Digital character animation control method, modeling method, device and storage medium

Abstract

The application discloses a digital role animation control method, a modeling method, equipment and a storage medium. The method comprises the following steps: acquiring an animation switching instruction corresponding to a digital role, the animation switching instruction comprising a target bone parameter vector corresponding to a target bone node; processing the target bone parameter vector and a current bone parameter vector based on a parameterized bone model to determine a bone transformation matrix; transforming a current vertex attribute corresponding to a target mesh vertex based on the bone transformation matrix and a first bone skinning weight to determine a target vertex attribute corresponding to the target mesh vertex; transforming a current Gaussian attribute corresponding to a target Gaussian point based on the bone transformation matrix and a second bone skinning weight corresponding to the target Gaussian point to determine a target Gaussian attribute corresponding to the target Gaussian point; and determining a target animation image based on the target vertex attribute corresponding to the target mesh vertex and the target Gaussian attribute corresponding to the target Gaussian point. The method can guarantee the synchronization and high detail fidelity of action control.

Description

Technical Field

[0001] This invention relates to the field of digital character animation control technology, and in particular to a digital character animation control method, modeling method, device and storage medium. Background Technology

[0002] With the widespread application of virtual digital characters in film, games, and virtual reality, how to realistically drive the body movements and facial expressions of digital characters has become a key technology.

[0003] Current digital character animations typically employ a skeletal-driven mesh skinning method. This method uses a pre-defined skeletal model to control the movement of mesh vertices to achieve limb movements. Facial expressions or other specific body parts require separate deformation units for driving. This separate driving approach easily leads to asynchrony or incoordination between limb movements and specific body part movements, reducing the realism of the animation. Furthermore, the surface details and textures of current digital characters usually rely on high-precision polygon meshes or textures, significantly increasing the computational and rendering burden, impacting animation control efficiency and detail fidelity. Summary of the Invention

[0004] This invention provides a digital character animation control method, modeling method, device, and storage medium to solve the problems of asynchronous motion control and low detail fidelity in existing digital character animation control processes.

[0005] A method for controlling digital character animation, comprising:

[0006] Obtain the animation switching instruction corresponding to the digital character. The animation switching instruction includes the target bone parameter vector corresponding to the target bone node. The target bone node includes bone joint nodes and specific part control nodes.

[0007] Based on the parameterized skeleton model corresponding to the digital character, the target skeleton parameter vector and the current skeleton parameter vector corresponding to the target skeleton node are processed to determine the skeleton transformation matrix;

[0008] Based on the bone transformation matrix and the first bone skinning weight corresponding to the target mesh vertex, the current vertex attribute corresponding to the target mesh vertex is transformed to determine the target vertex attribute corresponding to the target mesh vertex. The target mesh vertex is the mesh vertex in the mesh model corresponding to the digital character that has a binding relationship with the target bone node. The first bone skinning weight is the proportion by which the bone node drives the mesh vertex to move.

[0009] Based on the bone transformation matrix and the second bone skinning weight corresponding to the target Gaussian point, the current Gaussian attribute corresponding to the target Gaussian point is transformed to determine the target Gaussian attribute corresponding to the target Gaussian point. The target Gaussian point is the Gaussian point in the Gaussian point cloud model corresponding to the digital character that has a binding relationship with the target bone node. The second bone skinning weight is the proportion by which the bone node drives the Gaussian point to move.

[0010] The target animation image is determined based on the target vertex attributes corresponding to the target mesh vertices and the target Gaussian attributes corresponding to the target Gaussian points.

[0011] Preferably, the step of transforming the current Gaussian attribute corresponding to the target Gaussian point based on the bone transformation matrix and the second bone skinning weight corresponding to the target Gaussian point, to determine the target Gaussian attribute corresponding to the target Gaussian point, includes:

[0012] Based on the bone transformation matrix and the second bone skinning weight corresponding to the target Gaussian point, the current Gaussian attribute corresponding to the target Gaussian point is transformed to determine the initial Gaussian attribute corresponding to the target Gaussian point. The initial Gaussian attribute is the state of the Gaussian point position, covariance matrix and rendering attribute directly calculated by the bone transformation.

[0013] Based on the elastic network corresponding to the target Gaussian point, the initial Gaussian property corresponding to the target Gaussian point is oscillated and simulated to determine the target Gaussian property corresponding to the target Gaussian point.

[0014] Preferably, the step of transforming the current Gaussian attribute corresponding to the target Gaussian point based on the bone transformation matrix and the second bone skinning weight corresponding to the target Gaussian point, to determine the target Gaussian attribute corresponding to the target Gaussian point, includes:

[0015] Based on the bone transformation matrix and the second bone skinning weight corresponding to the target Gaussian point, the current Gaussian attribute corresponding to the target Gaussian point is transformed to determine the initial Gaussian attribute corresponding to the target Gaussian point. The initial Gaussian attribute is the state of the Gaussian point position, covariance matrix and rendering attribute directly calculated by the bone transformation.

[0016] Based on pre-set physical constraints and a physics engine, collision detection is performed on the initial Gaussian properties corresponding to the target Gaussian point to adjust and determine the target Gaussian properties corresponding to the target Gaussian point.

[0017] Preferably, determining the target animation image based on the target vertex attributes corresponding to the target mesh vertices and the target Gaussian attributes corresponding to the target Gaussian points includes:

[0018] Rasterization rendering is performed on the target vertex attributes corresponding to the target mesh vertices to determine the mesh rendering image;

[0019] Rasterization rendering is performed on the target Gaussian attributes corresponding to the target Gaussian point to determine the Gaussian rendered image;

[0020] The mesh rendering image and the Gaussian rendering image are fused together to determine the target animation image.

[0021] Preferably, the target Gaussian properties include the number of renders and the spatial expansion radius;

[0022] The step of rasterizing and rendering the target Gaussian attributes corresponding to the target Gaussian point to determine the Gaussian rendered image includes:

[0023] Obtain the rendering distance corresponding to the target Gaussian point;

[0024] Based on the rendering distance corresponding to the target Gaussian point, determine the number of rendering operations and the spatial expansion radius corresponding to the target Gaussian point;

[0025] Rasterization rendering is performed based on the number of renders corresponding to the target Gaussian point and the spatial expansion radius to determine the Gaussian rendered image.

[0026] Preferably, the target Gaussian attributes include the target position and the target orientation;

[0027] The step of rasterizing and rendering the target Gaussian attributes corresponding to the target Gaussian point to determine the Gaussian rendered image includes:

[0028] The position deviation is determined based on the target position corresponding to the target Gaussian point and the current position of the target Gaussian point;

[0029] The position deviation is processed based on the inertial physics equations to determine the inertial position corresponding to the target Gaussian point;

[0030] Based on the inertial position corresponding to the target Gaussian point, the target Gaussian property corresponding to the target Gaussian point is corrected to determine the corrected Gaussian property corresponding to the target Gaussian point;

[0031] Rasterization rendering is performed on the modified Gaussian properties corresponding to the target Gaussian point to determine the Gaussian rendered image.

[0032] Preferably, the target Gaussian attribute includes the target color;

[0033] The step of rasterizing and rendering the target Gaussian attributes corresponding to the target Gaussian point to determine the Gaussian rendered image includes:

[0034] Based on the target bone parameter vector corresponding to the target bone node, determine the driving state corresponding to the target bone node;

[0035] Based on the driving state corresponding to the target skeletal node, determine the color correction coefficient corresponding to the driving state;

[0036] The target color corresponding to the target Gaussian point is corrected based on the color correction coefficient, and the corrected color corresponding to the target Gaussian point is determined.

[0037] Rasterization rendering is performed based on the corrected color corresponding to the target Gaussian point to determine the Gaussian rendered image.

[0038] Preferably, the digital character is a humanoid character, and the specific part control node includes a facial expression control node;

[0039] Alternatively, the digital character may be an animal character, and the specific part control nodes may include tail control nodes and / or ear control nodes;

[0040] Alternatively, the digital character may be a cartoon character, and the specific part control node may be an exaggerated expression control node.

[0041] A digital character modeling method, comprising:

[0042] Obtain the bone parameter vector corresponding to the skeletal structure, and construct a parameterized skeletal model of the digital character based on the bone parameter vector corresponding to the skeletal structure. The skeletal structure includes multiple bone nodes, and the bone nodes include bone joint nodes and specific part control nodes.

[0043] The parameterized skeletal model of the digital character is bound to the mesh model of the digital character to determine the skeletal mesh model corresponding to the digital character; wherein, the skeletal mesh model includes multiple first bound bone nodes corresponding to all mesh vertices and a first bone skinning weight corresponding to each first bound bone node;

[0044] The Gaussian point cloud model of the digital character is bound to the corresponding skeletal mesh model of the digital character to determine the corresponding Gaussian skeletal mesh model of the digital character. The Gaussian skeletal mesh model is used to realize the animation control of the digital character. The Gaussian point cloud model includes multiple Gaussian points carrying a spatial expansion radius, and the Gaussian skeletal mesh model includes multiple second bound bone nodes corresponding to all Gaussian points and a second bone skinning weight corresponding to each second bound bone node.

[0045] Preferably, the step of binding the parametric skeletal model of the digital character to the mesh model of the digital character, and determining the skeletal mesh model corresponding to the digital character, includes:

[0046] Determine the first bound bone node corresponding to each mesh vertex of the mesh model of the digital character, and determine the initial bone skinning weight corresponding to the first bound bone node;

[0047] Based on the anatomical structure of the digital character, the initial bone skinning weights corresponding to the first bound bone node are smoothed to determine the first bone skinning weights corresponding to the first bound bone node.

[0048] The skeletal mesh model corresponding to the digital character is determined based on the first bound bone node corresponding to all mesh vertices and the first bone skinning weight corresponding to each of the first bound bone nodes.

[0049] Preferably, the step of binding the Gaussian point cloud model of the digital character with the corresponding skeletal mesh model of the digital character to determine the Gaussian skeletal mesh model of the digital character includes:

[0050] Bind the Gaussian point cloud model of the digital character to the skeletal mesh model corresponding to the digital character, and determine the associated mesh vertex corresponding to each Gaussian point in the Gaussian point cloud model.

[0051] The adjacent bone nodes corresponding to the associated mesh vertices are determined as the second bound bone nodes corresponding to the Gaussian points, and the initial bone skinning weights corresponding to the second bound bone nodes are determined.

[0052] Based on the anatomical structure of the digital character, the initial bone skinning weights corresponding to the second bound bone node are smoothed to determine the second bone skinning weights corresponding to the second bound bone node.

[0053] Based on the second bound bone nodes corresponding to all mesh vertices and the second bone skinning weights corresponding to each second bound bone node, the Gaussian bone mesh model corresponding to the digital character is determined.

[0054] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described digital character animation control method; or, the processor executes the computer program to implement the above-described digital character modeling method.

[0055] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described digital character animation control method; or, when executed by a processor, the computer program implements the above-described digital character modeling method.

