Three-dimensional model generation method and device, electronic equipment and storage medium

Abstract

The invention provides a three-dimensional model generation method and device, electronic equipment and a storage medium, and relates to the technical field of animation creation. The method comprises the steps of determining initial structure information of a target model in response to an input instruction; receiving an adjustment instruction input by the input device for the initial structure information, and updating the initial structure information according to the adjustment instruction to obtain target structure information; determining target structure information of the target model; according to the target parameter information of each component, generating three-dimensional grid data corresponding to each component; and according to the target structure information and the three-dimensional grid data corresponding to each component, performing aggregation to generate model data of the target model. By introducing the concept of the structure information, the model is converted into the intermediate structure representation which can be understood and edited by the user, so that the initial structure information is accurately adjusted according with the willingness of the user based on the user operation, the obtained target structure information lays a foundation for realizing the controllable model, and the accuracy of generating the model is improved.

Description

Technical Field

[0001] This application relates to the field of animation creation technology, and more specifically, to a method, apparatus, electronic device, and storage medium for generating three-dimensional models. Background Technology

[0002] In modern game development, film and animation, and other digital content industries, the production of 3D models is not only a core element of creative realization but also the most labor-intensive and time-consuming step in the entire production process. As the industry's demand for high-quality, diverse visual content continues to grow, traditional manual modeling methods are no longer sufficient to meet the increasingly complex project requirements and tight delivery cycles.

[0003] Currently, artificial intelligence content generation technology is often used to create models, such as predicting model information by analyzing a large amount of existing model data based on deep learning models.

[0004] However, the methods described above generate a fixed, overall model that lacks downstream editability, resulting in poor flexibility in model creation. Summary of the Invention

[0005] The purpose of this application is to address the shortcomings of the prior art by providing a method, apparatus, electronic device, and storage medium for generating three-dimensional models, so as to improve the efficiency, controllability, and flexibility of three-dimensional model production and enhance model quality.

[0006] To achieve the above objectives, the technical solutions adopted in the embodiments of this application are as follows: In a first aspect, embodiments of this application provide a method for generating a three-dimensional model, including: In response to input commands, the initial structural information of the target model is determined, including the identifiers of each component constituting the target model and the initial parameter information of each component; Receive adjustment instructions for the initial structure information input via an input device, update the initial structure information according to the adjustment instructions, and obtain target structure information; Determine the target structure information of the target model; the target structure information includes: target parameter information of each component; the target parameter information includes: target three-dimensional boundary volume information and abstract morphological encoding of the component; the target three-dimensional boundary volume information includes the three-dimensional bounding box parameters of the component, which are used to indicate the size, position and orientation of the component in three-dimensional space; the initial abstract morphological encoding is used to indicate the category, morphological attributes and style attributes of the component; Based on the target parameter information of each component, generate the corresponding three-dimensional mesh data for each component; Based on the target structure information and the three-dimensional mesh data corresponding to each component, the model data of the target model is aggregated and generated.

[0007] Secondly, embodiments of this application also provide a three-dimensional model generation apparatus, including: a determining module and a generating module; The determining module is used to respond to input instructions and determine the initial structural information of the target model. The initial structural information includes the identifiers of each component constituting the target model and the initial parameter information of each component. The generation module is configured to receive an adjustment instruction for the initial structural information input through an input device, update the initial structural information according to the adjustment instruction, and obtain the target structural information. The determining module is used to determine the target structure information of the target model. The target structure information includes: target parameter information of each component; the target parameter information includes: target three-dimensional boundary volume information of the component and abstract morphological encoding; the target three-dimensional boundary volume information includes the three-dimensional bounding box parameters of the component, which are used to indicate the size, position and orientation of the component in three-dimensional space; the initial abstract morphological encoding is used to indicate the category, morphological attributes and style attributes of the component. The generation module is used to generate three-dimensional mesh data corresponding to each component based on the target parameter information of each component. The generation module is used to aggregate and generate model data of the target model based on the target structure information and the three-dimensional mesh data corresponding to each component.

[0008] Thirdly, embodiments of this application provide an electronic device, including: a processor, a storage medium, and a bus. The storage medium stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the storage medium via the bus, and the processor executes the machine-readable instructions to perform the three-dimensional model generation method provided in the first aspect.

[0009] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, performs the three-dimensional model generation method as provided in the first aspect.

[0010] The beneficial effects of this application are: This application provides a method, apparatus, electronic device, and storage medium for generating a 3D model, comprising: responding to an input command to determine the initial structural information of a target model; receiving an adjustment command for the initial structural information input through an input device, updating the initial structural information according to the adjustment command to obtain target structural information; determining the target structural information of the target model; generating 3D mesh data corresponding to each component based on the target structural information; and generating model data of the target model based on the target structural information and the 3D mesh data corresponding to each component. This solution introduces the concept of structural information, converting the model into a user-understandable and editable intermediate structural representation, breaking the traditional invisible generation process. Users can perform editing operations on the model based on the structural information, thereby allowing user intervention between the two stages of generating initial structural information and generating 3D mesh data of components. Users can visually review and edit the generated initial structural information, and adjust it. Through user intervention, the automatically generated initial structural information can be precisely adjusted to meet user wishes, thereby achieving controllability in model creation and improving the accuracy of the generated model. By generating the target structure information of the target model, an intermediate structure representation that can be understood and seen by the user can be obtained, laying the foundation for the user to achieve model controllability during the model generation process. In the process of generating the three-dimensional mesh data of the components, by generating the three-dimensional mesh data corresponding to each component separately, the components of the model can be decoupled, which facilitates the subsequent individual adjustment of the components and secondary creation, thereby improving the model creation efficiency and application value.

[0011] Secondly, by decoupling the 3D mesh generation tasks for each component and processing each component using an independent component refinement network, the learning difficulty of each component refinement network can be significantly reduced, improving the geometric quality and structural rationality of the final model. Furthermore, by executing the component refinement network tasks in parallel, the generation efficiency of complex models can be greatly improved.

[0012] In addition, based on the independent generation of each component, it can support independent adjustment of any component without the need for expensive recalculation of the entire model. This on-demand update feature provides key technical support for agile development and rapid iterative design processes. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1A flowchart illustrating a three-dimensional model generation method provided in an embodiment of this application; Figure 2 A flowchart illustrating another method for generating a three-dimensional model provided in an embodiment of this application; Figure 3 A flowchart illustrating another method for generating a three-dimensional model provided in this application embodiment; Figure 4 A flowchart illustrating another method for generating a three-dimensional model provided in an embodiment of this application; Figure 5 A flowchart illustrating another method for generating a three-dimensional model provided in this application embodiment; Figure 6 A flowchart illustrating another method for generating a three-dimensional model provided in an embodiment of this application; Figure 7 A flowchart illustrating another method for generating a three-dimensional model provided in an embodiment of this application; Figure 8 A flowchart illustrating another method for generating a three-dimensional model provided in this application embodiment; Figure 9 A flowchart illustrating another method for generating a three-dimensional model provided in this application embodiment; Figure 10 This is a schematic diagram of the core process of a three-dimensional model generation method provided in an embodiment of this application; Figure 11 A schematic diagram of the core process of another three-dimensional model generation method provided in the embodiments of this application; Figure 12 This is a schematic diagram of a three-dimensional model generation device provided in an embodiment of this application; Figure 13 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0016] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0017] It should be noted that the term "comprising" will be used in the embodiments of this application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0018] While existing AI-powered content generation technologies can automate the creation of 3D models to some extent, they generally have serious limitations, making it difficult to efficiently utilize the generated model assets in actual production processes.

[0019] They generally suffer from the following core technological defects: 1. "Black box" generation, lacking controllable intermediate representation: Most methods encapsulate the complex generation process in an end-to-end neural network. Users cannot review or intervene in the intermediate states such as structural planning and component layout during the generation process, resulting in a "blind box" effect in the output results, which makes it difficult to meet precise design requirements.

[0020] 2. Structural monolithicity and lack of downstream editability: Existing technologies tend to generate a topologically fixed monolithic mesh, with no independent geometric definitions between its internal components. It lacks modularity and disassembly, which severely limits its value in industrial processes that require secondary creation, physical simulation, or interactive applications.

[0021] 3. Implicit structural associations and insufficient logical constraints: Even if some cutting-edge research attempts to generate multi-part models, the associations between the parts are usually learned in an implicit latent space. The lack of clear and interpretable spatial relationships and logical semantic guidance between the parts makes it difficult to guarantee the structural rationality of the generated results, and does not support secondary editing of the part layout.

[0022] Based on this, this solution provides a 3D model generation method. By generating the structural information of the model, the complex model generation task can be decoupled into two logically clear stages. It can not only autonomously generate a 3D model with a reasonable structure composed of multiple independent components, but also give users control capabilities, allowing users to make precise adjustments to the macroscopic structure of the model during the 3D model generation process, thereby generating a high-quality 3D model that truly meets the user's needs.

[0023] Figure 1 A flowchart illustrating a three-dimensional model generation method provided in this application embodiment; as shown Figure 1 As shown, the method includes: S101. Respond to the input command and determine the initial structural information of the target model.

[0024] The initial structural information includes the identifiers of each component that makes up the target model and the initial parameter information of each component.

[0025] It can respond to user input commands, which can indicate relevant information about the target model to be built. Based on the input commands, the initial structural information of the target model can be determined.

