Three-dimensional model generation method and apparatus, computer device, and storage medium

Abstract

A three-dimensional model generation method. The method comprises: obtaining prompt information for describing an object category, and determining a target network model that matches the object category indicated by the prompt information (202), the target network model comprising a first sub-network model and a second sub-network model; obtaining a first generative intensity, a second generative intensity, and a generative seed, wherein the first generative intensity indicates a quantity of iteration rounds of performing first processing by the first sub-network model, the second generative intensity indicates a quantity of iteration rounds of performing second processing by the second sub-network model, and the generative seed is a random number generated randomly, wherein each round of the first processing comprises at least one instance of first encoding and decoding processing, each round of the second processing comprises at least one instance of second encoding and decoding processing, and the first encoding and decoding processing and the second encoding and decoding processing each comprise at least encoding processing and decoding processing (204); by means of the first sub-network model and on the basis of the first generative intensity and the generative seed, performing at least one round of the first processing, to obtain a three-dimensional rough model (206); and by means of the second sub-network model and on the basis of the second generative intensity and the three-dimensional rough model, performing at least one round of the second processing, to obtain a three-dimensional object model that meets a preset high resolution condition and pertains to the object category (208).

Description

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

[0001] Related applications

[0002] This application claims priority to Chinese patent application number 2023111765434, filed on September 12, 2023, entitled “Three-dimensional model generation method, device, computer equipment and storage medium”, the entire text of which is hereby incorporated by reference. Technical Field

[0003] The present application relates to the field of computer technology, and in particular to a three-dimensional model generation method, apparatus, computer equipment, storage medium, and computer program product. Background Art

[0004] With the development of computer vision, 3D model generation has become an important topic in the field of AIGC (Artificial Intelligence Generated Content). It aims to create corresponding 3D models based on user input information and has broad application prospects in games, film and television, virtual reality, 3D (3D) printing and other fields.

[0005] Currently, 3D model generation is in its infancy. In practical applications, information related to the desired 3D model, such as image or voxel information, is typically input into pre-trained models such as Generative Adversarial Nets (GANs) or autoregressive models. These models are then randomly generated using these models. While these generative models can create 3D models, the quality and accuracy of the resulting models are often poor.

[0006] Summary of the Invention

[0007] The present application provides a three-dimensional model generation method, apparatus, computer equipment, computer-readable storage medium, and computer program product.

[0008] In a first aspect, the present application provides a three-dimensional model generation method, which is executed by a computer device, comprising:

[0009] Obtain prompt information describing an object category, and determine a target network model that matches the object category indicated by the prompt information; the target network model includes a first sub-network model and a second sub-network model;

[0010] Obtain a first generation strength, a second generation strength, and a generation seed; the first generation strength indicates an iterative round for the first sub-network model to perform the first processing, the second generation strength indicates an iterative round for the second sub-network model to perform the second processing, and the generation seed is a randomly generated random number; wherein each round of the first processing includes at least one first encoding and decoding processing, and each round of the second processing includes at least one second encoding and decoding processing, and the first encoding and decoding processing and the second encoding and decoding processing each include at least encoding processing and decoding processing;

[0011] Obtaining a three-dimensional rough model by performing at least one round of first processing based on the first generation strength and the generation seed through the first sub-network model; and

[0012] Through the second sub-network model, at least one round of second processing is performed based on the second generated strength and the three-dimensional rough model to obtain a three-dimensional object model that meets a preset high-resolution condition and belongs to the object category.

[0013] In a second aspect, the present application further provides a three-dimensional model generation device, comprising:

[0014] a network matching module, configured to obtain prompt information describing an object category and determine a target network model that matches the object category indicated by the prompt information; the target network model includes a first sub-network model and a second sub-network model;

[0015] An information acquisition module is configured to obtain a first generation strength, a second generation strength, and a generation seed; the first generation strength indicates the number of iterations of the first processing performed by the first sub-network model, the second generation strength indicates the number of iterations of the second processing performed by the second sub-network model, and the generation seed is a randomly generated random number; wherein each round of the first processing includes at least one first encoding and decoding process, and each round of the second processing includes at least one second encoding and decoding process, and both the first encoding and decoding process and the second encoding and decoding process include at least encoding and decoding processes;

[0016] a first model generating module, configured to obtain a three-dimensional rough model by performing at least one round of first processing on the first sub-network model based on the first generation strength and the generation seed;

[0017] The second model generation module is used to perform at least one round of second processing through the second sub-network model and based on the second generation strength and the three-dimensional rough model to obtain a three-dimensional object model that meets the preset high-resolution condition and belongs to the object category.

[0018] In a third aspect, the present application also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-mentioned three-dimensional model generation method when executing the computer program.

[0019] In a fourth aspect, the present application also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above-mentioned three-dimensional model generation method are implemented.

[0020] In a fifth aspect, the present application also provides a computer program product, comprising a computer program, which implements the steps of the above-mentioned three-dimensional model generation method when executed by a processor.

[0021] The details of one or more embodiments of the present application are set forth in the accompanying drawings and the description below. Other features, objects, and advantages of the present application will become apparent from the description, drawings, and claims. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the conventional technology, the following briefly introduces the drawings required for use in the embodiments or the conventional technology descriptions. Obviously, the drawings described below are merely embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the disclosed drawings without any creative work.

[0023] FIG1 is a diagram illustrating an application environment of a three-dimensional model generation method according to an embodiment;

[0024] FIG2 is a schematic diagram of a process for generating a three-dimensional model in one embodiment;

[0025] FIG3 is a flowchart of a target network model acquisition step in one embodiment;

[0026] FIG4 is a schematic flow chart of an object category determination step in one embodiment;

[0027] FIG5 is a schematic diagram of a flow chart of any first encoding and decoding process in one embodiment;

[0028] FIG6 is a flowchart of generating a three-dimensional rough model in one embodiment;

[0029] FIG7 is a model structure diagram of the first sub-network model in one embodiment;

[0030] FIG8 is a flowchart of feature fusion in one embodiment;

[0031] FIG9 is a flowchart of a model training process in one embodiment;

[0032] FIG10 is a schematic diagram of the architecture of a method for generating a three-dimensional model in one embodiment;

[0033] FIG11 is a block diagram of a three-dimensional model generating apparatus according to an embodiment;

[0034] FIG12 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0036] The three-dimensional model generation method provided in the embodiment of the present application can be applied to the application environment shown in Figure 1. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or it can be placed on the cloud or other network servers. The terminal 102 and the server 104 can be used alone to execute the three-dimensional model generation method in this application, and the terminal 102 and the server 104 can be used together to execute the three-dimensional model generation method in this application. Taking the terminal 102 and the server 104 cooperating to execute the present application as an example, when generating a three-dimensional model, the user can edit the prompt information related to the required three-dimensional model through the terminal 102 in combination with the three-dimensional model required by himself, and send the prompt information to the server 104. Server 104 can obtain the prompt information sent by terminal 102 and determine a target network model that matches the object category indicated by the prompt information, the target network model including a first sub-network model and a second sub-network model. Server 104 obtains a first coding feature of a first generation strength and a second coding feature of a generation seed. Server 104 performs a first encoding and decoding process based on the first and second coding features on the first sub-network model to obtain a three-dimensional rough model. Server 104 obtains a third coding feature of a second generation strength and obtains a fourth coding feature obtained by encoding the three-dimensional rough model. Server 104 performs a second encoding and decoding process based on the third and fourth coding features on the second sub-network model to obtain a three-dimensional object model that meets a preset high-resolution condition and belongs to the object category. Server 104 can feed the three-dimensional object model back to terminal 102 for display via terminal 102.

[0037] Among them, the terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, Internet of Things devices, portable wearable devices, intelligent voice interaction devices, smart home appliances, vehicle-mounted terminals, aircraft, etc. Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart vehicle-mounted devices, etc. Portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. Server 104 can be implemented as a standalone server or a server cluster consisting of multiple servers. The server can be implemented as a standalone server or a server cluster consisting of multiple servers, and can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The terminal and server can be directly or indirectly connected via wired or wireless communication. Embodiments of the present invention can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, smart transportation, assisted driving, etc.

[0038] Artificial Intelligence (AI) involves theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI also encompasses the study of the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making.

[0039] Artificial intelligence (AI) technology is a comprehensive discipline encompassing a wide range of fields, encompassing both hardware and software technologies. Foundational AI technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technologies primarily encompass computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0040] The following is a detailed introduction to the 3D model generation method of this application:

[0041] In an exemplary embodiment, as shown in FIG2 , a method for generating a three-dimensional model is provided. The method is described by taking an example of applying the method to a computer device (the computer device may specifically be the terminal 102 or the server 104 in FIG1 ). The method for generating a three-dimensional model includes the following steps:

[0042] Step 202: Obtain prompt information describing the object category and determine a target network model that matches the object category indicated by the prompt information; the target network model includes a first sub-network model and a second sub-network model.

[0043] The prompt information is information describing the object category, specifically an object identifier and / or information describing the object's attributes. Furthermore, the prompt information can be text, or identifiers such as letters, numbers, or feature codes, without limitation. The object category represents the category to which the 3D object model to be generated belongs.

[0044] The target network model is a network model built based on artificial intelligence algorithms and can be used to generate 3D object models. This model can be constructed using various algorithms, including supervised and unsupervised learning algorithms. When using supervised learning algorithms, the target network model can be constructed using various generative models, including diffusion generative models, variational autoencoder models, and adversarial generative models.

[0045] The target network model includes a first sub-network model and a second sub-network model. The first sub-network model is used to generate a 3D rough model. The output of the first sub-network model serves as input data for the second sub-network model. This allows the second sub-network model to further optimize the 3D rough model generated by the first sub-network model, thereby obtaining a higher-quality 3D object model and improving 3D model generation accuracy.

[0046] Specifically, when the computer device obtains the prompt information sent by the terminal, it can perform category judgment on the prompt information, determine the object category referred to by the prompt information, and further obtain a target network model that matches the object category, and the target network model includes a first sub-network model and a second sub-network model.

[0047] In some embodiments, the computer device may use a word segmentation tool to segment the prompt information and determine the various types of words in the prompt information, such as nouns, entity words, non-nouns, and stop words. Furthermore, the computer device may remove stop words and non-noun words (stop words are words without practical meaning, such as pronouns, particles, adjectives, and adverbs), retain entity words and nouns, such as object names and dimensions, which have practical meaning, and use the retained entity words and nouns as keywords. The computer device may then perform category analysis based on the determined keywords to determine the category of the object referred to by the prompt information.

[0048] In some embodiments, when the computer device performs category analysis on the enhanced information, it is carried out through an algorithm that can realize object category recognition. The algorithm used can be a similarity analysis algorithm, a decision tree algorithm, a support vector machine algorithm and other algorithms.

[0049] In some embodiments, the computer device can query and obtain the target network model to which the object category belongs based on the object category indicated by the determined prompt information. For example, the computer device can pre-build a corresponding network model library for various types of network models and the object categories corresponding to each type of network model, and the network model library records the correspondence between network models and object categories. Then, when the computer device obtains the target network model based on the object category, it can query it from the network model library according to actual needs. In some embodiments, when the computer device fails to find the target network model that matches the object category, it can output feedback information indicating that the search failed to the user, so that the user can adjust the prompt information based on the feedback information; the computer device can also re-select keywords for category judgment, determine a new object category, and obtain the target network model based on the new object category.

[0050] Step 204: Obtain a first generation strength, a second generation strength, and a generation seed; the first generation strength indicates the iterative rounds of the first processing performed by the first sub-network model, the second generation strength indicates the iterative rounds of the second processing performed by the second sub-network model, and the generation seed is a randomly generated random number; wherein each round of the first processing includes at least one first encoding and decoding processing, and each round of the second processing includes at least one second encoding and decoding processing, and both the first encoding and decoding processing and the second encoding and decoding processing include at least encoding processing and decoding processing.

