A three-dimensional digital asset intelligent generation data processing method and system

Abstract

The application discloses a kind of three-dimensional digital asset intelligent generation data processing method and system, the method includes: receiving multi-source data, respectively mapping as each source data corresponding feature vector, all feature vectors are weighted and fused, and generate structured latent variable vector.According to structured latent variable vector, generate sparse space unit set, establish local latent variable on each unit, calculate structure activation result according to local latent variable, and then determine the effective space unit set participating in three-dimensional structure construction.In effective space unit, generate material attribute vector according to corresponding local latent variable, and form three-dimensional asset intermediate result with structure activation result, combined with structured latent variable, execute local change detection, obtain to-be-updated space unit set.Perform local incremental update to to-be-updated space unit set, obtain updated three-dimensional asset result, carry out standardization packaging processing, to generate three-dimensional digital asset data package.

Description

Technical Field

[0001] This invention belongs to the field of three-dimensional asset processing technology, and in particular relates to a method and system for intelligent generation and processing of three-dimensional digital assets. Background Technology

[0002] With the continuous upgrading of digital content production methods, 3D digital assets have been widely used in product display, spatial content construction, digital exhibitions, interactive material production, and virtual object modeling. Unlike traditional 2D images, 3D digital assets not only contain geometric shape information but also multiple attribute data such as color, normal, roughness, metallicity, and transparency. Therefore, the generation process of 3D digital assets is essentially a data processing process involving multi-source data coupling, multi-stage reasoning, and multiple output results.

[0003] The existing 3D digital asset generation technology faces the following three types of technical bottlenecks.

[0004] Bottleneck 1: It is difficult to balance generation quality and structural maneuverability. While methods based on implicit fields or neural rendering intermediate representations can achieve high visual fidelity, the generated results often lack stable explicit structural representations, making it difficult to directly convert them into editable and reusable standard 3D asset data. On the other hand, methods based on explicit meshes or voxels, although structurally clear, are prone to loss of detail when dealing with complex boundaries, thin parts, and closed internal structures, leading to a decline in generation quality.

[0005] Bottleneck 2: Over-coupling of geometric results and material properties. Existing methods typically tightly bind the processing of material properties to geometric results, encoding lighting, color, and surface properties together into a single output channel. This results in 3D digital assets lacking independent material parameter channels after output, making it difficult to support consistent editing, independent reconstruction, and subsequent standardized processing, thus affecting asset reuse.

[0006] Bottleneck 3: Lack of local incremental update capability. When input prompts change locally or constraints are adjusted, existing 3D generation systems typically still regenerate the entire asset, resulting in a lengthy data processing chain, high computational resource consumption, and high result update latency. This makes it difficult to meet the requirements of large-scale 3D digital asset production scenarios for processing efficiency and continuous iteration capabilities.

[0007] Therefore, how to construct a data processing method that can simultaneously output high-quality 3D asset results, maintain structural tractability and material separability, and support local incremental updates is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention proposes a method for intelligent generation and data processing of three-dimensional digital assets, comprising:

[0009] Receive multi-source data, map the multi-source data respectively to generate feature vectors corresponding to each source data, and weight and fuse all the feature vectors to generate a structured latent variable vector, wherein the structured latent variable vector is used for global semantic and structural control.

[0010] A sparse spatial unit set is generated based on the structured latent variable vector, and local latent variables are established on each sparse spatial unit. The structure activation result of each sparse spatial unit is calculated based on the local latent variables, and the effective spatial unit set participating in the construction of the three-dimensional structure is determined based on the structure activation result. The sparse spatial unit set is a subset composed of voxel units predicted to be occupied by the target geometric structure after the target three-dimensional space is divided into voxel meshes of fixed resolution. The structure activation result is whether the current sparse spatial unit is occupied by the three-dimensional structure.

[0011] On the set of effective spatial units, a material attribute vector is generated according to the corresponding local latent variables, and together with the structure activation result, a three-dimensional asset intermediate result is formed. Based on the three-dimensional asset intermediate result and the structured latent variables, local change detection is performed to obtain the set of spatial units to be updated.

[0012] A local incremental update process is performed on the set of spatial units to be updated to obtain the updated 3D asset result. The updated 3D asset result is then standardized and encapsulated to generate a 3D digital asset data package.

[0013] Furthermore, the multi-source data includes: text prompt data, image reference data, structural constraint data, and historical asset prior data;

[0014] Generating the corresponding feature vectors includes: encoding the text prompt data through a text semantic mapping function to obtain a text semantic feature vector; encoding the image reference data through a visual reference mapping function to obtain a visual reference feature vector; encoding the structural constraint data through a structural constraint mapping function to obtain a structural constraint feature vector; and encoding the historical asset prior data through a prior pattern mapping function to obtain a prior pattern feature vector.

[0015] Furthermore, all the aforementioned feature vectors are weighted and fused to generate a structured latent variable vector, including:

[0016]

[0017] in, For the first The structured latent variable vector corresponding to each 3D digital asset generation task For the first The text prompt data for a 3D digital asset generation task For the first The image reference data for a 3D digital asset generation task For the first The structural constraint data for a three-dimensional digital asset generation task. For the first The historical asset prior data for each 3D digital asset generation task For the text semantic mapping function, For the visual reference mapping function, The structural constraint mapping function, The prior mode mapping function; The fusion weights are the semantic feature vectors of the text. The fusion weights of the visual reference feature vectors, The fusion weights are the structural constraint feature vectors. is the prior adjustment coefficient of the prior mode feature vector.

