A drivable three-dimensional digital human representation method and device capable of progressive transmission

Abstract

The application discloses a drivable three-dimensional digital human representation method and device capable of progressive transmission, and aims at an input portrait video frame sequence, generates a digital human template grid of each frame through a face key point tracking method, performs implicit adaptive subdivision on the generated template grid, binds three-dimensional Gaussians in a local coordinate system of each level of triangular face, and parameters include residual rotation, scale and mean value; based on a screen space gradient, triggers triangular face subdivision and newly added Gaussians, forms a multi-level tree structure on each face sheet of the template grid; carries out differentiable rendering supervision and position and scale regular optimization on multiple levels; calculates face importance according to rendering contribution, progressively transmits and loads the newly added Gaussians in the order of face importance, and realizes continuous level rendering output of the drivable digital human. The method and device can realize progressive loading and rendering of the digital human, and improve the realism, interactive usability and resource efficiency of the digital human.

Description

Technical Field

[0001] This invention relates to the field of computer vision and graphics technology, and in particular to a progressively transferable, drivable three-dimensional digital human representation method and apparatus. Background Technology

[0002] In recent years, computer vision and graphics have driven the application of virtual digital humans in scenarios such as immersive communication, virtual live streaming, film and television production, driven generation, virtual try-on, and metaverse. Facing terminal devices and real-time multi-party interaction, users expect "instant viewing": even with limited network bandwidth or computing power, they should first see a drivable 3D human image, which then gradually becomes clearer as data flows in; simultaneously, they expect stability and no jitter under changes in expression and posture, which places higher demands on the streaming capabilities and robustness of digital human representation, training, and rendering.

[0003] Current digital humans are mostly packaged as complete assets, lacking support for incremental loading based on importance or region granularity, and also lacking a continuously adjustable mechanism for controlling the level of detail. In practice, transmitting high-capacity assets at once often leads to prolonged blank periods or stuttering under limited bandwidth; discrete level switching is prone to noticeable jumps. More critically, when expressions and postures change, if the Gaussian layers at each level lack consistent local binding, the movement of coarse and fine layers becomes asynchronous, easily resulting in geometric misalignment and temporal jitter, compromising driveability and cross-layer stability. An ideal streaming solution should support: rapid placement of coarse layers with extremely small volumes to achieve immediate visibility, followed by the addition of new Gaussian layers without replacing existing content as data flows in, continuously improving rendering quality while maintaining consistent binding to ensure temporal smoothness. Existing methods generally lack this integrated design of continuous level of detail control and progressive loading, making it impossible to simultaneously achieve fast start-up speed, progressive quality improvement, and drive stability under low bandwidth or low computing power conditions.

[0004] Existing uniform subdivision or uniform expansion often results in overly dense Gaussians in large, smooth areas, while the density is insufficient in high-frequency areas such as the eye area and hairline, leading to an imbalance in resource allocation: on the one hand, the number of Gaussians and the size of the assets expand rapidly, significantly increasing bandwidth consumption and loading time; on the other hand, critical detail areas remain undersampled, making it difficult to meet the requirements of high-quality reconstruction. Redundancy in Gaussian scale can also cause a surge in rendering overhead, further amplifying storage and transmission bottlenecks in multi-terminal or multi-user concurrent scenarios. This equal-resolution subdivision strategy makes it difficult to simultaneously pursue high quality and high efficiency under limited bandwidth and storage conditions, lacking differentiated control over detail areas and smooth areas.

[0005] In view of this, the present invention is hereby proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a progressively transferable, driveable 3D digital human representation method and apparatus to solve the aforementioned technical problems in the prior art. The method and apparatus of this invention can achieve progressive loading and rendering of digital humans, enabling rapid start-up with a small file size, smooth quality improvement as data flows in, while reducing redundant Gaussian data and transmission volume, significantly improving the realism, interactivity, and resource efficiency of the digital human.