[0051] The generation intensity is information indicating the number of iterations of the model. In some embodiments, the greater the generation intensity, the more iterations of the model, and the better the final three-dimensional model.

[0052] The first generation strength is used to indicate the number of iterations of the first processing performed by the first sub-network model. The number of first processing iterations can be one or more. The first generation strength can be determined based on factors such as the quality of the 3D model to be generated and the generation efficiency.

[0053] The second generation strength is used to indicate the number of iterations of the second processing performed by the second sub-network model. The number of second processings can be one or more. The second generation strength can be determined based on the quality of the 3D model to be generated and the generation efficiency.

[0054] The generation seed is a randomly generated random number, which can be a randomly generated natural number, specifically a natural number randomly selected from a set range.

[0055] The first process may include at least one first codec process, which includes at least encoding and decoding processes. The second process may include at least one second codec process, which includes at least encoding and decoding processes. The first process and the second process may be the same or different, and the first codec process and the second codec process may be the same or different.

[0056] Specifically, the computer device can select a natural number from the natural number range according to the set natural number range, and determine the first generation intensity, the second generation intensity and the generation seed according to the selected natural number; the first generation intensity, the second generation intensity and the generation seed can also be determined by the natural number carried in the prompt information, and the computer device can obtain the first generation intensity, the second generation intensity and the generation seed by parsing the prompt information; the computer device can also combine the object category corresponding to the target network model, and directly query from the database according to the object category to obtain the first generation intensity, the second generation intensity and the generation seed. The database can pre-store the object category of each type of object and the association between the first generation intensity, the second generation intensity and the generation seed.

[0057] In some embodiments, the natural number carried in the prompt information may specify only one or two of the first generation strength, the second generation strength, and the generation seed. In this case, the computer device may select a natural number according to a set natural number range and assign values ​​to the unspecified data in the first generation strength, the second generation strength, and the generation seed based on the selected natural number.

[0058] In some embodiments, the computer device may obtain the first generation intensity, the second generation intensity, and the generation seed according to a natural number range. The natural number range may include the first generation intensity range, the second generation intensity range, and the generation seed range. The first generation intensity range, the second generation intensity range, and the generation seed range may be the same or different. For example, the set first generation intensity range may be 0-20, the second generation intensity range may be 20-50, and the generation seed range may be 100-150. The first generation intensity obtained by the computer device may be 13, the second generation intensity may be 30, and the generation seed may be 90.

[0059] Step 206 : Perform at least one round of first processing on the first sub-network model based on the first generation strength and the generation seed to obtain a three-dimensional rough model.

[0060] The first encoding and decoding process includes at least encoding and decoding processes. The first process may include at least one first encoding and decoding process. During the first process, the first generation strength may be used to determine the first encoding feature required for the current round of the first encoding and decoding process. In the first round of the first process, the second encoding feature required for the first first encoding and decoding process may be determined by generating a seed.

[0061] The three-dimensional rough model is a three-dimensional model obtained by performing at least one round of first processing on the first sub-network model. The three-dimensional rough model can form the approximate shape of the three-dimensional object model to be generated.

[0062] Specifically, the computer device may determine, based on the first generation strength, a first encoding feature required for the first encoding and decoding process in each round of the first processing, and, based on the generation seed, determine a second encoding feature required for the first encoding and decoding process in the first round of the first processing. The first sub-network model may then be used to perform encoding and decoding on the first and second encoding features to obtain a three-dimensional rough model.

[0063] In some embodiments, during the process of the computer device performing encoding and decoding processing through the first sub-network model, when the number of first processing times is multiple, and each first processing includes multiple first encoding and decoding processes, the first encoding feature is updated with the round of the first processing, that is, the first encoding feature of the same round is the same, and the first encoding features of different rounds are different. The second encoding feature is updated with the changes in the first encoding and decoding process, that is, the second encoding feature of different first encoding and decoding processes in the same round is different. In the same round, the second encoding feature of the later first encoding and decoding process is determined by the output of the previous first encoding and decoding process, and when the previous first encoding and decoding process is the first round, it is determined by the generated seed.

[0064] In some embodiments, the computer device may encode the first generation strength, determine a first encoding feature, and encode the generation seed to obtain a second encoding feature required for the first encoding and decoding process in the first round of the first process. The computer device may encode based on a linear neural network, a three-dimensional convolutional network, etc., and may also use other equivalent operators to replace the linear neural network or the three-dimensional convolutional network, without limitation herein.

[0065] Step 208 : Perform at least one round of second processing using the second sub-network model and based on the second generated strength and the three-dimensional rough model to obtain a three-dimensional object model that meets a preset high-resolution condition and belongs to the object category.

[0066] The second encoding and decoding process also includes at least encoding and decoding processes. The second process may include at least one second encoding and decoding process. During the second process, the second generated intensity may be used to determine the third encoding feature required for the current round of the second encoding and decoding process. The three-dimensional rough model may be used to determine the fourth encoding feature required for the first second encoding and decoding process in the first round of the second process.

[0067] The preset high-resolution condition is a condition used to determine whether the resolution of the generated three-dimensional object model meets the model quality requirements. The preset high-resolution condition can be set according to a set resolution threshold. That is, when the resolution of the generated three-dimensional object model reaches the resolution threshold, it means that the preset high-resolution condition is met.

[0068] Specifically, the computer device may determine, based on the second generation intensity, a third encoding feature required for the second encoding and decoding process in each round of the second processing, and, based on the three-dimensional rough model, determine a fourth encoding feature required for the first second encoding and decoding process in the first round of the second processing. The third and fourth encoding features are then encoded and decoded using the second sub-network model to obtain a three-dimensional object model that meets a preset high-resolution condition and belongs to the object category.

[0069] In some embodiments, the computer device may encode the second generated intensity to determine a third encoding feature, and may also encode the three-dimensional rough model to determine a fourth encoding feature required for the first second encoding and decoding process in the first round of the second process. The computer device may encode based on a linear neural network, a three-dimensional convolutional network, or the like, and may also use other equivalent operators to replace the linear neural network or the three-dimensional convolutional network, without limitation herein.

[0070] In some embodiments, when a computer device performs a second encoding and decoding process using a second sub-network model, if the second process is performed multiple times and each second process includes multiple second encoding and decoding processes, the third encoding feature is updated with each round of the second process, i.e., the third encoding feature is the same in the same round and different in different rounds. The fourth encoding feature is updated with changes in the second encoding and decoding process, i.e., the fourth encoding feature is different in different second encoding and decoding processes in the same round. In the same round, the fourth encoding feature of a later second encoding and decoding process is determined by the output of a previous second encoding and decoding process, and when the previous second encoding and decoding process is the first time, it is determined by the three-dimensional rough model.

[0071] In the above-mentioned three-dimensional model generation method, when prompt information is obtained, a target network model that matches the object category indicated by the prompt information can be searched, and then a high-precision three-dimensional object model of the category required by the user can be generated using the found target network model. In the process of generating the three-dimensional object model using the target network model, a first generation strength, a second generation strength, and a generation seed are obtained, and then the first generation strength and the generation seed are subjected to at least one round of first processing based on the first sub-network model in the target network model. The first processing includes at least one encoding and decoding processing, so that a three-dimensional rough model can be obtained. The three-dimensional rough model determines the approximate shape of the three-dimensional object model to be generated. Further, the second generation strength and the three-dimensional rough model are subjected to at least one round of second processing based on the second sub-network model in the target network model. Each round of second processing includes at least one second encoding and decoding processing, so as to obtain a three-dimensional object model that meets the preset high-resolution conditions and belongs to the object category. The three-dimensional object model is a model obtained after optimization based on the three-dimensional rough model. In this way, the model generation process is divided into two stages: coarse model generation and fine model generation. The coarse model generation stage is responsible for generating the approximate shape of the model, and the fine model generation stage is responsible for optimizing the generation results of the three-dimensional model. This can greatly improve the generation quality of the three-dimensional object model, that is, improve the generation accuracy of the three-dimensional object model.

[0072] In some embodiments, determining a target network model that matches the object category indicated by the prompt information includes: determining the object category indicated by the prompt information; obtaining a target network model that matches the object category from a network model pool; the network model pool includes multiple trained network models for generating three-dimensional object models of various categories, and the target network model is used to generate a three-dimensional object model belonging to the object category indicated by the prompt information.

[0073] It should be noted that the network model pool stores a variety of pre-trained network models for generating various categories of three-dimensional object models. The target network model is a network model selected from the network model pool that matches the prompt indication information and can be used to generate a three-dimensional object model belonging to the object category indicated by the prompt information.

[0074] Specifically, the computer device can perform a similarity analysis based on the prompt information and preset category information to determine the object category. The category information is information related to objects that match the network models stored in the network model pool. When performing the similarity analysis, the computer device can calculate cosine similarity, edit distance, or similarity coefficient to obtain similarity analysis results, and determine the object category based on the similarity analysis results. The computer device can then obtain a target network model that matches the object category from the network model pool based on the object category indicated in the similarity analysis results.

[0075] In some embodiments, the category information may be an object name, an object identifier, or other information used to describe an object in the network model pool that matches a stored network model.

[0076] In some embodiments, when constructing a target network model, the computer device may construct a corresponding network model for each object type. For example, for a chair, a network model corresponding to the chair may be constructed; for a table, a network model corresponding to the table may be constructed; and for a game character, a network model corresponding to the game character may be constructed. Furthermore, for each constructed network model of an object type, the computer device may store it in a network model pool or a pre-set model library. The network model pool stores the association between object types and network models. When a 3D model is actually needed, it can be directly retrieved from the network model pool or model library based on the object type.

[0077] In some embodiments, FIG3 shows a flowchart of a process for obtaining a target network model in a 3D model generation method. First, a user may enter prompt information, which provides a modeling prompt. For example, the user may enter the prompt information "a chair." Upon receiving the prompt information, the computer device performs category matching based on the prompt information to determine the object category in the prompt information. Based on the object category, the computer device searches for a target network model that matches the object category from a network model pool. If the computer device determines that a 3D object model of the object category cannot be generated, it returns a default prompt, such as "The system currently does not support objects of this category." If a 3D object model of the object category can be generated, the computer device may obtain a target network model that matches the object category from the network model pool. The network model pool includes multiple trained network models for generating 3D object models of various categories. For example, categories 1 through N each have corresponding 3D model generation networks. Based on the target network model, the computer device generates a 3D object model of the object category indicated by the prompt information. The generated 3D object model is then returned to the user.

[0078] In the above embodiment, the computer device can determine the object category based on the prompt information, and directly obtain the matching target network model from the network model pool according to the determined object category, so as to generate a three-dimensional object model belonging to the object category indicated by the prompt information based on the target network model, which not only improves the accuracy of three-dimensional model generation, but also improves the efficiency of three-dimensional model generation.

[0079] In some exemplary embodiments, as shown in FIG4 , the step of determining the object category indicated by the prompt information includes steps 402 to 408 .

[0080] Step 402: Obtain feature vectors of respective preset multiple object category names to obtain a first feature set.

[0081] The first feature set is composed of feature vectors of each of a plurality of preset object category names. For each object category name, feature extraction can be performed to obtain a corresponding feature vector.

[0082] Specifically, the computer device may perform vectorized representation processing on each object category name to obtain a feature vector. The vectorized representation can be achieved using a language model. After processing by the language model, the object category name can be converted into a vector representation, and the computer device can then obtain a first feature set based on the vector representation.

[0083] In some embodiments, when a computer device performs vectorized representation processing on an object category name, it can be based on a language model. For example, the language model can be BERT (Bidirectional Encoder Representations from Transformers), GPT3 (Generative Pre-trained Transformer 3), GPT4 (Generative Pre-trained Transformer 4), Clip (Contrastive Language-Image Pre-Training), etc., which is not limited to this embodiment of the present application.