[0018] Furthermore, establishing local latent variables on each of the sparse spatial units includes:

[0019]

[0020] in, For the first The task of generating three-dimensional digital assets in the first... The local latent variables on the sparse spatial units. For the first Position encoding of each sparse spatial unit. For the first The structural mask information corresponding to each of the sparse spatial units. These are the local fusion coefficients of the structured latent variable vector. The local fusion coefficients for the location encoding are... is the local fusion coefficient of the structural mask information.

[0021] Furthermore, calculating the structural activation results of each sparse spatial unit based on the local latent variables includes:

[0022]

[0023] in, For the first The task of generating three-dimensional digital assets in the first... Structure activation results on sparse spatial units It is the Sigmoid activation function. Predict the weight vector for the structure. For the first Geometric constraints of a sparse spatial unit, For the first Neighborhood consistency term of a sparse spatial unit The structural prediction adjustment coefficient for the local latent variables. The structural prediction adjustment coefficient for the geometric constraint term. The structural prediction adjustment coefficient of the neighborhood consistency term is used to activate the structural structure. The sparse spatial units that exceed a preset activation threshold are determined as valid structural units, and all the valid structural units constitute the set of valid spatial units.

[0024] Furthermore, generating the material attribute vector based on the corresponding local latent variables includes:

[0025]

[0026] in, For the first The task of generating three-dimensional digital assets in the first... The material attribute vector on each of the effective structural units. For the first The task of generating three-dimensional digital assets in the first... The local latent variables on each of the effective structural units. The basic color generation function outputs the first color. The basic color components of each of the effective structural units. This is the roughness generation function, and its output value is the first... The roughness component of each effective structural unit, This is the metallicity generation function, and its output value is the first... The metallicity component of each of the effective structural units, This is the transparency generation function, and its output value is the first... The transparency component of each of the effective structural units, The material component weights of the basic color components. The material component weight of the roughness component. The material component weight of the metallicity component. The material component weight of the transparency component.

[0027] Furthermore, performing local incremental update processing on the set of spatial units to be updated includes:

[0028]

[0029] in, For the first The task of generating three-dimensional digital assets in the first... The incremental reprocessing cost on each of the sparse spatial units. For the first The task of generating three-dimensional digital assets in the first... The change in the local latent variable on each of the sparse spatial units. It is an L1 norm. For the first The task of generating three-dimensional digital assets in the first... The L1 norm of the change of the local latent variable on each of the sparse spatial units. For the first The priority weight of each sparse spatial unit, For the first A three-dimensional digital asset generation task The result error on each of the sparse space units. For the first The task of generating three-dimensional digital assets in the first... Processing latency estimation on each of the sparse spatial units This is the adjustment coefficient for the error term in the result. To handle the adjustment factor of the time delay estimation term, Sparse spatial cells that exceed a preset cost threshold are identified as the set of spatial cells to be updated.

[0030] This invention also proposes a three-dimensional digital asset intelligent generation data processing system, comprising:

[0031] The fusion module is used to receive multi-source data, map the multi-source data respectively to generate feature vectors corresponding to each source data, and perform weighted fusion of all the feature vectors to generate a structured latent variable vector, wherein the structured latent variable vector is used for global semantic and structural control.

[0032] The unit construction module is used to generate a sparse spatial unit set based on the structured latent variable vector, establish local latent variables on each sparse spatial unit, calculate the structure activation result of each sparse spatial unit based on the local latent variables, and determine the effective spatial unit set participating in the construction of the three-dimensional structure based on the structure activation result. The sparse spatial unit set is a subset composed of voxel units predicted to be occupied by the target geometric structure after the target three-dimensional space is divided into a voxel mesh of fixed resolution. The structure activation result is whether the current sparse spatial unit is occupied by the three-dimensional structure.

[0033] The update module is used to generate a material attribute vector on the effective spatial unit set according to the corresponding local latent variables, and form a three-dimensional asset intermediate result together with the structure activation result. Based on the three-dimensional asset intermediate result and the structured latent variables, local change detection is performed to obtain the spatial unit set to be updated.

[0034] The encapsulation module is used to perform local incremental update processing on the set of spatial units to be updated to obtain the updated 3D asset result, and to perform standardized encapsulation processing on the updated 3D asset result to generate a 3D digital asset data package.

[0035] Furthermore, the multi-source data includes: text prompt data, image reference data, structural constraint data, and historical asset prior data;

[0036] Generating the corresponding feature vectors includes: encoding the text prompt data through a text semantic mapping function to obtain a text semantic feature vector; encoding the image reference data through a visual reference mapping function to obtain a visual reference feature vector; encoding the structural constraint data through a structural constraint mapping function to obtain a structural constraint feature vector; and encoding the historical asset prior data through a prior pattern mapping function to obtain a prior pattern feature vector.

[0037] Furthermore, all the aforementioned feature vectors are weighted and fused to generate a structured latent variable vector, including:

[0038]

[0039] in, For the first The structured latent variable vector corresponding to each 3D digital asset generation task For the first The text prompt data for a 3D digital asset generation task For the first The image reference data for a 3D digital asset generation task For the first The structural constraint data for a three-dimensional digital asset generation task. For the first The historical asset prior data for each 3D digital asset generation task For the text semantic mapping function, For the visual reference mapping function, The structural constraint mapping function, The prior mode mapping function; The fusion weights are the semantic feature vectors of the text. The fusion weights of the visual reference feature vectors, The fusion weights are the structural constraint feature vectors. is the prior adjustment coefficient of the prior mode feature vector.

[0040] Compared with the prior art, the present invention has the following advantages and technical effects:

[0041] 1. Improved consistency of multi-source inputs. This invention uses a structured latent variable computation model to uniformly encode four types of heterogeneous inputs: text prompts, image references, structural constraints, and historical priors. It also achieves feature fusion within the same latent variable space, thus solving the problem of result bias caused by the lack of consistent representation between different input channels in existing methods.