[0007] The objective of this invention is achieved through the following technical solution: A progressively transferable, driveable 3D digital human representation method, the method comprising: Step 1: For the input sequence of human portrait video frames, generate a digital human template mesh for each frame using a facial landmark tracking method; Step 2: Perform implicit adaptive subdivision on the generated template mesh to form a multi-level triangular face hierarchy; Step 3: Bind a 3D Gaussian to the local coordinate system of each triangular face, with parameters including residual rotation, scale and mean, driven by the face center and face scale to maintain drivability and cross-layer consistency. Step 4: Triangle subdivision and Gaussian addition are triggered based on screen space gradient to form a multi-level tree structure on each facet of the template mesh; Step 5: Perform differentiable rendering supervision and position and scale regularization optimization on multiple levels; Step 6: Calculate the importance of the surface based on the rendering contribution, and progressively transfer and load new Gaussians in order of surface importance to achieve continuous layer rendering output that can drive digital humans.

[0008] A progressively transferable, drivable 3D digital human representation device, the device comprising: The digital human template mesh generation unit is used to generate a digital human template mesh for each frame of an input human portrait video frame sequence using a facial key point tracking method. Multi-level triangular face generation unit is used to perform implicit adaptive subdivision on the generated template mesh to form a multi-level triangular face hierarchy; Gaussian binding units are used to bind 3D Gaussians to the local coordinate system of each triangular face. The parameters include residual rotation, scale and mean, which are driven by the face center and face scale to maintain drivability and cross-layer consistency. Multi-level tree structure generation unit is used to trigger triangular face subdivision and add Gaussian based on screen space gradient, forming a multi-level tree structure on each facet of the template mesh; The training optimization unit is used for differentiable rendering supervision and position and scale regularization optimization across multiple levels. The driveable digital human output unit is used to calculate the importance of surfaces based on their rendering contribution, and progressively transmit and load new Gaussians in order of surface importance to achieve continuous-level rendering output of the driveable digital human.

[0009] Compared with existing technologies, the above-mentioned methods and devices can realize the progressive loading and rendering of digital humans, enabling rapid playback with a small size, smooth quality improvement as data flows in, and reducing redundant Gaussian data and transmission volume. This significantly improves the realism, interactive usability, and resource efficiency of digital humans, and can be applied to scenarios such as immersive communication, film and television production, driver generation, virtual human / virtual try-on, and XR metaverse. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A flowchart illustrating the progressively transmittable, driveable 3D digital human representation method provided in an embodiment of the present invention; Figure 2 These are renderings of various levels of detail in embodiments of the present invention; Figure 3 This is a rendering of the most detailed level and a visualization of the subdivided levels in the embodiments of the present invention. Detailed Implementation

[0012] 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 a part of the embodiments of the present invention, and not all of them, and do not constitute a limitation on the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0013] First, the following explanations are provided for the terms that may be used in this article: The term "and / or" means that either or both can be achieved simultaneously. For example, X and / or Y means that it includes both "X" or "Y" as well as the three cases of "X and Y".

[0014] The terms "comprising," "including," "containing," "having," or other similar semantic descriptions should be interpreted as non-exclusive inclusion. For example, including a technical feature element (such as raw material, component, ingredient, carrier, dosage form, material, size, part, component, mechanism, device, step, process, method, reaction conditions, processing conditions, parameter, algorithm, signal, data, product or article of manufacture, etc.) should be interpreted as including not only the expressly listed technical feature element, but also other technical feature elements that are not expressly listed and are well-known in the art.

[0015] The term "composed of" excludes any technical features not expressly listed. When used in a claim, it closes the claim to exclude all technical features other than those expressly listed, except for associated conventional impurities. If the term appears only in a clause of a claim, it limits the claim to the elements expressly listed in that clause; elements recited in other clauses are not excluded from the overall claim.

[0016] The technical solution provided by this invention will be described in detail below. Contents not described in detail in the embodiments of this invention are prior art known to those skilled in the art. Where specific conditions are not specified in the embodiments of this invention, they shall be performed according to conventional conditions in the art or conditions recommended by the manufacturer. Reagents or instruments used in the embodiments of this invention whose manufacturers are not specified are all conventional products that can be purchased commercially.

