Design editor platform for dynamic real-time rendering

Abstract

The invention relates to the field of design editor platforms, in particular to a dynamic real-time rendering design editor platform which comprises a dynamic grid body module, an outline and scene management module, a real-time cooperation module and a universal class library. The dynamic grid body module realizes efficient processing of a dynamic grid through input preprocessing, dynamic simplification, detail enhancement, topology repair and LOD self-adaptive management; the outline and scene management module realizes organization and scheduling of a complex scene by means of hierarchical management and cross-platform resource management; the real-time cooperation module utilizes an operation log algorithm to solve editing conflicts and ensure multi-terminal consistency; the general class library comprises a math library and a geometric library, and the scene requirements of graphic rendering, physical simulation and the like are met. The rendering efficiency and the collaboration experience are improved, and the method has a wide application prospect.

Description

Technical Field

[0001] The present invention relates to the field of design editors, and in particular to a design editor platform with dynamic real-time rendering. Background Art

[0002] In the field of graphic design and real-time rendering, traditional design editors face many technical bottlenecks and are unable to meet the needs of efficient editing and collaboration in complex scenarios.

[0003] Existing solutions lack intelligent optimization mechanisms when processing large-scale dynamic meshes. Directly rendering high-precision meshes can overload computing resources, while manually adjusting levels of detail or simplifying meshes requires significant manpower and struggles to balance visual accuracy and performance. Furthermore, traditional methods rely on manual repair for topological errors and detail loss in meshes, resulting in a low degree of automation.

[0004] Existing editors rely on repetitive development for multi-platform compatibility and lack unified underlying engine support. Regarding performance optimization, compute-intensive tasks fail to fully utilize the GPU's parallel computing capabilities, resulting in large frame rate fluctuations in complex scenes and difficulty achieving real-time, high-quality rendering. In collaborative editing scenarios, traditional synchronization mechanisms are unable to effectively resolve concurrent operation conflicts, easily leading to data inconsistencies or incorrect editing results. To address these shortcomings, a technical solution is now provided. Summary of the Invention

[0005] In order to solve the technical problems raised by the above background technology, the present invention provides a design editor platform with dynamic real-time rendering.

[0006] The purpose of the present invention can be achieved through the following technical solutions:

[0007] The present invention is a design editor platform for dynamic real-time rendering, comprising a dynamic grid module, an outline and scene management module, a real-time collaboration module and a general class library.

[0008] The dynamic mesh module receives dynamic mesh data in real time and extracts geometric features from the mesh data as model input. It uses a graph convolutional neural network to predict vertex importance scores and performs edge collapse based on the predicted scores, preferentially removing low-importance areas while retaining high-curvature features. The specific process is as follows:

[0009] The dynamic mesh module is provided with an input preprocessing unit, a dynamic simplification unit, a detail enhancement and topology repair unit, and LOD adaptive management;

[0010] In the dynamic mesh module, input preprocessing is the starting point of the process. It is responsible for format conversion and feature extraction of the received dynamic mesh data, providing a foundation for subsequent processing.

[0011] Receive dynamic mesh data in real time, including vertex coordinates Triangle information T = {v1, v2, v3} and normal vector Where n is the number of vertices and i is the vertex number. To fully utilize the parallel computing capabilities of the GPU, the data is converted into a tensor format using PyTorch Tensor hardware. The vertex coordinate data is converted into a shape of [n,3], where 3 corresponds to the x, y, and z coordinate dimensions, and n has the same meaning as above. The triangle information is converted into a shape of [m,3], where m is the number of triangles, and the three elements in each row represent the three vertex indices of a triangle.

[0012] For the data that has completed the format conversion, the geometric features of the mesh are extracted as the model input. The geometric features include mean curvature, edge density, and normal vector deviation. The mean curvature is calculated using the discrete differential geometry method, and its calculation logic is as follows: Get the average curvature H, where N(i) represents the set of adjacent vertices of vertex vi, α ij and β ij is the angle corresponding to vi and vj, Ai is the area of ​​the Voronoi region of vertex vi, and the curvature value of each vertex is obtained by calculation to form the curvature tensor HC;

[0013] For edge density, we count the number of adjacent edges of each vertex to obtain the edge density tensor EC, specifically: E = {vi, vj} is an undirected edge set, that is, each edge connects two vertices; for each vertex vi, we traverse the edge set E and count the number of edges containing vi, which is the edge density tensor;