[0084] In some embodiments, the computer device can perform vectorized representation processing on each object name according to the BERT model. For example, the network model stored in the network model pool supports 1, 2...N object categories, where N is a positive integer greater than or equal to 1. The object category names of each object category can form an object category name list, and the object category name list is: C = {C1, C2, ..., C N}, C k Indicates the name of the k-th object category supported, such as "chair".

[0085] The computer device can extract the feature vector corresponding to each object category name based on the BERT model. Specifically, the BERT model can be marked as M nlp , then C k The calculation formula of the corresponding eigenvector can be expressed as: F k =M nlp (C k );

[0086] Among them, F kAs a high-dimensional floating-point vector, the computer device can calculate the feature vector of each object category name through the feature vector calculation formula, thereby obtaining the feature vector set F = {F1, F2, ..., F N}.

[0087] Step 404 : vectorize the prompt information to determine at least one feature vector related to the prompt information, and obtain a second feature set.

[0088] According to the prompt information, only one feature vector or multiple feature vectors may be extracted. The second feature set is obtained by extracting at least one feature vector from the prompt information.

[0089] Specifically, the computer device may process the prompt information to determine information for vectorized representation and obtain a second feature set. When processing the prompt information, the computer device may extract at least one keyword from the prompt information, and then perform vectorized representation processing based on the extracted at least one keyword to determine a feature vector corresponding to the at least one keyword. The computer device may also directly perform vectorized representation processing on the prompt information to obtain a feature vector corresponding to the prompt information.

[0090] In some embodiments, when extracting keywords from prompt information, the computer device may use word segmentation to extract keywords. In addition, other methods such as Deep Text (a text understanding engine based on deep learning) and content classification may also be used to extract keywords from prompt information.

[0091] In some embodiments, the prompt information is vectorized to determine at least one feature vector related to the prompt information to obtain a second feature set, including: performing word segmentation processing on the prompt information to obtain multiple word units in the prompt information; and extracting the feature vectors of the prompt information and each word unit respectively to obtain the second feature set.

[0092] The word unit is a single word segmented after the prompt information is segmented. The word segmented can be a noun, an entity word, a non-noun or other type of word.

[0093] Specifically, the computer device can use a word segmentation tool to perform word segmentation processing on the prompt information to obtain multiple word units, and perform vectorization representation processing based on the obtained word units to obtain the feature vectors of the word units, and perform vectorization representation processing based on the prompt information to obtain the feature vector of the prompt information, and further obtain a second feature set based on the obtained feature vectors.

[0094] In some embodiments, the prompt information obtained by the computer device may be a short sentence S. The computer device performs word segmentation processing on the short sentence S to obtain a plurality of word units corresponding to the short sentence S. For example, the obtained short sentence may be composed of Z words. The input short sentence may be represented as S = {S1, S2, ..., S Z}. The computer device can further use a language model, such as a BERT model, to extract S, S1, S2, ..., S Z The corresponding feature vectors are used to obtain the second feature set, which is marked as F′={F′0, F′1…, F′ Z}, where F′0=M nlp (S), F j ′=M nlp (S j ), 1≤j≤Z.

[0095] Step 406 : Calculate the similarity between each feature vector in the second feature set and each feature vector in the first feature set.

[0096] The similarity is a parameter used to characterize the similarity between the feature vectors in the first feature set and the feature vectors in the second feature set.

[0097] Specifically, the computer device may calculate the similarity between each feature vector in the second feature set and each feature vector in the first feature set.

[0098] As mentioned above, the computer device can perform vectorized representation processing on the object category name to obtain the first feature set F, and perform vectorized representation processing on the prompt information to obtain the second feature set F′. For any feature vector F in F′ j ′, the computer device can calculate its and any eigenvector F in F k The similarity Q jk :

[0099] Where 0≤Q jk ≤1, the number of similarities depends on the number of feature vectors in the first feature combination F and the second feature set F′.

[0100] Step 408 : Determine the object category to which the feature vector corresponding to the feature vector satisfying the preset similarity condition and having the greatest similarity in the first feature set belongs as the object category indicated by the prompt information.

[0101] The preset similarity condition is used to determine whether a target network model matching the object category indicated by the indication information exists in the network model pool based on similarity. When setting the preset similarity condition, it can be adaptively adjusted based on the quality requirements of the 3D model to be generated, the type of language model, and other factors.

[0102] Specifically, the computer device may select a target similarity from the calculated similarities and, based on the target similarity, determine whether a preset similarity condition is satisfied. If the preset similarity condition is satisfied, the computer device may determine the object category to which the feature vector corresponding to the maximum similarity in the first feature set belongs as the object category indicated by the prompt information. The target similarity selected by the computer device may be the maximum similarity among the similarities, or a similarity among the similarities whose value is greater than a set threshold, etc. The method for selecting the target similarity is not limited herein.

[0103] In some embodiments, the computer device may determine the maximum similarity among the multiple similarities obtained as the target similarity. The preset similarity condition may be determined based on a preset similarity threshold set based on the type of language model, that is, when the target similarity reaches the preset similarity threshold, it indicates that the preset similarity condition is satisfied. For example, the preset similarity threshold may be set to 0.97. If the target similarity selected by the computer device is 0.98, it indicates that the preset similarity condition is satisfied.

[0104] In other embodiments, the computer device may select a similarity that reaches a threshold as the target similarity, and determine that the preset similarity condition is satisfied when some of the selected similarities reach a preset similarity threshold. For example, the similarities selected by the computer device may be 0.7, 0.95, and 0.98, and the preset similarity threshold is 0.94. If there are two similarities that are greater than the preset similarity threshold, the computer device may determine that the preset similarity condition is satisfied.

[0105] In the above embodiment, the computer device obtains the first feature set and the second feature set, and calculates the similarity of the feature vectors in the first feature set and the second feature set, which can accurately confirm the object category indicated by the prompt information, ensure the quality of the three-dimensional object model, and improve the accuracy of the three-dimensional object model generation.

[0106] In some exemplary embodiments, step 206, that is, the step of obtaining a three-dimensional rough model by performing at least one round of first processing through the first sub-network model and based on the first generation strength and the generation seed, includes: determining a first number according to the first generation strength, performing a first processing of the first number through the first sub-network model, and obtaining the three-dimensional rough model based on the output of the last first processing; wherein each round of first processing includes at least two first encoding and decoding processes.

[0107] Among them, the first number is the number of times the first processing is performed through the first sub-network model. Generally speaking, the larger the value of the first number, the more times the first processing is performed through the first sub-network model, and the better the quality of the obtained three-dimensional rough model and the higher the accuracy.

[0108] Specifically, the computer device determines the first number of times to perform the first processing based on the first generation strength, so that the first sub-network model can be iteratively processed according to the first number, and the output of the last first processing is determined as a three-dimensional rough model.

[0109] In some embodiments, the first generation intensity may be a numerical value, and the computer device may directly use the first generation intensity as the first number, or obtain the first number by performing a mathematical operation on the first generation intensity. The mathematical operation performed here may be doubling, adding a preset value, etc. The preset value may be 1, 2, or other numerical values.

[0110] In one possible implementation, for example, if the first generation strength is 1, the computer device may add 1 to 1 to obtain a first number 2. When the first number is 2, it indicates that the first sub-network model needs to perform the first processing twice, i.e., the first processing and the second processing, and each first processing includes at least two first encoding and decoding processes.

[0111] In some embodiments, the computer device may perform convolution processing on the output of the last first processing to transform the dimension of the feature vector output of the last first processing to obtain a three-dimensional rough model, so that the obtained three-dimensional rough model can meet the input requirements of the second sub-network model and improve the processing accuracy of the subsequent second sub-network model. When performing convolution processing on the output, the computer device may perform processing based on three-dimensional convolution, depthwise separable convolution, etc., or may use other deep learning operators instead.

[0112] In some embodiments, each round of the first processing includes at least two first encoding and decoding processes. Referring to FIG5 , the first encoding and decoding process in any current round includes the following steps:

[0113] Step 502: Obtain a first coding feature in the first encoding and decoding process. The first coding feature is the same for each round in the current round and is determined according to the round corresponding to the current round.

[0114] Step 504 , obtaining the second coding feature in the first encoding and decoding process; wherein the first second coding feature of the first round is determined according to the generated seed, and the second coding feature of the non-first round is the output of the previous first encoding and decoding process.

[0115] Step 506 : Based on the first coding feature and the second coding feature, perform at least one hierarchical coding process according to different spatial resolution levels to obtain a first coding result.

[0116] The first sub-network model can include at least one coding layer, and the hierarchical coding process is performed on each coding layer. Each coding layer has a corresponding output spatial resolution. The voxel spatial resolution can be different or the same for different coding layers. The voxel spatial resolution can be adaptively set based on the accuracy requirements of the 3D model generation and the server's computing resources.

[0117] Step 508: Perform at least one fusion process based on the first encoding processing result and the first encoding feature to obtain a first intermediate processing result.

[0118] At least one connection layer may be provided between the encoding layer and the decoding layer of the first sub-network model, and each connection layer may be used to connect the encoding layer and the decoding layer. The first intermediate processing result is determined by a fusion result obtained by fusing the second encoding processing result and the third encoding feature based on the connection layer.

[0119] Step 510: Perform at least one hierarchical decoding process according to different spatial resolution levels based on the first intermediate processing result and the outputs of each hierarchical encoding process to obtain the output of the first encoding and decoding process.

[0120] The first sub-network model may further include at least one decoding layer, and hierarchical decoding processing is performed on each decoding layer. Different decoding layers also correspond to different spatial resolution levels, that is, a voxel spatial resolution is set for each decoding layer. The voxel spatial resolutions for different decoding layers can be different or the same.

[0121] Specifically, during any first encoding / decoding process of the current round, the computer device may determine the first encoding feature based on the round corresponding to the current round. Regarding the second encoding feature, if the current round is the first round and the first first encoding / decoding process in the first round, the computer device may determine it by generating a seed; if the current round is not the first round, the computer device may determine it based on the output of the previous first encoding / decoding process.

[0122] Furthermore, for any first encoding and decoding process of the current round, the computer device can perform hierarchical encoding processing on the first encoding feature and the second encoding feature through at least one encoding layer with spatial resolution levels in the first sub-network model to obtain a first encoding processing result. Based on the first encoding processing result, the computer device can perform at least one feature fusion processing on the first encoding feature and the first encoding processing result through the connection layer in the first sub-network model to obtain a first intermediate processing result. Furthermore, the computer device can perform hierarchical decoding processing on the first intermediate processing result and the output of each hierarchical encoding processing through at least one decoding layer with spatial resolution levels in the first sub-network model to obtain the first encoding processing result.

[0123] In some embodiments, the voxel spatial resolution corresponding to the coding layer decreases as the level increases. The first sub-network model may include a preset number of coding layers, such as 5 coding layers. In the order of hierarchical coding, the voxel spatial resolutions may be 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 10 3 ,32 3 ,16 3 ,8 3 ,4 3 The voxel spatial resolution corresponding to the decoding layer will increase with the increase of the level. The first sub-network model can include a preset number of decoding layers. For 5 layers of decoding layers, the voxel spatial resolution can be set to 4 or 5 in the order of hierarchical decoding. 3 ,8 3 ,16 3 ,32 3 ,64 3 .

[0124] In some embodiments, reference is made to FIG6 , which shows a flowchart for obtaining a three-dimensional rough model. In FIG6 , a computer device may perform N first processes based on a first subnetwork model, where N is greater than or equal to 1. The first first process may include two first encoding and decoding processes, and the Nth first process may also include two first encoding and decoding processes. For the first first encoding and decoding process, the input is a first encoding feature obtained in round T, determined based on the first generation strength, and a second encoding feature determined based on the generation seed S. The output is encoding and decoding output 1. During the second first encoding and decoding process of the first process, the first encoding feature remains unchanged, and the second encoding feature is determined by encoding and decoding output 1 of the same round of first encoding and decoding. For the Nth first process, the input data for the first first encoding and decoding process is the first encoding feature determined based on round TK, and the second encoding feature determined based on the previous first encoding and decoding output. The output is encoding and decoding output M, where T and K are both greater than or equal to 0, and K is less than or equal to T. During the second first encoding and decoding process, the first encoding feature remains unchanged, and the second encoding feature is determined by encoding and decoding output M of the same round of first encoding and decoding. The computer device can perform a three-dimensional convolution operation on the codec output N to obtain a three-dimensional rough model.