[0017] like Figure 1 The diagram shown is a flowchart illustrating a progressively transferable, driveable 3D digital human representation method provided in an embodiment of the present invention. The method includes: Step 1: For the input sequence of human portrait video frames, generate a digital human template mesh for each frame using a facial landmark tracking method; In this step, a video face tracking method based on a parameterized face model is adopted to extract the frame sequence {It|t=1,…,N} for each frame of the input portrait video frame sequence and perform face key point detection; By optimizing the fit, the corresponding expression parameters, pose parameters, and shape parameters are obtained, which then drive the template mesh to generate the face geometry for each frame.

[0018] Step 2: Perform implicit adaptive subdivision on the generated template mesh to form a multi-level triangular face hierarchy; In this step, implicit adaptive subdivision defines child vertices using barycentric coordinates. , is represented as: (1) Where V i V j Vk These are the three vertices of the parent triangle. Let these be the coordinates of the barycenter, and the initial values ​​be set to... , During training, the simplex is constrained by non-negative simplex with a sum of 1, allowing child vertices to adaptively move within the parent triangle to match local scale and deformation.

[0019] Step 3: Bind a 3D Gaussian to the local coordinate system of each triangular face, with parameters including residual rotation, scale and mean, driven by the face center and face scale to maintain drivability and cross-layer consistency. In this step, Gaussian elements are bound to the local coordinate system of each triangular face, and the rotation of the Gaussian elements is... ,scale ,Location Rotation matrix of the corresponding triangle ,scale Center of gravity express: (2) Where ΔR is the residual rotation of the Gaussian element relative to the parent triangle; ΔS is the residual scale; and Δμ is the residual position offset. All three are learnable parameters. This allows the Gaussian elements to change with the position and shape of the triangular facets, maintaining consistent drive.

[0020] Step 4: Triangle subdivision and Gaussian addition are triggered based on screen space gradient to form a multi-level tree structure on each facet of the template mesh; In this step, the current finest layer leaf is checked with a fixed iteration step size. If the screen spatial gradient exceeds a threshold... Then, the finest layer of leaf surface is subdivided and Gaussian bonds are applied to the sub-surfaces until the preset maximum depth is reached, forming a multi-level tree structure.

[0021] Step 5: Perform differentiable rendering supervision and position and scale regularization optimization on multiple levels; In this step, the loss function of the optimization process includes color error. Gaussian metaposition regularization Gaussian meta-scale regularization Total loss The formula is as follows: (3) Where λpos and λscale are the corresponding loss weight coefficients, respectively; During training, the model is optimized by minimizing the difference between the rendered result and the input image, specifically including: Color error Reconstruction accuracy is measured by calculating the L1 error and SSIM loss between the rendered RGB image and the ground truth image. During training, the rendering results at each level are supervised, as shown below: (4) Where S is the set of levels involved in supervision; For hierarchical indexes; For the first The weights of each layer; L1 is the L1 error between the rendered image and the real image; L SSIM λs represents the structural similarity loss; λs is the weighting coefficient balancing L1 error and SSIM loss. Should Simultaneously constraining color consistency and structural similarity at multiple levels ensures that the rendering results, from coarse to fine, match the real image. Among them, L1 error emphasizes pixel-level color matching, while SSIM loss focuses on local contrast and texture structure. The combination of the two suppresses color cast while preserving details. Gaussian metaposition regularization Assuming the Gaussian should fit the local coordinates of its parent triangle, the threshold... Small deviations within the triangular facet scale are allowed; deviations exceeding this scale are penalized using the L2 norm, expressed as: (5) This represents the position of the Gaussian element; it is assumed that the Gaussian should remain close to the local coordinates of its bound parent triangle, for example, the Gaussian at the tip of the nose should not drift to the cheek. Optimization during training may cause the Gaussian to drift away from the parent triangle, leading to mismatches. Small deviations within the parent triangle scale are tolerated with a threshold, while L2 penalty is applied to displacements exceeding the threshold to prevent Gaussian drift into the wrong region, ensuring consistent position and stable timing during driving. Gaussian meta-scale regularization Assuming the Gaussian surface should adapt to the scale of its parent triangle, the threshold... Small deviations are allowed; deviations exceeding this limit are penalized by the L2 norm, expressed as: (6) This represents the scale of the Gaussian element; if the Gaussian element is too large relative to the parent triangle, a small rotation of the triangle will be amplified into noticeable jitter. When the scale is above 0.6 × parent triangle, excessive scale is suppressed to reduce jitter caused by template mesh movement; when the scale is below the threshold, no pressure is applied to avoid Gaussian forced excessive shrinkage, which would require more Gaussian primitives to represent the current region and thus affect rendering efficiency. This balance maintains visual smoothness while controlling rendering efficiency.