[0042] 2. Improved computational efficiency. This invention utilizes a sparse latent variable grid mechanism to perform local latent variable calculations and subsequent inference only on the effective spatial region, skipping invalid spatial units and reducing overall computational resource consumption.

[0043] 3. Improved material separability. This invention calculates each material attribute component separately using independent basic color, roughness, metallicity, and transparency generation functions, enabling the output 3D digital assets to have independent material parameter channels and supporting independent editing, reconstruction, and standardization of each attribute.

[0044] 4. Improved Iteration Efficiency. This invention utilizes an incremental optimization mechanism to reprocess only the affected sparse spatial units when local changes occur in the input, avoiding repetitive generation operations at the whole asset level and significantly reducing processing latency during multiple iterations. Attached Figure Description

[0045] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0046] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention;

[0047] Figure 2 This is a system structure diagram of Embodiment 2 of the present invention. Detailed Implementation

[0048] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0049] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0050] The three-dimensional digital assets described in this invention refer to a collection of digital geometric and material data generated by a computer system and stored in a standardized file format, which can be directly accessed in a three-dimensional rendering engine or three-dimensional content creation software. Specifically, the three-dimensional digital assets described in this invention include the following two independent data components:

[0051] The first category: Geometric data components. These are three-dimensional geometric structures represented as polygonal meshes, specifically including: an array of vertex coordinates, where each vertex is represented by three-dimensional floating-point coordinates. The structured mesh data consists of: a facet index array, with triangular faces as the basic units, each triangular facet consisting of three vertex indices; a vertex normal vector array, describing the normal direction of the surface at each vertex; and a UV coordinate array, describing the mapping relationship between each vertex and the two-dimensional texture coordinate system. and normal data It can be stored in .obj, .fbx or .glb format on a computer-readable storage medium.

[0052] The second category: Material data components. This refers to a set of surface attribute parameters based on the Physically Based Rendering (PBR) specification, specifically including: basic color and texture data. For resolution not lower than A bitmap in RGBA format for each pixel, where each pixel stores the diffuse color value corresponding to its surface location; normal texture data (included in the normal data). The data is an RGB format bitmap with the same resolution as the basic color texture data, with each channel encoding the normal vector component in the tangent space; roughness and metallicity parameter data. As a dual-channel bitmap with the same resolution as the base color texture data, the R channel stores roughness values ​​(range). The larger the value, the higher the surface micro-roughness; the G channel stores the metallicity value (range). (The higher the value, the stronger the metallic reflective properties); and the transparency parameter, stored in the A channel of the basic color texture data, with a range of... .

[0053] The above two types of data components and asset description information (Including asset name, generation task number, and 3D bounding box size parameters) together constitute a standardized 3D digital asset data package. It can be directly imported and used by 3D content creation software such as Unity Engine, Unreal Engine, Blender, and Maya, and can be applied to 3D product display, virtual scene construction, digital human clothing modeling, and interactive virtual reality content production scenarios.

[0054] Example 1

[0055] like Figure 1 As shown in the figure, this embodiment proposes a data processing method for intelligent generation of three-dimensional digital assets, which specifically includes the following steps:

[0056] Step S1: Receive multi-source data, map the multi-source data respectively to generate feature vectors corresponding to each source data, and weight and fuse all the feature vectors to generate a structured latent variable vector.

[0057] Specifically, the multi-source data includes: text prompt data, image reference data, structural constraint data, and historical asset prior data;

[0058] The structured latent variable vector representation (SLV) maps multi-source input data (text, images, structural constraints, and historical priors) to a unified high-dimensional space, and then obtains a compact vector representation (e.g., dimension 1024) through weighted fusion. As the global semantic and structural control hub of the entire generation process, it undertakes the following functions: Semantic unification: aligning heterogeneous inputs to the same latent space, eliminating the expression gap between multi-source data; Structural guidance: providing global contextual information for the subsequent generation of sparse spatial units and the establishment of local latent variables; Consistency guarantee: serving as an "original reference benchmark" for detecting local changes during incremental updates (as described in step S3). Generating the corresponding feature vectors includes: encoding the text prompt data using a text semantic mapping function to obtain a text semantic feature vector; encoding the image reference data using a visual reference mapping function to obtain a visual reference feature vector; encoding the structural constraint data using a structural constraint mapping function to obtain a structural constraint feature vector; and encoding the historical asset prior data using a prior pattern mapping function to obtain a prior pattern feature vector.

[0059] Specifically, the weighted fusion of all the aforementioned feature vectors to generate a structured latent variable vector includes:

[0060] (1)

[0061] in, For the first The structured latent variable vector corresponding to each 3D digital asset generation task For the first The text prompt data for a 3D digital asset generation task For the first The image reference data for a 3D digital asset generation task For the first The structural constraint data for a three-dimensional digital asset generation task. For the first The historical asset prior data for each 3D digital asset generation task For the text semantic mapping function, For the visual reference mapping function, The structural constraint mapping function, The prior mode mapping function; The fusion weights are the semantic feature vectors of the text. The fusion weights of the visual reference feature vectors, The fusion weights are the structural constraint feature vectors. is the prior adjustment coefficient of the prior mode feature vector.