[0056] This invention provides a digital character animation control method, modeling method, device, and storage medium. It enables unified control of limb movements corresponding to skeletal joint nodes and specific movements corresponding to control nodes of specific body parts, ensuring synchronized movement control and improving the naturalness and coherence of character animation. Furthermore, by introducing a Gaussian point cloud model as a surface detail representation, the digital character retains rich texture and geometric details during animation, achieving high detail fidelity. The synchronous change of the Gaussian point cloud model with the parametric skeletal model ensures that details in the Gaussian point cloud model update dynamically as skeletal nodes move, achieving real-time detail updates. This also improves the versatility and scalability of the solution. Attached Figure Description

[0057] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. 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.

[0058] Figure 1 This is a flowchart of a digital character animation control method in an embodiment of the present invention;

[0059] Figure 2 This is a flowchart of a digital character modeling method in an embodiment of the present invention;

[0060] Figure 3 This is a diagram illustrating the overall architecture for controlling the animation of a digital character's Gaussian skeletal mesh model.

[0061] Figure 4 This is a comparison diagram between the Gaussian point cloud model and the traditional discrete point cloud in an embodiment of the present invention;

[0062] Figure 5 This is a schematic diagram of the synchronous deformation of a mesh model and a Gaussian point cloud model under skeleton-driven operation.

[0063] Figure 6 This is a schematic diagram of an adaptation architecture for digital roles. Detailed Implementation

[0064] To make the technical problems solved, the technical solutions, and the beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0065] This invention provides a digital character animation control method. Taking a computer device as an example, this digital character animation control method is applicable to a computer device with data processing capabilities. Based on the received animation switching instructions, it can quickly control the digital character to achieve synchronous driving of limb movements and specific part movements. Furthermore, a Gaussian point cloud model with a binding relationship with skeletal nodes is used to render the surface details and textures of the digital character to ensure the efficiency and quality of animation switching.

[0066] For ease of understanding, the following explanations and interpretations of some terms involved in the embodiments of the present invention are provided.

[0067] Digital characters typically refer to digital figures or images created using digital technology and existing in a virtual environment, encompassing virtual humanoid characters, animal characters, and cartoon characters.

[0068] Parametric skeletal models are the core underlying structure for driving the motion of digital characters. They define the hierarchical relationships, lengths, joint angles, and motion constraints of bones through a set of adjustable parameters, thereby achieving precise control over the posture and movements of digital characters. A parametric skeletal model includes multiple skeletal nodes and corresponding skeletal parameter vectors. Skeletal nodes include skeletal joint nodes and specific body part control nodes. Skeletal joint nodes are used to control limb movements. Specific body part control nodes are nodes in the digital character used to control the function of specific body parts; they are the control nodes for movements other than limb movements. The skeletal parameter vector is a quantified and structured mathematical expression of the key parameters such as the motion state, hierarchical relationships, and constraint attributes of skeletal nodes in the parametric skeletal model.

[0069] The mesh model is the core carrier of its visual shape. It is essentially a polygonal mesh structure composed of a large number of mesh vertices, edges, and faces. It is also the basic object for skeletal binding, material rendering, and animation-driven processes.

[0070] A Gaussian point cloud model is a collection of numerous Gaussian points, representing the combination and arrangement of multiple Gaussian points in three-dimensional space. A Gaussian point is a parametric probabilistic model in three-dimensional space, its core function being to describe the geometric and appearance features of a local region through a Gaussian distribution. The sampling density of the Gaussian point cloud model is adaptively adjusted according to the complexity of the digital character's surface details: the sampling density for densely detailed areas such as the face and hands is 500–1000 points / cm², while the sampling density for areas such as the torso and limbs is 100–300 points / cm². The initial spatial expansion radius of each Gaussian point is set to 0.1–1.0 mm, and color information is obtained through high-resolution texture mapping sampling or fitting of multi-view photogrammetry data.

[0071] This invention provides a digital character animation control method, which is illustrated using a computer device as an example. Figure 1 As shown, the digital character animation control methods include:

[0072] S11: Obtain the animation switching instructions corresponding to the digital character. The animation switching instructions include the target bone parameter vector corresponding to the target bone node. The target bone node includes bone joint nodes and specific part control nodes.

[0073] S12: Based on the parameterized skeleton model corresponding to the digital character, process the target skeleton parameter vector and the current skeleton parameter vector corresponding to the target skeleton node to determine the skeleton transformation matrix;

[0074] S13: Based on the bone transformation matrix and the first bone skinning weight corresponding to the target mesh vertex, transform the current vertex attribute corresponding to the target mesh vertex to determine the target vertex attribute corresponding to the target mesh vertex. The target mesh vertex is the mesh vertex in the mesh model corresponding to the digital character that has a binding relationship with the target bone node. The first bone skinning weight is the proportion of the bone node driving the mesh vertex to move.

[0075] S14: Based on the bone transformation matrix and the second bone skinning weight corresponding to the target Gaussian point, transform the current Gaussian attribute corresponding to the target Gaussian point to determine the target Gaussian attribute corresponding to the target Gaussian point. The target Gaussian point is the Gaussian point in the Gaussian point cloud model corresponding to the digital character that has a binding relationship with the target bone node. The second bone skinning weight is the proportion of the movement of the Gaussian point caused by the bone node.

[0076] S15: Determine the target animation image based on the target vertex attributes corresponding to the target mesh vertices and the target Gaussian attributes corresponding to the target Gaussian points.

[0077] Among them, the animation switching instruction is used to switch between animated images.

[0078] Here, the target skeleton node refers to the skeleton node that needs to be adjusted in this operation; it can be a joint node or a control node for a specific part. The target skeleton parameter vector refers to the skeleton parameter vector that needs to be adjusted in this operation. The current skeleton parameter vector refers to the skeleton parameter vector at the current moment.

[0079] As an example, in step S11, during the animation control of the digital character, the computer device needs to obtain the animation switching instruction of the digital character. This animation switching instruction includes target bone parameter vectors corresponding to one or more target bone nodes. For example, when the digital character is a humanoid character, when it is necessary to switch the limb movements and / or facial expressions of the current animation image, the bone parameter vectors of target bone nodes such as bone joint nodes and / or facial expression control nodes can be updated to determine the target bone parameter vectors corresponding to the target bone nodes.

[0080] In this example, the target bone parameter vector is input in the form of a normalized parameter vector. Taking a humanoid character as an example, the bone parameter vectors corresponding to the bone joint nodes use quaternions to represent rotation (to avoid gimbal lock issues) and Cartesian coordinates to represent displacement. The bone parameter vectors corresponding to the facial expression control nodes use normalized coefficients in the range of 0-1 to represent the intensity of each expression (e.g., a coefficient of 0.8 for 'smile' represents a smile with 80% intensity). The parameter vectors support keyframe interpolation, achieving smooth transition animations between adjacent keyframes through linear interpolation or Bezier interpolation.

[0081] In this example, the computer device obtains animation switching instructions in a highly flexible manner. In addition to using traditional keyframe interpolation or preset instructions to generate changes in skeletal parameters, this does not limit the input source. This solution is compatible with a variety of new animation control methods, such as, but not limited to, motion capture data, AI-generated animation, and user interaction control.

[0082] For example, motion capture data can be used to generate animation switching commands. Specifically, real-time captured human movements or facial expressions are mapped into target skeletal parameter vectors and input into the system to drive the digital character to perform corresponding actions synchronously. Taking a humanoid character as an example, because this solution unifies the driving of the body and face, it can ensure the synchronous update of full-body movements and facial details when using motion capture. This is very suitable for scenarios in virtual anchors or movie special effects where humanoid characters are driven by actors' performances.

[0083] For example, AI-generated animation switching commands specifically utilize AI technology to automatically generate character animations based on text, speech, or other high-level instructions. Specifically, AI can generate single-frame target skeletal parameter vectors, or it can convert the AI-generated output into a series of target skeletal parameter vectors (e.g., generating lip-shape changes and facial expression parameters based on speech, or generating continuous action sequences based on textual narratives), determined autonomously according to actual needs. Regardless of the source of the animation commands, the final driving layer is a unified skeletal parameter vector control. Therefore, AI-generated content can be naturally integrated with Gaussian point cloud detail rendering to achieve high-fidelity digital character animation driven by artificial intelligence.

[0084] For example, user interaction can generate animation switching commands. Specifically in XR (Extended Reality) or games, it's common for users to directly control character movements via controllers, gestures, and facial expressions. This solution supports such real-time interactive input because any real-time input signal can be converted into adjustments to skeletal node parameters. For instance, when a player smiles in a VR device, the system captures this facial expression and converts it into parameters for the humanoid character's facial control nodes. The corresponding facial mesh and Gaussian point cloud details will immediately change synchronously, thereby enhancing the user's immersive experience.

[0085] In this example, as long as the target skeleton parameter vector can be generated, regardless of whether the animation instructions come from a preset keyframe sequence, a real-time motion capture device, user interaction control, or automatic generation by an AI algorithm, the method of the present invention can receive and drive the corresponding digital character animation.

[0086] As an example, in step S12, the computer device can input the target bone parameter vector corresponding to the target bone node into the parameterized bone model corresponding to the digital character, and use the parameterized bone model to transform the target bone parameter vector corresponding to the target bone node and its current bone parameter vector to determine the bone transformation matrix between the two. This bone transformation matrix is ​​the transformation matrix for switching from the current bone parameter vector corresponding to the target bone node to its corresponding target bone parameter vector, and it can be a 4×4 homogeneous transformation matrix, including rotation and translation.

[0087] The first bone skinning weight is the proportion by which a bone node drives the movement of a mesh vertex; this first bone skinning weight is a pre-set bone skinning weight. Current vertex attributes refer to the properties of a mesh vertex in the current animation image, which may include position or other attributes. Target vertex attributes refer to the attributes that control the switching of mesh vertices, specifically the attributes that need to be switched to in the target animation image.

[0088] As an example, in step S13, after determining the target skeletal node, the computer device needs to identify the mesh vertices that are bound to the target skeletal node from the mesh model corresponding to the digital character, designate them as target mesh vertices, and obtain the current vertex attributes corresponding to the target mesh vertices; the target vertex attributes corresponding to the target mesh vertices can be obtained through the linear skinning formula. Linear calculation yields, where, For target vertex attributes (e.g., target vertex position). This refers to the attributes of the current vertex (e.g., the current vertex position). Let the first bone skinning weight be the weight of the i-th target bone node. Let i be the target transformation matrix corresponding to the i-th target skeletal node, and then... It can quickly determine the target vertex attributes corresponding to the target mesh vertices that are bound to the target bone node when the target bone node is moving.

[0089] In this example, the mesh model corresponding to the digital character is pre-bound to the parameterized skeleton model corresponding to the digital character to form a skeleton mesh model. Each mesh vertex in the skeleton mesh model is bound to one or more bone nodes. The binding relationship here is characterized by the fact that when the bone node moves, the mesh vertex bound to it moves accordingly. The proportion of the movement of each bone node and the mesh vertex is determined as the first bone skinning weight.