[0125] In the above embodiment, the computer device performs at least one round of first processing using the first sub-network model and based on the first generation strength and the generation seed to obtain a 3D rough model. The 3D rough model can reflect the general shape of the 3D object model to be generated. Subsequent optimization can then be performed based on the generated 3D rough model to improve the accuracy of 3D model generation.

[0126] In some embodiments, the step of determining the first coding feature of the current round includes: determining the value to be encoded in the current round according to the round number of the current round; expanding the value to be encoded into a feature vector of a preset dimension; performing at least one linear transformation on the feature vector of the preset dimension to obtain the first coding feature.

[0127] The value to be encoded is the encoding data determined by the round number of the current round, and specifically, the value corresponding to the round number is determined as the value to be encoded.

[0128] Specifically, the computer device determines a value to be encoded based on the current round, and performs dimension expansion on the value to be encoded to expand the value to be encoded into a high-dimensional feature vector. The dimension of the feature vector can be preset, such as 32 dimensions, 64 dimensions, 128 dimensions, or any other dimension. The computer device can further perform at least one linear transformation on the high-dimensional feature vector to obtain a first encoding feature.

[0129] In some embodiments, when encoding a numerical value to be encoded into a high-dimensional feature vector, the computer device may expand it according to the following formula:

[0130] in, is the 16-dimensional floating point vector obtained through training, sin and cos are sine and cosine functions respectively. is a 33-dimensional feature vector. After further passing through two layers of linear neural network, a high-dimensional feature vector is obtained.

[0131] Among them, SiLU is the activation function, linear1 and linear2 are R 33 →R 128 and R 128 →R 128 Linear neural network, high-dimensional feature vector This is the first coding feature.

[0132] In the above embodiment, the computer device determines the value to be encoded in the current round based on the round number of the current round, expands the value to be encoded into a feature vector of a preset dimension, and performs at least one linear transformation on the feature vector of the preset dimension, so that the obtained first encoding feature meets the input requirements of the first sub-network model and improves the model processing accuracy of the first sub-network model.

[0133] In some embodiments, the step of determining the first second coding feature in the first round includes: using a Gaussian function to generate a first tensor of a preset shape based on a generated seed; performing three-dimensional convolution processing on the first tensor of the preset shape to obtain a second coding feature in a three-dimensional space.

[0134] Specifically, the first tensor is used to describe the shape of the generated seed in the multidimensional space. For the first round of determining the second coding feature, the computer device can generate the seed coding 3D model generation space to obtain the second coding feature.

[0135] In some embodiments, when encoding a 3D model generation space, the computer device may specifically use a Gaussian function to generate a floating-point tensor of shape [1, T0, T0, T0] based on a generated seed. All elements are between 0 and 1. Furthermore, the computer device can process floating point tensors. After a layer of three-dimensional convolutional network, the feature encoding of the generated space is obtained

[0136] Among them, CNN 3Dis a standard three-dimensional convolution operation, BN is batch normalization, is a floating-point tensor of shape [32, T0, T0, T0], T0 = 64, This is the second coding feature for the first time in the first round.

[0137] In the above implementation, the computer device uses a Gaussian function to generate a first tensor of a preset shape based on a generated seed, performs three-dimensional convolution processing on the first tensor of the preset shape, and obtains a second coding feature in a three-dimensional space. This can make the obtained second coding feature meet the input requirements of the first sub-network model and improve the model processing accuracy of the first sub-network model.

[0138] It should be noted that, referring to FIG7 , FIG7 is an illustration of the model structure of the first sub-network model. The specific encoding process, fusion process, and decoding process will be introduced later. Among them, in the first sub-network model shown in FIG7 , is the first coding feature, is the second coding feature. The D0, D1, D2, D3, and D4 modules are encoding layers, M0, M1, M2, and M3 are connection layers, and the U0, U1, U2, U3, and U4 modules are decoding layers. It is worth noting that the D4, M0, M1, M2, M3, and U0 modules do not have sample operations, and the rest of the operations are the same; the Sample operation in the U1, U2, U3, and U4 modules represents the upsampling operation of the three-dimensional convolution; the Sample operation in the D0, D1, D2, D3, and D4 modules represents the downsampling operation of the three-dimensional convolution. Assume that the operation result obtained by module U0 is Based on this result, a layer of standard 3D convolution is connected to output the final 3D rough model.

[0139] In some embodiments, based on the first coding feature and the second coding feature, at least one hierarchical coding process is performed according to different spatial resolution levels to obtain a first coding process result, including: obtaining the input model feature of the current coding layer; wherein the input model feature of the first coding layer is the second coding feature, and the input model feature of the non-first coding layer is the output of the previous coding layer; fusing the first coding feature and the input model feature of the current coding layer to obtain a first fusion result, and downsampling the first fusion result to obtain the output of the current coding layer to complete the hierarchical coding process of the current coding layer; continuing the hierarchical coding process of the next coding layer until the last layer is processed and the first fusion result of the last coding layer is used as the output of the last coding layer; the output of the last coding layer is the first coding process result.

[0140] For any first encoding / decoding process in any current round, a fusion process is performed based on at least one coding layer. During the fusion process, the input data of the coding layer may include input model features and first coding features. The input model features of the first coding layer are second coding features determined by the generated seed, and the input model features of non-first coding layers are determined by the output of the previous coding layer.

[0141] Specifically, the computer device can set a feature fusion module in the coding layer, and perform feature fusion processing on the input model features and the first coding features based on the feature fusion module of the coding layer to obtain a first fusion result. The first fusion result is further down-sampled to obtain the output of the current coding layer. The down-sampling process can save computing resources and improve the data processing efficiency of the computer device. When multiple coding layers are included, the computer device will perform hierarchical coding processing on the next coding layer until the feature fusion processing is completed on the last coding layer, and the output of the last coding layer is obtained. The first fusion result output of the last coding layer is determined as the first coding processing result.

[0142] In some embodiments, the feature fusion module may be a computing module set based on a feature fusion algorithm, and the feature fusion algorithm may include ResNet (residual network), FPN (feature pyramid networks, target detection algorithm), and FSSD (Feature Fusion Single Shot Multibox Detector, improved lightweight feature fusion algorithm).

[0143] In some embodiments, the computer device can perform feature fusion processing based on ResNet. Please refer to Figure 8, which is a flowchart of the fusion processing based on ResNet in any first encoding and decoding process. The input data of ResNet may include input model features. and the first coding feature The first coding feature can be Perform a three-dimensional convolution operation to obtain the three-dimensional model convolution operation result F1 corresponding to the first coding feature. The input model feature can be processed by the B0 unit. A three-dimensional convolution operation is performed to obtain a three-dimensional model convolution operation result F0 corresponding to the input model feature. The computer device may further perform feature fusion on F0 and F1 to obtain an initial fusion result B2 of F1 and F2, and perform a three-dimensional convolution operation on the input model feature through the B2 unit to obtain a three-dimensional model convolution operation result F3 corresponding to the input model feature. The service may fuse the initial fusion result B2 with the three-dimensional model convolution operation result F3 to obtain a first fusion result. The computer device may downsample the first fusion result to obtain the output model feature, i.e., the output of the current coding layer.

[0144] In some embodiments, the ResNet module involves the following computational process:

[0145] in, For tensor addition operations that follow the tensor propagation mechanism, SiLU is the activation function, linear is the linear neural network, and BN is batch normalization.

[0146] In the above embodiment, the computer device can encode the input model features and the first encoding features through at least one encoding layer of the first sub-network model to obtain a first fusion result of the encoding layer, and further perform downsampling processing on the first fusion result to obtain the output of the current encoding layer. Through downsampling processing, computing resources can be saved and the data processing efficiency of the computer device can be improved.

[0147] In some embodiments, based on the first encoding processing result and the first encoding feature, at least one fusion process is performed to obtain a first intermediate processing result, including: obtaining the input model feature in the current fusion process; wherein, the input model feature in the first fusion process is the first encoding processing result, and the input model feature in the non-first fusion process is the output of the previous fusion process; fusing the first encoding feature and the input model feature in the current fusion process to obtain a second fusion result to complete the current fusion process; continuing the next fusion process until the last fusion process is completed, and using the second fusion result of the last fusion process as the first intermediate processing result.

[0148] Among them, for any first encoding and decoding process in any current round, fusion processing will be performed based on at least one connection layer. When fusion processing is performed based on the connection layer, the input data of the connection layer may include input model features and first encoding features. The input model features of the first connection layer are the first encoding processing results, and the input model features of non-first connection layers are determined by the output after fusion processing of the previous connection layer.

[0149] Specifically, the computer device can set a feature fusion module in the connection layer, and perform feature fusion processing on the input model features and the first coding features based on the feature fusion module of the connection layer to obtain a first intermediate processing result. When multiple connection layers are included, the computer device will perform fusion processing on the next connection layer when completing the current fusion processing of the current connection layer, until the feature fusion of the last connection layer is completed, and the output of the fusion processing of the last connection layer is obtained, and the second fusion result of the output of the last connection layer is used as the first intermediate processing result. When the feature fusion module based on the connection layer performs fusion processing on the input model features and the first coding features, the computer device may not perform sampling processing on the output of the fusion processing, so that feature processing can be performed at a specific resolution, thereby improving the accuracy of feature processing.

[0150] In some embodiments, the feature fusion module provided in the connection layer may use the same feature fusion algorithm as the feature fusion module provided in the coding layer, or may use different feature fusion algorithms. For example, the feature fusion module provided in the connection layer is a calculation module based on ResNet.

[0151] In the above embodiment, the computer device can fuse the input model features and the first encoding features through at least one connection layer of the first sub-network model to obtain a second fusion result of the connection layer. Sampling processing may not be performed on the second fusion result, so that feature processing can be performed at a specific resolution through the connection layer, thereby improving the accuracy of feature processing.

[0152] In some embodiments, based on the first intermediate processing result and the output of each hierarchical encoding process, at least one hierarchical decoding process is performed according to different spatial resolution levels to obtain the output of the first encoding and decoding process of the time, including: obtaining the input model features of the current decoding layer; wherein the input model features of the first decoding layer are the first intermediate processing results, and the input model features of the non-first decoding layer include the output of the previous decoding layer and the output of the encoding layer corresponding to the same resolution; the first encoding features and the input model features of the current decoding layer are fused to obtain a third fusion result, and the third fusion result is upsampled to obtain the output of the current decoding layer to complete the hierarchical decoding process of the current decoding layer; the hierarchical decoding process of the next decoding layer is continued until the last layer is processed and the third fusion result of the last decoding layer is used as the output of the last decoding layer; the output of the last decoding layer is the output of the first encoding and decoding process of the time.

[0153] For any first encoding / decoding process in any current round, fusion processing is performed based on at least one decoding layer. During the fusion processing, the input data of the decoding layer may include input model features and first encoding features. The input model features of the first decoding layer are the first intermediate processing results. The input model features of non-first encoding layers include the output of the previous decoding layer and the output of the encoding layer corresponding to the same resolution.

[0154] Specifically, the computer device can set a feature fusion module at the decoding layer, and perform feature fusion processing on the input model features and the first encoding features based on the feature fusion module of the decoding layer to obtain a third fusion result. The third fusion result is further up-sampled to obtain the output of the current decoding layer. The up-sampling process can increase the resolution and improve the accuracy of three-dimensional model generation. When multiple decoding layers are included, the computer device will perform hierarchical decoding processing on the next decoding layer until the feature fusion processing is completed on the last decoding layer, and the output of the last decoding layer is obtained. The third fusion result output by the last decoding layer is determined as the output of the first encoding and decoding process at that time.