[0022] Step 6: Calculate the importance of the surface based on the rendering contribution, and progressively transfer and load new Gaussians in order of surface importance to achieve continuous layer rendering output that can drive digital humans.

[0023] In this step, the importance of the surface This is obtained by accumulating the rendering contribution of Gaussian bound to a face, and is represented as: (7) Among them G i Let g be the set of Gaussian elements bound to the i-th triangular face; g is the Gaussian element index; p is the pixel index; α g,p T represents the opacity of Gaussian pixel g at pixel p; g,p Let g be the cumulative transmittance of Gaussian pixel g at pixel p; Before transmitting data or rendering, assign face importance to all faces. And the order of the hierarchy; Triangles with lower hierarchical levels and their associated Gaussian elements are transmitted first, while Gaussian elements of the same hierarchical level are transmitted first based on their importance. Higher primitives enable the achievement of quality comparable to the most detailed head rendering, even with a small amount of transmission.

[0024] like Figure 2 The following are renderings of various levels of detail in embodiments of the present invention, demonstrating the rendering effect of the method. Figure 2 The image shows the face rendering results at four detail levels: Level 0, Level 1.5, Level 2.5, and Level 4. Each level is accompanied by a comparison image of a magnified local area. It can be seen that as the level increases, the facial details in the rendered result gradually become richer, especially in high-frequency detail areas such as the mouth, eyes, and hair. The rendering quality of higher levels is significantly better than that of lower levels, indicating that the multi-level representation of this method can effectively improve rendering accuracy step by step from coarse to fine.

[0025] like Figure 3 The rendering of the finest level and the visualization of the subdivision level in the embodiments of the present invention demonstrate the advantage of the adaptive subdivision method in saving the number of primitives. Figure 3 Each group contains two images: the rendering result and a corresponding visualization of the subdivision level. The subdivision level visualization uses different gray levels to represent different subdivision depths. It can be seen that high-frequency detail areas of the face (such as eyes, lips, and nose) are adaptively subdivided to deeper levels, while flat areas (such as cheeks and forehead) only require shallower levels to achieve good rendering results. This significantly reduces the total number of Gaussian pixels required, effectively improving transmission and rendering efficiency while maintaining rendering quality.

[0026] Based on the above method, embodiments of the present invention also provide a progressively transferable, driveable three-dimensional digital human representation device, the device comprising: The digital human template mesh generation unit is used to generate a digital human template mesh for each frame of an input human portrait video frame sequence using a facial key point tracking method. Multi-level triangular face generation unit is used to perform implicit adaptive subdivision on the generated template mesh to form a multi-level triangular face hierarchy; Gaussian binding units are used to bind 3D Gaussians to the local coordinate system of each triangular face. The parameters include residual rotation, scale and mean, which are driven by the face center and face scale to maintain drivability and cross-layer consistency. Multi-level tree structure generation unit is used to trigger triangular face subdivision and add Gaussian based on screen space gradient, forming a multi-level tree structure on each facet of the template mesh; The training optimization unit is used for differentiable rendering supervision and position and scale regularization optimization across multiple levels. The driveable digital human output unit is used to calculate the importance of surfaces based on their rendering contribution, and progressively transmit and load new Gaussians in order of surface importance to achieve continuous-level rendering output of the driveable digital human.

[0027] The specific implementation process of the functions and roles of each unit in the above device is detailed in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0028] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method.

[0029] This invention also provides a computer storage medium storing a plurality of instructions adapted for loading and executing the method by a processor.

[0030] It is worth noting that the contents not described in detail in the embodiments of the present invention belong to the prior art known to those skilled in the art.