[0062] Step S2: Generate a sparse spatial unit set based on the structured latent variable vector, and establish local latent variables on each sparse spatial unit. In one specific embodiment, the sparse spatial unit set is generated as follows: First, the target three-dimensional space is divided into a voxel mesh with a fixed resolution (e.g., The structured latent variable vector is input into the occupancy prediction network, and the occupancy probability is output for each voxel unit. Voxel units with an occupancy probability greater than a preset threshold (e.g., 0.3) are selected as sparse spatial units, thus forming a sparse spatial unit set. Alternatively, invalid region markers in the structural constraint data and historical prior point cloud density can be combined for correction, ultimately generating an index set K. The structural activation results of each sparse spatial unit are calculated based on the local latent variables, and the effective spatial unit set participating in the 3D structure construction is determined based on the structural activation results.

[0063] Specifically, a sparse spatial cell set refers to a subset consisting only of voxel cells predicted to be occupied by the target geometry after dividing the target's 3D space into a fixed-resolution voxel mesh (e.g., 64×64×64). Sparse spatial cell sets can reduce computational cost by skipping inference on blank or invalid regions and processing only valid cells; focus on details by concentrating computational resources on key regions such as structural boundaries and thin walls; and support local updates by using cells as the smallest granularity, allowing incremental updates to process only affected cells.

[0064] Establishing local latent variables on each of the sparse spatial units includes:

[0065] (2)

[0066] in, For the first The task of generating three-dimensional digital assets in the first... The local latent variables on the sparse spatial units. For the first Position encoding of each sparse spatial unit. For the first The structural mask information corresponding to each of the sparse spatial units. These are the local fusion coefficients of the structured latent variable vector. The local fusion coefficients for the location encoding are... is the local fusion coefficient of the structural mask information.

[0067] Specifically, the structure activation result is a probability value between (0,1), representing the likelihood that the current sparse spatial unit is occupied by a 3D geometric structure, i.e., whether the unit belongs to a valid surface or interior region of the target model. The structure activation result is used to filter valid structural units (activation value > 0.5) to determine which units participate in geometry construction; it also serves as the input for subsequent material generation and mesh extraction (Marching Cubes).

[0068] The structural activation results of each sparse spatial unit are calculated based on the local latent variables, including:

[0069] (3)

[0070] in, For the first The task of generating three-dimensional digital assets in the first... Structure activation results on sparse spatial units It is the Sigmoid activation function. Predict the weight vector for the structure. For the first Geometric constraints of a sparse spatial unit, For the first Neighborhood consistency term of a sparse spatial unit The structural prediction adjustment coefficient for the local latent variables. The structural prediction adjustment coefficient for the geometric constraint term. The structural prediction adjustment coefficient of the neighborhood consistency term is used to activate the structural structure. The sparse spatial units exceeding a preset activation threshold are determined as valid structural units, and all valid structural units constitute the set of valid spatial units. In one specific implementation, the neighborhood consistency term is calculated as follows: for each sparse spatial unit, its neighboring units in the 3D mesh are selected; the L2 distance between the local latent variable of the current unit and the local latent variable of the neighboring units is calculated, and the average value is taken for all neighbors as the neighborhood consistency term to constrain the continuity of the local structure.

[0071] Step S3: On the set of effective spatial units, a material attribute vector is generated according to the corresponding local latent variables, and together with the structure activation result, a three-dimensional asset intermediate result is formed. Based on the three-dimensional asset intermediate result and the structured latent variables, local change detection is performed to obtain a set of spatial units to be updated. In one specific implementation, local change detection is achieved by comparing the projection difference of the structured latent variables on each sparse spatial unit before and after the update: the difference norm of the local latent variables before and after the update is calculated for each sparse spatial unit. When the difference exceeds a preset threshold, it is determined that the unit has changed and it is added to the set of spatial units to be updated.

[0072] Specifically, generating the material attribute vector based on the corresponding local latent variables includes:

[0073] (4)

[0074] in, For the first The task of generating three-dimensional digital assets in the first... The material attribute vector on each of the effective structural units. For the first The task of generating three-dimensional digital assets in the first... The local latent variables on each of the effective structural units. The basic color generation function outputs the first color. The basic color components of each of the effective structural units. This is the roughness generation function, and its output value is the first... The roughness component of each effective structural unit, This is the metallicity generation function, and its output value is the first... The metallicity component of each of the effective structural units, This is the transparency generation function, and its output value is the first... The transparency component of each of the effective structural units, The material component weights of the basic color components. The material component weight of the roughness component. The material component weight of the metallicity component. The material component weight of the transparency component.

[0075] Step S4: Perform local incremental update processing on the set of spatial units to be updated to obtain the updated 3D asset result. Then, perform standardized encapsulation processing on the updated 3D asset result to generate a 3D digital asset data package.

[0076] Specifically, performing local incremental update processing on the set of spatial units to be updated includes:

[0077] (6)

[0078] in, For the first The task of generating three-dimensional digital assets in the first... The incremental reprocessing cost on each of the sparse spatial units. For the first The task of generating three-dimensional digital assets in the first... The change in the local latent variable on each of the sparse spatial units. It is an L1 norm. This value is used to quantify the magnitude of the change in the unit before and after the update. The larger the value, the more severely the unit is affected by the input change, and therefore the higher the incremental reprocessing cost. The higher the value, the more likely it is to be selected as a unit to be updated. For the first The priority weight of each sparse spatial unit, For the first A three-dimensional digital asset generation task The result error on each of the sparse space units. For the first The task of generating three-dimensional digital assets in the first... Processing latency estimation on each of the sparse spatial units This is the adjustment coefficient for the error term in the result. To handle the adjustment factor of the time delay estimation term, Sparse spatial cells that exceed a preset cost threshold are identified as the set of spatial cells to be updated.