[0026] It is worth noting that the initial structural information may include the identifiers of each component that makes up the target model and the parameter information of each component. However, the initial structural information may not be accurate enough, and it can be adjusted later according to user needs.

[0027] The initial parameter information for each component may include: the initial three-dimensional boundary volume information and abstract morphology encoding of the component; the initial three-dimensional boundary volume information includes the component's three-dimensional bounding box parameters, which are used to indicate the component's size, position and orientation in three-dimensional space; the abstract morphology encoding is used to indicate the component's category, morphological attributes and style attributes.

[0028] In some embodiments, the initial structural information of the target model can be automatically generated based on the input instructions and the relevant information of the target model indicated by the input instructions. This automation can be achieved with the help of intelligent tools such as AI.

[0029] S102. Receive adjustment instructions for the initial structure information input through the input device, update the initial structure information according to the adjustment instructions, and obtain the target structure information.

[0030] To improve the accuracy of the structural information of the obtained target model, this embodiment also provides a user editing function, which allows users to adjust the initial structural information according to their own needs in order to obtain more accurate target structural information.

[0031] In other words, this solution uses a human-machine collaborative model, where AI is responsible for efficient preliminary planning and detailed generation, while users correct and guide the AI's planning results, ultimately achieving a perfect combination of AI technology and user creation.

[0032] In some embodiments, users can input adjustment commands through input devices such as keyboards, mice, and microphones. Of course, they can also input adjustment commands through controls in the user interface or through touch input.

[0033] S103. Determine the target structure information of the target model.

[0034] The target structure information includes: the identifiers of each component constituting the target model and the target parameter information of each component. The target parameter information includes: the target 3D boundary volume information and abstract morphological encoding of the component; the target 3D boundary volume information includes the 3D bounding box parameters of the component, which are used to indicate the size, position and orientation of the component in 3D space; the abstract morphological encoding is used to indicate the category, morphological attributes and style attributes of the component.

[0035] The target structure information and the initial structure information have the same data structure, including: the identifier of each component and the target parameter information of each component. The meaning and concept of the target parameter information are similar to those of the initial parameter information.

[0036] The target model can refer to the three-dimensional model to be constructed, and the target structural information can represent the structural blueprint that describes the macroscopic structure of the target model.

[0037] In this embodiment, the target structure information may include the identifiers of each component constituting the target model, that is, indicating which core components the target model is composed of; in addition, it also includes target parameter information for each component, which indicates the position of each component in three-dimensional space, as well as rough geometric planning information such as size and orientation.

[0038] The 3D boundary volume information of a component is used to describe the complete information required for a component in 3D space. It can refer to the 3D bounding box parameters of the component, usually represented by a vector, which includes: Position: The coordinates (x, y, z) of the center point of the bounding box.

[0039] Size / Extents: The half-length, half-width, and half-height (dx, dy, dz) of the bounding box along the three axes of its own coordinate system.

[0040] Orientation: Defines the rotational information of the bounding box orientation, usually represented by a quaternion (qw, qx, qy, qz) because it is more stable than Euler angles and does not have gimbal lock issues.

[0041] These parameters together precisely define the area occupied by each component in three-dimensional space.

[0042] The origin and meaning of abstract morphological encoding: Source: The abstract morphological encoding is generated by a separate, pre-trained part-level variational autoencoder (Part-VAE). This Part-VAE is trained on a large-scale dataset containing tens of thousands of independent, normalized 3D part models (e.g., chair legs, aircraft engines, car wheel hubs, etc. from autonomous driving datasets).

[0043] Encoding and decoding process: VAE's encoder E vae It receives a 3D mesh of a component as input and compresses (maps) it into a d-dimensional latent vector. This vector is the abstract morphological code L. i (In practice, the mean vector of the VAE output is usually taken). This compression process forces the model to ignore microscopic vertex position details and instead learn and capture the most essential, high-level semantic information of the components. In contrast, the VAE decoder D... vae Then it is possible to obtain a potential vector L i Starting from this point, the corresponding 3D mesh of the component is reconstructed. The goal of training is to minimize the difference between the original mesh and the reconstructed mesh.

[0044] The meaning of abstract morphological encoding: It is precisely because of this "compression-reconstruction" training paradigm that the latent vector L i The captured "high-level semantics" is not single-dimensional information, but a complex, decoupled set of features. Specifically, abstract morphological encoding includes: Category identity: Specific dimensions or regions in the vector clearly distinguish whether the component is a "wheel," "wing," or "backrest." This is its most basic semantic information.

[0045] Style characteristics: The aesthetic style of the component can be encoded, such as whether it is "sci-fi streamlined" or "industrial rugged", "minimalist" or "glamorous".

[0046] Topological and structural properties: The basic topological structure of a component, such as "slender", "flat", "hollow" or "symmetric", is also implied in the vector.

[0047] Abstract form encoding typically does not directly contain material, texture, or color information. These visual attributes can be assigned in subsequent rendering processes, thereby further decoupling form generation from appearance design and enhancing the flexibility of this solution in industrial processes.

[0048] For example, for a target model "quadruped robot", the obtained target structure information can include component identifiers such as torso, head and four legs, as well as the 3D bounding box parameters of each component and information such as the basic topology and style of each component.

[0049] By representing the target model with such target structural information, the structural information of the target model is macroscopically represented by structured data, which can overcome the traditional black-box processing method. This allows users to intuitively grasp intermediate states such as structural planning and component layout, providing a basis for subsequent parallel processing of components and user intervention and adjustment.

[0050] S104. Generate the corresponding three-dimensional mesh data for each component based on the target parameter information of each component.

[0051] Based on the obtained target structure information, the system enters the detail shaping stage, which can generate the three-dimensional mesh data corresponding to each component independently by refining the network of the components.

[0052] The system can use the target parameter information of each component as constraints to generate a high-precision 3D mesh with rich details and conforming to the design style within the spatial range. Each component can generate a corresponding 3D mesh, and the 3D mesh data records the information of the 3D mesh in the form of data.

[0053] In this embodiment, by independently generating the three-dimensional mesh data corresponding to each component, the target model can be effectively decomposed into independent parts, which facilitates subsequent independent modification and secondary creation of some components.

[0054] S105. Based on the target structure information and the corresponding 3D mesh data of each component, aggregate and generate the model data of the target model.

[0055] Based on the target structure information and the corresponding 3D mesh data of each component, the 3D meshes of each component can be aggregated to form a complete target model with topological independence between components, thereby obtaining the model data of the target model, which records the model information of the target model.

[0056] The obtained target model data can be imported into the game engine to load and display the target model in the game scene.

[0057] In summary, the 3D model generation method provided in this embodiment includes: responding to input instructions to determine the initial structural information of the target model; receiving adjustment instructions for the initial structural information input through an input device, updating the initial structural information according to the adjustment instructions to obtain target structural information; determining the target structural information of the target model; generating 3D mesh data corresponding to each component based on the target structural information; and generating model data of the target model based on the target structural information and the 3D mesh data corresponding to each component. This solution introduces the concept of structural information, converting the model into a user-understandable and editable intermediate structural representation, breaking the traditional invisible generation process. Users can perform editing operations on the model based on the structural information, thereby allowing user intervention between the two stages of generating initial structural information and generating 3D mesh data of components. Users can visually review and edit the generated initial structural information and adjust it. Through user intervention, the automatically generated initial structural information can be precisely adjusted to meet user wishes, thereby achieving controllability in model creation and improving the accuracy of the generated model. By generating the target structure information of the target model, an intermediate structure representation that can be understood and seen by the user can be obtained, laying the foundation for the user to achieve model controllability during the model generation process. In the process of generating the three-dimensional mesh data of the components, by generating the three-dimensional mesh data corresponding to each component separately, the components of the model can be decoupled, which facilitates the subsequent individual adjustment of the components and secondary creation, thereby improving the model creation efficiency and application value.

[0058] Figure 2 This is a flowchart illustrating another three-dimensional model generation method provided in this application embodiment; optionally, in step S102, receiving an adjustment instruction for the initial structure information input through an input device, updating the initial structure information according to the adjustment instruction, and obtaining the target structure information may include: S201. Generate and display the initial model based on the initial structure information.

[0059] In some embodiments, the initial structural information can be visualized to generate an initial model, which is then displayed to the user through a user interface. The initial model refers to a rough model generated based on the initial structural information.

[0060] S202. Receive adjustment instructions for the initial model input through the input device, update the initial structure information according to the adjustment data, and obtain the target structure information.

[0061] The user interface provides an interactive interface to the user, through which the user can be given adjustment instructions for the initial model. Based on the obtained adjustment data, the initial structural information can be updated to obtain the target structural information.

[0062] Of course, you can also use the above-mentioned methods to input adjustment commands via keyboard, mouse, etc.

[0063] This embodiment allows user intervention between the two stages of generating initial structural information and generating 3D mesh data of components. Users can visually review and edit the generated initial structural information and adjust it. Through user intervention, the automatically generated initial structural information can be precisely adjusted to meet the user's wishes, thereby achieving controllability in model creation and improving the accuracy of the generated model.

[0064] The reason why user intervention is possible is due to the introduction of the concept of structural information, which transforms the model into an intermediate structural representation that users can understand and edit. This breaks the traditional invisible generation process, allowing users to edit the model based on the structural information.