[0155] In some embodiments, the feature fusion module provided in the connection layer may use the same feature fusion algorithm as the feature fusion modules provided in the encoding layer and the intermediate layer, or may use different feature fusion algorithms. For example, the feature fusion module provided in the decoding layer is a calculation module based on ResNet.

[0156] In the above embodiment, the computer device can decode the input model features and the first encoding features through at least one decoding layer of the first sub-network model to obtain a third fusion result of the decoding layer, and further perform upsampling processing on the third fusion result to obtain the output of the current encoding layer. Through upsampling processing, the resolution can be increased and the accuracy of three-dimensional model generation can be improved.

[0157] In some embodiments, at least one round of second processing is performed through a second sub-network model and based on the second generation intensity and the three-dimensional rough model to obtain a three-dimensional object model that meets a preset high-resolution condition and belongs to an object category, including: determining a second number based on the second generation intensity, performing a second processing of the second number through the second sub-network model, and obtaining a three-dimensional object model based on the output of the last second processing.

[0158] Among them, the second number is the number of times the second processing is performed through the second sub-network model. Generally speaking, the larger the value of the second number, the more times the second processing is performed through the second sub-network model, and the better the quality of the obtained three-dimensional rough model and the higher the accuracy.

[0159] Specifically, the computer device determines the second number of times to perform the second processing based on the second generation strength, so that the second sub-network model can be iteratively processed according to the second number, and the output of the last second processing is determined as the three-dimensional object model.

[0160] In some embodiments, the computer device determines, based on the second generation strength, that the second number is 3, indicating that the first sub-network model needs to be subjected to the second processing three times, including the first second processing, the second second processing, and the third second processing, and each second processing includes at least two second encoding and decoding processes. The computer device may obtain the three-dimensional object model based on the output of the third second processing.

[0161] In some embodiments, the second generation intensity can be a numerical value, and the computer device can directly use the second generation intensity as the second number, or perform a mathematical operation on the second generation intensity to obtain the second number. The mathematical operation performed here can include doubling, adding a preset value, etc. The preset value can be, for example, 1, 2, or other numerical values.

[0162] In one possible implementation, for example, if the second generation strength is 1, the computer device may add 1 to 1 to obtain a second number 2. When the second number is 2, it indicates that the second sub-network model needs to perform the second processing twice, i.e., including the first processing for the first time and the second processing for the second time, and each second processing includes at least two first encoding and decoding processes.

[0163] In some embodiments, the computer device may perform format conversion processing on the output of the last second processing to obtain a triangular mesh model that meets high-resolution conditions and belongs to an object category. The computer device may determine the obtained triangular mesh model as a three-dimensional object model.

[0164] In some embodiments, a three-dimensional object model is obtained based on the output of the last second processing, including: performing three-dimensional convolution processing on the output of the last second processing to obtain the output of the second sub-network model; converting the output of the second sub-network model into a triangular mesh model to obtain a three-dimensional object model belonging to the object category; the three-dimensional object model meets the preset high-resolution conditions.

[0165] Specifically, the computer device can perform a three-dimensional convolution process on the output of the last second processing to transform the dimension of the feature vector of the last second processing output to obtain the output of the second sub-network model. The computer device can then use a surface drawing algorithm, such as a contour connection method, a cube algorithm, a decomposition cube algorithm, and a marching cube algorithm to convert the output of the second sub-network model into a triangular mesh model, thereby obtaining a final three-dimensional object model that meets the preset high-resolution conditions based on the triangular mesh model. When the computer device performs convolution processing on the second processing output, it can be processed based on three-dimensional convolution, depth-separable convolution, etc., or it can be replaced by other deep learning operators.

[0166] In the above embodiment, the computer device performs three-dimensional convolution processing on the output of the last second processing to obtain the output of the second sub-network model, and then converts the output of the second sub-network model into a triangular mesh model, and obtains a three-dimensional object model based on the three-dimensional mesh model, so as to help users better observe and understand specific object concepts.

[0167] In some embodiments, each round of second processing includes at least two second encoding and decoding processes, and each second encoding and decoding process in any current round includes the following steps: obtaining the third encoding feature in the second encoding and decoding process; wherein, the third encoding feature of each time in the current round is the same and is determined according to the round corresponding to the current round; obtaining the fourth encoding feature in the second encoding and decoding process; wherein, the fourth encoding feature for the first time in the first round is determined according to the three-dimensional rough model, and the fourth encoding feature for the first time in the non-first round is the output of the previous second encoding and decoding process; based on the third encoding feature and the fourth encoding feature, at least one hierarchical encoding process is performed according to different spatial resolution levels to obtain the second encoding process result; based on the second encoding process result and the third encoding feature, at least one fusion process is performed to obtain the second intermediate processing result; based on the second intermediate processing result and the output of each hierarchical encoding process, at least one hierarchical decoding process is performed according to different spatial resolution levels to obtain the output of the second encoding and decoding process.

[0168] The second sub-network model may include at least one coding layer, and the hierarchical coding process is to perform coding processing on each coding layer. For each coding layer, there will be a corresponding output spatial resolution level. The output spatial resolution level is the voxel spatial resolution set for each coding layer. For different coding layers, the voxel spatial resolution can be different or the same. The setting of the voxel spatial resolution can be adaptively set based on the generation accuracy requirements of the 3D model and the computing resources of the computer equipment.

[0169] The second sub-network model may further include at least one decoding layer, wherein the hierarchical decoding process is performed on each decoding layer. Different decoding layers also correspond to a corresponding spatial resolution, and the voxel spatial resolutions of different decoding layers may be different or the same.

[0170] At least one connection layer may be provided between the encoding layer and the decoding layer of the second sub-network model, and each connection layer may be used to connect the encoding layer and the decoding layer. The second intermediate processing result is determined by fusing the second encoding processing result and the third encoding feature based on the connection layer.

[0171] Specifically, the computer device can determine a second number based on the second generation strength and set a round for performing the second processing based on the second number. During any second encoding / decoding process of the current round, the computer device can determine the third encoding feature based on the round corresponding to the current round. Regarding the fourth encoding feature, if the current round is the first round and the second encoding / decoding process is the first in the first round, the computer device can determine it using a three-dimensional rough model; if it is not the first in the first round, the computer device can determine it based on the output of the previous second encoding / decoding process.

[0172] Then, for any second encoding and decoding process of the current round, the computer device may perform hierarchical encoding processing on the third encoding feature and the fourth encoding feature using at least one encoding layer having spatial resolution levels in the second sub-network model to obtain a second encoding processing result. Based on the second encoding processing result, the computer device may perform at least one feature fusion processing on the third encoding feature and the second encoding processing result using a connection layer in the second sub-network model to obtain a second intermediate processing result.

[0173] Furthermore, the computer device may perform hierarchical decoding processing on the second intermediate processing result and the output of each hierarchical encoding process by providing at least one decoding layer with spatial resolution levels in the second sub-network model, thereby obtaining a second encoding processing result. The spatial resolution level of the encoding layer may match the resolution level of the decoding layer.

[0174] In some embodiments, the voxel spatial resolution corresponding to the coding layer decreases as the level increases. The second sub-network model may include a preset number of coding layers, such as 4 coding layers. In the order of hierarchical coding, the voxel spatial resolutions may be 128, ... 3 ,64 3 ,32 3 ,16 3 ,8 3The voxel spatial resolution corresponding to the decoding layer will increase with the increase of the level. The second sub-network model can include a preset number of decoding layers. For 5 layers of decoding layers, the voxel spatial resolution can be set to 8 or 100 in the order of hierarchical decoding. 3 ,16 3 ,32 3 ,64 3 ,128 3 The voxel spatial resolution of the second sub-network model can be set higher than the voxel spatial resolution of the first sub-network model, so that a more accurate three-dimensional object model can be obtained.

[0175] In some embodiments, the steps for determining the fourth coding feature for the first time in the first round include: upsampling the three-dimensional rough model to obtain a second tensor; taking the elements in the second tensor whose values ​​are greater than a preset value as the first element, and taking the elements in the second tensor other than the first element as the second element; replacing the value of the first element with a randomly generated random number to obtain an updated second tensor; performing three-dimensional convolution processing on the updated second tensor to obtain the fourth coding feature in the three-dimensional space; wherein, the second element in the updated second tensor does not participate in the three-dimensional convolution processing.

[0176] The second tensor is used to represent the shape of the three-dimensional rough model in multidimensional space. The preset value is a pre-set value used to determine whether each element in the second tensor participates in the three-dimensional convolution process. The random number is a Gaussian parameter generated based on a random seed, such as a Gaussian floating-point number, and the random number can range between 0 and 1.

[0177] Specifically, for the determination of the fourth coding feature for the first time in the first round, the computer device can upsample through a three-dimensional rough model to obtain a second tensor, take the elements in the second tensor whose values ​​are greater than a preset value as the first element, take the elements in the second tensor other than the first element as the second element, replace the value of the first element with a randomly generated random number, obtain an updated second tensor, perform three-dimensional convolution processing on the updated second tensor, and obtain the fourth coding feature in the three-dimensional space.

[0178] In some embodiments, the computer device may use the following formula to upsample the three-dimensional rough model:

[0179] Among them, UpSample is the upsampling operation in three-dimensional Euclidean space, and Rand performs the following operations on all elements of the tensor:

[0180] Among them, gr(s c ) represents the random seed s cA random Gaussian floating-point number between 0 and 1 is generated, and -1 indicates that no operation is performed in the corresponding area. When an element in the second tensor is greater than 0, the computer device can use the element as the first element and replace the first element based on the random Gaussian floating-point number. The computer device can also determine the element in the second tensor that is less than or equal to 0 as the second element and replace the second element with -1. After completing the replacement process of the elements in the second tensor, the computer device can use the same operation process as that for calculating the second coding feature to calculate the fourth coding feature based on the random Gaussian floating-point number in the second tensor.

[0181] In this embodiment, the computer device obtains a second tensor by upsampling through a three-dimensional rough model, takes the elements in the second tensor whose values ​​are greater than a preset value as the first element, and takes the elements in the second tensor other than the first element as the second element. The value of the first element is replaced by a randomly generated random number, thereby giving Gaussian noise to the first element and encoding it, which can effectively save computing resources.

[0182] In some embodiments, based on the third coding feature and the fourth coding feature, at least one hierarchical coding process is performed according to different spatial resolution levels to obtain a second coding process result, including: obtaining the input model feature of the current coding layer; wherein the input model feature of the first coding layer is the fourth coding feature, and the input model feature of the non-first coding layer is the output of the previous coding layer; fusing the third coding feature and the input model feature of the current coding layer to obtain a fourth fusion result, and downsampling the fourth fusion result to obtain the output of the current coding layer to complete the hierarchical coding process of the current coding layer; continuing the hierarchical coding process of the next coding layer until the last layer is processed and the fourth fusion result of the last coding layer is used as the output of the last coding layer; the output of the last coding layer is the second coding process result.

[0183] For any second encoding / decoding process in any current round, fusion processing is performed based on at least one coding layer. During the fusion processing, the input data of the coding layer may include input model features and third coding features. The input model features of the first coding layer are fourth coding features determined by the three-dimensional rough model. The input model features of non-first coding layers are determined by the output of the previous coding layer.

[0184] Specifically, the computer device can set a feature fusion module in the coding layer, and perform feature fusion processing on the input model features and the third coding features based on the feature fusion module of the coding layer to obtain a fourth fusion result. The fourth fusion result is further downsampled to obtain the output of the current coding layer. The downsampling process can save computing resources and improve the data processing efficiency of the computer device. When multiple coding layers are included, the computer device will perform hierarchical coding processing on the next coding layer until the feature fusion processing is completed on the last coding layer, and the output of the last coding layer is obtained. The fourth fusion processing of the output of the last coding layer is determined as the second coding processing result.