[0079] Example 2

[0080] like Figure 2 As shown in the figure, this embodiment proposes a three-dimensional digital asset intelligent generation data processing system, which specifically includes the following modules:

[0081] The fusion module is used to receive multi-source data, map the multi-source data respectively, generate feature vectors corresponding to each source data, and perform weighted fusion of all the feature vectors to generate a structured latent variable vector.

[0082] The unit construction module is used to generate a sparse spatial unit set based on the structured latent variable vector, establish local latent variables on each sparse spatial unit, calculate the structure activation result of each sparse spatial unit based on the local latent variables, and determine the effective spatial unit set participating in the three-dimensional structure construction based on the structure activation result.

[0083] The update module is used to generate a material attribute vector on the effective spatial unit set according to the corresponding local latent variables, and form a three-dimensional asset intermediate result together with the structure activation result. Based on the three-dimensional asset intermediate result and the structured latent variables, local change detection is performed to obtain the spatial unit set to be updated.

[0084] The encapsulation module is used to perform local incremental update processing on the set of spatial units to be updated to obtain the updated 3D asset result, and to perform standardized encapsulation processing on the updated 3D asset result to generate a 3D digital asset data package.

[0085] Since the system technical solution of this embodiment 2 is based on the technical solution of embodiment 1, it will not be described again.

[0086] Example 3

[0087] In this embodiment, a three-dimensional digital asset intelligent generation data processing method is deployed on a computing platform in the form of a computer program. The computing platform includes a model processing node composed of at least one graphics processing unit (GPU), a result cache node composed of memory or a solid-state drive, and an asset output node composed of a file system. Intermediate calculation results from each step are stored in GPU memory in tensor form.

[0088] The computing platform runs a 3D asset generation service program. After receiving a generation request, the program sequentially calls the corresponding neural network modules for each step and performs the corresponding calculations. Finally, it writes the 3D digital asset data package to a specified directory in the file system and stores it in .glb or .fbx format.

[0089] The three-dimensional space is pre-divided into Each sparse spatial unit is arranged in a uniform three-dimensional voxel mesh, with a voxel mesh resolution set to [value missing]. ,Right now The side length of each voxel corresponds to the bounding box of the target asset. In practical applications, the voxel mesh resolution can be adjusted according to the structural complexity of the target asset. to Values ​​within the specified range. Each sparse spatial unit. Corresponding position code The sine-cosine position encoding function is used to index the cell based on its coordinates in the 3D mesh. Calculated.

[0090] Set of valid structural unit indices The set of material property vectors corresponding to each effective structural unit The data is serialized into a binary tensor file and stored in the result cache node for incremental comparison and encapsulation calls in step S4.

[0091] Example 4

[0092] In this embodiment, the text semantic mapping function Implemented using a text encoder based on the Transformer architecture, it takes a UTF-8 encoded text prompt string (maximum 512 tokens) as input and outputs a dimension. floating-point eigenvectors The encoder's parameters are pre-trained on a corpus containing 3D asset description text and fine-tuned during the overall system training phase.

[0093] Visual reference mapping function Implemented by a visual Transformer, with a receiving resolution of at least [missing information]. The RGB format reference image of the pixels is used as input, and the output dimension is... floating-point eigenvectors Extract color distribution, boundary contours, and local texture features. When the generation request does not include a reference image, Set to zero vector and fuse weights In this case, set it to zero.

[0094] Structural constraint mapping function Implemented by a three-layer fully connected network, it receives a numerical structural constraint parameter vector as input. This structural constraint parameter vector includes: bounding box 3D dimension parameters (3D floating-point numbers), morphological contour keypoint coordinates (up to 32 3D keypoints, totaling 96 dimensions), and local part constraint markers (up to 16 binary markers, totaling 16 dimensions). All parameters are concatenated to form a 115-dimensional input vector, and the output dimension is... eigenvectors .

[0095] Prior pattern mapping function Implemented using a PointNet-style point cloud feature extraction network, it receives a structural mesh sampled point cloud of historical 3D assets (1024 sample points, each containing 3D coordinates and a normal vector, with input dimensions...). As input, output dimension Prior feature vectors When the generated request does not contain prior historical asset data, Set to zero vector, prior adjustment coefficient In this case, set it to zero.

[0096] Fusion weights and prior adjustment coefficient Satisfying constraints The determination is made through end-to-end supervised learning during the system training phase.

[0097] The following explanation uses a specific generation task as an example: When the text prompt for the generation request is "Metallic material, sharp edges, hollow structure", and a reference image of the target object (resolution) is provided, (pixels), and the structural constraint data includes the bounding box size. When marking cavities: Extract the high metallicity semantic features corresponding to "metallic material", the high curvature boundary semantic features corresponding to "sharp edge" and the internal cavity topological semantic features corresponding to "hollow structure", and output a 1024-dimensional text semantic feature vector; Further extract color distribution and boundary contour features from the reference image; The bounding box size and cavity markings are encoded into 1024-dimensional structural constraint feature vectors; the four types of feature vectors are linearly superimposed according to equation (1) to generate a 1024-dimensional structured latent variable vector. .

[0098] Example 5

[0099] In this embodiment, the three-dimensional space is pre-divided into Each sparse spatial unit It has a corresponding position code and structural mask information Location encoding Describes the coordinate position of sparse spatial units in three-dimensional space, and structural mask information. This describes whether the cell is within a valid structural region. The structural mask information is generated jointly by structural constraint data and historical asset prior data: first, invalid regions in the structural constraint data are mapped to a voxel grid to obtain an initial mask; then, the mask is corrected by combining the point cloud distribution in the historical asset prior data and using a spatial density estimation method; finally, binary structural mask information is generated to indicate whether the corresponding sparse spatial cell participates in subsequent calculations.