[0065] Figure 3 A flowchart illustrating another three-dimensional model generation method provided in this application embodiment; optionally, in the above steps, determining the initial structural information of the target model in response to an input command may include: S301. Parse the input command and obtain the input data.

[0066] The input data includes image information of the target model and / or textual description information of the target model.

[0067] In some embodiments, user input instructions can be received, parsed, and used to obtain input data. The input data may include image information of the target model, such as a conceptual diagram of the target model designed by the user; the input data may also include textual description information of the target model. Of course, the input data may contain only one of image information and textual description information, or it may contain both image information and textual description information.

[0068] Optionally, for the received image information, before feeding it into the encoder, normalization preprocessing can be performed, including scaling the image to a preset resolution required by the model, such as scaling it to 224. 224.

[0069] In addition, images can be cropped and padded to ensure consistent input dimensions.

[0070] Before the received text description information is fed into the encoder, a tokenizer can be used to convert the text information into a sequence of integers, with each integer corresponding to a token ID in the vocabulary.

[0071] For example, if the text description is "Hello world!", after word segmentation, we get "["Hello", "world","!"]". Both "Hello" and "world" are treated as tokens, and these tokens are then mapped to unique integer numbers, i.e., Token IDs. This mapping is stored in a dictionary called the vocabulary.

[0072] For example, a vocabulary list might look like this:

[0073] Therefore, the input text description "Hello world!" will be converted into "[100, 200, 300]", which is the Token ID sequence.

[0074] S302. Perform feature encoding processing on the input data to obtain global feature information.

[0075] Then, a feature encoder can be used to perform feature encoding on the preprocessed input data to obtain global feature information, which can be a global feature vector.

[0076] S303. Based on global feature information, a pre-trained layout prediction network is used to determine the initial structural information of the target model.

[0077] By taking global feature information as input and using a pre-trained layout prediction network, the global feature vector can be deterministically mapped to a structured, symbolic intermediate representation that is decoupled from upstream and downstream geometric generation, thus obtaining the initial structural information.

[0078] Figure 4 This is a flowchart illustrating another three-dimensional model generation method provided in an embodiment of this application; optionally, in step S302, feature encoding processing is performed on the input data to obtain global feature information, including: S401. A visual encoder is used to perform feature encoding on the image information of the target model to generate a global image feature vector.

[0079] Optionally, the visual encoder can employ a visual model pre-trained on a massive unlabeled image dataset through self-supervised learning. Self-supervised learning enables it to capture the general structural and textural features of objects, rather than just classification labels. By processing the image information through the visual encoder, a d-dimensional token can be obtained as the global image feature vector E. img .

[0080] S402. Use a text encoder to perform feature encoding on the text description information of the target model to generate a global text feature vector.

[0081] A text encoder can be trained using a multimodal model paired with a visual encoder. It takes as input an integer sequence of segmented text description information and outputs a global text feature vector E, also d-dimensional. text .

[0082] Paired training refers to training a multimodal model using paired datasets. That is, each training data point contains an image and its corresponding text description. For example, a photo of a dog. During training, the model sees both the image and the text simultaneously, and its goal is to learn to map the image content and text semantics to similar locations in the same high-dimensional feature space. In this way, the model learns to understand the relationship between images and text. Therefore, the "pairing" here refers to the "image-text" pairing characteristic of the training data itself.

[0083] S403. Based on the global image feature vector and the global text feature vector, perform multimodal feature fusion processing to obtain global feature information.

[0084] To achieve robust processing of image and text input, this embodiment employs an attention-based fusion module. This module can be structured as a small decoder layer.

[0085] The global image feature vector E obtained above can be used as a reference. img and global text feature vector E text As key and value inputs, a learnable query vector Q is also introduced. fusion By calculating Attention(Q) fusion , [E img E text ],[E img E text This yields a final global semantic feature vector C that integrates information from both modalities. global This refers to the global feature information mentioned above.

[0086] This method outperforms simple concatenation or weighted averaging because it allows the model to dynamically determine the weights of the two modalities based on the context. The global feature information remains d-dimensional.

[0087] Optionally, in step S303, determining the initial structural information of the target model using a pre-trained layout prediction network based on global feature information may include: inputting global feature information into the layout prediction network, and having the layout prediction network determine the initial structural information of the target model through a preset iterative denoising process.

[0088] In some embodiments, the generated global feature information can be input into the layout prediction network, which then processes and analyzes the information to predict the initial structural information of the target model.

[0089] In the layout prediction network training phase, massive amounts of pre-annotated 3D model data with component structures (i.e., the bounding box and category of each component) can be used as training data. By learning from this data, the network learns which components typically make up a specific concept (such as "car", "chair", "airplane") and the typical spatial arrangement relationships between these components. For example, it learns that a "car" typically has one large "body" component and four small "wheel" components, and that the "wheels" are usually symmetrically distributed on both sides of the lower part of the "body".

[0090] In practical use, the user's input "a car" is transformed into global feature information by a feature encoder. This global feature information acts like a "general instruction" to the network. When generating a layout, the layout prediction network uses this global feature information as a strong constraint, searches the "knowledge base" for the most likely structural layout that best matches the semantics of "car," and outputs it as the initial structural information.

[0091] Optionally, the layout prediction network determines the initial structural information of the target model through a preset iterative denoising process, including: the layout prediction network determines the noise data of the current time step based on the denoised features of the current time step and global feature information; the initial denoised features are obtained by sampling random noise; based on the denoised features of the current time step and the noise data of the current time step, the denoised features of the current time step are denoised to obtain the denoised features of the next time step, and the process is repeated until the time step ends, and the features at the end of the time step are used as the initial structural information of the target model.

[0092] In this embodiment, the layout prediction network can be implemented based on the conditional diffusion model. The main structure of the layout prediction network adopts the conditional diffusion architecture, and the traditional U-Net backbone network in the diffusion model is replaced with a series of encoder / decoder blocks. This design is more suitable for processing serialized, variable-length outputs.

[0093] The initial structural information we need S= It is considered as a sequence of length N, where each component From its boundary volume parameter information "B i "and abstract morphological encoding "L i "Composition. Before being sent into the network, each..." Flattened and concatenated into a "dp"-dimensional vector, where "dp" is composed of "B i The dimensions of "" (e.g., 7-dimensional, containing 3 position coordinates and 4 rotation quaternions) and "L" i The vectors of all N components are concatenated to form a complete vector "X0" that represents the clean initial structural information.

[0094] Conditional injection in layout prediction networks: Global Feature Information: Through cross-attention injection, in each encoding / decoding block, in addition to the standard self-attention layer, a cross-attention layer is added. Its queries come from the intermediate representation of the component, while the keys and values ​​come from the global feature information.

[0095] Time step condition: The time step t (from T to 0) in the diffusion process is encoded into a time embedding E by a sine wave. t It is incorporated into the encoding / decoding block in an additive or adaptive layer normalization manner.

[0096] Forward diffusion (built during training): Defines a forward diffusion method from real data. (This can refer to the actual structural information used when training the network) to pure noise A Markov chain. At any time step t, we have:

[0097] in, It is a preset noise variance scheduling table, from which it can be derived Closed-form solution:

[0098] in, , yes The accumulation.

[0099] Inverse denoising (inference-time generation): Training a neural network To predict the noise applied at time step t The training objective is to minimize:

[0100] Sampling generation: from random Gaussian noise Initially, using a sampler, iterate T times, gradually from... Calculate until clean That is, the predicted initial structural information. Finally, the layout prediction network outputs a clear initial structural information containing N effective components.

[0101] The essence of generating initial structural information through layout prediction networks is to deterministically map unstructured global feature information onto a structured, symbolic intermediate representation that is decoupled from downstream geometric generation.

[0102] Figure 5 This is a flowchart illustrating another method for generating a three-dimensional model provided in an embodiment of this application; optionally, in step S102, receiving adjustment instructions for the initial model input through an input device, updating the initial structural information based on the adjustment data, and obtaining target structural information includes: S501: Receive operation instructions for the target component in the initial model and obtain operation data.

[0103] In some embodiments, the initial model can be shown to the user by visualizing the initial structural information. The user can adjust the initial model by moving, scaling, or rotating the bounding box of a component.

[0104] Typically, in memory, the initial structural information is maintained as a data structure that can be dynamically added, deleted, modified, and retrieved, such as a hash table or a dynamic array, where each element... Each component has a unique identifier.

[0105] The system maintains an operation history stack, recording every editing operation performed by the user, so as to facilitate undoing and redoing.

[0106] When a user interacts with a target component in the initial model on the interface, such as by dragging, the front end calculates the transformation matrix in the world coordinate system through ray casting and 3D transformation. , and obtain the operation data.

[0107] S502. Based on the identifier of the target component, read the initial three-dimensional boundary volume information of the target component.

[0108] Based on the identifier of the target component being operated on, the backend can read the current 3D boundary volume information of the target component from the data structure. Here, it is assumed that the current 3D boundary volume information is the same as the initial 3D boundary volume information B. i .

[0109] S503. Based on the initial three-dimensional boundary volume information of the target component and the operation data, obtain the target three-dimensional boundary volume information of the target component.

[0110] The backend can first process the initial 3D boundary volume information B i Convert to a matrix Then, based on the received operation data sum matrix The target component's three-dimensional boundary volume information is calculated: = .

[0111] S504. Update the three-dimensional boundary volume information of the target component to the initial structural information of the target model to obtain the target structural information.