[0090] The second bone skinning weight is the proportion by which a bone node drives the movement of a Gaussian point; this second bone skinning weight is a pre-set bone skinning weight. The current Gaussian attribute refers to the attribute of the Gaussian point in the current animation image. The target Gaussian attribute refers to the attribute that needs to be controlled when switching the Gaussian point, specifically the attribute that needs to be switched to in the target animation image. The Gaussian attribute here can include, but is not limited to, position, spatial expansion radius, orientation, color, density, or other attributes.

[0091] As an example, in step S14, after determining the target skeleton node, the computer device needs to identify Gaussian points that are bound to the target skeleton node from the Gaussian point cloud model corresponding to the digital character, designate them as target Gaussian points, and obtain the current Gaussian attribute corresponding to the target Gaussian point; the target Gaussian attribute corresponding to the target Gaussian point can be obtained through the linear skinning formula. Perform linear calculations, where, For target Gaussian properties (e.g., target Gaussian position). This represents the current Gaussian property (e.g., the current Gaussian position). The second bone skinning weight for the i-th target bone node. Let i be the target transformation matrix for the i-th target skeletal node, and then... It can quickly determine the target Gaussian attributes of the target Gaussian points that are bound to the target skeletal node when the target skeletal node is moving.

[0092] In this example, the Gaussian point cloud model corresponding to the digital character is bound to the parametric skeletal model corresponding to the digital character. This binding can be direct or indirect, via the mesh model of the digital character. Each Gaussian point in the Gaussian point cloud model is bound to one or more skeletal nodes. This binding relationship is characterized by the Gaussian point moving along with the skeletal node. The proportion by which each skeletal node causes the Gaussian point to move is determined as the second skeletal skinning weight. Because Gaussian points have spatial expansion properties, their shape (the distribution range described by the covariance matrix) also needs to be adjusted according to the scaling or rotation of the skeleton. For example, when the skeleton undergoes non-uniform scaling, the radius of the Gaussian point can be scaled proportionally; when the skeleton rotates, the distribution direction of the Gaussian point needs to rotate synchronously to maintain the correct orientation relative to the surface. After the above calculations, the new position and shape of each Gaussian point are determined, realizing the Gaussian point cloud model's tracking of skeletal movement.

[0093] For example, when a digital character is a humanoid character, its target skeletal nodes can include skeletal joint nodes and / or facial expression control nodes. Facial expression control nodes are used to control changes in facial expressions. For instance, by adjusting the skeletal parameter vector of the "open mouth" facial expression control node, the rotation of the mandible and the separation and movement of several Gaussian points around the mouth are controlled, thereby simultaneously affecting the target mesh vertices (causing the jaw to open and the lips to separate geometrically) and target Gaussian points (causing the stretching and changes in lip texture and wrinkles) that are associated with it. Similarly, the expression of raising the eyebrows is driven by facial expression control nodes such as the forehead and brow bone. When the eyebrows are raised, the target mesh vertices that are bound to the facial expression control nodes corresponding to the eyebrows rise, and the eyebrow details (such as the distribution of wrinkles and hair) in the target Gaussian points move upwards and bend accordingly. Throughout this process, all these changes are uniformly regulated by a set of target skeletal parameter vectors (including expression coefficients used to control the facial expression control nodes), ensuring that the detailed movement of facial expressions and the detailed changes corresponding to Gaussian points are coordinated and synchronized, without any lag between details and mesh deformation.

[0094] As an example, in step S15, the computer device can render the target vertex attributes corresponding to the target mesh vertices and the target Gaussian attributes corresponding to the target Gaussian points to update the animation and obtain the updated target animation image. In this example, the target mesh vertices are skinned onto the target bone nodes, and the target Gaussian points are also skinned onto the target bone nodes. When rendering the target vertex attributes corresponding to the target mesh vertices and the target Gaussian attributes corresponding to the target Gaussian points, it is possible to achieve synchronous updates of the target mesh vertices and the target Gaussian cloud based on a set of target bone parameter vectors corresponding to the target bone nodes. This avoids the problem of misalignment or pose mismatch caused by any two being out of sync, thereby ensuring the overall quality of the target animation image.

[0095] During animation playback or interactive driving, a set of skeletal parameters (target bone parameter vectors corresponding to target bone nodes) are input, including rotational displacement parameters of each joint of the body and control parameters for facial expressions (such as expression coefficients). This simultaneously drives the deformation of the mesh model and the Gaussian point cloud model: mesh vertices are calculated to new positions according to traditional skeletal skinning, achieving large-scale deformation of the body and face; concurrently, each Gaussian point in the Gaussian point cloud model, based on its bone binding weights, is calculated using the bone transformation matrix of the corresponding joint to obtain a new position and orientation under the current skeletal posture. In this way, the mesh vertices and Gaussian points of the digital character deform synchronously, ensuring that facial expressions and body movements change collaboratively driven by a unified skeleton. For example, when the skeletal parameters specify raising the arm and showing a smile, the arm vertex of the mesh model will raise, and the mesh vertex corresponding to the facial expression will produce a smiling expression; at the same time, the Gaussian points in the Gaussian point cloud model corresponding to the arm muscle texture and the Gaussian points for facial details will adjust their positions and shapes following the skeletal movement, achieving consistency between muscle bulges and facial wrinkle changes and the overall movement. Through this integrated driving mechanism, facial expressions and body movements are no longer driven independently, avoiding the asynchronous and uncoordinated situation between the face and body in traditional methods.

[0096] In the process of skeletal-driven digital character movement, this invention employs a strategy of coordinated control of joint rotation and facial expression parameters to ensure that the details of the Gaussian point cloud model are updated in real time along with the movement of the skeletal mesh model. Specifically, when the bone nodes in the parametric skeletal model cause deformation of the mesh vertices in the mesh model, each Gaussian point in the Gaussian point cloud model adjusts its position in real time according to the transformation of its bound bone node, while maintaining its distribution shape relative to the local surface. For example, if the facial bones control the mouth to open, the Gaussian points at the lips will be stretched and dispersed, presenting the details of the lip texture changing as it opens; if the body joints bend, the Gaussian points at the corresponding skin folds will converge or stretch according to the change in the joint angle, dynamically reflecting the changes in wrinkles and muscle direction. At the same time, the texture color, normal direction, and other attributes carried in the Gaussian point cloud model can also be updated in conjunction with the bone pose to match new lighting angles or changes in surface orientation, thereby ensuring that the rendering result of the Gaussian point cloud model maintains consistent lighting and texture effects with the mesh surface. Through the above-mentioned coordinated update mechanism, this invention can provide high-fidelity and smooth, synchronized facial expression details and limb movement performance during animation. In other words, even when a digital character moves at high speed or makes complex facial expressions continuously, the details represented by the Gaussian point cloud model remain strictly synchronized with the bone-driven mesh deformation, without any lag or distortion. This method ensures smooth animation while preserving a great deal of detail in facial textures and other details, making digital character animations more realistic and vivid.

[0097] By combining the above technical solutions, this invention establishes binding relationships between bone nodes and mesh vertices, as well as between bone nodes and Gaussian points, through a parametric skeletal model, a mesh model, and a Gaussian point cloud model. This enables synchronized driving of limb movements and specific body part movements in digital characters, achieving high-precision animation display. Compared with existing technologies, this invention has the following advantages:

[0098] (1) The limb movements corresponding to the skeletal joint nodes and the specific movements corresponding to the control nodes of specific parts are controlled in a unified manner. For example, the same set of skeletal parameters is used to drive the limb movements and facial expressions to change synchronously, avoiding the asynchrony or conflict that may occur when driving them separately, ensuring that the movements of each part of the digital character are coordinated and consistent, and improving the naturalness and coherence of the character animation.

[0099] (2) Introducing a Gaussian point cloud model as a representation of surface details allows digital characters to retain rich textures and geometric details during animation, resulting in high detail fidelity. Compared to a pure mesh approach, this method does not require an excessive increase in the number of mesh polygons to achieve a level of detail comparable to high-density models, while also reducing the demand for rendering resources.

[0100] (3) By introducing the synchronous changes of the Gaussian point cloud model and the parametric bone model, the details in the Gaussian point cloud model are updated dynamically when the bone nodes move, so as to achieve the effect of real-time detail update and solve the defect that traditional texture mapping or offline deformation cannot reflect the changes in motion details in real time. Whether it is a fast limb movement or a subtle change in facial expression, the details can be adjusted in real time in response to the bone drive, ensuring that the animation process is smooth and distortion-free.

[0101] (4) Improve the universality and scalability of the solution. This invention adopts a universal description method, does not rely on a specific brand or existing model, and the implementation method is applicable to different types of digital characters. By describing the essence of the "Gaussian point cloud model" and defining the skeletal driving function, this technical solution has a wide range of applications. It can be used for the animation control of digital characters with real-life proportions, as well as cartoon characters and animal characters, and can be compatible and extended with other animation technology modules (such as physical simulation, gaze tracking, etc.).

[0102] In one embodiment, step S14 is to transform the current Gaussian attribute corresponding to the target Gaussian point based on the bone transformation matrix and the second bone skinning weight corresponding to the target Gaussian point, and determine the target Gaussian attribute corresponding to the target Gaussian point, including A141-A142.

[0103] A141: Based on the bone transformation matrix and the second bone skinning weight corresponding to the target Gaussian point, the current Gaussian attribute corresponding to the target Gaussian point is transformed to determine the initial Gaussian attribute corresponding to the target Gaussian point. The initial Gaussian attribute is the position of the Gaussian point, the covariance matrix and the state of the rendering attribute obtained directly from the bone transformation.

[0104] A142: Based on the elastic network corresponding to the target Gaussian point, the initial Gaussian property corresponding to the target Gaussian point is oscillated and simulated to determine the target Gaussian property corresponding to the target Gaussian point.

[0105] As an example, in step A141, after determining the target skeletal node, the computer device needs to identify Gaussian points that are bound to the target skeletal node from the Gaussian point cloud model corresponding to the digital character, designate these points as the target Gaussian points, and obtain the current Gaussian attributes corresponding to the target Gaussian points; then, using the linear skinning formula... Perform linear calculations to determine the initial Gaussian properties corresponding to the target Gaussian point. Specifically, the initial Gaussian properties here are the basic state of the Gaussian point position, covariance matrix, and rendering properties (color, normal direction, etc.) obtained directly from the bone transformation.

[0106] As an example, in step A142, the computer device performs oscillation simulation on the initial Gaussian properties corresponding to the target Gaussian point based on the elastic network pre-established for the target Gaussian point, and determines the target Gaussian properties corresponding to the target Gaussian point. Specifically, it can calculate the restoring force and damping force corresponding to each target Gaussian point based on the initial position of the target Gaussian point, the spring stiffness and damping coefficient connected by the virtual elastic network, etc.; then, based on the restoring force and damping force corresponding to each target Gaussian point, it determines the net external force of each target Gaussian point; based on the net external force and the mass of the Gaussian point, it obtains the acceleration through Newton's second law, and uses semi-implicit Euler integral to update the velocity and position of the Gaussian point, forming a single-frame oscillating displacement; after the skeleton movement stops, the target position driving the target Gaussian point no longer changes, but some Gaussian points need to perform cyclic calculations due to residual vibration until the velocity is lower than the convergence threshold, completing the decay process of the secondary shaking, determining the target Gaussian properties corresponding to the target Gaussian point, and presenting the visual effect of secondary shaking.

