static clothing and dynamic clothing with the existing state-of-the-art text-guided 3D generation methods. Using the method of the present invention, 3D clothing is generated from the text generated by randomly selecting 100 language models, and is respectively compared with the existing state-of-the-art 3D generation methods: Shap-E (Related Technology 1), ProlificDreamer (Related Technology 2), LGM (Related Technology 3), GaussianDreamer (Related Technology 4) and DreamGaussian (Related Technology 5). Six texts are randomly selected from 100 texts for display, and the comparison effects are generated as shown in Figure 4 shown. The generation method of the present invention shows the highest overall quality and the finest details in terms of geometric shape and texture details. In addition, it can be seen from Figure 4 that in the case of the same text input, the existing related technologies may result in incomplete clothing or additional human body parts being generated, while the method of the present invention is more robust in the task of clothing generation and can stably generate 3D clothing assets that conform to the text description.

[0110] As shown in Figure 5 shown, the present invention also compares the generation of virtual character clothing for dynamic clothing with HumanGaussian ( Figure 5 referred to as related technology in it). When integrating with the body, HumanGaussian often generates blurred hand and head shadows on the clothes and performs poorly in dealing with loose clothing. Two sets of template mesh sequences are used, and 4 frames are randomly selected from the total 180 rendered images for analysis by the method of the present invention. By linearly blending and binding the entire virtual character with a parametric human model, the animation generated by HumanGaussian will produce many artifacts, especially in the knee area. Compared with other clothing generation methods, the method of the present invention maintains a higher overall geometric integrity of the clothing, and is closer to reality in terms of texture appearance, which is more obvious in the task of generating complex loose clothing. In addition, it supports more free user personalized clothing generation and can achieve realistic clothing animation.

[0111] As mentioned above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of various equivalent modifications or substitutions, and these modifications or substitutions should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A three-dimensional clothing generation method based on a Gaussian model, characterized in that: The method comprises: Step 1: parse the data of human body parameters, clothing type and human body texture from the text prompt words, and use these data to initialize the decoupled clothing 3D Gaussian model; Step 2: Take the data of human body parameters, clothing type and human body texture as point cloud input to construct a single 3D Gaussian model, which includes a clothing vertex set and a body vertex set; Step 3: Input the clothing vertex set and body vertex set in step 2 into the differentiable renderer, and output the color rendering image and depth rendering image of the clothing and body respectively; Step 4: Input the text prompt words and the rendering image of step 3 into the optimized single 3D Gaussian model to generate 3D clothing.

2. A three-dimensional clothing generation method based on a Gaussian model according to claim 1, characterized in that: The step 1 comprises: 1-1: Parse the data of human body parameters, clothing type and human body texture from the text prompt words, use these data to initialize the decoupled clothing 3D Gaussian model, select the parameterized human body model according to the human body parameters, and obtain the clothing mesh and human body mesh according to the clothing type and parameterized human body model; 1-2: According to the vertex set of the clothing type, uniform sampling is performed according to the ratio of the clothing mesh and the human body mesh to obtain the clothing vertex set and the body vertex set; 1-3: Uniformly scale all clothing vertex sets, set mean color, and do not apply rotation; 1-4: Uniformly scale all body vertex sets, without applying rotation, sample the color of the input texture, and assign the corresponding color to each body point.

3. The method for generating three-dimensional clothing based on a Gaussian model according to claim 1, characterized in that: The process of constructing a single three-dimensional Gaussian model includes: 2-1: Set two sets of corresponding labels for the clothing vertex set and the body vertex set respectively; 2-2: Merge the clothing vertex set and the body vertex set, and perform uniform spatial alignment and scaling on the merged vertex set; 2-3: According to the clothing vertex set, body vertex set and their corresponding labels, set the single 3D Gaussian model parameters and gradient mask; 2-4: Use stochastic gradient descent to estimate the parameters of the three-dimensional Gaussian to fit the scene.

4. The method for generating three-dimensional clothing based on a Gaussian model according to claim 1, characterized in that: The single 3D Gaussian model performs point-based alpha blending calculation on the depth values ​​d(p) of N ordered Gaussian instances overlapping on the pixel to obtain: Among them, d i is the depth value of the i-th Gaussian, α i is the opacity of the ith Gaussian point, is the value defined by the i-th Gaussian at the query point p.

5. The method for generating three-dimensional clothing based on a Gaussian model according to claim 1, characterized in that: Use a differentiable renderer to output color renderings and depth renderings for clothing and human bodies, respectively. Specifically: Among them, x b and d b They are the color rendering and depth map of the human body, x c and d c They are the color rendering and depth map of the clothing. is the human body part of a single 3D Gaussian model, is the clothing part of a single 3D Gaussian model, is the human body part of the differentiable renderer, It is the clothing part of the differentiable renderer.

6. The method for generating three-dimensional clothing based on a Gaussian model according to claim 1, characterized in that: Step 3 also includes: estimating human posture from the color rendering image using the DWPose detection network.

7. The method for generating three-dimensional clothing based on a Gaussian model according to claim 6, characterized in that: The position update of the three-dimensional Gaussian points in the single three-dimensional Gaussian model follows exponential decay, and the mean absolute error loss and the structural similarity loss are used in the training; The single 3D Gaussian model is geometrically constrained using an adaptive clothing clipping strategy.

8. The method for generating three-dimensional clothing based on a Gaussian model according to claim 1, characterized in that: The single 3D Gaussian model uses a bidirectional fractional distillation loss to simultaneously supervise the renderings of clothing and human bodies, specifically: Among them, λ1 and λ2 are the coefficients that balance the weights between the RGB image and the depth image, x b-t ,d b-t and They are the color rendering, depth map and initial noise of the human body at time t, x c-t ,d c-t and They are the color rendering, depth map and initial noise of the clothing at time step t; is the loss function for human body rendering; is the loss function of a single clothing rendering; E[*] is the mathematical expectation of the image at time step t, is the mathematical expectation of the human body rendering at time step t, is the mathematical expectation of the human body depth map at time step t; is the mathematical expectation of the clothing rendering at time step t; is the mathematical expectation of the clothing depth map at time step t; w t To control the noise scheduling, it is determined by the sampler of the diffusion model; ∈ φ is the noise predictor; P c is the three-dimensional Gaussian point set of the clothing part; y is the text input; θ is the training parameter.

9. The method for generating three-dimensional clothing based on a Gaussian model according to claim 1, characterized in that: An adaptive clothing clipping strategy is introduced to train a single 3D Gaussian model with geometric constraints.

10. The method for generating three-dimensional clothing based on a Gaussian model according to claim 1, characterized in that: The method further comprises step 5: registering the template grid sequence of the 3D clothing to obtain the animated 3D clothing, specifically: 5-1, using the iterative closest point algorithm to achieve preliminary alignment between the three-dimensional clothing and the first clothing mesh; 5-2: Find the nearest clothing mesh point of the three-dimensional clothing Gaussian point; 5-3: Use the deformation matrix between the vertices of the clothing mesh to calculate the transformation of the Gaussian points of the three-dimensional clothing, and finally output the animated three-dimensional clothing.

Images