[0014] Normal vector deviation measures the difference between the normal vector of a vertex and the average normal vector of the adjacent faces of the vertex, reflecting the smoothness of the mesh surface at the vertex. For vertex vi, calculate the normal vector set of its adjacent faces κ=1....,g, where g is the number of adjacent faces, κ is the number of adjacent faces, and the average normal vector of the adjacent faces is calculated The calculation logic is: The angle θ between the vertex normal vector and the average normal vector is calculated according to the vector dot product formula. The calculation logic is: Then use 1-cosθ to get the normal vector deviation. The larger the value, the greater the deviation and the less smooth the surface.

[0015] Combine the curvature tensor, edge density tensor, and normal vector deviation into the input feature tensor F;

[0016] Mesh simplification uses graph convolutional neural network GCN as the core model architecture, transforming the feature tensor F of each vertex vi i lAt layer l, the feature of adjacent vertices is aggregated and updated. The calculation logic is: Where σ is the activation function, W l is the weight matrix b of layer l l is the bias vector, cij is the normalization constant, and the rest of the meanings are the same as above; by stacking multiple layers of GCN, the topological and geometric features of the mesh are gradually extracted, and the importance score s of each vertex is finally output. i ;

[0017] Vertex importance score s predicted by GCN i , perform edge folding operation to achieve dynamic simplification, the edge folding operation is to collapse an edge (v a ,v b ) is collapsed into a new vertex v new , and update the triangle face associated with the edge; give priority to the edge connecting the low importance vertices, that is, according to the vertex importance score s i Calculate edge importance scores Select ab The lowest edge is folded; curvature constraint is introduced in the edge folding process, for the edge (v a ,v b ), the average curvature of its two end vertices If the edge is greater than the preset threshold, the edge folding operation is restricted to obtain the simplified mesh feature tensor. It should be noted that the introduction of curvature constraint is to avoid the loss of important geometric features. By continuously performing edge folding operations, the number of polygons is gradually reduced to achieve mesh simplification.

[0018] The detail enhancement and topology restoration unit uses the U-Net structure to achieve generative detail recovery. U-Net is a symmetrical encoder-decoder network structure. The encoder extracts the features of the simplified mesh through convolution and downsampling operations, and the decoder restores the feature maps into high-precision normal maps and displacement maps through deconvolution and upsampling operations. Specifically, the simplified mesh feature tensor is input, and a low-resolution feature representation is obtained through multi-layer convolution and downsampling of the encoder. In the decoder, the resolution is gradually restored through deconvolution operations, and the features of the corresponding levels in the encoder are integrated with jump connections. Finally, high-precision normal map tensors and displacement map tensors are output. Using surface subdivision technology, the mesh is subdivided according to the displacement map tensor to restore details.

[0019] Automatic topology optimization is achieved through reinforcement learning to resolve topological errors such as holes and non-manifold edges. Specifically, the topological state of the mesh is defined as the environment state, and topological repair operations are defined as actions, including adding vertices, merging edges, and splitting triangles. Actions are selected based on the current environment state, and after executing the action, the environment state is updated and a reward is obtained. It should be noted that the reward function is designed to encourage the repair of topological errors. For example, a successful hole repair is rewarded with +10, and a new topological error penalty of -5 is introduced. Through continuous training, the hardware learns the optimal strategy, that is, selecting the best repair action based on different topological states, to achieve automatic topological optimization.

[0020] The LOD adaptive management unit trains a lightweight CNN model to predict the screen space error to achieve dynamic switching of LOD levels. The input of the CNN model is the camera distance d and the image features of the viewing angle coverage area. Finally, the predicted screen space error value ε is output. Specifically, the actual error of the meshes at different LOD levels on the screen is used as a label. The actual error is obtained by calculating the pixel difference between the rendered image of the meshes at different LOD levels and the original high-resolution mesh rendered image. The mean square error calculation logic is: Among them, B p is the number of image pixels, is the color difference of the hth pixel in the original high-resolution image, is the color value of the hth pixel in the current LOD level image. By minimizing the loss between the predicted error and the actual error, a CNN model that can accurately predict the screen space error is trained; in the calculation process, according to the predicted screen space error value ε and the preset error threshold ε threshold For comparison, when ε>ε threshold When ε<ε, switch to high LOD level. threshold When the LOD level is low, it switches to the low LOD level to achieve dynamic adaptive switching of LOD levels;