[0100] System traverses all The system skips the calculation of local latent variables for sparse spatial cells that are marked as invalid regions in the structural constraint data or historical asset prior data. For valid sparse spatial cells, their local latent variables are calculated according to equation (2). Local fusion coefficient , , Determined during the training phase.

[0101] Taking the generation of 3D assets with hollow structures as an example, the sparse spatial units corresponding to the hollow interior region are marked as invalid regions in the structural constraint data. The system skips the calculation of these units and only calculates the local latent variables of the sparse spatial units corresponding to the outer wall structure, thereby reducing the overall computational load.

[0102] Example 6:

[0103] In this embodiment, the system calculates the structure activation value for all sparse spatial cells in the sparse latent variable grid according to equation (3). The Sigmoid activation function normalizes the computation results to... The interval, where the output value represents the probability that the sparse spatial unit is occupied by a 3D geometric structure. Sparse spatial units with structural activation values ​​exceeding a preset activation threshold of 0.5 are identified as valid structural units and participate in 3D shape construction. Structural prediction weight vector. and structural prediction adjustment coefficient During the system training phase, supervised learning was used to determine that the training data consisted of a 3D asset dataset containing voxelized occupancy annotations of real 3D models.

[0104] For each effective structural unit, calculate the material property vector according to equation (4). Four generating functions , , , All are implemented using a four-layer multilayer perceptron network (MLP), with each layer containing 1024 neurons and ReLU activation function, using local latent variables. For input. Specifically, Output a 3D floating-point vector representing the RGB color value of the corresponding surface location for that element, with the range of each component. ; Output a 1D floating-point scalar representing the surface micro-roughness, with a range of... A larger value indicates a rougher surface. Output a 1D floating-point scalar representing the metallicity, with a range of... A larger value indicates a stronger metal reflective property; Output a 1D floating-point scalar representing transparency and range. A value of 0 indicates complete opacity. Material component weight. satisfy This is determined during the system training phase.

[0105] Geometric material coupling constraint terms Calculate according to formula (5), as follows:

[0106] The geometric material coupling constraint terms include:

[0107] (5)

[0108] in, For the first Geometric material coupling constraints for a 3D digital asset generation task For the first The local gradient change at each of the effective structural units in a three-dimensional digital asset generation task is the set of structural activation results, which comprises the structural activation results of all the effective structural units. For the first The local gradient change at each of the effective structural units of a 3D digital asset generation task, which is composed of the material attribute vectors of all the effective structural units. For the first The target reference result for the three-dimensional digital asset generation task For the first The current generation results of a 3D digital asset generation task. As the first constraint weight, As the second constraint weight, It is an L1 norm. This is the square of the L2 norm. This constraint reduces distortion caused by misalignment between structural and material boundaries, resulting in more stable generated results fed into subsequent incremental optimization processes.

[0109] Structure activation value three-dimensional scalar field The construction: All The structural activation value of a sparse spatial unit According to each unit (In this embodiment) 3D coordinate index in voxel mesh Fill in the corresponding positions to form the shape as follows A three-dimensional floating-point array, i.e., a three-dimensional scalar field of structure activation values. In this scalar field, the value at the corresponding position of the effective structural element is its structural activation value. (scope The value at the corresponding position of an ineffective structural element (i.e., a spatial location not occupied by a geometric structure) is... (scope ) or the location marked as an invalid region (set to zero directly).

[0110] Material property three-dimensional vector field Construction: Combining each effective structural unit Material property vector (Includes 3D basic color components, 1D roughness component, 1D metallicity component, and 1D transparency component, for a total of 6 dimensions) Indexed by the corresponding 3D coordinates Fill in the corresponding positions, and set the positions of non-effective structural units to zero vectors, forming a shape of A four-dimensional floating-point array, i.e., a three-dimensional vector field of material properties. .

[0111] In one specific implementation, the system can first determine the resolution. Low-resolution structural results are generated on a coarse-grained voxel mesh to determine the main body contour, and then the resulting low-resolution structural results are upsampled to... Resolution, based on the main body outline, involves refining the material properties, and finally... Perform local corrections to the structural results at the specified resolution.

[0112] Example 7

[0113] In this embodiment, when a user requests partial modification based on the existing generated result, taking the example that the user only requests to adjust the surface material of the generated object from a frosted effect to a high-reflectivity effect: the system re-executes step S1 on the modified text prompt data to generate an updated structured latent variable vector. Subsequently, the changes in local latent variables were calculated for each sparse spatial unit. Since only the material description changes, the sparse spatial units associated with the surface material... Large numerical values ​​of sparse spatial units that are independent of geometry The value is close to zero.

[0114] Calculate the incremental reprocessing cost of each sparse spatial unit according to equation (6). Result error The error is obtained by comparing the difference between the current generated result of the sparse spatial unit and the corresponding target reference result; in supervised training scenarios, the result error is calculated by the difference between the current generated result and the target 3D asset, including voxel occupancy error and material attribute error; in unsupervised or inference stages, the result error is estimated by the difference between the current generated result and the historical generated result or by structural consistency indicators. Processing latency estimation. The estimation is based on the historical processing latency records of the sparse spatial unit. Sparse spatial units whose incremental reprocessing cost exceeds a preset cost threshold are identified as regions to be reprocessed (forming a set of spatial units to be updated). The system only re-executes the material attribute vector calculation for these units, while the original structural activation values ​​and material results remain unchanged for sparse spatial units with geometrically related structures. For spatial units to be updated, only the local latent variable calculation and material attribute generation steps are re-executed, while the structural activation results and material attributes of unchanged sparse spatial units remain unchanged, thereby achieving local incremental updates.