[0185] In the above embodiment, the computer device can encode the input model features and the third encoding features through at least one encoding layer of the second sub-network model to obtain a fourth fusion result of the encoding layer, and further perform downsampling processing on the fourth fusion result to obtain the output of the current encoding layer. Through downsampling processing, computing resources can be saved and the data processing efficiency of the computer device can be improved.

[0186] In some embodiments, based on the second encoding processing result and the third encoding feature, at least one fusion process is performed to obtain a second intermediate processing result, including: obtaining the input model feature in the current fusion process; wherein, the input model feature in the first fusion process is the second encoding processing result, and the input model feature in the non-first fusion process is the output of the previous fusion process; fusing the third encoding feature and the input model feature in the current fusion process to obtain a fifth fusion result to complete the current fusion process; continuing the next fusion process until the last fusion process is completed, and taking the fifth fusion result of the last fusion process as the second intermediate processing result.

[0187] Among them, for any second encoding and decoding process in any current round, fusion processing will be performed based on at least one connection layer. When fusion processing is performed based on the connection layer, the input data of the connection layer may include input model features and third encoding features. The input model features of the first connection layer are the results of the second encoding processing, and the input model features of the non-first connection layer are determined by the output after the fusion processing of the previous connection layer.

[0188] Specifically, the computer device can set a feature fusion module in the connection layer, and perform feature fusion processing on the input model features and the third coding features based on the feature fusion module of the connection layer to obtain a second intermediate processing result. When multiple connection layers are included, the computer device will perform fusion processing on the next connection layer when completing the current fusion processing of the current connection layer, until the feature fusion of the last connection layer is completed, and the output of the fusion processing of the last connection layer is obtained, and the fifth fusion result output of the last connection layer is used as the second intermediate processing result. When the feature fusion module based on the connection layer performs fusion processing on the input model features and the third coding features, the computer device may not perform sampling processing on the output of the fusion processing, so that feature processing can be performed at a specific resolution, thereby improving the accuracy of feature processing.

[0189] In the above embodiment, the computer device can fuse the input model features and the third encoding features through at least one connection layer of the second sub-network model to obtain a fifth fusion result of the connection layer. Sampling processing may not be performed on the fifth fusion result, so that feature processing can be performed at a specific resolution, thereby improving the accuracy of feature processing.

[0190] In some embodiments, based on the second intermediate processing result and the output of each hierarchical encoding process, at least one hierarchical decoding process is performed according to different spatial resolution levels to obtain the output of the second encoding and decoding process of the time, including: obtaining the input model features of the current decoding layer; wherein the input model features of the first decoding layer are the second intermediate processing results, and the input model features of the non-first decoding layer include the output of the previous decoding layer and the output of the encoding layer corresponding to the same resolution; the third encoding features and the input model features of the current decoding layer are fused to obtain a sixth fusion result, and the sixth fusion result is upsampled to obtain the output of the current decoding layer to complete the hierarchical decoding process of the current decoding layer; the hierarchical decoding process of the next decoding layer is continued until the last layer is processed and the sixth fusion result of the last decoding layer is used as the output of the last decoding layer; the output of the last decoding layer is the output of the second encoding and decoding process of the time.

[0191] For any second encoding and decoding process in any current round, fusion processing is performed based on at least one decoding layer. During the fusion processing, the input data of the decoding layer may include the input model features and the third encoding features. The input model features of the first decoding layer are the second intermediate processing results. The input model features of non-first encoding layers include the output of the previous decoding layer and the output of the encoding layer corresponding to the same resolution.

[0192] Specifically, the computer device can set a feature fusion module at the decoding layer, and perform feature fusion processing on the input model features and the third encoding features based on the feature fusion module of the decoding layer to obtain a sixth fusion result. The sixth fusion result is further up-sampled to obtain the output of the current decoding layer. The up-sampling process can increase the resolution and improve the accuracy of three-dimensional model generation. When multiple decoding layers are included, the computer device will perform hierarchical decoding processing on the next decoding layer until the feature fusion processing is completed on the last decoding layer, and the output of the last decoding layer is obtained. The sixth fusion result output of the last decoding layer is determined as the output of the second encoding and decoding process at that time.

[0193] In the above embodiment, the computer device can decode the input model features and the third encoding features through at least one decoding layer of the second sub-network model to obtain a third fusion result of the decoding layer, and further perform upsampling processing on the third fusion result to obtain the output of the current encoding layer. Through upsampling processing, the resolution can be increased and the accuracy of three-dimensional model generation can be improved.

[0194] In some embodiments, the training steps of the first sub-network model include: obtaining multiple three-dimensional sample models belonging to the object category, and normalizing each three-dimensional sample model to a preset vector space to obtain a standardized sample; obtaining first training data corresponding to a first resolution based on the standardized sample conversion; adding noise to the first training data to obtain a first input sample; determining a first training iteration number, and determining the first intensity sample in any round of training of the first sub-network model by the first training iteration number; performing a first iterative training on the first sub-network model in the network model to be trained using the first training data, the first input sample and the first intensity sample, and obtaining a trained first sub-network model at the end of the training.

[0195] The 3D sample models used in the model training phase can be models of various object types. They are composed of 3D parameters such as vertices and connecting lines that describe the 3D properties of the object. Multiple 3D sample models can be used for each object type.

[0196] The preset vector space is a Euclidean space, and each three-dimensional sample model is normalized to the Euclidean space to obtain a normalized sample. The first training data of the first resolution is obtained by transforming the normalized sample.

[0197] Specifically, for each object category, the computer device can obtain multiple three-dimensional sample models of each object category, perform standardization on any three-dimensional sample model, standardize the three-dimensional sample model into the Euclidean space, and obtain a standardized triangular mesh. The computer device can convert the standardized triangular mesh into first training data of a first resolution. Furthermore, the computer device can add various types of noise to the first training data to obtain a first input sample. The computer device can then perform a first iterative training on the first sub-network model in the network model to be trained based on the determined first intensity sample, the first training data, and the first input sample, and terminate the training when the training stop condition is reached to obtain a first sub-network model that has been trained. Among them, the training stop condition can specifically be that a preset number of iterations is reached, the error between the predicted value and the target value is less than an error threshold, or the training time reaches a preset time, etc., and the embodiments of the present application do not limit this.

[0198] It should be noted that during each training process, the network model can be iterated at least twice, that is, at least two encoding and decoding processes are performed, which can effectively improve the robustness of the model.

[0199] In some embodiments, the network model to be trained may be a generative model, such as a diffusion model, a variational autoencoder model, and the like.

[0200] In some embodiments, the first resolution and format of the first training data can be adaptively adjusted according to actual 3D model generation requirements, scenarios, etc. For example, the first resolution can be 128 ^ 3, or 64 ^3 The first training data can be data represented based on a signed distance field or data represented based on a voxel space.

[0201] In some embodiments, in order to obtain the first training data of the first resolution, the computer device may use the Mesh2SDF (Mesh Distance Field) algorithm to convert the normalized triangular mesh into an implicit resolution of 128 ^ 3 signed distance field.

[0202] In some embodiments, as shown in FIG9 , it is a flowchart illustrating a training process of training the first sub-network model.

[0203] The computer device obtains first training data and adds various types of noise to the first training data to obtain a first input sample. Furthermore, the server may randomly select a natural number and use the selected natural number as a first training iteration number. The server may use the first input sample and the first training iteration number as input data and the first training data as a label to perform a first iterative training on the network model to be trained, thereby obtaining a first output. Furthermore, the server may re-input the first output and the first training iteration number into the network model to perform a second iterative training.

[0204] In some embodiments, the training steps of the second sub-network model include: obtaining multiple three-dimensional sample models belonging to the object category, and normalizing each three-dimensional sample model to a preset vector space to obtain a standardized sample; obtaining second training data corresponding to a second resolution based on the standardized sample conversion; the second resolution is greater than the first resolution, and the first resolution is the resolution of the first training data used to train the first sub-network model; adding noise to the second training data to obtain a second input sample; determining the second training iteration number, and determining the second intensity sample in any round of second sub-network model training process by the second training iteration number; performing second iterative training on the second sub-network model in the network model to be trained through the second training data, the second input sample and the second intensity sample, and obtaining a trained second sub-network model at the end of the training.

[0205] The standardized samples are obtained by normalizing each three-dimensional sample model to the Euclidean space, and the second training data of the second resolution is obtained by transforming the standardized samples.

[0206] In some embodiments, to obtain the second training data of the second resolution, the resolution of 128 3 The SDF (Signed Distance Field) is converted to 64 3 The voxel space V, at this time each voxel V in the voxel space i There are 8 SDF values, which can be expressed as Then V i The strategy for determining whether it is on the triangular mesh surface is as follows: If It is believed that V i Located on the corresponding triangular mesh surface, and the voxel is mapped to 1; otherwise, it is considered that V i Does not belong to the triangulated surface and should have the voxel mapped to 0.

[0207] Specifically, for each object category, the computer device can obtain multiple three-dimensional sample models of each object category, standardize any three-dimensional sample model, standardize the three-dimensional sample model into the Euclidean space, and obtain a standardized triangular mesh. The computer device can convert the standardized triangular mesh into second training data of a second resolution. Furthermore, the computer device can add various types of noise to the second training data to obtain a second input sample. The computer device can then perform a second iterative training on the second sub-network model in the network model to be trained based on the determined second intensity sample, the second training data, and the second input sample, and end the training when the training stop condition is reached to obtain a trained second sub-network model. Among them, the training stop condition can specifically be reaching a preset number of iterations, the error between the predicted value and the target value is less than the error threshold, or the training time reaches a preset time, etc., and the embodiments of the present application do not limit this.

[0208] In some embodiments, the resolution and format of the second training data can be adaptively adjusted according to actual 3D model generation requirements, scenarios, etc. For example, the second resolution can be 128 ^ 3, or 64 ^3 , as long as it is lower than the first resolution. The second training data can be data represented based on the signed distance field or data represented based on the voxel space.

[0209] In the above embodiment, the server can obtain multiple three-dimensional sample models belonging to the object category, and transform the obtained multiple three-dimensional sample models to obtain second training data corresponding to the second resolution, so as to perform second iterative training on the second sub-network model in the network model to be trained based on the second training data, the second input sample and the second intensity sample, which can improve the quality of three-dimensional model generation and significantly improve the success rate of three-dimensional model generation.

[0210] In one embodiment, the present application further provides an application scenario, in which the above-mentioned three-dimensional model generation method is applied. Specifically, the application of the three-dimensional model generation method in this application scenario is as follows:

[0211] Figure 10 shows the 3D model generation network architecture of the 3D model generation method provided in this application. As can be seen from Figure 10, this application involves two networks: a coarse model generation network and a fine model generation network. The coarse model generation network is the first sub-network model, and the fine model generation network is the second sub-network model. A computer device can determine a 3D coarse model based on the coarse model generation network and optimize the 3D coarse model based on the fine model generation network to obtain a high-quality 3D object model.

[0212] The coarse model generation network and the fine model generation network are pre-trained models. During the training phase, computer equipment can collect a large number of three-dimensional models, classify the data and label them, and construct training data for each category of objects. Then, a three-dimensional model generation network is trained separately for each category of objects. The training process is divided into two stages: coarse model generation and fine model generation. The generation networks in both stages can be designed and trained based on the diffusion model to improve the quality of three-dimensional model generation. Using the diffusion generation model to train three-dimensional model generation not only improves the quality of three-dimensional model generation, but also significantly improves the success rate of three-dimensional model generation. Among them, the training of the two-stage generation networks can be carried out simultaneously. Specifically, the loss function used in the training process is:

[0213] Among them, G1 is Gaussian distribution, G2 is uniform distribution, β(t) = e-10t2-10-4, f corresponds to the coarse model generation network and the fine model generation network. In the coarse model generation training, x0 corresponds to a resolution of 64 3 In the voxel space of fine model generation training, x0 corresponds to a resolution of 128 3 The signed distance field, t is the set number of training iterations, is the last prediction result in a training process, x t is the input sample after adding noise. During the model training process, there is a 50% probability that Assign f(x t ,0,t); another 50% probability, Assigns a tensor with the same shape as x0 but with all elements set to 0.