[0021] The outline and scene management module implements scene graph hierarchical management and cross-platform resource management to achieve efficient organization and dynamic scheduling of complex scenes, solving data management and cross-platform collaboration issues in large-scale scene editing. The specific process is as follows:

[0022] The outline and scene management module is equipped with a hierarchical management unit and a cross-platform resource management unit;

[0023] The hierarchical management unit adopts a tree node structure The nodes It includes the transformation matrix, entity properties of rendering components and physical components. The edge φ defines the parent-child hierarchical relationship. Each node supports dynamic binding of child nodes to form an extensible hierarchical architecture. The node grouping mechanism is used to perform unified transformations on the point set. The changes include translation, rotation and scaling. The matrix cascade optimization algorithm is used. Its calculation logic is as follows: Where T global is the global change matrix of the node in the coordinate system, Represents the local transformation matrix of the node at the cth layer parent node, T local The local transformation matrix of the current node, t represents the depth of the node hierarchy, and c represents the hierarchy number. It supports dynamic creation or disconnection of parent-child relationships at runtime. The binding process triggers the scene graph reconstruction algorithm, automatically updates the global transformation matrix of the affected nodes, and synchronizes it to the rendering pipeline and physics engine.

[0024] The cross-platform resource management unit builds a unified resource access interface, which includes transparent access to APL, local files, cloud storage, and engine internal resources. The hardware implementation includes protocol adapters and cache managers.

[0025] The real-time collaboration module resolves editing conflicts based on an operation log algorithm. When multiple users edit a scene simultaneously, the algorithm merges and adjusts the operations of different users based on the order and logical relationships of the operations to ensure that each user ultimately sees the same scene state. Specifically:

[0026] The editing operation is abstracted into a five-tuple operation model, which is mathematically expressed as: O =<opTYPE,entityID,ψ,Q,D> ,where optype represents the set of operation types, including move, rotate and delete, entityID is the unique identifier of the target entity, ψ is the operation timestamp, Q is the operation parameter, and D is the globally unique version number;

[0027] When two operations O1 and O2 act on the same entity, the conflict is eliminated by the operation transformation function. It is divided into two scenarios for processing. Scenario 1: When the move operation conflicts, the operation logic is: A = <Move,ec,ψ A , ΔV A ,D A >, user A moves the entity ec by ΔV A , where ec represents an editing operation, ΔV A represents the version vector of user A, O B = <Move,ec,ψ B , ΔV B ,D B >, user B moves ec at the same time, if ψ A <ψ B, then the client executes the logic locally: first apply O A , and then O B Convert to relative to O A The increment ΔV BA =ΔV B -ΔV A ,Server execution logic: ΔV final =V initial +ΔV A +ΔV BA , V initial Indicates the state before execution, ΔV final Indicates the status after execution;

[0028] Scenario 2: Use version vector verification to avoid outdated operations during mutually exclusive operations. Define the version vector of entity e as ΔV e , when the client sends an operation O, the server checks: O ≥ ΔV e +1, if true, update O=ΔV e If the operation fails, it is considered as an outdated operation and ignored or an error is returned.

[0029] The Lamport timestamp is introduced to establish a partial order relation of operations, which is defined as: O1→O2 if and only if According to the order of creating entities first and then editing, through this partial order relationship, the server can reorder concurrent operations into a causally consistent sequence, and the final state satisfies: O∈ξ, the execution order respects the → relationship, where ξ is the combination of operation logs, ensuring that the final state of all clients is equivalent to the execution result of a certain causally legal sequence.

[0030] The general class library is equipped with a mathematical library and a geometric library. The mathematical library includes quaternion rotation, Euler angle conversion, affine transformation matrix construction, vector matrix operations and vectorized operations. Its target scenarios mainly include graphics rendering, physical simulation and animation systems; the geometric library includes mesh operations, surface subdivision, collision detection and geometric measurement. Its target scenarios mainly include terrain and building modeling generation, real-time rendering optimization and physical interaction.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] Hierarchical management and matrix optimization: The tree-based scene graph architecture supports dynamic parent-child binding. Combined with the matrix cascade optimization algorithm, it reduces the time complexity of batch transformation operations, reduces the time required to group and translate millions of nodes, and improves editing fluency. The adaptive scene graph pruning algorithm dynamically adjusts the traversal order based on visibility and editing status, improving the efficiency of traversing related nodes and reducing CPU utilization, enabling real-time preview of complex scenes.