[0031] In summary, the embodiments of the present invention have the following advantages compared with traditional three-dimensional digital human representation methods: (1) Progressive transmission and continuous detail level control: This invention loads Gaussian content in a streaming manner based on rendering contribution. Old content is not replaced, only new content is added, forming continuous detail level control. It can still present a driving digital human in real time under low bandwidth, and smoothly improve the quality under high bandwidth, avoiding the jump of discrete level switching. For multi-terminal / multi-user concurrent scenarios, it can truncate or continue loading as needed, improving the streaming interactive experience; (2) Adaptive resource allocation: implicit subdivision driven by screen space gradient is used to concentrate Gaussian density in high-frequency areas such as eyes, mouth and hairline, while keeping smooth areas sparse; the number of Gaussians and model volume are significantly reduced under the same quality, reducing transmission / storage / loading costs, avoiding the contradiction of "over-dense - undersampling" in uniform subdivision, and making it easier to implement when bandwidth is limited; (3) Cross-layer driveability and stability: Local Gaussian binding of surfaces ensures that each layer moves in coordination with the template deformation. Cross-layer alignment under changes in expression / posture reduces misalignment and jitter. Position / scale regularization suppresses drift and amplification effects. Multi-layer supervision keeps coarse layers usable and fine layers refined. The timing is smooth during progressive loading of inference, improving realism and interactive usability.

[0032] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included 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. The information disclosed in the background section is intended only to enhance the understanding of the overall background technology of the present invention and should not be construed as an admission or implication in any way that such information constitutes prior art known to those skilled in the art.

Claims

1. A progressively transferable, driveable three-dimensional digital human representation method, characterized in that, The method includes: Step 1: For the input sequence of human portrait video frames, generate a digital human template mesh for each frame using a facial landmark tracking method; Step 2: Perform implicit adaptive subdivision on the generated template mesh to form a multi-level triangular face hierarchy; Step 3: Bind a 3D Gaussian to the local coordinate system of each triangular face, with parameters including residual rotation, scale and mean, driven by the face center and face scale to maintain drivability and cross-layer consistency. Step 4: Triangle subdivision and Gaussian addition are triggered based on screen space gradient to form a multi-level tree structure on each facet of the template mesh; Step 5: Perform differentiable rendering supervision and position and scale regularization optimization on multiple levels; Step 6: Calculate the importance of the surface based on the rendering contribution, and progressively transfer and load new Gaussians in order of surface importance to achieve continuous layer rendering output that can drive digital humans.

2. The progressively transmittable, driveable three-dimensional digital human representation method according to claim 1, characterized in that, In step 1, a video face tracking method based on a parameterized face model is specifically adopted to extract the frame sequence {It|t=1,…,N} for each frame of the input portrait video frame sequence and perform face key point detection. By optimizing the fit, the corresponding expression parameters, pose parameters, and shape parameters are obtained, which then drive the template mesh to generate the face geometry for each frame.

3. The progressively transmittable, driveable three-dimensional digital human representation method according to claim 1, characterized in that, In step 2, implicit adaptive subdivision defines subvertices using barycentric coordinates. , is represented as: （1） Where V i V j V k These are the three vertices of the parent triangle. Let these be the coordinates of the barycenter, and the initial values ​​be set to... , During training, the simplex is constrained by non-negative simplex with a sum of 1, allowing child vertices to adaptively move within the parent triangle to match local scale and deformation.

4. The progressively transmittable, driveable three-dimensional digital human representation method according to claim 1, characterized in that, In step 3, Gaussian elements are bound to the local coordinate system of each triangular face, and the rotation of the Gaussian elements is performed. ,scale ,Location Rotation matrix of the corresponding triangle ,scale Center of gravity express: （2） Where ΔR is the residual rotation of the Gaussian element relative to the parent triangle; ΔS is the residual scale; and Δμ is the residual position offset. All three are learnable parameters. This allows the Gaussian elements to change with the position and shape of the triangular facets, maintaining consistent drive.

5. The progressively transmittable, driveable three-dimensional digital human representation method according to claim 1, characterized in that, In step 4, the current finest layer leaf surface is checked with a fixed iteration step size. If the screen spatial gradient exceeds the threshold... Then, the finest layer of leaf surface is subdivided and Gaussian bonds are applied to the sub-surfaces until the preset maximum depth is reached, forming a multi-level tree structure.