[0107] In this example, Gaussian point cloud models are well-suited for representing the volume and deformation of soft tissue. For instance, soft areas like a character's abdomen or cheeks may exhibit slight secondary swaying after skeletal movement ceases. This approach can simulate inertial delay effects using a simplified spring-damped model, or it can be combined with muscle or soft tissue physics simulations. Specifically, an elastic network (such as a virtual spring net) can be established between Gaussian points. After the initial displacement is provided by the skeleton, the oscillation convergence of the elastic network over several frames simulates the jittering of fat or muscle to achieve a more realistic effect. Similarly, this extension can be integrated into existing frameworks: the skeleton provides the target position, the actual position of the Gaussian points is calculated through physics simulation, and then used for rendering output. Therefore, Gaussian point cloud models can be easily coupled with soft body physics simulations to achieve richer and more nuanced dynamic effects.

[0108] In one embodiment, step S14 is to transform the current Gaussian attribute corresponding to the target Gaussian point based on the bone transformation matrix and the second bone skinning weight corresponding to the target Gaussian point, and determine the target Gaussian attribute corresponding to the target Gaussian point, including B141-B142.

[0109] B141: Based on the bone transformation matrix and the second bone skinning weight corresponding to the target Gaussian point, transform the current Gaussian attribute corresponding to the target Gaussian point to determine the initial Gaussian attribute corresponding to the target Gaussian point. The initial Gaussian attribute is the Gaussian point position, covariance matrix and rendering attribute state directly calculated by the bone transformation.

[0110] B142: Based on pre-set physical constraints and a physics engine, perform collision detection and adjustment on the initial Gaussian properties corresponding to the target Gaussian point to determine the target Gaussian properties corresponding to the target Gaussian point.

[0111] As an example, in step B141, after determining the target skeletal node, the computer device needs to identify Gaussian points that are bound to the target skeletal node from the Gaussian point cloud model corresponding to the digital character, designate these points as the target Gaussian points, and obtain the current Gaussian attributes corresponding to the target Gaussian points; then, using the linear skinning formula... Perform linear calculations to determine the initial Gaussian properties corresponding to the target Gaussian point. Specifically, the initial Gaussian properties here are the basic state of the Gaussian point position, covariance matrix, and rendering properties (color, normal direction, etc.) obtained directly from the bone transformation.

[0112] As an example, in step B142, after determining the initial Gaussian properties corresponding to the target Gaussian point, the computer device can introduce physical constraints, apply gravity to all Gaussian points, and calculate the displacement compensation amount under the action of gravity. Then, in the physics engine, a collider is constructed for the character's body, clothing itself, and environmental obstacles, and an intersection judgment is made for the target Gaussian point. If penetration occurs, the collision displacement correction amount is calculated according to preset collision rules (such as bounce coefficient and penetration correction amount). Then, the displacement compensation amount under the action of gravity and the collision displacement correction amount are merged to determine the target compensation amount. Finally, the initial Gaussian properties corresponding to the target Gaussian point are corrected based on the target compensation amount to determine the target Gaussian properties corresponding to the target Gaussian point.

[0113] In this embodiment, the Gaussian point cloud model can be combined with a physics engine to realize collision responses between the digital character and its own body or external environment. For example, Gaussian point clouds can be applied to the clothing or hair areas of the character. Based on the movement driven by the skeleton, physical constraints and a physics engine are introduced to make these Gaussian points subject to gravity and collisions, thereby simulating fluttering clothing or swaying hair. This requires adding collision detection after the skeleton movement is updated and making secondary adjustments to the positions of the relevant Gaussian points. In addition, the volumetric expansion characteristics of Gaussian point clouds help avoid clothing "clipping" phenomena, because Gaussian points are distributed in volumetric form, which can more smoothly handle potential geometric penetration during collision detection.

[0114] In this embodiment, it can be integrated with existing physics engines or animation system modules (such as cloth simulation modules, rigid body collision modules, etc.): On the one hand, this solution is compatible with other animation technology modules (such as physics simulation extensions), demonstrating its openness; on the other hand, in specific implementations, reasonable interfaces or coordination mechanisms can be designed to allow bone driving and physics simulation to perform their respective functions while maintaining synchronization, thereby not disrupting the consistency of the animation. Through such modular integration, the application scope of the method of this invention can be further expanded and the realistic effect of the animation can be enhanced.

[0115] In one embodiment, step S15, namely determining the target animation image based on the target vertex attributes corresponding to the target mesh vertices and the target Gaussian attributes corresponding to the target Gaussian points, includes:

[0116] S151: Perform rasterization rendering on the target vertex attributes corresponding to the target mesh vertices to determine the mesh rendering image;

[0117] S152: Perform rasterization rendering on the target Gaussian attributes corresponding to the target Gaussian point to determine the Gaussian rendered image;

[0118] S153: Perform fusion processing on the mesh rendering image and the Gaussian rendering image to determine the target animation image.

[0119] As an example, in step S151, after obtaining the target vertex attributes corresponding to the target mesh vertices, the computer device needs to form a triangle to be rendered based on these attributes. Then, a rasterization operation is performed on the triangle to be rendered, specifically projecting it onto screen space to determine multiple mesh pixels corresponding to the screen space. Finally, each mesh pixel is shading-processed to determine its color, depth, or other details, forming a mesh rendering image. This mesh rendering image is an image re-rendered based on the target mesh vertices. In this example, rasterization rendering based on the target mesh vertices with changing vertex attributes helps ensure the rendering efficiency of the mesh rendering image.

[0120] As an example, in step S152, after obtaining the target Gaussian attributes corresponding to the target Gaussian point, the computer device needs to project the target Gaussian point onto the screen space to determine the two-dimensional projection space corresponding to the target Gaussian point. This two-dimensional projection space can be circular or elliptical, determined by the scaling and orientation of the target Gaussian point. Next, the contribution of each target Gaussian point to the screen pixels is calculated. Finally, based on the contribution of the target Gaussian point to the screen pixels, depth testing and blending processing are performed to handle the occlusion relationship between the target Gaussian point and adjacent Gaussian points, determining the pixel color, depth, or other details corresponding to each target Gaussian point, thus forming a Gaussian rendered image. This Gaussian rendered image is an image re-rendered based on the target Gaussian point. In this example, rasterization rendering is performed based on the target Gaussian point whose Gaussian attributes have changed, which helps to ensure the rendering efficiency of the Gaussian rendered image.

[0121] As an example, in step S153, after determining the mesh rendering image and the Gaussian rendering image, the computer device can perform a fusion process on the mesh rendering image and the Gaussian rendering image to determine the target animation image. In one example, the mesh rendering image can be used as the bottom layer image, and the Gaussian rendering image (usually with a transparency effect) can be used as the top layer image. The two can be superimposed to enhance the detailed features of the final target animation image.

[0122] In this embodiment, since the Gaussian point cloud model, as a point-based primitive, does not involve complex topology recalculation, it is suitable for GPU parallel processing. This allows for parallel processing of mesh model rendering and Gaussian point cloud model rendering, meaning that parallel rasterization rendering based on splats is performed directly after skeletal transformation, ensuring rendering efficiency and stability while achieving scheduling synchronization. Merging the mesh-rendered image and the Gaussian-rendered image to determine the target animation image ensures the overall quality of the final target animation image. In contrast, if an extremely high polygon count mesh is used to attempt to achieve the same level of detail, the skeletal drive calculation and rasterization become significantly more burdened. This example achieves a trade-off between performance and effect by attaching skeletal drives to a reasonable number of target Gaussian points. In testing, for a Gaussian skeletal mesh model formed by binding a mesh model containing tens of thousands of mesh vertices and a Gaussian point cloud model containing hundreds of thousands of Gaussian points, real-time frame rate animation updates and rendering can still be achieved on modern GPUs, demonstrating the feasibility and practicality of this example.

[0123] Steps S151-S153 describe in detail how to perform rasterization rendering on the transformed target mesh vertices and target Gaussian points separately, and then fuse the rendered mesh image and Gaussian image to determine the target animation image. In particular, the rendering process for Gaussian points provides specific steps for projecting the Gaussian points as elliptical splats, calculating pixel contribution values, and performing transparency blending. These implementation steps essentially provide a pseudocode-like explanation, making it clear how each step is executed and ensuring the correct implementation of the animation frame update process.

[0124] In one embodiment, the target Gaussian properties include the number of renders and the spatial expansion radius;

[0125] Step S152, namely, rasterizing the target Gaussian attribute corresponding to the target Gaussian point to determine the Gaussian rendering image, including A1521-A1523;

[0126] A1521: Get the rendering distance corresponding to the target Gaussian point;

[0127] A1522: Based on the rendering distance corresponding to the target Gaussian point, determine the number of renders and the spatial expansion radius corresponding to the target Gaussian point;

[0128] A1523: Perform rasterization rendering based on the number of renders corresponding to the target Gaussian point and the spatial expansion radius to determine the Gaussian rendered image.

[0129] As an example, in step A1521, during the rasterization rendering process of the target Gaussian point, the computer device needs to obtain the distance between the target Gaussian point and the camera (virtual camera) and determine it as the rendering distance corresponding to the target Gaussian point. This rendering distance is used to determine whether the rendered scene corresponding to the target Gaussian point is a foreground scene or a background scene. Generally, when the rendering distance corresponding to the target Gaussian point is greater than a preset distance, the rendered scene can be determined to be a background scene; conversely, when the rendering distance corresponding to the target Gaussian point is not greater than the preset distance, the rendered scene can be determined to be a foreground scene.

[0130] As an example, in step A1522, when the computer device determines the rendering distance corresponding to the target Gaussian point and the corresponding rendering scene, it adopts a rendering adjustment strategy corresponding to the rendering scene, adaptively updating the number of renders and the spatial expansion radius corresponding to the target Gaussian point, and applying the updated parameters to the subsequent Gaussian point rasterization rendering process. Here, the number of renders corresponding to the target Gaussian point refers to the actual number of target Gaussian points participating in rasterization rendering. The spatial expansion radius refers to the radius of the coverage area of ​​a single Gaussian point in three-dimensional space, determining the shape and size of the Gaussian point projected onto the screen space. In this example, when the rendering scene is a distant scene, the number of renders corresponding to the target Gaussian point and / or the spatial expansion radius can be reduced to optimize system performance; when the rendering scene is a close-up scene, the number of renders corresponding to small target Gaussian points and / or the spatial expansion radius can be increased to obtain more detailed features and improve detail fidelity.