[0112] Finally, the target's three-dimensional boundary volume information The data is decomposed into new translation, rotation, and scaling parameters, and the initial 3D boundary volume information of the target component stored in the data structure is updated.

[0113] Optionally, this method also includes updating the initial model based on the target three-dimensional boundary volume information of the target component.

[0114] In some embodiments, a lightweight rendering engine can be used to build the interactive window. Whenever the initial parameter information of a component in the initial structural information is detected to be modified, the system can update only the transformation properties of the rendering object corresponding to the modified component, thereby updating the initial model instead of reloading the entire scene, thus achieving a smoother screen update.

[0115] Figure 6 A flowchart illustrating another three-dimensional model generation method provided in this application embodiment; optionally, in step S102, receiving adjustment instructions for the initial model input through an input device, updating the initial structural information based on the adjustment data, and obtaining target structural information includes: S601, Receive component addition operation for the initial model, obtain the addition position information corresponding to the component to be added and the abstract form code specified for the component to be added.

[0116] In some embodiments, users can also perform operations such as adding or deleting components. When a user adds a new component's bounding box to the initial model, the system does not know what this new bounding box refers to. At this time, the user needs to select a specified abstract form code for the component to be added from the component library.

[0117] In other words, users can first specify the location where the component will be added, and then select the specified abstract form code from the component library.

[0118] S602. Update the initial structure information with the addition position information of the component to be added and the specified abstract form code as the target parameter information of the component to be added, and obtain the target structure information.

[0119] The system can add the user-specified abstract morphology code as the abstract morphology code of the component to be added to the initial structural information of the target model. At the same time, the addition position information of the component to be added is also added to the initial structural information to update the initial structural information.

[0120] For example, if a user selects the abstract form code for "external storage box", it is equivalent to the user telling the system, "Please generate a component for me at location A, according to the style of 'external storage box'".

[0121] Figure 7 A flowchart illustrating another three-dimensional model generation method provided in this application embodiment; optionally, in step S102, receiving adjustment instructions for the initial model input through an input device, updating the initial structural information based on the adjustment data, and obtaining target structural information includes: S701: Receive confirmation operation for the target component in the initial model and obtain the initial abstract form code of the target component.

[0122] Similarly, users can also adjust the initial abstract form code of a component. Based on the displayed initial model, users can select a target component and then read its initial abstract form code according to its identifier.

[0123] S702: Receive confirmation operation for the abstract form code of a specified component in the component library, replace the initial abstract form code of the target component with the abstract form code of the specified component, and obtain the abstract form code of the target component.

[0124] The component library stores target parameter information for various pre-generated components.

[0125] Since the initial abstract form code of the target component is generated by the layout prediction network, if the user is not satisfied, they can select an abstract form code from the component library that is similar to the target component but more complex or refined to update the initial abstract form code of the target component.

[0126] For example, if a user feels that the "hub" component in the generated initial model is too simple or too rough, they can select the "hub" component and then select a more complex "heavy-duty engineering hub" abstract form code from the component library to update the initial abstract form code of the "hub" component.

[0127] Optionally, in step S102, the three-dimensional mesh data corresponding to each component is generated according to the target parameter information of each component, including: according to the target parameter information of each component, the component refinement network corresponding to each component is used to generate the three-dimensional mesh data corresponding to each component in parallel.

[0128] After obtaining the target structure information of the target model, the system enters the detail shaping stage. The target parameter information of each component in the target model can be distributed to the independent component refinement network corresponding to each component. Each component refinement network is like a focused digital craftsman, receiving the target parameter information of the component it is assigned to, and using this as a constraint, generating a high-precision 3D mesh of the component with rich details that conforms to the design style within the corresponding spatial range, thus obtaining the 3D mesh data of the component.

[0129] It is worth noting that each component corresponds to an independent component refinement network, and all component refinement networks run in parallel. As a result, the refinement process of all components is executed in parallel, which greatly improves the efficiency of generating model data for the target model.

[0130] Since each component refinement network only processes one component, the learning difficulty of each component refinement network is significantly reduced, improving the geometric quality and structural rationality of the final generated target model.

[0131] Optionally, based on the target parameter information of each component, a component refinement network corresponding to each component is used to generate the three-dimensional mesh data corresponding to each component in parallel. This includes: inputting the target parameter information of each component and the global feature information corresponding to the input command into the component refinement network corresponding to each component, and having each component refinement network generate the three-dimensional mesh data corresponding to each component in parallel based on the target parameter information and global feature information of the corresponding component.

[0132] The component refinement network can generate 3D meshes for each component based on the previously generated global feature information and the target structure information of the target model, thus obtaining 3D mesh data.

[0133] The training process of the component refinement network involves learning how to generate a corresponding 3D mesh within a given bounding box under the dual guidance of global feature information and target structure information. In this way, the details and identity of the components are preserved, while the overall style remains harmonious and unified.

[0134] Figure 8 This is a flowchart illustrating another 3D model generation method provided in an embodiment of this application; optionally, each component refinement network generates 3D mesh data corresponding to each component in parallel based on the target parameter information and global feature information of the corresponding component, including: S801. The first component refinement network determines the bounding box information of the first component in three-dimensional space based on the three-dimensional boundary volume information of the first component.

[0135] This embodiment takes the generation of 3D mesh data for a first component as an example. The first component can be any component that constitutes the target model, and the refinement network for the first component is an independent refinement network corresponding to the first component.

[0136] The target parameter information of the first component in the target structure information of the target model can be distributed to the first component refinement network. The first component refinement network can determine the bounding box information of the first component in three-dimensional space based on the three-dimensional boundary volume information of the first component. The bounding box information here includes the position, size, orientation and other information of the bounding box.

[0137] S802. The first component refinement network generates the three-dimensional mesh data corresponding to the first component within the range indicated by the bounding box information, based on the abstract morphological encoding of the first component and global feature information.

[0138] Based on the bounding box information of the first component in three-dimensional space, the local spatial range of the first component in three-dimensional space can be determined. Using this as a constraint, and based on the global feature information of the target model and the abstract morphological encoding of the first component, a high-precision three-dimensional mesh of the component with rich details and conforming to the design style can be generated within the constrained local spatial range.

[0139] Optionally, in step S802, generating three-dimensional mesh data corresponding to the first component within the range indicated by the bounding box information includes: generating a symbolic distance field data for characterizing the geometry of the first component within the range indicated by the bounding box information; and extracting and generating three-dimensional mesh data corresponding to the first component based on the symbolic distance field data using a preset mesh extraction algorithm.

[0140] To obtain high-quality geometric details, instead of directly generating the 3D mesh data for the corresponding parts, an implicit representation of the signed distance field is first generated, and then the final 3D mesh data is extracted using algorithms such as Marching Cubes.

[0141] The component refinement network is also implemented through a conditional diffusion model, which is trained to generate 3D geometry, typically using implicit representations to achieve higher quality.

[0142] Geometric representation: Using the Signed Distance Field (SDF), which is a function It returns the shortest distance (signed, negative inside and positive outside) from any point in space to the surface of the object. The model predicts a discretized SDF mesh.

[0143] Implementation details of conditional injection: Abstract morphological encoding L iAs a global condition, it is injected into the SDF prediction network through cross-attention or AdaLNd, which affects the overall shape of the generated SDF and determines whether to generate a "wheel" or a "seat".

[0144] Application of boundary information constraints: 3D boundary volume information B i The effect is manifested in the post-processing stage. The network always generates an SDF mesh within a canonical cube of [-1,1]. During aggregation, it is only in this canonical space that the B-type mesh is generated. i The defined transformation is mapped to world space.

[0145] Conversion from SDF to Mesh: After the diffusion process generates a clean SDF mesh, a preset algorithm can be used to extract isosurfaces, i.e., surfaces with SDF=0, thus generating a triangular mesh with vertices and faces. The generated mesh is further processed through post-processing operations such as smoothing and simplification to optimize its topology and face count, meeting the performance requirements of the game engine.

[0146] The above methods can significantly improve the quality of the generated 3D mesh data.

[0147] Optionally, based on the target parameter information of each component, a component refinement network corresponding to each component is used to generate the three-dimensional mesh data corresponding to each component in parallel, including: if the initial structural information of the target component in the initial model is detected to be updated, the component refinement network corresponding to the target component is triggered, and the three-dimensional mesh data corresponding to the target component is regenerated based on the updated target parameter information of the target component.

[0148] In some embodiments, when a user submits modifications to the parameter information of a target component according to the above embodiments, the system will detect the data before and after the modification. When it detects that the parameter information (3D boundary volume information or abstract morphological encoding) of the target component has changed, it will trigger the component refinement network corresponding to the target component to regenerate the 3D mesh data of the target component based on the modified target parameter information. For components whose parameters have not changed, the system can directly retrieve the previously generated 3D mesh data from the cache.

[0149] This triggering mechanism avoids a large number of repeated generation processes, greatly improving the iterative efficiency of design modifications.

[0150] By distributing the component refinement task to multiple independent component refinement networks, highly parallel processing is achieved, significantly improving the generation efficiency of complex models. More innovatively, this architecture supports the independent regeneration, replacement, or deformation of any component without requiring costly recalculation of the entire model. When a user modifies only a local design, the system only needs to update the corresponding component. This "on-demand update" feature provides crucial technical support for agile development and rapid iterative design processes, representing an engineering advantage unmatched by existing end-to-end generation methods.