[0214] During the application phase, the user's input modeling text is first matched to the most appropriate category of 3D model generation network. A 3D object model is then generated based on the corresponding 3D model generation network and fed back to the user. Compared to previous completely random 3D model generation methods, the computer device can receive text prompts and obtain 3D models of the same category based on these prompts, improving the quality and success rate of 3D model generation.

[0215] When a computer device generates a three-dimensional object model based on a three-dimensional model generation network, it can determine a three-dimensional coarse model based on a coarse-mode generation network in the three-dimensional model generation network. Specifically, when generating a three-dimensional coarse model, the coarse-mode generation network can be used to perform at least one round of first processing based on a first generation strength and a generation seed to obtain a three-dimensional coarse model. When generating a three-dimensional object model, at least one round of second processing can be performed based on the fine-mode generation network, a second generation strength, and the three-dimensional coarse model to obtain a three-dimensional object model that meets preset high-resolution conditions and belongs to an object category. This method of generating three-dimensional object models can easily support new object category generation tasks and allows each category model generation network to be independently maintained and updated, thereby enhancing the maintainability of the system.

[0216] The three-dimensional model generation method provided in this application can be applied in a variety of scenarios. For example, in games involving three-dimensional models, it can help game makers generate the three-dimensional models they want as game assets, and based on this, create more complex characters, scenes and special effects, thereby speeding up the development of games and improving game production efficiency; it can help players generate three-dimensional models of interest and import them into games for use, thereby enhancing user participation. In product design and manufacturing, it can help designers quickly generate large quantities of three-dimensional models and test and modify them based on the generated models. These models can then be used to guide the production process, such as manufacturing physical products through 3D printing. In virtual and augmented reality, a large number of three-dimensional models can be quickly generated to enrich virtual world scenes, thereby creating more realistic virtual reality (VR) and augmented reality (AR) experiences. In education and training, three-dimensional models of objects of interest to learners can be generated to help learners better observe and understand specific object concepts and enhance learning fun.

[0217] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0218] Based on the same inventive concept, embodiments of the present application also provide a 3D model generation device for implementing the aforementioned 3D model generation method. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more 3D model generation device embodiments provided below can be found in the above-described limitations of the 3D model generation method and will not be further elaborated here.

[0219] In an exemplary embodiment, as shown in FIG11 , a three-dimensional model generation apparatus 1100 is provided, comprising: a network matching module 1102 , an information acquisition module 1104 , a first model generation module 1106 , and a second model generation module 1108 , wherein:

[0220] The network matching module 1102 is configured to obtain prompt information describing an object category and determine a target network model that matches the object category indicated by the prompt information; the target network model includes a first sub-network model and a second sub-network model.

[0221] The information acquisition module 1104 is used to obtain the first generation strength, the second generation strength, and the generation seed.

[0222] The first model generation module 1106 is configured to perform at least one round of first processing based on the first generation strength and the generation seed through the first sub-network model to obtain a three-dimensional rough model; wherein each round of first processing includes at least one first encoding and decoding process.

[0223] The second model generation module 1108 is used to perform at least one round of second processing through the second sub-network model and based on the second generation strength and the three-dimensional rough model to obtain a three-dimensional object model that meets the preset high-resolution conditions and belongs to the object category; wherein each round of second processing includes at least one second encoding and decoding process.

[0224] In one embodiment, the network matching module 1102 is also used to determine the object category indicated by the prompt information; obtain a target network model that matches the object category from the network model pool; the network model pool includes multiple trained network models for generating three-dimensional object models of various categories, and the target network model is used to generate a three-dimensional object model belonging to the object category indicated by the prompt information.

[0225] In one embodiment, the network matching module 1102 includes a category determination module; the category determination module is used to obtain the feature vectors of each of a plurality of preset object category names to obtain a first feature set; vectorize the prompt information to determine at least one feature vector related to the prompt information to obtain a second feature set; calculate the similarity between each feature vector in the second feature set and each feature vector in the first feature set; and determine the object category to which the feature vector in the first feature set corresponding to the feature vector that meets the preset similarity condition and has the greatest similarity belongs as the object category indicated by the prompt information.

[0226] In one embodiment, the network matching module 1102 further includes a feature set determination module; the feature set determination module is used to perform word segmentation processing on the prompt information to obtain multiple word units in the prompt information; and extract the feature vectors of the prompt information and each word unit respectively to obtain a second feature set.

[0227] In one embodiment, the first model generation module 1106 is further used to determine a first number based on the first generation strength, perform a first processing of the first number through the first sub-network model, and obtain a three-dimensional rough model based on the output of the last first processing; wherein, each round of first processing includes at least two first encoding and decoding processes, and any first encoding and decoding process in any current round includes the following steps: obtaining a first encoding feature in the first encoding and decoding process of the current round; wherein, the first encoding feature of each time in the current round is the same and is determined according to the round corresponding to the current round; obtaining a second encoding feature in the first encoding and decoding process of the current round; wherein, the first second encoding feature of the first round is determined according to the generation seed, and the second encoding feature of the non-first round is the output of the previous first encoding and decoding process; based on the first encoding feature and the second encoding feature, at least one hierarchical encoding process is performed according to different spatial resolution levels to obtain a first encoding process result; based on the first encoding process result and the first encoding feature, at least one fusion process is performed to obtain a first intermediate processing result; based on the first intermediate processing result and the output of each hierarchical encoding process, at least one hierarchical decoding process is performed according to different spatial resolution levels to obtain the output of the first encoding and decoding process of the current round.

[0228] In one embodiment, the three-dimensional model generation device also includes a first coding feature determination module; the first coding feature determination module is used to determine the value to be encoded in the current round according to the round number of the current round; expand the value to be encoded into a feature vector of a preset dimension; perform at least one linear transformation on the feature vector of the preset dimension to obtain the first coding feature.

[0229] In one embodiment, the three-dimensional model generation device also includes a second coding feature determination module; the second coding feature determination module is used to use a Gaussian function to generate a first tensor of a preset shape based on a generation seed; and perform three-dimensional convolution processing on the first tensor of the preset shape to obtain a second coding feature in a three-dimensional space.

[0230] In one embodiment, the first model generation module 1106 is also used to obtain the input model features of the current coding layer; wherein, the input model features of the first coding layer are the second coding features, and the input model features of the non-first coding layer are the output of the previous coding layer; the first coding features and the input model features of the current coding layer are fused to obtain a first fusion result, and the first fusion result is downsampled to obtain the output of the current coding layer to complete the hierarchical coding processing of the current coding layer; the hierarchical coding processing of the next coding layer is continued until the last layer is processed and the first fusion result of the last coding layer is used as the output of the last coding layer; the output of the last coding layer is the first coding processing result.

[0231] In one embodiment, the first model generation module 1106 is also used to obtain the input model features in the current fusion process; wherein, the input model features in the first fusion process are the first encoding processing results, and the input model features in the non-first fusion process are the output of the previous fusion process; the first encoding features and the input model features in the current fusion process are fused to obtain a second fusion result to complete the current fusion process; the next fusion process is continued until the last fusion process is completed, and the second fusion result of the last fusion process is used as the first intermediate processing result.

[0232] In one embodiment, the first model generation module 1106 is also used to obtain the input model features of the current decoding layer; wherein, the input model features of the first decoding layer are the first intermediate processing results, and the input model features of the non-first decoding layer include the output of the previous decoding layer and the output of the encoding layer corresponding to the same resolution; the first encoding features and the input model features of the current decoding layer are fused to obtain a third fusion result, and the third fusion result is downsampled to obtain the output of the current decoding layer to complete the hierarchical decoding processing of the current decoding layer; the hierarchical decoding processing of the next decoding layer is continued until the last layer is processed and the third fusion result of the last decoding layer is used as the output of the last decoding layer; the output of the last decoding layer is the output of the first encoding and decoding processing of the current time.

[0233] In one embodiment, the second model generation module 1108 is further configured to determine a second number based on the second generation strength, perform a second processing on the second number through a second sub-network model, and obtain a three-dimensional object model based on the output of the last second processing; wherein each round of the second processing includes at least two second encoding and decoding processes, and each second encoding and decoding process in any current round includes the following steps: obtaining a third encoding feature in the second encoding and decoding process of the current round; wherein the third encoding feature of each current round is the same and is determined according to the round corresponding to the current round; obtaining a fourth encoding feature in the second encoding and decoding process of the current round; wherein the fourth encoding feature for the first time in the first round is determined according to the three-dimensional rough model, and the fourth encoding feature for the first time in a non-first round is the output of the previous second encoding and decoding process; based on the third encoding feature and the fourth encoding feature, at least one hierarchical encoding process is performed according to different spatial resolution levels to obtain a second encoding process result; based on the second encoding process result and the third encoding feature, at least one fusion process is performed to obtain a second intermediate processing result; based on the second intermediate processing result and the output of each hierarchical encoding process, at least one hierarchical decoding process is performed according to different spatial resolution levels to obtain the output of the second encoding and decoding process of the current round.

[0234] In one embodiment, the three-dimensional model generation device also includes a fourth coding feature determination module; the fourth coding feature determination module is used to upsample the three-dimensional rough model to obtain a second tensor; the elements in the second tensor whose values ​​are greater than a preset value are used as the first element, and the elements in the second tensor other than the first element are used as the second element; the value of the first element is replaced by a randomly generated random number to obtain an updated second tensor; three-dimensional convolution processing is performed on the updated second tensor to obtain a fourth coding feature in the three-dimensional space; wherein the second element in the updated second tensor does not participate in the three-dimensional convolution processing.

[0235] In one embodiment, the second model generation module 1108 is also used to perform three-dimensional convolution processing on the output of the last second processing to obtain the output of the second sub-network model; convert the output of the second sub-network model into a triangular mesh model to obtain a three-dimensional object model belonging to the object category; the three-dimensional object model meets the preset high-resolution conditions.

[0236] In one embodiment, the three-dimensional model generation device also includes a first sub-network model training module; the first sub-network model training module is used to obtain multiple three-dimensional sample models belonging to the object category, and standardize each three-dimensional sample model to a preset vector space to obtain a standardized sample; based on the standardized sample conversion, first training data corresponding to the first resolution is obtained, and noise is added to the first training data to obtain a first input sample; the first training iteration number is determined, and the first intensity sample in any round of first sub-network model training process is determined by the first training iteration number; the first sub-network model in the network model to be trained is trained for the first iterative training through the first training data, the first input sample and the first intensity sample, and the trained first sub-network model is obtained at the end of the training.

[0237] In one embodiment, the three-dimensional model generation device also includes a second sub-network model training module; the second sub-network model training module is used to obtain multiple three-dimensional sample models belonging to the object category, and standardize each three-dimensional sample model to a preset vector space to obtain a standardized sample; based on the standardized sample conversion, second training data corresponding to a second resolution is obtained; the second resolution is greater than the first resolution, and the first resolution is the resolution of the first training data used to train the first sub-network model; noise is added to the second training data to obtain a second input sample; the second training iteration number is determined, and the second intensity sample in any round of second sub-network model training is determined by the second training iteration number; the second sub-network model in the network model to be trained is subjected to second iterative training through the second training data, the second input sample and the second intensity sample, and a trained second sub-network model is obtained at the end of the training.

[0238] Each module in the above-mentioned 3D model generation device can be implemented in whole or in part through software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in the computer device in the form of software, so that the processor can call and execute the corresponding operations of each module.

[0239] In an exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be shown in FIG12 . The computer device includes a processor, a memory, an input / output interface (I / O), and a communication interface. The processor, the memory, and the input / output interface are connected via a system bus, and the communication interface is connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store relevant data for generating a three-dimensional model. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a three-dimensional model generation method is implemented.