[0131] As an example, in step A1523, after obtaining the updated number of renders and spatial expansion radius corresponding to the target Gaussian point, the computer device needs to perform rasterization processing on parameters including but not limited to the number of renders and spatial expansion radius corresponding to the target Gaussian point based on a pre-set Gaussian rasterization rendering process, calculate the pixel contribution of the Gaussian distribution, and determine its final Gaussian rendered image.

[0132] In this embodiment, the application scenario of dynamically adjusting the number of Gaussian points rendered and the spatial expansion radius based on the rendering distance (i.e., near and far viewing distance) corresponding to the target Gaussian point is described. For example, the number of Gaussian points involved in rendering is reduced to optimize performance when rendering distant scenes, while the density of Gaussian points is increased to improve detail accuracy when rendering close-up scenes. This embodiment demonstrates how to dynamically switch the level of detail (LOD) based on the viewing distance in practical applications, achieving a balance between performance and image detail, thereby enriching the application of this solution in performance optimization.

[0133] In one embodiment, the target Gaussian properties include the target position and the target orientation;

[0134] Step S152, namely, rasterizing the target Gaussian attribute corresponding to the target Gaussian point to determine the Gaussian rendering image, including B1521-B152;

[0135] B1521: Determine the position deviation based on the target position corresponding to the target Gaussian point and the current position of the target Gaussian point;

[0136] B1522: Based on the inertial physics equations, the position deviation is processed to determine the inertial position corresponding to the target Gaussian point;

[0137] B1523: Based on the inertial position corresponding to the target Gaussian point, the target Gaussian property corresponding to the target Gaussian point is corrected, and the corrected Gaussian property corresponding to the target Gaussian point is determined.

[0138] B1524: Perform rasterization rendering on the modified Gaussian properties corresponding to the target Gaussian point to determine the Gaussian rendered image.

[0139] The inertial physics equations are pre-defined equations used to simulate the hysteresis of Gaussian points tracking skeletal motion, producing inertial effects such as natural rebound or tremors. As an example, the inertial physics equations include, but are not limited to, the spring-damped model (MSD). The current position of the target Gaussian point refers to its location within the current animation image. The target position corresponding to the target Gaussian point is the position to which the target skeletal node needs to be switched after it is driven.

[0140] As an example, in step B1521, the computer device can determine the positional deviation between the target position corresponding to the target Gaussian point and the current position of the target Gaussian point. This positional deviation reflects the positional deviation of the target Gaussian point in the two consecutive animation frames that need to be controlled. The larger the positional deviation, the stronger its inertial driving force.

[0141] As an example, in step B1522, the computer device can process the position deviation using a pre-set inertial physics equation, for example, by using a spring-damped model to process the position deviation, first determining the acceleration corresponding to the position deviation; then updating the velocity based on the acceleration; and finally determining the inertial position of the target Gaussian point based on the updated velocity and the current position of the target Gaussian point.

[0142] This inertial position can be understood as using inertial physics equations to predict the current position of the target Gaussian point under the inertial action driven by the skeleton.

[0143] As an example, in step B1523, after determining the inertial position corresponding to the target Gaussian point, the computer device can correct the target Gaussian attribute corresponding to the target Gaussian point based on the inertial position to determine the corrected Gaussian attribute corresponding to the target Gaussian point. For example, when the target Gaussian attribute includes the target position and the target orientation, the inertial position can be determined as the corrected position; the target orientation can be corrected based on the orientation displacement corresponding to the inertial position to determine the corrected orientation; the corrected position and the corrected orientation can be determined as the corrected Gaussian attribute corresponding to the target Gaussian point.

[0144] As an example, in step B1524, after obtaining the modified Gaussian attribute corresponding to the target Gaussian point, the computer device needs to perform rasterization processing on the modified Gaussian attribute corresponding to the target Gaussian point based on the pre-set Gaussian rasterization rendering process to determine its final Gaussian rendered image.

[0145] In this embodiment, physical simulation can be superimposed on the skeleton drive to simulate the inertial driving force driven by the limb movements and specific movements of specific parts during the process of switching from the current Gaussian attribute corresponding to the target Gaussian point to the target Gaussian attribute, and the adjusted inertial position, so that the details of the Gaussian point cloud model respond to physical effects such as inertia and collision. The physical simulation parameters and skeleton parameters work together without destroying the synchronous deformation relationship between the mesh and the point cloud.

[0146] This embodiment illustrates a specific application scenario of Gaussian point inertial delay, which corresponds to the physical phenomena of character movement in reality: for example, when a digital character suddenly stops moving rapidly, the soft parts of the body will still sway slightly, or the clothing will still swing for a short period of time after the movement ends, etc. This embodiment uses a spring-damped model to simulate this inertial delay following effect, making the animation more realistic and natural, and making the animation show the natural rebound or trembling effect after the action stops.

[0147] In one embodiment, the target Gaussian property includes the target color;

[0148] Step S152, namely, rasterizing the target Gaussian attribute corresponding to the target Gaussian point to determine the Gaussian rendering image, including C1521-C1524;

[0149] C1521: Determine the driving state of the target bone node based on the target bone parameter vector corresponding to the target bone node;

[0150] C1522: Determine the color correction coefficient corresponding to the driving state based on the driving state corresponding to the target bone node;

[0151] C1523: Correct the target color corresponding to the target Gaussian point based on the color correction coefficient, and determine the corrected color corresponding to the target Gaussian point;

[0152] C1524: Perform rasterization rendering based on the corrected color corresponding to the target Gaussian point to determine the Gaussian rendered image.

[0153] As an example, in step C1521, the computer device can determine the driving state corresponding to the target bone node based on the target bone parameter vector corresponding to the target bone node and the current bone parameter vector corresponding to the target bone node. The driving state is used to reflect the type of state that the bone node needs to drive the digital character to change its state. It can be a movement posture or an emotional change, etc.

[0154] As an example, in step C1522, after obtaining the driving state corresponding to the target skeletal node, the computer device can query a pre-set state color coefficient mapping table based on the driving state to determine the color correction coefficient corresponding to the driving state. This state color coefficient mapping table is used to characterize the mapping relationship between different driving states and their corresponding correction colors. The color correction coefficient is the coefficient used for color correction under the driving state.

[0155] As an example, in step C1523, after determining the color correction coefficient, the computer device needs to input the color correction coefficient to the target color corresponding to the target Gaussian point into the color correction function corresponding to the driving state for processing, so as to determine the corrected color corresponding to the target Gaussian point.

[0156] As an example, in step C1524, after obtaining the corrected color corresponding to the target Gaussian point, the computer device needs to perform rasterization processing on the corrected color corresponding to the target Gaussian point based on the pre-set Gaussian rasterization rendering process to determine its final Gaussian rendered image.

[0157] In this embodiment, the color information of the Gaussian point cloud model supports dynamic updates. It can be adjusted in combination with the driving state corresponding to the target skeletal node in the digital character (such as reddening due to injury, facial discoloration caused by emotions). The color update is performed synchronously with the deformation driven by the skeleton to ensure visual consistency.

[0158] This embodiment illustrates how driving states (such as emotions and the degree of force) affect the color effects of Gaussian point cloud rendering. For example, when a character is angry, their face turns red, and when injured, their skin darkens. These effects can be achieved by mapping the driving states to the color parameters of Gaussian points. This embodiment expands the expressive dimensions of this solution, demonstrating that it can not only handle geometric deformation details but also reflect the dynamic changes in material appearance (color), making digital characters more vivid and realistic.

[0159] In one embodiment, the digital character is a humanoid character, and the control nodes for specific parts include facial expression control nodes;

[0160] Alternatively, the digital character may be an animal character, with specific part control nodes including tail control nodes and / or ear control nodes;

[0161] Alternatively, the digital character can be a cartoon character, and the control nodes for specific body parts can be exaggerated facial expression control nodes.

[0162] For example, when the digital character is a humanoid character, the parametric skeletal model corresponding to the humanoid character includes skeletal joint nodes such as the torso and limbs (such as the spine, shoulders, elbows, hips, knees, etc.) and several facial expression control nodes (such as eyebrow raising, eye opening and closing, mouth opening and closing, etc.). Here, the facial expression control nodes are control nodes for specific parts. The skeletal joint nodes and facial expression control nodes are used to realize the synchronous control of the humanoid character's limb movements and facial expressions.

[0163] For example, when the digital character is an animal character, the corresponding parametric skeletal model of the animal character includes skeletal joint nodes (torso joints, limb joints and head joints) as well as control nodes for specific parts such as tails or ears. Through these skeletal nodes, the synchronous control of the animal character's limb movements and the detailed movements of its tail / ears is realized.

[0164] For example, when the digital character is a cartoon character, the corresponding parametric skeletal model of the cartoon character includes skeletal joint nodes and exaggerated expression control nodes. The skeletal joint nodes and exaggerated expression control nodes are used to realize the synchronous control of the cartoon character's limb movements and exaggerated expressions.

[0165] In this example, the skeletal nodes of the parametric skeletal model support custom extensions, exhibiting good versatility and scalability. For animal characters, limb joint chains, tail joint chains, and ear control nodes can be added; for cartoon characters, exaggerated expression control nodes (such as eye zooming and mouth stretching) can be added. The extended skeletal nodes still employ unified weight binding rules and transformation matrix calculation logic. For different forms of digital characters, "specific part control nodes" are introduced, such as facial expression control for humanoid characters, tail / ear control for animal characters, and exaggerated expression control for cartoon characters. Figure 6 The extended architecture diagram shown illustrates how adding corresponding nodes to the basic skeletal model and adjusting parameters such as Gaussian point cloud sampling density and radius can adapt to different types of characters. This means that regardless of whether the digital character is humanoid, animal, or cartoon, this method can be applied by extending the skeletal nodes and adjusting the Gaussian point cloud parameters, demonstrating cross-character portability. Furthermore, this example uses a general description method, independent of specific brands or existing models, and has a wide range of applicability. This is crucial in practical applications because digital characters come in various forms, and only highly versatile solutions can be widely adopted.

[0166] Based on the aforementioned versatility, this solution can also be extended to other uncommon digital character types, such as virtual creatures or monsters with unconventional skeletal structures, and is even applicable to scenarios involving synchronized multi-character animation. From a technical perspective, as long as corresponding bone definitions and weight bindings are provided for different digital characters, this solution is also applicable to fantasy creatures (such as creatures with multiple arms or tails) and animation control driven by multiple characters in parallel (i.e., multiple characters each having their own independent bone-mesh-Gaussian model). By extending the corresponding control nodes on the basic bone model and adjusting the Gaussian point cloud parameter settings, this solution can seamlessly adapt to such extended scenarios, further demonstrating its universality and support for multi-character extensions.

[0167] This invention provides a digital character modeling method, which is illustrated using an example of its application in a computer device (hereinafter referred to as a computer device). Figure 2 As shown, digital character modeling methods include:

[0168] S21: Obtain the bone parameter vector corresponding to the skeletal structure, and construct a parameterized skeletal model of the digital character based on the bone parameter vector corresponding to the skeletal structure. The skeletal structure includes multiple bone nodes, and the bone nodes include bone joint nodes and control nodes for specific parts.