[0151] Figure 9 This is a flowchart illustrating another method for generating a 3D model provided in an embodiment of this application; optionally, in step S103, model data of the target model is aggregated and generated based on the target structure information and the 3D mesh data corresponding to each component, including: S901. Based on the target three-dimensional boundary volume information of each component, determine the transformation matrix information of each component from local space to world space.

[0152] This embodiment is the final assembly stage, responsible for integrating the generated scattered parts into a structured, deliverable product, that is, obtaining the model data of the target model.

[0153] Optionally, an automated script or service can be used to iterate through the results of all completed component refinement tasks, with each task result corresponding to the 3D mesh data of a component, that is, the local spatial mesh data of the component.

[0154] S902. Based on the transformation matrix information of each component from local space to world space and the corresponding three-dimensional mesh data of each component, determine the target three-dimensional mesh data of each component in world space.

[0155] For each component, its corresponding 3D boundary volume information can be read.

[0156] Using 3D boundary volume information, a 4 4 affine transformation matrix M i , affine transformation matrix M i Applying the three-dimensional mesh data of a component to each vertex yields the target three-dimensional mesh data of the component in world space.

[0157] That is, the generated 3D mesh data of the component in local space is converted to world space.

[0158] S903. Based on the target 3D mesh data of each component in world space and the hierarchical relationship between each component, aggregate and generate model data of the target model; the model data is used to import into the game engine to load the target model.

[0159] A 3D file format library that supports scene graphs can be used to create a root node within the file. For each part, a new child node can be created, attaching the part's target 3D mesh data in world space to the child node, and naming the child node according to the part's identifier or semantics. All child nodes corresponding to parts are direct children of the root node, ensuring they are parallel and independent in hierarchy.

[0160] Ultimately, a single, but internally structured model data is generated, which can be used in game engines or animation software to be correctly parsed, and its internal component hierarchy will be fully preserved.

[0161] Figure 10 This is a schematic diagram of the core process of a three-dimensional model generation method provided in the embodiments of this application. The specific implementation of each step has been described in detail in the above embodiments and will not be repeated here.

[0162] Figure 11 This is a schematic diagram of the core process of another three-dimensional model generation method provided in the embodiments of this application. Figure 11 The focus is on demonstrating the relevant processes for user intervention in adjusting structural information, and the specific implementation methods will not be elaborated further.

[0163] Figure 11 This demonstrates a closed-loop interaction process: 1. Initial structural information is first rendered into a user-visible 3D bounding box by the visualization rendering engine; 2. The user interacts with the data in the 3D interactive window, and these actions are captured by the interaction event listener; 3. The listener parses the user's low-level input (such as user actions) into high-level, meaningful operation instructions; 4. The operation instructions are sent to the structural information operation module, which is responsible for safely adding, deleting, and modifying data structures stored in memory; 5. The modified data immediately triggers a re-rendering by the visualization engine, and the user can instantly see the effects of their changes in the window; 6. Only after the user completes all editing and confirms it is the final target structural information passed to the component refinement network. This closed loop ensures intuitive operation and real-time feedback.

[0164] The following is a specific example to illustrate the overall solution: Step 1 - Set Global Style: The artist provides a reference image with a specified style (e.g., rugged, modular, dark gray paint with yellow warning stripes) and enters the text "a heavy six-wheeled mining truck, modular design". The system then integrates and processes this information into global feature information.

[0165] Step 2 - Preliminary Planning: Global feature information is input into the layout prediction network, which generates initial structural information containing 8 core components. The initial structural information includes: 1 cockpit, 1 large cargo box, and 6 wheel modules. Each component contains its bounding box parameters and abstract morphological encoding in 3D space.

[0166] Step 3 - Human-Machine Collaboration: The system displayed the eight semi-transparent bounding boxes in a 3D view. After reviewing them, the artist felt that the cargo boxes planned by the layout prediction network were too small and did not conform to the "heavy mining" setting. He selected the bounding box of the cargo box and stretched it a certain length. He also wanted to add some functional modules, so he manually added two new, smaller bounding boxes on both sides of the vehicle body and assigned them abstract form codes for external storage boxes from the system's component library. After adjustments, a target structure containing 10 components, conforming to the design intent, was generated.

[0167] Step 4 - Parallel Generation of 3D Meshes and On-Demand Modification: The system distributes the target parameter information corresponding to each component in the generated target structure information to the corresponding component refinement network. Each component refinement network executes in parallel to generate the corresponding 3D mesh data. If the artist feels that the generated wheel hub style is too simple, they can select six wheel components, assign them a more complex abstract form code of "heavy-duty engineering hub," and then regenerate the 3D mesh data for the six wheel components, while leaving the other components unchanged.

[0168] Step 5 - Final Delivery: The model aggregator assembles the 3D meshes of all 10 final confirmed parts and outputs a model data. This model data is directly imported into the game engine. Game designers can easily design vehicle modification systems for players based on these individual parts. Players can change different styles of tires, cargo boxes, or add different storage boxes. Technical artists can also easily set independent health and destruction effects for each part.

[0169] Ultimate Success: 1. Highly Controllable: The final model perfectly matches the designer's overall layout intentions; 2. Rapid Iteration: Modifications to local designs are reduced from hours of manual modeling to minutes of "on-demand regeneration"; 3. Native Support for Gameplay Design: The generated model naturally supports modularity and custom gameplay, greatly empowering the downstream game design process.

[0170] In summary, the 3D model generation method provided in this embodiment includes: responding to input instructions to determine the initial structural information of the target model; receiving adjustment instructions for the initial structural information input through an input device, updating the initial structural information according to the adjustment instructions to obtain target structural information; determining the target structural information of the target model; generating 3D mesh data corresponding to each component based on the target structural information; and generating model data of the target model based on the target structural information and the 3D mesh data corresponding to each component. This solution introduces the concept of structural information, converting the model into a user-understandable and editable intermediate structural representation, breaking the traditional invisible generation process. Users can perform editing operations on the model based on the structural information, thereby allowing user intervention between the two stages of generating initial structural information and generating 3D mesh data of components. Users can visually review and edit the generated initial structural information and adjust it. Through user intervention, the automatically generated initial structural information can be precisely adjusted to meet user wishes, thereby achieving controllability in model creation and improving the accuracy of the generated model. By generating the target structure information of the target model, an intermediate structure representation that can be understood and seen by the user can be obtained, laying the foundation for the user to achieve model controllability during the model generation process. In the process of generating the three-dimensional mesh data of the components, by generating the three-dimensional mesh data corresponding to each component separately, the components of the model can be decoupled, which facilitates the subsequent individual adjustment of the components and secondary creation, thereby improving the model creation efficiency and application value.

[0171] Secondly, by decoupling the 3D mesh generation tasks for each component and processing each component using an independent component refinement network, the learning difficulty of each component refinement network can be significantly reduced, improving the geometric quality and structural rationality of the final model. Furthermore, by executing the component refinement network tasks in parallel, the generation efficiency of complex models can be greatly improved.

[0172] In addition, based on the independent generation of each component, it can support independent adjustment of any component without the need for expensive recalculation of the entire model. This on-demand update feature provides key technical support for agile development and rapid iterative design processes.

[0173] The following describes the apparatus, device, and storage medium used to execute the three-dimensional model generation method provided in this application. The specific implementation process and technical effects are described above and will not be repeated below.

[0174] Figure 12This is a schematic diagram of a three-dimensional model generation device provided in an embodiment of this application. The functions implemented by this three-dimensional model generation device correspond to the steps performed by the above-described method. This device can be understood as the aforementioned server, or the server's processor, or as a component independent of the aforementioned server or processor that implements the functions of this application under the control of the server, such as... Figure 12 As shown, the device may include: a determining module 120 and a generating module 121; The determination module 120 is used to respond to input instructions and determine the initial structural information of the target model. The initial structural information includes the identifiers of each component that constitutes the target model and the initial parameter information of each component. The generation module 121 is used to receive adjustment instructions for the initial structure information input through the input device, update the initial structure information according to the adjustment instructions, and obtain the target structure information; The determination module 120 is used to determine the target structure information of the target model; the target structure information includes: target parameter information of each component; the target parameter information includes: target 3D boundary volume information of the component and abstract morphological encoding; the target 3D boundary volume information includes the 3D bounding box parameters of the component, which are used to indicate the size, position and orientation of the component in 3D space; the abstract morphological encoding is used to indicate the category, morphological attributes and style attributes of the component; The generation module 121 is used to generate the three-dimensional mesh data corresponding to each component based on the target parameter information of each component. The generation module 121 is used to aggregate and generate model data of the target model based on the target structure information and the three-dimensional mesh data corresponding to each component.

[0175] Optionally, the determining module 120 is specifically used to determine the initial structural information of the target model in response to input instructions; Receive adjustment instructions for the initial structure information, update the initial structure information according to the adjustment instructions, and obtain the target structure information.

[0176] Optionally, the generation module 120 is specifically used to generate and display the initial model based on the initial structural information; It receives adjustment instructions for the initial model input through the input device, updates the initial structural information based on the adjustment data, and obtains the target structural information. Optionally, the determining module 120 is specifically used to parse the input instructions and obtain the input data; the input data includes the image information of the target model and / or the text description information of the target model; The input data is processed by feature encoding to obtain global feature information; Based on global feature information, a pre-trained layout prediction network is used to determine the initial structural information of the target model.