6. The progressively transmittable, driveable three-dimensional digital human representation method according to claim 3, characterized in that, In step 5, the loss function of the optimization process includes color error. Gaussian metaposition regularization Gaussian meta-scale regularization Total loss The formula is as follows: （3） Where λpos and λscale are the corresponding loss weight coefficients, respectively; During training, the model is optimized by minimizing the difference between the rendered result and the input image, specifically including: Color error Reconstruction accuracy is measured by calculating the L1 error and SSIM loss between the rendered RGB image and the ground truth image. During training, the rendering results at each level are supervised, as shown below: （4） Where S is the set of levels involved in supervision; For hierarchical indexes; For the first The weights of each layer; L1 is the L1 error between the rendered image and the real image; L SSIM λs represents the structural similarity loss; λs is the weighting coefficient balancing L1 error and SSIM loss. Should Simultaneously constraining color consistency and structural similarity at multiple levels ensures that the rendering results, from coarse to fine, match the real image. Among them, L1 error emphasizes pixel-level color matching, while SSIM loss focuses on local contrast and texture structure. The combination of the two suppresses color cast while preserving details. Gaussian metaposition regularization Assuming the Gaussian should fit the local coordinates of its parent triangle, the threshold... Small deviations within the triangular facet scale are allowed; deviations exceeding this scale are penalized using the L2 norm, expressed as: （5） Indicates the position of the Gaussian element; Small deviations within the parent triangle scale are tolerated with a threshold, while L2 penalty is applied to displacements exceeding the threshold to prevent Gaussian drift into the wrong region, ensuring consistent position and stable timing during driving. Gaussian meta-scale regularization Assuming the Gaussian surface should adapt to the scale of its parent triangle, the threshold... Small deviations are allowed; deviations exceeding this limit are penalized by the L2 norm, expressed as: （6） The scale of the Gaussian element; Suppress excessively large scales above 0.6 × parent triangle to reduce jitter caused by template mesh movement; do not apply pressure below the threshold to avoid forced excessive shrinkage of Gaussian meshes.

7. The progressively transmittable, driveable three-dimensional digital human representation method according to claim 1, characterized in that, In step 6, the importance of the surface This is obtained by accumulating the rendering contribution of Gaussian bound to a face, and is represented as: （7） Among them G i Let g be the set of Gaussian elements bound to the i-th triangular face; g is the Gaussian element index; p is the pixel index; α g,p T represents the opacity of Gaussian pixel g at pixel p; g,p Let g be the cumulative transmittance of Gaussian pixel g at pixel p; Before transmitting data or rendering, assign face importance to all faces. And the order of the hierarchy; Triangles with lower hierarchical levels and their associated Gaussian elements are transmitted first, while Gaussian elements of the same hierarchical level are transmitted first based on their importance. Higher primitives enable the achievement of quality comparable to the most detailed head rendering, even with a small amount of transmission.

8. A progressively transmittable, driveable three-dimensional digital human representation device, characterized in that, The device includes: The digital human template mesh generation unit is used to generate a digital human template mesh for each frame of an input human portrait video frame sequence using a facial key point tracking method. Multi-level triangular face generation unit is used to perform implicit adaptive subdivision on the generated template mesh to form a multi-level triangular face hierarchy; Gaussian binding units are used to bind 3D Gaussians to the local coordinate system of each triangular face. The parameters include residual rotation, scale and mean, which are driven by the face center and face scale to maintain drivability and cross-layer consistency. Multi-level tree structure generation unit is used to trigger triangular face subdivision and add Gaussian based on screen space gradient, forming a multi-level tree structure on each facet of the template mesh; The training optimization unit is used for differentiable rendering supervision and position and scale regularization optimization across multiple levels. The driveable digital human output unit is used to calculate the importance of surfaces based on their rendering contribution, and progressively transmit and load new Gaussians in order of surface importance to achieve continuous-level rendering output of the driveable digital human.

9. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 7.

10. A computer storage medium, characterized in that, The computer storage medium stores a plurality of instructions adapted for loading by a processor and executing the method of any one of claims 1 to 7.

Images