[0169] S22: Bind the parameterized skeletal model of the digital character to the mesh model of the digital character to determine the skeletal mesh model corresponding to the digital character; wherein, the skeletal mesh model includes multiple first-bound bone nodes corresponding to all mesh vertices and the first bone skinning weight corresponding to each first-bound bone node;

[0170] S23: Bind the Gaussian point cloud model of the digital character to the corresponding skeletal mesh model of the digital character, and determine the Gaussian skeletal mesh model corresponding to the digital character. The Gaussian skeletal mesh model is used to realize the animation control of the digital character. The Gaussian point cloud model includes multiple Gaussian points with spatial expansion radius, and the Gaussian skeletal mesh model includes multiple second bound bone nodes corresponding to all Gaussian points and the second bone skinning weight corresponding to each second bound bone node.

[0171] As an example, in step S21, during the modeling of the digital character, it is necessary to first obtain the skeletal structure of the digital character and the corresponding bone parameter vector. Based on the bone parameter vector, the parameterized skeletal model corresponding to the digital character is determined. The skeletal structure includes multiple bone nodes organized in a hierarchical tree structure. Each bone node has adjustable transformation parameters (translation, rotation, scaling, etc.), which are specifically achieved by adjusting its corresponding bone parameter vector. In this example, the parameterized skeletal model of the digital character can describe any pose and expression of the digital character using the bone parameter vectors of different skeletal characters. For example, given a set of joint angles and expression coefficients, the pose configuration of the entire skeleton can be determined.

[0172] As an example, in step S22, after obtaining the parametric skeletal model corresponding to the digital character, the computer device needs to obtain the mesh model of the digital character. The mesh model includes multiple mesh vertices, and its topological structure is a structure formed by triangular facets. Next, the parametric skeletal model corresponding to the digital character is bound to the mesh model of the digital character to form the skeletal mesh model corresponding to the digital character. The skeletal mesh model includes multiple first-bound bone nodes corresponding to all mesh vertices and a first bone skinning weight corresponding to each first-bound bone node to achieve traditional linear skinning (LBS) deformation. Here, the first-bound bone node is used to represent the bone node that will cause the mesh vertex to move, and the first bone skinning weight refers to the proportion by which the finally determined first-bound bone node causes the mesh vertex to move.

[0173] In this example, for any mesh vertex, there are K first-bound bone nodes, and the first bone skinning weight for each first-bound bone node is... The sum of the skinning weights of the first bones corresponding to the K first bound bone nodes is 1, that is... This indicates that the mesh vertex is controlled by K first-bound bone nodes. Limiting the sum of the skinning weights of all corresponding first bones to 1 can prevent excessive movement or misalignment of the mesh vertex. For example, if mesh vertex A is bound to two first-bound bone nodes, the thigh and the calf, then when the thigh bone rotates and lifts, mesh vertex A will partially lift accordingly. This, combined with the influence of the calf bone, achieves continuous and smooth skin deformation.

[0174] As an example, in step S23, after acquiring the skeletal mesh model corresponding to the digital character, the computer device needs to pre-sample or generate a series of Gaussian points carrying spatial expansion radii. Each Gaussian point can explicitly represent the surface details and textures of the digital character, specifically through the position and color of the Gaussian point. Here, the position of each Gaussian point corresponds to a specific feature location on the surface of the digital character; the color of the Gaussian point comes from high-resolution texture maps or multi-view camera measurement data. Each Gaussian point carries a spatial expansion radius and shape, specifically a spherical or ellipsoidal Gaussian distribution, considered as a small spot covering a local surface area. A Gaussian point cloud model formed by a large collection of such Gaussian points can approximately reconstruct the surface details and textures of the digital character. It is worth noting that the Gaussian point cloud model is not simply attached to the mesh vertices, but rather an independent explicit surface representation, not directly related to the mesh topology of the skeletal mesh model, but can be aligned with spatial coordinates to cover the mesh surface of the skeletal mesh model. In this example, each Gaussian point carries a spatial expansion radius, meaning it is not an ideal point without dimensions, but rather exhibits a volume diffusion effect based on a Gaussian function in three-dimensional space, used to represent the details and color information within a certain range of the target surface.

[0175] In this example, the computer device needs to bind the Gaussian point cloud model of the digital character to its corresponding skeletal mesh model to incorporate the Gaussian point cloud model into the animation system of the skeletal mesh model, forming a Gaussian skeletal mesh model corresponding to the digital character. This Gaussian skeletal mesh model includes multiple second-bound bone nodes corresponding to all Gaussian point cloud models and a second-bone skinning weight corresponding to each second-bound bone node. Here, the second-bound bone nodes represent the bone nodes that will cause the Gaussian point cloud model to move, and the second-bone skinning weight refers to the proportion by which the final determined second-bound bone node causes the Gaussian point cloud model to move. In this example, any Gaussian point cloud model corresponds to L second-bound bone nodes, and the second-bone skinning weight corresponding to each second-bound bone node is... The sum of the second bone skinning weights corresponding to the L second bound bone nodes is 1, that is... This indicates that the Gaussian point cloud model is controlled by L second-bound bone nodes. Limiting the sum of all corresponding second-bone skinning weights to 1 prevents excessive movement or misalignment of the Gaussian point cloud model. In this example, each Gaussian point is associated with one or more second-bound bone nodes based on its position within the digital character, and the skinning weights corresponding to each second-bound bone node are determined. This ensures that when any second-bound bone node moves, its corresponding Gaussian point cloud model undergoes an affine transformation, causing its position to shift and its shape to change.

[0176] For example, Gaussian points located on the cheeks can be primarily bound to the head bones and partially bound to the neck bones, so that the Gaussian points move with the head when it rotates, and are slightly moved when the neck moves slightly. Similarly, Gaussian points distributed on the skin surface of the upper arm can be bound to the upper arm and forearm bones, with different secondary bone skinning weights assigned based on their distance from the bone nodes to reflect the degree of deformation of skin details when the elbow is bent. After binding, each Gaussian point can be considered to carry a set of weight data corresponding to the bones, stored in the point cloud data structure.

[0177] In this embodiment, since the parametric skeleton model includes skeletal joint nodes and control nodes for specific parts, the Gaussian skeleton mesh model can synchronously drive the limb movements and specific parts of the digital character's movements during digital character animation control through a set of skeletal parameter vectors corresponding to the parametric skeleton model. Because the Gaussian skeleton mesh model binds the parametric skeleton model, the mesh model, and the Gaussian point cloud model, establishing binding relationships between skeletal nodes and mesh vertices, as well as between skeletal nodes and Gaussian points, the set of skeletal parameter vectors corresponding to the parametric skeleton model enables synchronized display of movements and high-precision animation during digital character animation control.

[0178] In one embodiment, step S22, which involves binding the parameterized skeletal model of the digital character to the mesh model of the digital character to determine the skeletal mesh model corresponding to the digital character, includes:

[0179] S221: Determine the first bound bone node corresponding to each mesh vertex of the digital character's mesh model, and determine the initial bone skinning weights corresponding to the first bound bone node;

[0180] S222: Based on the anatomical structure of the digital character, smooth the initial bone skinning weights corresponding to the first bound bone node, and determine the first bone skinning weights corresponding to the first bound bone node.

[0181] S223: Determine the skeletal mesh model corresponding to the digital character based on the first bound bone node corresponding to all mesh vertices and the first bone skinning weight corresponding to each first bound bone node.

[0182] The anatomical structure of digital characters does not refer to the medical anatomy of real organisms, but rather to the digital simulation and structured design of the character's external outline, muscle distribution, and joint movement patterns.

[0183] As an example, in step S221, the computer device can determine one or more first bound bone nodes corresponding to each mesh vertex in the mesh model according to pre-set bound bone node determination rules. For example, if the pre-set number of first bound bone nodes is x, the x bone nodes that are spatially closest to the mesh vertex can be determined as its corresponding first bound bone nodes to avoid interference from distant bone nodes and prevent misalignment caused by distant bone nodes. Next, the computer device determines the initial bone skinning weight of each first bound bone node to the mesh vertex based on the degree of influence of each first bound bone node on the mesh vertex. The initial bone skinning weight corresponding to the first bound bone node refers to the proportion by which the initially determined first bound bone node drives the mesh vertex to move. Specifically, the computer device processes parameters such as the distance between the mesh vertex and the first bound bone node and / or the influence radius of the first bound bone node based on a pre-set weight determination program to determine the initial bone skinning weight. Generally, the smaller the distance between the mesh vertex and the first bound bone node, the higher its corresponding initial bone skinning weight; conversely, the larger the distance, the lower its corresponding initial bone skinning weight. In this example, each bone node is pre-configured with its corresponding influence radius. Mesh vertices will only be assigned initial bone skinning weights if they are within the influence radius of the first bound bone node.

[0184] As an example, in step S222, after determining the first bound bone nodes and their corresponding initial bone skinning weights for all mesh vertices, the computer device needs to smooth the initial bone skinning weights for the first bound bone nodes based on the anatomical structure of the digital character. Specifically, this requires combining the anatomical structure of the digital character, such as joint movement patterns and muscle change patterns, to smooth the initial bone skinning weights. For example, it is necessary to smooth the joint transition areas of bending joints such as knees and elbows, and determine their corresponding smooth bone skinning weights to make the movement of each bone node more natural. Then, it is necessary to perform weight normalization processing on the smooth bone skinning weights of all first bound bone nodes corresponding to the same mesh vertex, and determine the first bone skinning weights of all first bound bone nodes corresponding to that mesh vertex, so that the sum of all first bone skinning weights corresponding to that mesh vertex is 1, to avoid excessive or misaligned movement of the mesh vertex.

[0185] As an example, in step S223, the computer device can determine the skeletal mesh model corresponding to the digital character based on the first bone skinning weights of the first bound bone nodes corresponding to all mesh vertices. This skeletal mesh model includes multiple mesh vertices and their corresponding first bound bone nodes' first bone skinning weights, so that in the subsequent animation control process, the new positions of different mesh vertices can be obtained through linear hybrid bone transformation, specifically determined by the bone parameter vector of the first bound bone node. When the bone parameter vector of any bone node changes, the position of the mesh vertex it affects will also change, so as to realize the linkage change between the mesh vertex and its corresponding first bound bone node.

[0186] In one embodiment, step S23, which involves binding the Gaussian point cloud model of the digital character to the corresponding skeletal mesh model of the digital character, and determining the Gaussian skeletal mesh model corresponding to the digital character, includes:

[0187] S231: Bind the Gaussian point cloud model of the digital character to the skeletal mesh model corresponding to the digital character, and determine the associated mesh vertex corresponding to each Gaussian point in the Gaussian point cloud model.

[0188] S232: Determine the adjacent bone nodes corresponding to the associated mesh vertices as the second bound bone nodes corresponding to the Gaussian points, and determine the initial bone skinning weights corresponding to the second bound bone nodes.

[0189] S233: Based on the anatomical structure of the digital character, smooth the initial bone skinning weights corresponding to the second bound bone node, and determine the second bone skinning weights corresponding to the second bound bone node.