[0177] Optionally, the determining module 120 is specifically used to perform feature encoding processing on the image information of the target model using a visual encoder to generate a global image feature vector; A text encoder is used to perform feature encoding on the text description information of the target model to generate a global text feature vector; Based on the global image feature vector and the global text feature vector, multimodal feature fusion processing is performed to obtain global feature information.

[0178] Optionally, the determination module 120 is specifically used to input global feature information into the layout prediction network, and the layout prediction network determines the initial structural information of the target model through a preset iterative denoising process.

[0179] Optionally, the determination module 120 is specifically used by the layout prediction network to determine the noise data of the current time step based on the noise features of the current time step and global feature information; the initial noise features are obtained by sampling random noise. Based on the noise features and noise data of the current time step, the noise features of the current time step are denoised to obtain the noise features of the next time step. This process is repeated until the time step ends, and the features at the end of the time step are used as the initial structural information of the target model.

[0180] Optionally, the determining module 120 is specifically used to receive operation instructions for the target component in the initial model and to obtain operation data; Based on the identifier of the target component, read the initial three-dimensional boundary volume information of the target component; Based on the initial three-dimensional boundary volume information of the target component and the operation data, the target three-dimensional boundary volume information of the target component is obtained; The three-dimensional boundary volume information of the target component is updated into the initial structural information of the target model to obtain the target structural information.

[0181] Optionally, it also includes: an update module; The update module is used to update the initial model based on the target 3D boundary volume information of the target component.

[0182] Optionally, the determination module 120 is specifically used to receive the component addition operation for the initial model, obtain the addition position information corresponding to the component to be added, and the abstract form code specified for the component to be added; The addition location information of the component to be added and the specified abstract form code are used as the target parameter information of the component to be added to update the initial structure information, thus obtaining the target structure information.

[0183] Optionally, the determining module 120 is specifically used to receive a confirmation operation for the target component in the initial model and obtain the initial abstract form code of the target component; The system receives a confirmation operation for the abstract form code of a specified component in the component library, replaces the initial abstract form code of the target component with the abstract form code of the specified component, and obtains the abstract form code of the target component. The component library stores target parameter information of various pre-generated components.

[0184] Optionally, the generation module 121 is specifically used to generate three-dimensional mesh data corresponding to each component in parallel by using the component refinement network corresponding to each component based on the target parameter information of each component.

[0185] Optionally, the generation module 121 is specifically used to input the target parameter information of each component and the global feature information corresponding to the input command into the component refinement network corresponding to each component, and the component refinement network generates the three-dimensional mesh data corresponding to each component in parallel according to the target parameter information and global feature information of the corresponding component.

[0186] Optionally, the generation module 121 is specifically used to determine the bounding box information of the first component in three-dimensional space by the first component refinement network based on the three-dimensional boundary volume information of the first component; The first component refinement network generates the 3D mesh data corresponding to the first component within the range indicated by the bounding box information, based on the abstract morphological encoding of the first component and global feature information.

[0187] Optionally, the generation module 121 is specifically used to generate a symbolic distance field data for characterizing the geometry of the first component within the range indicated by the bounding box information; and to extract and generate the three-dimensional mesh data corresponding to the first component based on the symbolic distance field data using a preset mesh extraction algorithm.

[0188] Optionally, the generation module 121 is specifically used to trigger the component refinement network corresponding to the target component if it is detected that the initial structural information of the target component in the initial model has been updated, and to regenerate the three-dimensional mesh data corresponding to the target component based on the updated target parameter information of the target component.

[0189] Optionally, the generation module 121 is specifically used to determine the transformation matrix information of each component from local space to world space based on the target three-dimensional boundary volume information of each component; Based on the transformation matrix information of each component from local space to world space and the corresponding three-dimensional mesh data of each component, the target three-dimensional mesh data of each component in world space is determined; Based on the target 3D mesh data of each component in world space and the hierarchical relationship between the components, the model data of the target model is aggregated and generated; the model data is used to import into the game engine to load the target model.

[0190] These modules can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), one or more digital signal processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs). Alternatively, when a module is implemented using processing element scheduler code, the processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. Furthermore, these modules can be integrated together as a system-on-a-chip (SOC).

[0191] The modules described above can be connected or communicate with each other via wired or wireless connections. Wired connections can include metal cables, optical fibers, hybrid cables, or any combination thereof. Wireless connections can include connections via LAN, WAN, Bluetooth, ZigBee, or NFC, or any combination thereof. Two or more modules can be combined into a single module, and any module can be divided into two or more units. Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the method embodiments, and will not be repeated here.

[0192] Figure 13 A schematic diagram of an electronic device provided in this application embodiment includes: a processor 801, a storage medium 802, and a bus 803. The storage medium 802 stores machine-readable instructions executable by the processor 801. When the electronic device runs a three-dimensional model generation method as described in the embodiment, the processor 801 communicates with the storage medium 802 via the bus 803. The processor 801 executes the machine-readable instructions to perform the following steps: In response to input commands, determine the initial structural information of the target model. The initial structural information includes the identifiers of each component that makes up the target model and the initial parameter information of each component. Receive adjustment instructions for the initial structure information input through the input device, update the initial structure information according to the adjustment instructions, and obtain the target structure information; Determine the target structure information of the target model; the target structure information includes: target parameter information of each component; the target parameter information includes: target 3D boundary volume information of the component and abstract morphological encoding; the target 3D boundary volume information includes the 3D bounding box parameters of the component, which are used to indicate the size, position and orientation of the component in 3D space; the abstract morphological encoding is used to indicate the category, morphological attributes and style attributes of the component; Based on the target parameter information of each component, generate the corresponding three-dimensional mesh data for each component; Based on the target structure information and the corresponding 3D mesh data of each component, the model data of the target model is aggregated and generated.

[0193] In one feasible implementation, when the processor 801 executes the adjustment instruction for the initial structure information received through the input device, updates the initial structure information according to the adjustment instruction, and obtains the target structure information, it is specifically used to: generate and display the initial model according to the initial structure information; It receives adjustment instructions for the initial model input through the input device, updates the initial structural information based on the adjustment data, and obtains the target structural information.

[0194] In a feasible implementation, when the processor 801 executes the response input instruction to determine the initial structural information of the target model, it is specifically used to: perform feature encoding processing on the input data to obtain global feature information; Based on global feature information, a pre-trained layout prediction network is used to determine the initial structural information of the target model.

[0195] In one feasible implementation, when the processor 801 performs feature encoding processing on the input data to obtain global feature information, it is specifically used to: perform feature encoding processing on the image information of the target model using a visual encoder to generate a global image feature vector. A text encoder is used to perform feature encoding on the text description information of the target model to generate a global text feature vector; Based on the global image feature vector and the global text feature vector, multimodal feature fusion processing is performed to obtain global feature information.

[0196] In a feasible implementation, when the processor 801 executes the process of determining the initial structural information of the target model using a pre-trained layout prediction network based on global feature information, it specifically performs the following: inputting global feature information into the layout prediction network, and having the layout prediction network determine the initial structural information of the target model through a preset iterative denoising process.

[0197] In a feasible implementation, when the processor 801 performs the initial structural information of the target model by the layout prediction network through a preset iterative denoising process, it is specifically used to: determine the noise data of the current time step by the layout prediction network based on the noise features of the current time step and global feature information; the initial noise features are obtained by sampling random noise; Based on the noise features and noise data of the current time step, the noise features of the current time step are denoised to obtain the noise features of the next time step. This process is repeated until the time step ends, and the features at the end of the time step are used as the initial structural information of the target model.

[0198] In one feasible implementation, when the processor 801 executes the following steps: receiving adjustment instructions for the initial model input via an input device, updating the initial structure information based on the adjustment data, and obtaining the target structure information, the processor 801 is specifically used to: receive operation instructions for the target component in the initial model and obtain operation data. Based on the identifier of the target component, read the initial three-dimensional boundary volume information of the target component; Based on the initial three-dimensional boundary volume information of the target component and the operation data, the target three-dimensional boundary volume information of the target component is obtained; The three-dimensional boundary volume information of the target component is updated into the initial structural information of the target model to obtain the target structural information.

[0199] In one feasible implementation, the processor 801 is also used to perform: updating the initial model based on the target three-dimensional boundary volume information of the target component.

[0200] In a feasible implementation, when the processor 801 executes the following: receiving adjustment instructions for the initial model input through the input device, updating the initial structure information according to the adjustment data, and obtaining the target structure information, the processor 801 is specifically used to: receive component addition operations for the initial model, obtain the addition position information corresponding to the component to be added, and the abstract form code specified for the component to be added. The addition location information of the component to be added and the specified abstract form code are used as the target parameter information of the component to be added to update the initial structure information, thus obtaining the target structure information.

[0201] In one feasible implementation, when the processor 801 executes the following: receiving user adjustment instructions for the initial model input through the input device, updating the initial structure information according to the adjustment data, and obtaining the target structure information, the processor 801 is specifically used to: receive a confirmation operation for the target component in the initial model and obtain the initial abstract form code of the target component. The system receives a confirmation operation for the abstract form code of a specified component in the component library, replaces the initial abstract form code of the target component with the abstract form code of the specified component, and obtains the abstract form code of the target component. The component library stores target parameter information of various pre-generated components.