[0033] Intelligent Mesh Optimization: An AI-driven topology optimization algorithm automatically identifies high-curvature features, reducing polygon counts when simplifying the mesh while maintaining visual accuracy error below 1%, addressing the inefficiencies of traditional manual optimization. Adaptive LOD management based on screen-space error dynamically switches mesh levels of detail, improving rendering performance in distant scenes while preserving close-up details. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. The following drawings are not intentionally scaled to the actual size, and the focus is on illustrating the main purpose of the present invention.

[0035] Figure 1 This is a principle block diagram of the present invention. DETAILED DESCRIPTION

[0036] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts also fall within the scope of protection of the present invention.

[0037] Please refer to Figure 1 As shown, the present invention is a design editor platform for dynamic real-time rendering, including a dynamic grid module, an outline and scene management module, a real-time collaboration module and a general class library.

[0038] The dynamic mesh module receives dynamic mesh data in real time and extracts geometric features from the mesh data as model input. It uses a graph convolutional neural network to predict vertex importance scores and performs edge collapse based on the predicted scores, preferentially removing low-importance areas while retaining high-curvature features. The specific process is as follows:

[0039] The dynamic mesh module is provided with an input preprocessing unit, a dynamic simplification unit, a detail enhancement and topology repair unit, and LOD adaptive management;

[0040] In the dynamic mesh module, input preprocessing is the starting point of the process. It is responsible for format conversion and feature extraction of the received dynamic mesh data, providing a foundation for subsequent processing.

[0041] Receive dynamic mesh data in real time, including vertex coordinates Triangle information T = {v1, v2, v3} and normal vector Where n is the number of vertices and i is the vertex number. To fully utilize the parallel computing capabilities of the GPU, the data is converted into a tensor format using PyTorch Tensor hardware. The vertex coordinate data is converted into a shape of [n,3], where 3 corresponds to the x, y, and z coordinate dimensions, and n has the same meaning as above. The triangle information is converted into a shape of [m,3], where m is the number of triangles, and the three elements in each row represent the three vertex indices of a triangle.

[0042] For the data that has completed the format conversion, the geometric features of the mesh are extracted as the model input. The geometric features include mean curvature, edge density, and normal vector deviation. The mean curvature is calculated using the discrete differential geometry method, and its calculation logic is as follows: Get the average curvature H, where N(i) represents the set of adjacent vertices of vertex vi, α ij and β ij is the angle corresponding to vi and vj, Ai is the area of ​​the Voronoi region of vertex vi, and the curvature value of each vertex is obtained by calculation to form the curvature tensor HC;

[0043] For edge density, we count the number of adjacent edges of each vertex to obtain the edge density tensor EC, specifically: E = {vi, vj} is an undirected edge set, that is, each edge connects two vertices; for each vertex vi, we traverse the edge set E and count the number of edges containing vi, which is the edge density tensor;

[0044] Normal vector deviation measures the difference between the normal vector of a vertex and the average normal vector of the adjacent faces of the vertex, reflecting the smoothness of the mesh surface at the vertex. For vertex vi, calculate the normal vector set of its adjacent faces κ=1....,g, where g is the number of adjacent faces, κ is the number of adjacent faces, and the average normal vector of the adjacent faces is calculated The calculation logic is: The angle θ between the vertex normal vector and the average normal vector is calculated according to the vector dot product formula. The calculation logic is: Then use 1-cosθ to get the normal vector deviation. The larger the value, the greater the deviation and the less smooth the surface.

[0045] Combine the curvature tensor, edge density tensor, and normal vector deviation into the input feature tensor F;

[0046] Mesh simplification uses graph convolutional neural network GCN as the core model architecture, transforming the feature tensor F of each vertex vi i l At layer l, the feature of adjacent vertices is aggregated and updated. The calculation logic is: Where σ is the activation function, W l is the weight matrix b of layer ll is the bias vector, cij is the normalization constant, and the rest of the meanings are the same as above; by stacking multiple layers of GCN, the topological and geometric features of the mesh are gradually extracted, and the importance score s of each vertex is finally output. i ;