[0190] S234: Determine the Gaussian bone mesh model corresponding to the digital character based on the second bound bone nodes corresponding to all Gaussian points and the second bone skinning weights corresponding to each second bound bone node.

[0191] As an example, in step S231, the Gaussian point cloud model of the digital character is bound to the corresponding skeletal mesh model of the digital character. Specifically, the Gaussian point cloud model is overlaid on the mesh surface of the skeletal mesh model using coordinate alignment. The associated mesh vertex corresponding to each Gaussian point in the Gaussian point cloud model is determined. Here, the associated mesh vertex is the mesh vertex corresponding to the position of the Gaussian point, which can be understood as the location of the center of the Gaussian point. In this example, the binding relationship between the Gaussian point and the mesh vertex is established through position alignment. Since the mesh vertices and bone nodes in the skeletal mesh model are bound together, the Gaussian point can also be bound to the bone nodes.

[0192] As an example, in step S232, the computer device can determine one or more adjacent bone nodes corresponding to each associated mesh vertex in the skeletal mesh model according to pre-set binding bone node determination rules, and identify them as the second bound bone nodes corresponding to the Gaussian point. For example, if the pre-set number of second bound bone nodes is y, the y adjacent bone nodes that are spatially closest to the associated mesh vertex can be identified as the second bound bone nodes corresponding to the Gaussian point aligned with its position, to avoid interference from distant bone nodes and prevent misalignment caused by distant bone nodes. Next, the computer device determines the initial bone skinning weight corresponding to each second bound bone node based on the degree of influence of each second bound bone node on the Gaussian point. The initial bone skinning weight corresponding to the second bound bone node refers to the proportion by which the initially determined second bound bone node drives the Gaussian point to move. Specifically, the computer device processes parameters such as the distance between the Gaussian point and the second bound bone node and / or the influence radius of the second bound bone node based on a pre-set weight determination program to determine the initial bone skinning weight corresponding to the second bound bone node. Generally, the smaller the distance between the Gaussian point and the second bound bone node, the higher its initial bone skinning weight; conversely, the larger the distance, the lower its initial bone skinning weight. In this example, each bone node is pre-configured with its corresponding influence radius, and the Gaussian point will only be assigned an initial bone skinning weight if it falls within the influence radius of the second bound bone node.

[0193] As an example, in step S233, after determining the second bound bone nodes and the initial bone skinning weights corresponding to all Gaussian points, the computer device needs to smooth the initial bone skinning weights corresponding to the second bound bone nodes based on the anatomical structure of the digital character. Specifically, it needs to combine the anatomical structure of the digital character, such as joint movement patterns and muscle change patterns, to smooth the initial bone skinning weights corresponding to the second bound bone nodes. For example, it needs to smooth the joint transition areas of bending joints such as knees and elbows, and determine their corresponding smooth bone skinning weights to make the movement of each bone node more natural. Then, it needs to perform weight normalization processing on the smooth bone skinning weights of all second bound bone nodes corresponding to the same Gaussian point, and determine the second bone skinning weights of all second bound bone nodes corresponding to that Gaussian point, so that the sum of all second bone skinning weights corresponding to that Gaussian point is 1, to avoid excessive or misaligned movement of the Gaussian point.

[0194] As an example, in step S234, the computer device can determine the Gaussian bone mesh model corresponding to the digital character based on the second bone skinning weights of the second bound bone nodes corresponding to all Gaussian points. This Gaussian bone mesh model includes multiple Gaussian points and their corresponding second bound bone nodes' second bone skinning weights, so that in the subsequent animation control process, the new positions of different Gaussian points can be obtained through linear hybrid bone transformation, specifically determined by the bone parameter vector of the second bound bone node. When the bone parameter vector of any bone node changes, the position of the Gaussian point it affects will also change, so as to realize the linkage change between the Gaussian point and its corresponding second bound bone node.

[0195] Figure 3 This is a diagram illustrating the overall architecture for controlling the animation of a digital character using a Gaussian skeletal mesh model. Figure 3 This demonstrates the complete workflow from parameter input to final rendering. Specifically: the input module receives standardized bone parameter vectors (i.e., the target bone parameter vectors corresponding to the target bone nodes), including quaternion rotations of body joints, Cartesian coordinate displacements, and facial expression coefficients in the 0-1 range; the core control module pre-constructs a hierarchical bone structure using a parametric bone model, and the hybrid model binding unit pre-implements the bone skinning weight binding between the mesh model and the Gaussian point cloud model; the animation-driven module reuses the same set of bone transformation matrices (i.e., bone transformation matrices), synchronously calculating the new positions of mesh vertices and the positions and covariance matrices of Gaussian points to ensure deformation consistency; the rendering adaptation module implements synchronized lighting properties and dynamic detail adjustments, and optional extension modules support physical simulation and multi-character adaptation. The overall architecture balances real-time performance and scalability. This architecture can achieve a single-frame computation time of ≤3 ms, meeting the real-time rendering requirements of 60 FPS (frames per second).

[0196] Figure 4 This is a comparison between the Gaussian point cloud model in this invention and a traditional discrete point cloud. The upper part shows the structure and effect of the traditional discrete point cloud, which consists of discrete points without dimensions and suffers from defects such as uneven point spacing, surface granularity, and obvious detail breaks. The lower part shows the Gaussian point cloud model used in this invention. Each Gaussian point has spatial expansion properties and is superimposed as semi-transparent spots to form a continuous and smooth surface with natural detail transitions. As can be seen from the comparison, the Gaussian point cloud model of this invention improves detail fidelity by more than 60% compared to the traditional point cloud without increasing the computational burden, and the surface continuity is significantly optimized.

[0197] Figure 5This is a schematic diagram illustrating the synchronous deformation of a mesh model and a Gaussian point cloud model under skeletal-driven conditions. Initially, the skeletal parameters are neutral values, the mesh vertices are in a neutral pose, and the Gaussian point cloud model conforms to the surface distribution of the mesh vertices. After inputting the target skeletal parameter vector (arm raised 30° + smile coefficient 0.8), the skeletal transformation matrix synchronously drives the deformation of the mesh vertices in the mesh model and the Gaussian points in the Gaussian point cloud model: the mesh vertices achieve the changes in the raised arm and smiling facial contours, while the corresponding muscle areas in the Gaussian point cloud model converge and facial wrinkles stretch. In the final state, the deformation of the mesh and the Gaussian point cloud model is seamless and without misalignment, achieving coordinated consistency between facial expressions and body movements, with a synchronization error ≤1ms. This example visually demonstrates the synchronous driving of body movements and facial expressions, as well as the effect of synchronous deformation of the mesh model and the Gaussian point cloud model. When the digital character raises its arm and smiles, the mesh model undergoes a wide range of deformations (such as raising the arm and raising the corners of the mouth), while details such as muscle bulges and facial wrinkles depicted by the Gaussian point cloud model are also adjusted accordingly. Moreover, the two are strictly synchronized without any timing delay. This proves that under unified skeletal drive, the large movements and detailed changes of the digital character can be coordinated and consistent, avoiding the desynchronization problems that may occur when driven separately in the past.

[0198] Figure 6 This is a schematic diagram of an adaptation architecture for digital characters. Based on a basic parametric skeletal model, it achieves multi-scene compatibility through character type adaptation units: when adapting to animal characters, it expands the limb joint chains and tail / ear control nodes, and densifies Gaussian point sampling in the fur area; when adapting to cartoon characters, it adds exaggerated expression nodes, increases the Gaussian point expansion radius, and expands the weight smoothing range; all adaptation types follow unified weight binding rules and driving logic to ensure consistency between hybrid model construction and animation driving. This extended architecture demonstrates that the invention can be widely applied to various digital characters such as humans, animals, and cartoons, reflecting good versatility.

[0199] In summary, this embodiment implements the Gaussian-mesh hybrid animation control method of the present invention through specific system modules, verifying its effectiveness in practical applications. The described method ensures both the consistency of animation driving and the full preservation and real-time display of the visual details of digital characters, and can be widely applied to scenarios such as real-time interactive animation of digital characters, virtual anchors, and film and television special effects. The above embodiments are only used to illustrate the technical solution of the present invention. All equivalent substitutions or modifications made within the spirit and principle of the present invention should be included within the protection scope of the present invention.

[0200] The digital character animation control method provided in this embodiment also has broad application prospects, including but not limited to the following scenarios:

[0201] (I) Virtual Live Streaming / Virtual Humanoid Anchor: This solution can be applied to real-time live streaming interactive scenarios with virtual anchors, ensuring highly realistic and synchronized expressions and movements of the virtual avatar, meeting the virtual anchor's requirements for realism and real-time performance. For example, during the live stream, motion capture equipment is used to acquire the anchor's body movements and facial expression data, driving the humanoid character's skeletal model in real time; the Gaussian point cloud rendering of this invention synchronously presents subtle details such as facial wrinkles and muscle textures without significantly burdening system performance. Viewers can see that the virtual anchor's subtle facial expressions and body movements are strictly synchronized, enhancing the immersion and realism during the interaction.

[0202] (II) Interactive Games / Virtual Reality (VR): This solution is highly suitable for real-time interactive games and VR scenarios. In games, digital characters require high-precision detail but are limited by the performance constraints of real-time rendering. This solution uses a hybrid rendering of mesh models and Gaussian point cloud models to provide rich detail in close-up scenes and automatically reduce detail density in distant scenes, perfectly meeting the game's requirements for a balance between image detail and frame rate. Simultaneously, unified skeletal driving avoids common animation desynchronization bugs, making the character's movements and expressions in the game more fluid and natural. In a VR environment, when a user observes a virtual character up close, the Gaussian point cloud ensures that details such as pores and wrinkles on the character's face are clearly visible and synchronized with changes in facial expressions in real time; when the user moves away, the system automatically reduces the number of Gaussian points rendered to maintain a high frame rate, thus achieving an ideal balance between immersion and performance. Therefore, whether in games or VR scenarios, this solution enables virtual characters to run smoothly while maintaining realistic details, greatly improving the user experience.

[0203] (III) Film and Television Special Effects / Digital Stand-ins: In film special effects production, this solution is used to create highly realistic digital character stand-ins or fantasy creatures. Films have extremely high requirements for the details of character appearance and movement, and the digital characters need to perfectly match the actors' actual performances. This solution enables the details of the digital stand-in character, such as muscle texture and skin wrinkles, to dynamically change in fine detail with the actor's movements, making the special effects visuals so realistic that it is difficult to distinguish them from reality. For example, motion capture is used to obtain the actor's facial expressions and body movements data to drive the skeletal animation of the digital character; then, a Gaussian point cloud model is used to render the subtle wrinkles, scars, and other texture changes on the digital character's skin, synchronizing with every detailed movement of the actor. The resulting humanoid character is indistinguishable from a real actor, whether in close-ups or fast-moving scenes, greatly improving the realism of film and television special effects.