[0115] Example 8

[0116] In this embodiment, the system encapsulates the structural activation results and material attribute vector sets of effective structural units according to the following specific steps to generate a standardized 3D digital asset data package. .

[0117] Structured grid data Extraction: All The structural activation value of a sparse spatial unit The resolution is A three-dimensional scalar field is generated. The Marching Cubes algorithm is applied to this scalar field with a preset activation threshold of 0.5 as the isosurface threshold to extract triangular polygonal meshes. The resulting mesh vertex coordinates are converted to the actual physical dimensions (in meters) of the target asset based on the voxel grid size. Mesh patches are stored as a list of triangular patches, typically with 5,000 to 50,000 vertices and 8,000 to 100,000 patches.

[0118] UV coordinate generation: The polygonal mesh extracted by the Marching Cubes algorithm is parametrically unfolded using a UV unfolding algorithm based on angle-preserving mapping, unfolding each triangular facet of the 3D surface without overlap. In the two-dimensional texture coordinate space, generate the UV coordinate array of each vertex.

[0119] Basic color texture data Generation: The basic color components (3D RGB floating-point vectors) of each effective structural unit are rasterized and projected onto the resolution according to the UV mapping relationship. A pixel-based RGBA format bitmap, with the RGB three channels storing diffuse color values ​​and the A channel storing the alpha component, generates basic color texture data. .

[0120] Normal data Generation: Calculate the vertex normal vectors based on the vertex coordinates and face normal vectors of the polygon mesh, and rasterize and project the vertex normal vectors to the resolution according to the UV mapping relationship. A pixel's RGB format bitmap, where the RGB three channels encode the normal vector in tangent space. , , Components, range of each component (Corresponding normal vector component range) Generate normal texture data, and construct normal data. .

[0121] Roughness and metallicity parameter data Generation: The roughness and metallicity components of each effective structural unit are rasterized and projected onto the resolution according to the UV mapping relationship. A dual-channel bitmap of pixels, with the R channel storing roughness values ​​and the G channel storing metallicity values, generates roughness and metallicity parameter data. .

[0122] Data packet encapsulation: Encapsulating structured grid data (Includes vertex coordinate array, patch index array, and UV coordinate array), basic color texture data Normal data Roughness and metallicity parameter data And asset description information and index information (Includes the generation task number) Text prompt summary, 3D bounding box size parameters The generated 3D digital assets (along with the file path index) are serialized and written to a .glb format file. This file conforms to the glTF 2.0 standard and can be directly imported into 3D content creation software such as Unity Engine, Unreal Engine, Blender, and Maya. It can be used in applications such as 3D product display (importing the generated 3D digital assets into the 3D display module of an e-commerce platform for user interaction and browsing) and virtual scene construction (importing the generated 3D digital assets into a virtual reality scene editor as scene objects).

[0123] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligently generating data processing for three-dimensional digital assets, characterized in that, include: Receive multi-source data, map the multi-source data respectively to generate feature vectors corresponding to each source data, and weight and fuse all the feature vectors to generate a structured latent variable vector, wherein the structured latent variable vector is used for global semantic and structural control. A sparse spatial unit set is generated based on the structured latent variable vector, and local latent variables are established on each sparse spatial unit. The structure activation result of each sparse spatial unit is calculated based on the local latent variables, and the effective spatial unit set participating in the construction of the three-dimensional structure is determined based on the structure activation result. The sparse spatial unit set is a subset composed of voxel units predicted to be occupied by the target geometric structure after the target three-dimensional space is divided into voxel meshes of fixed resolution. The structure activation result is whether the current sparse spatial unit is occupied by the three-dimensional structure. On the set of effective spatial units, a material attribute vector is generated according to the corresponding local latent variables, and together with the structure activation result, a three-dimensional asset intermediate result is formed. Based on the three-dimensional asset intermediate result and the structured latent variables, local change detection is performed to obtain the set of spatial units to be updated. A local incremental update process is performed on the set of spatial units to be updated to obtain the updated 3D asset result. The updated 3D asset result is then standardized and encapsulated to generate a 3D digital asset data package.

2. The intelligent generation data processing method for three-dimensional digital assets as described in claim 1, characterized in that, The multi-source data includes: text prompt data, image reference data, structural constraint data, and historical asset prior data; Generating the corresponding feature vectors includes: encoding the text prompt data through a text semantic mapping function to obtain a text semantic feature vector; encoding the image reference data through a visual reference mapping function to obtain a visual reference feature vector; encoding the structural constraint data through a structural constraint mapping function to obtain a structural constraint feature vector; and encoding the historical asset prior data through a prior pattern mapping function to obtain a prior pattern feature vector.

3. The intelligent data processing method for generating three-dimensional digital assets as described in claim 2, characterized in that, The structured latent variable vector is generated by weighting and fusing all the aforementioned feature vectors, including: ； in, For the first The structured latent variable vector corresponding to each 3D digital asset generation task For the first The text prompt data for a 3D digital asset generation task For the first The image reference data for a 3D digital asset generation task For the first The structural constraint data for a three-dimensional digital asset generation task. For the first The historical asset prior data for each 3D digital asset generation task For the text semantic mapping function, For the visual reference mapping function, The structural constraint mapping function, The prior mode mapping function; The fusion weights are the semantic feature vectors of the text. The fusion weights of the visual reference feature vectors, The fusion weights are the structural constraint feature vectors. is the prior adjustment coefficient of the prior mode feature vector.