[0202] In a feasible implementation, when the processor 801 generates the corresponding three-dimensional mesh data for each component based on the target parameter information of each component, it specifically performs the following: based on the target parameter information of each component, it uses the component refinement network corresponding to each component to generate the corresponding three-dimensional mesh data for each component in parallel.

[0203] In a feasible implementation, when the processor 801 executes the parallel generation of three-dimensional mesh data corresponding to each component using the component refinement network corresponding to each component based on the target parameter information of each component, it specifically performs the following: inputting the target parameter information of each component and the global feature information corresponding to the input instruction into the component refinement network corresponding to each component, and having each component refinement network generate the three-dimensional mesh data corresponding to each component in parallel based on the target parameter information and global feature information of the corresponding component.

[0204] In a feasible implementation, when the processor 801 executes the parallel generation of three-dimensional mesh data corresponding to each component by the refinement network of each component based on the target parameter information and global feature information of the corresponding component, it is specifically used to: determine the bounding box information of the first component in three-dimensional space by the first component refinement network based on the three-dimensional boundary volume information of the first component. The first component refinement network generates the 3D mesh data corresponding to the first component within the range indicated by the bounding box information, based on the abstract morphological encoding of the first component and global feature information.

[0205] In one feasible implementation, when the processor 801 generates three-dimensional mesh data corresponding to the first component within the range indicated by the bounding box information, it is specifically used to generate a symbolic distance field data for characterizing the geometry of the first component within the range indicated by the bounding box information. Based on the symbolic distance field data, a preset mesh extraction algorithm is used to extract and generate the three-dimensional mesh data corresponding to the first component.

[0206] In a feasible implementation, when the processor 801 executes the parallel generation of three-dimensional mesh data corresponding to each component by using the component refinement network corresponding to each component based on the target parameter information of each component, it specifically performs the following: if it detects that the initial structural information of the target component in the initial model has been updated, it triggers the component refinement network corresponding to the target component to regenerate the three-dimensional mesh data corresponding to the target component based on the updated target parameter information of the target component.

[0207] In a feasible implementation, when the processor 801 executes the model data of the target model by aggregating and generating the target model based on the target structure information and the three-dimensional mesh data corresponding to each component, it is specifically used to: determine the transformation matrix information of each component from local space to world space based on the target three-dimensional boundary volume information of each component. Based on the transformation matrix information of each component from local space to world space and the corresponding three-dimensional mesh data of each component, the target three-dimensional mesh data of each component in world space is determined; Based on the target 3D mesh data of each component in world space and the hierarchical relationship between the components, the model data of the target model is aggregated and generated; the model data is used to import into the game engine to load the target model.

[0208] The storage medium 802 stores program code, which, when executed by the processor 801, causes the processor 801 to perform various steps in the three-dimensional model generation method according to various exemplary embodiments of this application as described in the "Exemplary Methods" section above.

[0209] The processor 801 can be a general-purpose processor, such as a central processing unit (CPU), digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.

[0210] Storage medium 802, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. The memory can include at least one type of storage medium, such as flash memory, hard disk, multimedia card, card-type memory, random access memory (RAM), static random access memory (SRAM), programmable read-only memory (PROM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), magnetic memory, magnetic disk, optical disk, etc. Memory is any other medium capable of carrying or storing desired program code in the form of instructions or data structures that can be accessed by a computer, but is not limited thereto. Storage medium 802 in the embodiments of this application can also be a circuit or any other device capable of implementing storage functions for storing program instructions and / or data.

[0211] Optionally, embodiments of this application also provide a computer-readable storage medium storing a computer program, which is executed by a processor, and the processor performs the following steps: In response to input commands, determine the initial structural information of the target model. The initial structural information includes the identifiers of each component that makes up the target model and the initial parameter information of each component. Receive adjustment instructions for the initial structure information input through the input device, update the initial structure information according to the adjustment instructions, and obtain the target structure information; Determine the target structure information of the target model; the target structure information includes: target parameter information of each component; the target parameter information includes: target 3D boundary volume information of the component and abstract morphological encoding; the target 3D boundary volume information includes the 3D bounding box parameters of the component, which are used to indicate the size, position and orientation of the component in 3D space; the abstract morphological encoding is used to indicate the category, morphological attributes and style attributes of the component; Based on the target parameter information of each component, generate the corresponding three-dimensional mesh data for each component; Based on the target structure information and the corresponding 3D mesh data of each component, the model data of the target model is aggregated and generated.

[0212] In one feasible implementation, when the processor 801 executes the adjustment instruction for the initial structure information received through the input device, updates the initial structure information according to the adjustment instruction, and obtains the target structure information, it is specifically used to: generate and display the initial model according to the initial structure information; It receives adjustment instructions for the initial model input through the input device, updates the initial structural information based on the adjustment data, and obtains the target structural information.

[0213] In a feasible implementation, when the processor 801 executes the response input instruction to determine the initial structural information of the target model, it is specifically used to: perform feature encoding processing on the input data to obtain global feature information; Based on global feature information, a pre-trained layout prediction network is used to determine the initial structural information of the target model.

[0214] In one feasible implementation, when the processor 801 performs feature encoding processing on the input data to obtain global feature information, it is specifically used to: perform feature encoding processing on the image information of the target model using a visual encoder to generate a global image feature vector. A text encoder is used to perform feature encoding on the text description information of the target model to generate a global text feature vector; Based on the global image feature vector and the global text feature vector, multimodal feature fusion processing is performed to obtain global feature information.

[0215] In a feasible implementation, when the processor 801 executes the process of determining the initial structural information of the target model using a pre-trained layout prediction network based on global feature information, it specifically performs the following: inputting global feature information into the layout prediction network, and having the layout prediction network determine the initial structural information of the target model through a preset iterative denoising process.

[0216] In a feasible implementation, when the processor 801 performs the initial structural information of the target model by the layout prediction network through a preset iterative denoising process, it is specifically used to: determine the noise data of the current time step by the layout prediction network based on the noise features of the current time step and global feature information; the initial noise features are obtained by sampling random noise; Based on the noise features and noise data of the current time step, the noise features of the current time step are denoised to obtain the noise features of the next time step. This process is repeated until the time step ends, and the features at the end of the time step are used as the initial structural information of the target model.

[0217] In one feasible implementation, when the processor 801 executes the process of receiving adjustment instructions for the initial model input through the input device, updating the initial structure information according to the adjustment data, and obtaining the target structure information, it is specifically used to: receive operation instructions for the target component in the initial model and obtain operation data; Based on the identifier of the target component, read the initial three-dimensional boundary volume information of the target component; Based on the initial three-dimensional boundary volume information of the target component and the operation data, the target three-dimensional boundary volume information of the target component is obtained; The three-dimensional boundary volume information of the target component is updated into the initial structural information of the target model to obtain the target structural information.

[0218] In one feasible implementation, the processor 801 is also used to perform: updating the initial model based on the target three-dimensional boundary volume information of the target component.

[0219] In a feasible implementation, when the processor 801 executes the process of receiving adjustment instructions for the initial model input through the input device, updating the initial structure information according to the adjustment data, and obtaining the target structure information, it is specifically used to: receive a component addition operation for the initial model, obtain the addition position information corresponding to the component to be added, and the abstract form code specified for the component to be added. The addition location information of the component to be added and the specified abstract form code are used as the target parameter information of the component to be added to update the initial structure information, thus obtaining the target structure information.

[0220] In one feasible implementation, when the processor 801 executes the following: receiving adjustment instructions for the initial model input through the input device, updating the initial structure information according to the adjustment data, and obtaining the target structure information, the processor 801 is specifically used to: receive a confirmation operation for the target component in the initial model and obtain the initial abstract form code of the target component. The system receives a confirmation operation for the abstract form code of a specified component in the component library, replaces the initial abstract form code of the target component with the abstract form code of the specified component, and obtains the abstract form code of the target component. The component library stores target parameter information of various pre-generated components.

[0221] In a feasible implementation, when the processor 801 generates the corresponding three-dimensional mesh data for each component based on the target parameter information of each component, it specifically performs the following: based on the target parameter information of each component, it uses the component refinement network corresponding to each component to generate the corresponding three-dimensional mesh data for each component in parallel.

[0222] In a feasible implementation, when the processor 801 executes the parallel generation of three-dimensional mesh data corresponding to each component using the component refinement network corresponding to each component based on the target parameter information of each component, it specifically performs the following: inputting the target parameter information of each component and the global feature information corresponding to the input instruction into the component refinement network corresponding to each component, and having each component refinement network generate the three-dimensional mesh data corresponding to each component in parallel based on the target parameter information and global feature information of the corresponding component.

[0223] In a feasible implementation, when the processor 801 executes the parallel generation of three-dimensional mesh data corresponding to each component by the refinement network of each component based on the target parameter information and global feature information of the corresponding component, it is specifically used to: determine the bounding box information of the first component in three-dimensional space by the first component refinement network based on the three-dimensional boundary volume information of the first component. The first component refinement network generates the 3D mesh data corresponding to the first component within the range indicated by the bounding box information, based on the abstract morphological encoding of the first component and global feature information.

[0224] In one feasible implementation, when the processor 801 generates three-dimensional mesh data corresponding to the first component within the range indicated by the bounding box information, it is specifically used to generate a symbolic distance field data for characterizing the geometry of the first component within the range indicated by the bounding box information. Based on the symbolic distance field data, a preset mesh extraction algorithm is used to extract and generate the three-dimensional mesh data corresponding to the first component.