[0204] This solution can also be extended to many other fields, such as large-scale virtual human interaction in the metaverse, real-time virtual avatar display in online meetings, and high-precision humanoid character models in science and education simulations. Because this solution combines high versatility with high-fidelity animation effects, it can function effectively in all the aforementioned application scenarios, demonstrating broad application prospects.

[0205] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0206] This invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described digital character animation control method, or, when the processor executes the computer program, it implements the above-described digital character modeling method. To avoid repetition, these will not be described in detail here.

[0207] This invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the above-described digital character animation control method, or, when the computer program is executed by a processor, it implements the above-described digital character modeling method. To avoid repetition, further details are omitted here.

[0208] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for controlling digital character animation, characterized in that, include: Obtain the animation switching instruction corresponding to the digital character, wherein the animation switching instruction includes the target bone parameter vector corresponding to the target bone node; The target skeletal node refers to the skeletal node that needs to be adjusted this time. The target skeletal node includes skeletal joint nodes and specific part control nodes; the target skeletal parameter vector refers to the skeletal parameter vector that needs to be adjusted this time. Based on the parameterized skeleton model corresponding to the digital character, the target skeleton parameter vector and the current skeleton parameter vector corresponding to the target skeleton node are processed to determine the skeleton transformation matrix. The skeleton transformation matrix is ​​the transformation matrix that switches from the current skeleton parameter vector corresponding to the target skeleton node to its corresponding target skeleton parameter vector. Based on the bone transformation matrix and the first bone skinning weight corresponding to the target mesh vertex, the current vertex attribute corresponding to the target mesh vertex is transformed to determine the target vertex attribute corresponding to the target mesh vertex. The target mesh vertex is the mesh vertex in the mesh model corresponding to the digital character that has a binding relationship with the target bone node. The first bone skinning weight is the proportion by which the bone node drives the mesh vertex to move. Based on the bone transformation matrix and the second bone skinning weight corresponding to the target Gaussian point, the current Gaussian attribute corresponding to the target Gaussian point is transformed to determine the target Gaussian attribute corresponding to the target Gaussian point. The target Gaussian point is the Gaussian point in the Gaussian point cloud model corresponding to the digital character that has a binding relationship with the target bone node. The second bone skinning weight is the proportion by which the bone node drives the Gaussian point to move. The second bone skinning weight corresponding to each Gaussian point in the Gaussian point cloud model is determined in the following way: the Gaussian point cloud model of the digital character is overlaid on the mesh surface of the bone mesh model through coordinate alignment, so that the Gaussian point is bound to the mesh vertex, and the second bone skinning weight corresponding to each Gaussian point in the Gaussian point cloud model is determined. The associated mesh vertices are identified; adjacent bone nodes corresponding to the associated mesh vertices are determined as the second bound bone nodes corresponding to the Gaussian point; and the initial bone skinning weights corresponding to the second bound bone nodes are determined based on the distance between the Gaussian point and the second bound bone node and / or the influence radius of the second bound bone node; based on the anatomical structure of the digital character, the initial bone skinning weights corresponding to the second bound bone nodes are smoothed, and the smoothed bone skinning weights of all second bound bone nodes corresponding to the same Gaussian point are normalized to determine the second bone skinning weights corresponding to the second bound bone nodes, so that the sum of all second bone skinning weights corresponding to the same Gaussian point is 1; The target animation image is determined based on the target vertex attributes corresponding to the target mesh vertices and the target Gaussian attributes corresponding to the target Gaussian points.

2. The digital character animation control method according to claim 1, characterized in that, The step of transforming the current Gaussian attribute corresponding to the target Gaussian point based on the bone transformation matrix and the second bone skinning weight corresponding to the target Gaussian point, and determining the target Gaussian attribute corresponding to the target Gaussian point, includes: Based on the bone transformation matrix and the second bone skinning weight corresponding to the target Gaussian point, the current Gaussian attribute corresponding to the target Gaussian point is transformed to determine the initial Gaussian attribute corresponding to the target Gaussian point. The initial Gaussian attribute is the state of the Gaussian point position, covariance matrix and rendering attribute directly calculated by the bone transformation. Based on the elastic network corresponding to the target Gaussian point, the initial Gaussian property corresponding to the target Gaussian point is oscillated and simulated to determine the target Gaussian property corresponding to the target Gaussian point.

3. The digital character animation control method according to claim 1, characterized in that, The step of transforming the current Gaussian attribute corresponding to the target Gaussian point based on the bone transformation matrix and the second bone skinning weight corresponding to the target Gaussian point, and determining the target Gaussian attribute corresponding to the target Gaussian point, includes: Based on the bone transformation matrix and the second bone skinning weight corresponding to the target Gaussian point, the current Gaussian attribute corresponding to the target Gaussian point is transformed to determine the initial Gaussian attribute corresponding to the target Gaussian point. The initial Gaussian attribute is the state of the Gaussian point position, covariance matrix and rendering attribute directly calculated by the bone transformation. Based on pre-set physical constraints and a physics engine, collision detection is performed on the initial Gaussian properties corresponding to the target Gaussian point to adjust and determine the target Gaussian properties corresponding to the target Gaussian point.

4. The digital character animation control method according to claim 1, characterized in that, The step of determining the target animation image based on the target vertex attributes corresponding to the target mesh vertices and the target Gaussian attributes corresponding to the target Gaussian points includes: Rasterization rendering is performed on the target vertex attributes corresponding to the target mesh vertices to determine the mesh rendering image; Rasterization rendering is performed on the target Gaussian attributes corresponding to the target Gaussian point to determine the Gaussian rendered image; The mesh rendering image and the Gaussian rendering image are fused together to determine the target animation image.

5. The digital character animation control method according to claim 4, characterized in that, The target Gaussian properties include the number of renders and the spatial expansion radius; The step of rasterizing and rendering the target Gaussian attributes corresponding to the target Gaussian point to determine the Gaussian rendered image includes: Obtain the rendering distance corresponding to the target Gaussian point; Based on the rendering distance corresponding to the target Gaussian point, determine the number of rendering operations and the spatial expansion radius corresponding to the target Gaussian point; Rasterization rendering is performed based on the number of renders corresponding to the target Gaussian point and the spatial expansion radius to determine the Gaussian rendered image.

6. The digital character animation control method according to claim 4, characterized in that, The target Gaussian attributes include the target position and the target orientation; The step of rasterizing and rendering the target Gaussian attributes corresponding to the target Gaussian point to determine the Gaussian rendered image includes: The position deviation is determined based on the target position corresponding to the target Gaussian point and the current position of the target Gaussian point; The position deviation is processed based on the inertial physics equations to determine the inertial position corresponding to the target Gaussian point; Based on the inertial position corresponding to the target Gaussian point, the target Gaussian property corresponding to the target Gaussian point is corrected to determine the corrected Gaussian property corresponding to the target Gaussian point; Rasterization rendering is performed on the modified Gaussian properties corresponding to the target Gaussian point to determine the Gaussian rendered image.

7. The digital character animation control method according to claim 4, characterized in that, The target Gaussian attribute includes the target color; The step of rasterizing and rendering the target Gaussian attributes corresponding to the target Gaussian point to determine the Gaussian rendered image includes: Based on the target bone parameter vector corresponding to the target bone node, determine the driving state corresponding to the target bone node; Based on the driving state corresponding to the target skeletal node, determine the color correction coefficient corresponding to the driving state; The target color corresponding to the target Gaussian point is corrected based on the color correction coefficient, and the corrected color corresponding to the target Gaussian point is determined. Rasterization rendering is performed based on the corrected color corresponding to the target Gaussian point to determine the Gaussian rendered image.

8. The digital character animation control method according to any one of claims 1-7, characterized in that, The digital character is a humanoid character, and the specific part control node includes a facial expression control node; Alternatively, the digital character may be an animal character, and the specific part control nodes may include tail control nodes and / or ear control nodes; Alternatively, the digital character may be a cartoon character, and the specific part control node may be an exaggerated expression control node.

9. A digital character modeling method, characterized in that, include: Obtain the bone parameter vector corresponding to the skeletal structure, and construct a parameterized skeletal model of the digital character based on the bone parameter vector corresponding to the skeletal structure. The skeletal structure includes multiple bone nodes, and the bone nodes include bone joint nodes and specific part control nodes. The parameterized skeletal model of the digital character is bound to the mesh model of the digital character to determine the skeletal mesh model corresponding to the digital character; wherein, the skeletal mesh model includes multiple first bound bone nodes corresponding to all mesh vertices and a first bone skinning weight corresponding to each first bound bone node; The Gaussian point cloud model of the digital character is overlaid on the mesh surface of the skeletal mesh model through coordinate alignment, so that the Gaussian points and mesh vertices are bound together, and the associated mesh vertex corresponding to each Gaussian point in the Gaussian point cloud model is determined. The adjacent bone nodes corresponding to the associated mesh vertex are determined as the second bound bone nodes corresponding to the Gaussian point, and the initial bone skinning weights corresponding to the second bound bone nodes are determined based on the distance between the Gaussian point and the second bound bone node and / or the influence radius of the second bound bone node. Based on the anatomical structure of the digital character, the initial bone skinning weights corresponding to the second bound bone nodes are smoothed, and the smoothed bone skinning weights of all second bound bone nodes corresponding to the same Gaussian point are normalized to determine the second bone skinning weights corresponding to the second bound bone nodes, so that the sum of all second bone skinning weights corresponding to the same Gaussian point is 1. The second bone skinning weight is the proportion by which the bone node drives the Gaussian point to move. Based on the second bound bone nodes corresponding to all Gaussian points and the second bone skinning weights corresponding to each second bound bone node, a Gaussian bone mesh model corresponding to the digital character is determined. The Gaussian bone mesh model is used to realize the animation control of the digital character. The Gaussian bone mesh model includes multiple second bound bone nodes corresponding to all Gaussian points and the second bone skinning weights corresponding to each second bound bone node.

10. The digital character modeling method according to claim 9, characterized in that, The step of binding the parameterized skeletal model of the digital character to the mesh model of the digital character, and determining the skeletal mesh model corresponding to the digital character, includes: Determine the first bound bone node corresponding to each mesh vertex of the mesh model of the digital character, and determine the initial bone skinning weight corresponding to the first bound bone node; Based on the anatomical structure of the digital character, the initial bone skinning weights corresponding to the first bound bone node are smoothed, and the smoothed bone skinning weights of all first bound bone nodes corresponding to the same mesh vertex are normalized to determine the first bone skinning weights corresponding to the first bound bone node. The skeletal mesh model corresponding to the digital character is determined based on the first bound bone node corresponding to all mesh vertices and the first bone skinning weight corresponding to each of the first bound bone nodes.

11. A computer 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 computer program, it implements the digital character animation control method according to any one of claims 1 to 8; or, when the processor executes the computer program, it implements the digital character modeling method according to any one of claims 9 to 10.

12. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the digital character animation control method according to any one of claims 1 to 8; or, when the computer program is executed by the processor, it implements the digital character modeling method according to any one of claims 9 to 10.

Images