[0047] Vertex importance score s predicted by GCN i , perform edge folding operation to achieve dynamic simplification, the edge folding operation is to collapse an edge (v a ,v b ) is collapsed into a new vertex v new , and update the triangle face associated with the edge; give priority to the edge connecting the low importance vertices, that is, according to the vertex importance score s i Calculate edge importance scores Select ab The lowest edge is folded; curvature constraint is introduced in the edge folding process, for the edge (v a ,v b ), the average curvature of its two end vertices If the edge is greater than the preset threshold, the edge folding operation is restricted to obtain the simplified mesh feature tensor. It should be noted that the introduction of curvature constraint is to avoid the loss of important geometric features. By continuously performing edge folding operations, the number of polygons is gradually reduced to achieve mesh simplification.

[0048] The detail enhancement and topology restoration unit uses the U-Net structure to achieve generative detail recovery. U-Net is a symmetrical encoder-decoder network structure. The encoder extracts the features of the simplified mesh through convolution and downsampling operations, and the decoder restores the feature maps into high-precision normal maps and displacement maps through deconvolution and upsampling operations. Specifically, the simplified mesh feature tensor is input, and a low-resolution feature representation is obtained through multi-layer convolution and downsampling of the encoder. In the decoder, the resolution is gradually restored through deconvolution operations, and the features of the corresponding levels in the encoder are integrated with jump connections. Finally, high-precision normal map tensors and displacement map tensors are output. Using surface subdivision technology, the mesh is subdivided according to the displacement map tensor to restore details.

[0049] Automatic topology optimization is achieved through reinforcement learning to resolve topological errors such as holes and non-manifold edges. Specifically, the topological state of the mesh is defined as the environment state, and topological repair operations are defined as actions, including adding vertices, merging edges, and splitting triangles. Actions are selected based on the current environment state, and after executing the action, the environment state is updated and a reward is obtained. It should be noted that the reward function is designed to encourage the repair of topological errors. For example, a successful hole repair is rewarded with +10, and a new topological error penalty of -5 is introduced. Through continuous training, the hardware learns the optimal strategy, that is, selecting the best repair action based on different topological states, to achieve automatic topological optimization.

[0050] The LOD adaptive management unit trains a lightweight CNN model to predict the screen space error to achieve dynamic switching of LOD levels. The input of the CNN model is the camera distance d and the image features of the viewing angle coverage area. Finally, the predicted screen space error value ε is output. Specifically, the actual error of the meshes at different LOD levels on the screen is used as a label. The actual error is obtained by calculating the pixel difference between the rendered image of the meshes at different LOD levels and the original high-resolution mesh rendered image. The mean square error calculation logic is: Among them, B p is the number of image pixels, is the color difference of the hth pixel in the original high-resolution image, is the color value of the hth pixel in the current LOD level image. By minimizing the loss between the predicted error and the actual error, a CNN model that can accurately predict the screen space error is trained; in the calculation process, according to the predicted screen space error value ε and the preset error threshold ε threshold For comparison, when ε>ε threshold When ε<ε, switch to high LOD level. threshold When the LOD level is low, it switches to the low LOD level to achieve dynamic adaptive switching of LOD levels;

[0051] The outline and scene management module implements scene graph hierarchical management and cross-platform resource management to achieve efficient organization and dynamic scheduling of complex scenes, solving data management and cross-platform collaboration issues in large-scale scene editing. The specific process is as follows:

[0052] The outline and scene management module is equipped with a hierarchical management unit and a cross-platform resource management unit;

[0053] The hierarchical management unit adopts a tree node structure The nodes It includes the transformation matrix, entity properties of rendering components and physical components. The edge φ defines the parent-child hierarchical relationship. Each node supports dynamic binding of child nodes to form an extensible hierarchical architecture. The node grouping mechanism is used to perform unified transformations on the point set. The changes include translation, rotation and scaling. The matrix cascade optimization algorithm is used. Its calculation logic is as follows: Where T global is the global change matrix of the node in the coordinate system, Represents the local transformation matrix of the node at the cth layer parent node, T local The local transformation matrix of the current node, t represents the depth of the node hierarchy, and c represents the hierarchy number. It supports dynamic creation or disconnection of parent-child relationships at runtime. The binding process triggers the scene graph reconstruction algorithm, automatically updates the global transformation matrix of the affected nodes, and synchronizes it to the rendering pipeline and physics engine.