[0240] Those skilled in the art will understand that the structure shown in FIG12 is merely a block diagram of a portion of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different arrangement of components.

[0241] In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps of the above-mentioned three-dimensional model generation method when executing the computer program.

[0242] In one embodiment, a computer-readable storage medium is provided, storing a computer program thereon. When the computer program is executed by a processor, the computer program implements the steps of the above-described three-dimensional model generation method. In one embodiment, a computer program product is provided, including the computer program. When the computer program is executed by a processor, the computer program implements the steps of the above-described three-dimensional model generation method.

[0243] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0244] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processor involved in the various embodiments provided herein may be, but are not limited to, a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic unit, a data processing logic unit based on quantum computing, and the like.

[0245] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0246] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A three-dimensional model generation method, executed by a computer device, the method comprising: Acquire prompt information describing the object category, and determine a target network model that matches the object category indicated by the prompt information; The target network model includes a first sub-network model and a second sub-network model; Obtaining a first generation strength, a second generation strength, and a generation seed; The first generation strength indicates the iteration rounds of the first processing performed by the first sub-network model, the second generation strength indicates the iteration rounds of the second processing performed by the second sub-network model, and the generation seed is a randomly generated random number; wherein each round of the first processing includes at least one first encoding and decoding processing, and each round of the second processing includes at least one second encoding and decoding processing, and the first encoding and decoding processing and the second encoding and decoding processing both include at least encoding processing and decoding processing; Obtaining a three-dimensional rough model by performing at least one round of first processing based on the first generation strength and the generation seed through the first sub-network model; and Through the second sub-network model, at least one round of second processing is performed based on the second generated strength and the three-dimensional rough model to obtain a three-dimensional object model that meets the preset high-resolution condition and belongs to the object category.

2. According to the method of claim 1, the determining of the target network model matching the object category indicated by the prompt information comprises: Determining the object category indicated by the prompt information; Acquire a target network model matching the object category from a network model pool; The network model pool includes a plurality of trained network models for generating three-dimensional object models of various categories, and the target network model is used to generate a three-dimensional object model belonging to the object category indicated by the prompt information.

3. According to the method of claim 1 or 2, determining the object category indicated by the prompt information comprises: Obtaining feature vectors of respective preset multiple object category names to obtain a first feature set; Vectorizing the prompt information to determine at least one feature vector related to the prompt information to obtain a second feature set; Calculating the similarity between each feature vector in the second feature set and each feature vector in the first feature set; The object category to which the feature vector corresponding to the feature vector satisfying the preset similarity condition and having the greatest similarity in the first feature set belongs is determined as the object category indicated by the prompt information.

4. The method according to claim 3, wherein the vectorization of the prompt information to determine at least one feature vector related to the prompt information to obtain a second feature set comprises: Performing word segmentation processing on the prompt information to obtain multiple word units in the prompt information; The prompt information and feature vectors of each word unit are extracted respectively to obtain a second feature set.

5. The method according to any one of claims 1 to 4, wherein the first sub-network model is used and at least one round of first processing is performed based on the first generation strength and the generation seed to obtain a three-dimensional rough model, comprising: The first number of times is determined according to the first generation strength, the first number of first processings is performed through the first sub-network model, and a three-dimensional rough model is obtained according to the output of the last first processing; wherein each round of the first processing includes at least two first encoding and decoding processings, and any first encoding and decoding processing process in any current round includes the following steps: Obtaining a first coding feature in the first coding and decoding process of the current round; wherein the first coding feature of each round in the current round is the same and is determined according to the round corresponding to the current round; Obtaining the second coding feature in the current first coding and decoding process; wherein the first second coding feature of the first round is determined according to the generated seed, and the second coding feature of the non-first round is the output of the previous first coding and decoding process; Based on the first coding feature and the second coding feature, performing at least one hierarchical coding process according to different spatial resolution levels to obtain a first coding process result; Based on the first encoding processing result and the first encoding feature, perform at least one fusion processing to obtain a first intermediate processing result; According to the first intermediate processing result and the output of each hierarchical encoding process, at least one hierarchical decoding process is performed according to different spatial resolution levels to obtain the output of the first encoding and decoding process of that time.

6. The method according to claim 5, wherein the step of determining the first coding feature of the current round comprises: Determine the value to be encoded in the current round according to the round number of the current round; Expanding the to-be-encoded value into a feature vector of a preset dimension; Perform at least one linear transformation on the feature vector of the preset dimension to obtain a first coding feature.

7. The method according to claim 5 or 6, wherein the step of determining the second coding feature for the first time in the first round comprises: Using a Gaussian function to generate a first tensor of a preset shape based on the generated seed; A three-dimensional convolution process is performed on the first tensor of the preset shape to obtain a second encoding feature in the three-dimensional space.

8. The method according to any one of claims 5 to 7, wherein the step of performing at least one hierarchical coding process based on the first coding feature and the second coding feature according to different spatial resolution levels to obtain a first coding process result comprises: Obtaining input model features of the current coding layer; wherein the input model features of the first coding layer are the second coding features, and the input model features of the non-first coding layers are the outputs of the previous coding layer; Fusing the first coding feature and the input model feature of the current coding layer to obtain a first fusion result, and downsampling the first fusion result to obtain an output of the current coding layer to complete hierarchical coding processing of the current coding layer; Continue the hierarchical coding process of the next coding layer until the last layer is processed and the first fusion result of the last coding layer is used as the output of the last coding layer; the output of the last coding layer is the first coding process result.

9. The method according to any one of claims 5 to 8, wherein the step of performing at least one fusion process based on the first encoding processing result and the first encoding feature to obtain a first intermediate processing result comprises: Acquire the input model features in the current fusion process; wherein the input model features in the first fusion process are the first encoding processing results, and the input model features in the non-first fusion process are the output of the previous fusion process; Performing a fusion process on the first coding feature and the input model feature in the current fusion process to obtain a second fusion result to complete the current fusion process; The next fusion process is continued until the last fusion process is completed, and the second fusion result of the last fusion process is used as the first intermediate processing result.

10. The method according to any one of claims 5 to 9, wherein the step of performing at least one hierarchical decoding process according to different spatial resolution levels based on the first intermediate processing result and the outputs of each hierarchical coding process to obtain the output of the first coding and decoding process at that time comprises: Obtaining input model features of the current decoding layer; wherein the input model features of the first decoding layer are the first intermediate processing results, and the input model features of the non-first decoding layers include the output of the previous decoding layer and the output of the encoding layer corresponding to the same resolution; The first coding feature and the input model feature of the current decoding layer are fused to obtain a third fusion result, and the third fusion result is upsampled to obtain the output of the current decoding layer to complete the hierarchical decoding process of the current decoding layer; Continue the hierarchical decoding process of the next decoding layer until the last layer is processed, and the third fusion result of the last decoding layer is used as the output of the last decoding layer; the output of the last decoding layer is the output of the first encoding and decoding process at that time.

11. The method according to any one of claims 1 to 10, wherein the second sub-network model is used to perform at least one round of second processing based on the second generation strength and the three-dimensional rough model to obtain a three-dimensional object model that meets a preset high-resolution condition and belongs to the object category, comprising: The second number of times is determined according to the second generation strength, the second number of times of second processing is performed by the second sub-network model, and a three-dimensional object model is obtained according to the output of the last second processing; wherein each round of the second processing includes at least two second encoding and decoding processings, and each second encoding and decoding processing process in any current round includes the following steps: Obtain the third encoding feature in the second encoding and decoding process of the current round; wherein the third encoding feature of each time in the current round The signs are the same and are determined according to the round corresponding to the current round; Obtaining the fourth coding feature in the second coding and decoding process of the current time; wherein the fourth coding feature of the first time in the first round is determined according to the three-dimensional rough model, and the fourth coding feature of the non-first time in the first round is the output of the previous second coding and decoding process; Based on the third coding feature and the fourth coding feature, performing at least one hierarchical coding process according to different spatial resolution levels to obtain a second coding process result; Based on the second encoding processing result and the third encoding feature, perform at least one fusion processing to obtain a second intermediate processing result; According to the second intermediate processing result and the output of each hierarchical encoding process, at least one hierarchical decoding process is performed according to different spatial resolution levels to obtain the output of the second encoding and decoding process at that time.

12. The method according to claim 11, wherein the step of determining the fourth coding feature for the first time in the first round comprises: Upsampling the three-dimensional rough model to obtain a second tensor; The elements in the second tensor whose values ​​are greater than a preset value are used as first elements, and the elements in the second tensor other than the first elements are used as second elements; Replace the value of the first element by a randomly generated random number to obtain an updated second tensor; Performing three-dimensional convolution processing on the updated second tensor to obtain a fourth encoding feature in a three-dimensional space; wherein the second element in the updated second tensor does not participate in the three-dimensional convolution processing.

13. The method according to claim 11 or 12, wherein obtaining a three-dimensional object model according to the output of the last second processing comprises: Performing three-dimensional convolution processing on the output of the last second processing to obtain the output of the second sub-network model; The output of the second sub-network model is converted into a triangular mesh model to obtain a three-dimensional object model belonging to the object category; the three-dimensional object model meets a preset high-resolution condition.

14. According to the method according to any one of claims 1 to 13, the training step of the first sub-network model comprises: Acquire multiple three-dimensional sample models belonging to the object category, and standardize each three-dimensional sample model to a preset vector space to obtain a standardized sample; Obtain first training data corresponding to a first resolution based on the standardized sample conversion, and add noise to the first training data to obtain a first input sample; Determine a first training iteration number, and determine a first strength sample in any round of first sub-network model training process according to the first training iteration number; The first sub-network model in the network model to be trained is subjected to first iterative training by using the first training data, the first input sample and the first strength sample, and a trained first sub-network model is obtained at the end of the training.

15. According to the method according to any one of claims 1 to 14, the training step of the second sub-network model comprises: Acquire multiple three-dimensional sample models belonging to the object category, and standardize each three-dimensional sample model to a preset vector space to obtain a standardized sample; Obtain second training data corresponding to a second resolution based on the standardized sample conversion; the second resolution is greater than the first resolution, and the first resolution is the resolution of the first training data used to train the first sub-network model; Adding noise to the second training data to obtain a second input sample; Determine a second training iteration number, and determine a second strength sample in any round of second sub-network model training process according to the second training iteration number; The second sub-network model in the network model to be trained is subjected to second iterative training by using the second training data, the second input samples and the second strength samples, and a trained second sub-network model is obtained at the end of the training.

16. A three-dimensional model generating device, the device comprising: A network matching module, used to obtain prompt information describing the object category and determine a target network model that matches the object category indicated by the prompt information; The target network model includes a first sub-network model and a second sub-network model; An information acquisition module, used to acquire a first generation strength, a second generation strength, and a generation seed; The first life The generation strength indicates the iterative rounds of the first processing performed by the first sub-network model, the second generation strength indicates the iterative rounds of the second processing performed by the second sub-network model, and the generation seed is a randomly generated random number; wherein each round of the first processing includes at least one first encoding and decoding processing, and each round of the second processing includes at least one second encoding and decoding processing, and the first encoding and decoding processing and the second encoding and decoding processing both include at least encoding processing and decoding processing; A first model generation module, configured to obtain a three-dimensional rough model by performing at least one round of first processing based on the first generation strength and the generation seed through the first sub-network model; The second model generation module is used to perform at least one round of second processing through the second sub-network model and based on the second generation strength and the three-dimensional rough model to obtain a three-dimensional object model that meets the preset high-resolution condition and belongs to the object category.

17. A computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method according to any one of claims 1 to 15 when executing the computer program.

18. A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method according to any one of claims 1 to 15.

19. A computer program product, comprising a computer program, which, when executed by a processor, implements the steps of the method according to any one of claims 1 to 15.

Images