[0225] In a feasible implementation, when the processor 801 executes the parallel generation of three-dimensional mesh data corresponding to each component by using the component refinement network corresponding to each component based on the target parameter information of each component, it specifically performs the following: if it detects that the initial structural information of the target component in the initial model has been updated, it triggers the component refinement network corresponding to the target component to regenerate the three-dimensional mesh data corresponding to the target component based on the updated target parameter information of the target component.

[0226] In a feasible implementation, when the processor 801 executes the model data of the target model by aggregating and generating the target model based on the target structure information and the three-dimensional mesh data corresponding to each component, it is specifically used to: determine the transformation matrix information of each component from local space to world space based on the target three-dimensional boundary volume information of each component. Based on the transformation matrix information of each component from local space to world space and the corresponding three-dimensional mesh data of each component, the target three-dimensional mesh data of each component in world space is determined; Based on the target 3D mesh data of each component in world space and the hierarchical relationship between the components, the model data of the target model is aggregated and generated; the model data is used to import into the game engine to load the target model.

[0227] In this embodiment, the computer program, when run by the processor, can also execute other machine-readable instructions to perform other methods as described in the embodiments. For details on the specific execution steps and principles, please refer to the description of the embodiments, which will not be repeated here.

[0228] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0229] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0230] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in a combination of hardware and software functional units.

[0231] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for generating a three-dimensional model, characterized in that, include: In response to input commands, the initial structural information of the target model is determined, including the identifiers of each component constituting the target model and the initial parameter information of each component; Receive adjustment instructions for the initial structure information input via an input device, update the initial structure information according to the adjustment instructions, and obtain target structure information; Determine the target structure information of the target model; The target structure information includes: target parameter information for each component; the target parameter information includes: target three-dimensional boundary volume information and abstract morphological encoding of the component; the target three-dimensional boundary volume information includes the three-dimensional bounding box parameters of the component, used to indicate the size, position and orientation of the component in three-dimensional space; the abstract morphological encoding is used to indicate the category, morphological attributes and style attributes of the component; Based on the target parameter information of each component, generate the corresponding three-dimensional mesh data for each component; Based on the target structure information and the three-dimensional mesh data corresponding to each component, the model data of the target model is aggregated and generated.

2. The method according to claim 1, characterized in that, The step of receiving an adjustment instruction for the initial structural information input via an input device, updating the initial structural information according to the adjustment instruction, and obtaining target structural information includes: Based on the initial structural information, an initial model is generated and displayed; The system receives adjustment instructions for the initial model input via an input device, updates the initial structural information based on the adjustment data, and obtains the target structural information.

3. The method according to claim 1, characterized in that, The response input command determines the initial structural information of the target model, including: The input instruction is parsed to obtain input data; the input data includes image information of the target model and / or text description information of the target model. The input data is subjected to feature encoding processing to obtain global feature information; Based on the global feature information, a pre-trained layout prediction network is used to determine the initial structural information of the target model.

4. The method according to claim 3, characterized in that, The step of performing feature encoding processing on the input data to obtain global feature information includes: A visual encoder is used to perform feature encoding on the image information of the target model to generate a global image feature vector. A text encoder is used to perform feature encoding on the text description information of the target model to generate a global text feature vector; Based on the global image feature vector and the global text feature vector, multimodal feature fusion processing is performed to obtain the global feature information.

5. The method according to claim 3, characterized in that, The step of determining the initial structural information of the target model using a pre-trained layout prediction network based on the global feature information includes: The global feature information is input into the layout prediction network, which then determines the initial structure information of the target model through a preset iterative denoising process.

6. The method according to claim 5, characterized in that, The determination of the initial structural information of the target model by the layout prediction network through a preset iterative denoising process includes: The layout prediction network determines the noise data for the current time step based on the noise features at the current time step and the global feature information; the initial noise features are obtained by sampling random noise. Based on the noise features and noise data of the current time step, the noise features of the current time step are denoised to obtain the noise features of the next time step. This process is repeated until the time step ends, and the features at the end of the time step are used as the initial structural information of the target model.

7. The method according to claim 2, characterized in that, The step of receiving adjustment instructions for the initial model input via an input device, updating the initial structural information based on the adjustment data, and obtaining the target structural information includes: Receive operation instructions for the target component in the initial model and acquire operation data; Based on the identifier of the target component, read the initial three-dimensional boundary volume information of the target component; Based on the initial three-dimensional boundary volume information of the target component and the operation data, the target three-dimensional boundary volume information of the target component is obtained; The three-dimensional boundary volume information of the target component is updated into the initial structural information of the target model to obtain the target structural information.

8. The method according to claim 7, characterized in that, Also includes: The initial model is updated based on the target three-dimensional boundary volume information of the target component.

9. The method according to claim 2, characterized in that, The step of receiving adjustment instructions for the initial model input via an input device, updating the initial structural information based on the adjustment data, and obtaining the target structural information includes: Receive component addition operations for the initial model, obtain the addition position information corresponding to the component to be added and the abstract form code specified for the component to be added; The addition location information of the component to be added and the specified abstract form code are used as the target parameter information of the component to be added to update the initial structure information to obtain the target structure information.

10. The method according to claim 2, characterized in that, The step of receiving adjustment instructions for the initial model input via an input device, updating the initial structural information based on the adjustment data, and obtaining the target structural information includes: Receive confirmation operation for the target component in the initial model, and obtain the initial abstract form code of the target component; The system receives a confirmation operation for the abstract form code of a specified component in the component library, replaces the initial abstract form code of the target component with the abstract form code of the specified component, and obtains the abstract form code of the target component; the component library stores target parameter information of various pre-generated components.

11. The method according to claim 1, characterized in that, The step of generating 3D mesh data for each component based on the target parameter information of each component includes: Based on the target parameter information of each component, the corresponding component refinement network is used to generate the three-dimensional mesh data of each component in parallel.

12. The method according to claim 11, characterized in that, The step of generating 3D mesh data for each component in parallel, based on the target parameter information of each component and using the corresponding component refinement network, includes: The target parameter information of each component and the global feature information corresponding to the input command are input into the component refinement network corresponding to each component. The component refinement network generates the three-dimensional mesh data corresponding to each component in parallel based on the target parameter information of the corresponding component and the global feature information.

13. The method according to claim 12, characterized in that, The process of refining each component's network in parallel generating 3D mesh data for each component based on the target parameter information of the corresponding component and the global feature information includes: The first component refinement network determines the bounding box information of the first component in three-dimensional space based on the three-dimensional boundary volume information of the first component; The first component refinement network generates 3D mesh data corresponding to the first component within the range indicated by the bounding box information, based on the abstract morphological encoding of the first component and the global feature information.

14. The method according to claim 13, characterized in that, Generating the 3D mesh data corresponding to the first component within the range indicated by the bounding box information includes: Within the range indicated by the bounding box information, a symbolic distance field data is generated to characterize the geometry of the first component; Based on the symbol distance field data, a preset mesh extraction algorithm is used to extract and generate the three-dimensional mesh data corresponding to the first component.

15. The method according to claim 11, characterized in that, The step of generating 3D mesh data for each component in parallel, based on the target parameter information of each component and using the corresponding component refinement network, includes: If the initial structural information of the target component is detected to be updated, the component refinement network corresponding to the target component is triggered, and the three-dimensional mesh data corresponding to the target component is regenerated according to the updated target parameter information of the target component.

16. The method according to claim 1, characterized in that, The process of aggregating and generating model data for the target model based on the target structure information and the corresponding 3D mesh data of each component includes: Based on the target 3D boundary volume information of each component, determine the transformation matrix information of each component from local space to world space; Based on the transformation matrix information of each component from local space to world space and the corresponding three-dimensional mesh data of each component, the target three-dimensional mesh data of each component in world space is determined; Based on the target 3D mesh data of each component in world space and the hierarchical relationship between the components, model data of the target model is generated by aggregating the data; the model data is used to import the game engine to load the target model.

17. A three-dimensional model generation device, characterized in that, include: Determine the module and generate the module; The determining module is used to respond to input instructions and determine the initial structural information of the target model. The initial structural information includes the identifiers of each component constituting the target model and the initial parameter information of each component. The generation module is configured to receive an adjustment instruction for the initial structural information input through an input device, update the initial structural information according to the adjustment instruction, and obtain the target structural information. The determining module is used to determine the target structure information of the target model; The target structure information includes: target parameter information for each component; the target parameter information includes: target three-dimensional boundary volume information and abstract morphology encoding of the component; the target three-dimensional boundary volume information includes the three-dimensional bounding box parameters of the component, which are used to indicate the size, position and orientation of the component in three-dimensional space; the initial abstract morphology encoding is used to indicate the category, morphological attributes and style attributes of the component; The generation module is used to generate three-dimensional mesh data corresponding to each component based on the target parameter information of each component. The generation module is used to aggregate and generate model data of the target model based on the target structure information and the three-dimensional mesh data corresponding to each component.

18. An electronic device, characterized in that, include: The device includes a processor, a storage medium, and a bus, wherein the storage medium stores program instructions executable by the processor, and when the electronic device is running, the processor communicates with the storage medium via the bus, and the processor executes the program instructions to perform the three-dimensional model generation method as described in any one of claims 1 to 16.

19. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, performs the three-dimensional model generation method as described in any one of claims 1 to 16.

Images