[0054] The cross-platform resource management unit builds a unified resource access interface, which includes transparent access to APL, local files, cloud storage, and engine internal resources. The hardware implementation includes protocol adapters and cache managers.

[0055] The real-time collaboration module resolves editing conflicts based on an operation log algorithm. When multiple users edit a scene simultaneously, the algorithm merges and adjusts the operations of different users based on the order and logical relationships of the operations to ensure that each user ultimately sees the same scene state. Specifically:

[0056] The editing operation is abstracted into a five-tuple operation model, which is mathematically expressed as: O =<opTYPE,entityID,ψ,Q,D> ,where optype represents the set of operation types, including move, rotate and delete, entityID is the unique identifier of the target entity, ψ is the operation timestamp, Q is the operation parameter, and D is the globally unique version number;

[0057] When two operations O1 and O2 act on the same entity, the conflict is eliminated by the operation transformation function. It is divided into two scenarios for processing. Scenario 1: When the move operation conflicts, the operation logic is: A = <Move,ec,ψ A , ΔV A ,D A >, user A moves the entity ec by ΔV A , where ec represents an editing operation, ΔV A represents the version vector of user A, O B = <Move,ec,ψ B , ΔV B ,D B >, user B moves ec at the same time, if ψ A <ψ B, then the client executes the logic locally: first apply O A , and then O B Convert to relative to O A The increment ΔV BA =ΔV B -ΔV A ,Server execution logic: ΔV final =V initial +ΔV A +ΔV BA , V initial Indicates the state before execution, ΔV final Indicates the status after execution;

[0058] Scenario 2: Use version vector verification to avoid outdated operations during mutually exclusive operations. Define the version vector of entity e as ΔV e , when the client sends an operation O, the server checks: O ≥ ΔV e +1, if true, update O=ΔV e If the operation fails, it is considered as an outdated operation and ignored or an error is returned.

[0059] The Lamport timestamp is introduced to establish a partial order relation of operations, which is defined as: O1→O2 if and only if According to the order of creating entities first and then editing, through this partial order relationship, the server can reorder concurrent operations into a causally consistent sequence, and the final state satisfies: O∈ξ, the execution order respects the → relationship, where ξ is the combination of operation logs, ensuring that the final state of all clients is equivalent to the execution result of a certain causally legal sequence.

[0060] The general class library is equipped with a mathematical library and a geometric library. The mathematical library includes quaternion rotation, Euler angle conversion, affine transformation matrix construction, vector matrix operations and vectorized operations. Its target scenarios mainly include graphics rendering, physical simulation and animation systems; the geometric library includes mesh operations, surface subdivision, collision detection and geometric measurement. Its target scenarios mainly include terrain and building modeling generation, real-time rendering optimization and physical interaction.

[0061] The above is an illustration of the present invention and should not be considered as limiting thereof. Although several exemplary embodiments of the present invention have been described, it will be readily understood by those skilled in the art that many modifications may be made to the exemplary embodiments without departing from the novel teachings and advantages of the present invention. Therefore, all such modifications are intended to be included within the scope of the present invention as defined by the claims. It should be understood that the above is an illustration of the present invention and should not be considered as being limited to the specific embodiments disclosed, and modifications to the disclosed embodiments and other embodiments are intended to be included within the scope of the appended claims. The present invention is defined by the claims and their equivalents.

Claims

1. A design editor platform for dynamic real-time rendering, including a dynamic mesh module, an outline and scene management module, a real-time collaboration module, and a general class library, characterized by: The dynamic mesh module is equipped with an input preprocessing unit, a dynamic simplification unit, a detail enhancement and topology repair unit, and an LOD adaptive management unit. It receives dynamic mesh data in real time, extracts geometric features within the mesh data as model input, uses a graph convolutional neural network to predict vertex importance scores, and performs edge collapse based on the predicted scores, preferentially removing low-importance areas while retaining high-curvature features. The outline and scene management module is equipped with a hierarchical management unit and a cross-platform resource management unit; Scheduling of complex scenes through scene graph hierarchical management and cross-platform resource management; The hierarchical management unit adopts a tree node structure The nodes It includes the entity attributes of the transformation matrix, rendering components, and physical components. The edge φ defines the parent-child hierarchical relationship. Each node supports dynamic binding of child nodes to form an extensible hierarchical architecture. A unified transformation is performed on the point set through a node grouping mechanism. The changes include translation, rotation, and scaling. The matrix cascade optimization algorithm is used. The calculation logic is as follows: Where T global is the global change matrix of the node in the coordinate system, Represents the local transformation matrix of the node at the cth layer parent node, T local is the local transformation matrix of the current node, t represents the depth of the node hierarchy, and c represents the hierarchy number. It supports dynamic creation or disconnection of parent-child relationships at runtime. The binding process triggers the scene graph reconstruction algorithm, automatically updates the global transformation matrix of the affected nodes, and synchronizes it with the rendering pipeline and physics engine.