4. The intelligent generation data processing method for three-dimensional digital assets as described in claim 3, characterized in that, Establishing local latent variables on each of the sparse spatial units includes: ； in, For the first The task of generating three-dimensional digital assets in the first... The local latent variables on the sparse spatial units. For the first Position encoding of each sparse spatial unit. For the first The structural mask information corresponding to each of the sparse spatial units. These are the local fusion coefficients of the structured latent variable vector. The local fusion coefficients for the location encoding are... is the local fusion coefficient of the structural mask information.

5. The intelligent generation data processing method for three-dimensional digital assets as described in claim 4, characterized in that, The structural activation results of each sparse spatial unit are calculated based on the local latent variables, including: ； in, For the first The task of generating three-dimensional digital assets in the first... Structure activation results on sparse spatial units It is the Sigmoid activation function. Predict the weight vector for the structure. For the first Geometric constraints of a sparse spatial unit, For the first Neighborhood consistency term of a sparse spatial unit The structural prediction adjustment coefficient for the local latent variables. The structural prediction adjustment coefficient for the geometric constraint term. The structural prediction adjustment coefficient of the neighborhood consistency term is used to activate the structural structure. The sparse spatial units that exceed a preset activation threshold are determined as valid structural units, and all the valid structural units constitute the set of valid spatial units.

6. The intelligent generation data processing method for three-dimensional digital assets as described in claim 5, characterized in that, Generating a material attribute vector based on the corresponding local latent variables includes: ； in, For the first The task of generating three-dimensional digital assets in the first... The material attribute vector on each of the effective structural units. For the first The task of generating three-dimensional digital assets in the first... The local latent variables on each of the effective structural units. The basic color generation function outputs the first color. The basic color components of each of the effective structural units. This is the roughness generation function, and its output value is the first... The roughness component of each effective structural unit, This is the metallicity generation function, and its output value is the first... The metallicity component of each of the effective structural units, This is the transparency generation function, and its output value is the first... The transparency component of each of the effective structural units, The material component weights of the basic color components. The material component weight of the roughness component. The material component weight of the metallicity component. The material component weight of the transparency component.

7. The intelligent generation data processing method for three-dimensional digital assets as described in claim 4, characterized in that, Performing local incremental update processing on the set of spatial units to be updated includes: ； in, For the first The task of generating three-dimensional digital assets in the first... The incremental reprocessing cost on each of the sparse spatial units. For the first The task of generating three-dimensional digital assets in the first... The change in the local latent variable on each of the sparse spatial units. It is an L1 norm. For the first The task of generating three-dimensional digital assets in the first... L1 norm of the change of the local latent variable on each of the sparse spatial units. For the first The priority weight of each sparse spatial unit, For the first A three-dimensional digital asset generation task The result error on each of the sparse space units. For the first The task of generating three-dimensional digital assets in the first... Processing latency estimation on each of the sparse spatial units. This is the adjustment coefficient for the error term in the result. To handle the adjustment factor of the time delay estimation term, Sparse spatial cells that exceed a preset cost threshold are identified as the set of spatial cells to be updated.

8. A three-dimensional digital asset intelligent generation data processing system, characterized in that, include: The fusion module is used to receive multi-source data, map the multi-source data respectively, generate feature vectors corresponding to each source data, and perform weighted fusion of all the feature vectors to generate a structured latent variable vector, wherein the structured latent variable vector is used for global semantic and structural control. The unit construction module is used to generate a sparse spatial unit set based on the structured latent variable vector, establish local latent variables on each sparse spatial unit, calculate the structure activation result of each sparse spatial unit based on the local latent variables, and determine the effective spatial unit set participating in the construction of the three-dimensional structure based on the structure activation result. The sparse spatial unit set is a subset composed of voxel units predicted to be occupied by the target geometric structure after the target three-dimensional space is divided into a voxel mesh of fixed resolution. The structure activation result is whether the current sparse spatial unit is occupied by the three-dimensional structure. The update module is used to generate a material attribute vector on the effective spatial unit set according to the corresponding local latent variables, and form a three-dimensional asset intermediate result together with the structure activation result. Based on the three-dimensional asset intermediate result and the structured latent variables, local change detection is performed to obtain the spatial unit set to be updated. The encapsulation module is used to perform local incremental update processing on the set of spatial units to be updated to obtain the updated 3D asset result, and to perform standardized encapsulation processing on the updated 3D asset result to generate a 3D digital asset data package.

9. The intelligent data processing system for generating three-dimensional digital assets as described in claim 8, characterized in that, The multi-source data includes: text prompt data, image reference data, structural constraint data, and historical asset prior data; Generating the corresponding feature vectors includes: encoding the text prompt data through a text semantic mapping function to obtain a text semantic feature vector; encoding the image reference data through a visual reference mapping function to obtain a visual reference feature vector; encoding the structural constraint data through a structural constraint mapping function to obtain a structural constraint feature vector; and encoding the historical asset prior data through a prior pattern mapping function to obtain a prior pattern feature vector.

10. The intelligent data processing system for generating three-dimensional digital assets as described in claim 9, characterized in that, The structured latent variable vector is generated by weighting and fusing all the aforementioned feature vectors, including: ； in, For the first The structured latent variable vector corresponding to each 3D digital asset generation task For the first The text prompt data for a 3D digital asset generation task For the first The image reference data for a 3D digital asset generation task For the first The structural constraint data for a three-dimensional digital asset generation task. For the first The historical asset prior data for each 3D digital asset generation task For the text semantic mapping function, For the visual reference mapping function, The structural constraint mapping function, The prior mode mapping function; The fusion weights are the semantic feature vectors of the text. The fusion weights of the visual reference feature vectors, The fusion weights are the structural constraint feature vectors. is the prior adjustment coefficient of the prior mode feature vector.

Images