2. A dynamic real-time rendering design editor platform according to claim 1, characterized in that: The input preprocessing unit performs format conversion and feature extraction on the received dynamic mesh data, specifically: Receive dynamic mesh data in real time, including vertex coordinates Triangle information T = {v1, v2, v3} and normal vector Where n is the number of vertices and i is the vertex number; Using PyTorchTensor hardware, the vertex coordinate data is converted to the shape of [n, 3], where 3 corresponds to the x, y, and z coordinate dimensions respectively; the triangle information is converted to the shape of [m, 3], where m represents the number of triangles; For the data that has completed the format conversion, the geometric features of the mesh are extracted as the model input. The geometric features include mean curvature, edge density, and normal vector deviation. The mean curvature is calculated using the discrete differential geometry method, and its calculation logic is as follows: Get the average curvature H, where N(i) represents the set of adjacent vertices of vertex vi, α ij and β ij is the angle corresponding to vi and vj, Ai is the area of ​​the Voronoi region of vertex vi, and the curvature value of each vertex is obtained by calculation to form the curvature tensor HC; For edge density, we count the number of adjacent edges of each vertex to obtain the edge density tensor EC, specifically: E = {vi, vj} is an undirected edge set, that is, each edge connects two vertices; for each vertex vi, we traverse the edge set E and count the number of edges containing vi, which is the edge density tensor; Normal vector deviation measures the difference between the normal vector of a vertex and the average normal vector of the adjacent faces of the vertex, reflecting the smoothness of the mesh surface at the vertex. For vertex vi, calculate the normal vector set of its adjacent faces Where g is the number of adjacent faces, κ is the number of adjacent faces, and the average normal vector of the adjacent faces is calculated The calculation logic is: The angle θ between the vertex normal vector and the average normal vector is calculated according to the vector dot product formula. The calculation logic is: Then use 1-cosθ to get the normal vector deviation. The larger the value, the greater the deviation and the less smooth the surface. The curvature tensor, edge density tensor, and normal vector bias are combined into the input feature tensor F.

3. A dynamic real-time rendering design editor platform according to claim 2, characterized in that: The dynamic simplification unit uses a graph convolutional neural network to predict vertex importance scores, performs edge collapse based on the predicted scores, and preferentially removes low-importance regions while retaining high curvature features. Specifically: Mesh simplification uses graph convolutional neural network GCN as the core model architecture, transforming the feature tensor F of each vertex vi i l At layer l, the feature of adjacent vertices is aggregated and updated. The calculation logic is: Where σ is the activation function, W l is the weight matrix b of layer l l is the bias vector, cij is the normalization constant; by stacking multiple layers of GCN, the topological and geometric features of the mesh are gradually extracted, and the importance score s of each vertex is output. i ; Vertex importance score s predicted by GCN i , perform edge folding operation to achieve dynamic simplification, the edge folding operation is to collapse an edge (v a ,v b ) is collapsed into a new vertex v new , and update the triangle face associated with the edge; Prioritize edges connecting low-importance vertices and calculate the importance scores of edges Select ab The lowest edge is folded; curvature constraint is introduced in the edge folding process, for the edge (v a ,v b ), the average curvature of its two end vertices If it is greater than a preset threshold, the folding operation of the edge is restricted to obtain a simplified mesh feature tensor.

4. A dynamic real-time rendering design editor platform according to claim 3, characterized in that: The detail enhancement and topology restoration unit uses a U-Net structure to achieve generative detail restoration. The encoder extracts features of the simplified mesh through convolution and downsampling operations, and the decoder restores the feature maps into high-precision normal maps and displacement maps through deconvolution and upsampling operations. Specifically, the simplified mesh feature tensor is input, and a low-resolution feature representation is obtained through multi-layer convolution and downsampling of the encoder. In the decoder, the resolution is gradually restored through deconvolution operations, and the features of the corresponding levels in the encoder are integrated by jump connections. Finally, a high-precision normal map tensor and displacement map tensor are output. The mesh is subdivided according to the displacement map tensor using surface subdivision technology to restore details. The topological state of the mesh is defined as the environment state, and the topological repair operation is defined as an action, including adding vertices, merging edges, and splitting triangles; the action is selected according to the current environment state, and the environment state is updated after the action is executed, and a reward is obtained.

5. The dynamic real-time rendering design editor platform according to claim 4, characterized in that: The LOD adaptive management unit trains a lightweight CNN model to perform screen space error prediction to achieve dynamic switching of LOD levels, specifically: The input of the CNN model is the camera distance d and the image features of the viewing angle coverage area Finally, the predicted screen space error value ε is output. Specifically, the actual error of the meshes at different LOD levels on the screen is used as a label. The actual error is obtained by calculating the pixel difference between the rendered image of the meshes at different LOD levels and the original high-resolution mesh rendered image. The mean square error calculation logic is: Among them, B p is the number of image pixels, is the color difference of the hth pixel in the original high-resolution image, is the color value of the hth pixel in the current LOD level image. By minimizing the loss between the predicted error and the actual error, a CNN model that can accurately predict the screen space error is trained. In the calculation process, according to the predicted screen space error value ε and the preset error threshold ε threshold For comparison, when ε>ε threshold When ε<ε, switch to high LOD level. threshold When the LOD level is high, it switches to the low LOD level, thereby realizing dynamic adaptive switching of LOD levels.

6. The dynamic real-time rendering design editor platform according to claim 1, characterized in that: The cross-platform resource management unit builds a unified resource access interface, which includes transparent access to APL, local files, cloud storage and engine internal resources, and the hardware implementation includes a protocol adapter and a cache manager.

7. The dynamic real-time rendering design editor platform according to claim 1, characterized in that: The real-time collaboration module resolves editing conflicts based on the operation log algorithm. When multiple users edit a scene simultaneously, the operations of different users are merged and adjusted according to the order and logical relationship of the operations. Specifically: The editing operation is abstracted into a five-tuple operation model, which is mathematically expressed as: O =<opTYPE,entityID,ψ,Q,D> ,where optype represents the set of operation types, including move, rotate and delete, entityID is the unique identifier of the target entity, ψ is the operation timestamp, Q is the operation parameter, and D is the globally unique version number; When two operations O1 and O2 act on the same entity, the conflict is eliminated by the operation transformation function. It is divided into two scenarios for processing. Scenario 1: When the move operation conflicts, the operation logic is: A = <Move,ec,ψ A , ΔV A ,D A >, user A moves the entity ec by ΔV A , where ec represents an editing operation, ΔV A represents the version vector of user A, O B = <Move,ec,ψ B , ΔV B ,D B >, user B moves ec at the same time, if ψ A <ψ B , then the client executes the logic locally: first apply O A , and then O B Convert to relative to O A The increment ΔV BA =ΔV B -ΔV A ,Server execution logic: ΔV final =V initial +ΔV A +ΔV BA , V initial Indicates the state before execution, ΔV final Indicates the status after execution; Scenario 2: Use version vector verification to avoid outdated operations during mutually exclusive operations. Define the version vector of entity e as ΔV e , when the client sends an operation O, the server checks: O ≥ ΔV e +1, if true, update O=ΔV e If the operation fails, it is considered as an outdated operation and ignored or an error is returned. The Lamport timestamp is introduced to establish a partial order relation of operations, which is defined as: O1→O2 if and only if According to the order of creating entities first and then editing, through this partial order relationship, the server reorders concurrent operations into a causally consistent sequence, and the final state satisfies: O∈ξ, the execution order respects the → relationship, where ξ is the combination of operation logs.

8. The dynamic real-time rendering design editor platform according to claim 1, characterized in that: The general class library is provided with a mathematical library and a geometric library. The mathematical library includes quaternion rotation, Euler angle conversion, affine transformation matrix construction, vector matrix operation and vectorized operation, and its target scenes mainly include graphics rendering, physical simulation and animation systems; the geometric library includes mesh operation, surface subdivision, collision detection and geometric measurement, and its target scenes mainly include terrain and building modeling generation, real-time rendering optimization and physical interaction.

Images