Three-dimensional model adaptive lightweight method and system based on large model driving

Abstract

The invention discloses a three-dimensional model adaptive lightweight method and system based on large model driving, and the method comprises the steps: developing a three-dimensional geometric engine and a three-dimensional model lightweight algorithm set; the method comprises the following steps: establishing an automatic optimization module framework based on open source libraries such as Transform, and performing initialization setting of an automatic optimization module; performing data processing and three-dimensional model lightweight training based on a large model to form an automatic optimization module with complete functions; a three-dimensional model self-adaptive lightweight system based on large model driving is constructed, the three-dimensional model and user lightweight demand data are imported, a user imports the three-dimensional model into a three-dimensional geometry engine, and three-dimensional model self-adaptive lightweight processing and display are completed. According to the method, a large number of three-dimensional model data features are learned through a large model, the most suitable lightweight algorithm is intelligently selected according to the input model features, manual experience intervention is not needed, the adaptability and accuracy of lightweight processing are improved, and the method is suitable for three-dimensional models of various types and complexity.

Description

Technical Field

[0001] The present invention belongs to the technical field of three-dimensional model processing, and in particular relates to a three-dimensional model adaptive lightweighting method and system based on large model driving. Background Art

[0002] 3D model lightweighting refers to the optimization of 3D models to reduce their storage space and computing resource usage, thereby improving processing speed and system performance. In today's digital age, 3D models are widely used in many fields, including industrial manufacturing, architectural design, game development, and virtual reality. 3D models contain a large amount of data, including vertices, faces, and textures. However, as 3D models become increasingly complex, the amount of model data becomes enormous, posing significant challenges to their storage, transmission, and real-time rendering. Therefore, 3D model lightweighting technology is needed to reduce data volume and better adapt to various scenarios and needs.

[0003] Traditional 3D model lightweighting methods are often designed for specific models or application scenarios, lacking versatility and adaptability. Furthermore, existing methods often rely on manual selection and adjustment of lightweighting algorithm parameters, making them inefficient and empirically dependent. They fail to fully utilize the information contained in large amounts of model data and are unable to automatically match algorithm combinations based on natural language requirements, resulting in insufficient rendering frame rates for dynamic interactive scenes. Therefore, there is an urgent need to develop a technology that can intelligently and efficiently lightweight various 3D models. Summary of the Invention

[0004] The technical problem to be solved by this invention is to address the shortcomings of existing technologies and provide a method and system for adaptive lightweighting of 3D models based on a large model, which features a reasonable design, high processing efficiency, and improved versatility and adaptability in model processing. This method and system automatically learns the features of a large amount of 3D model data from a large model, intelligently selects and calls appropriate lightweighting algorithms, and achieves efficient lightweighting of 3D models. This solves the problems of traditional 3D model lightweighting, such as a single algorithm selection, low processing efficiency, and a lack of in-depth understanding of model features and dynamic adjustment processing.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A three-dimensional model adaptive lightweight method based on large model driving, characterized in that the method includes the following steps:

[0007] First, develop a 3D geometry engine and a set of 3D model lightweight algorithms. Select a set of suitable algorithm numbers from the set to create lightweight algorithm labels. Build an automated optimization module framework based on the Transformer open-source library and initialize the learning rate, batch size, number of attention heads, and number of iterations for the automated optimization module.

[0008] Second, data processing and lightweight training of 3D models based on large models are performed to form a fully functional automated optimization module, which includes a data management submodule, a large model training submodule, a result evaluation submodule, and an adaptive lightweight processing module;

[0009] Import a large amount of 3D model data of different types and complexities through the 3D geometry engine. Use the data management submodule to annotate each 3D model, record the type of 3D model, application scenario, original performance indicators, and expected lightweight performance indicators. Convert the 3D model data into a unified data format and extract the model's geometric and topological features to generate feature vectors.

[0010] The preprocessed 3D model data is divided into training set, validation set and test set.

[0011] Each sample in the training set consists of a feature vector of a 3D model and a corresponding lightweight algorithm label set. For each 3D model in the training set, the feature vector is input into the large model training submodule. The corresponding algorithm is called from the 3D model lightweight algorithm set in sequence according to the algorithm number set to perform 3D model lightweight processing, and the lightweight optimization capability of the large model is trained. The performance of the lightweight 3D model is evaluated through the result evaluation submodule, such as calculating the rendering frame rate and the reduction ratio of the 3D model data volume. The evaluation results are fed back to the large model training submodule. If the expected performance indicators do not meet the expected training requirements, the training reaches the set number of iterations, or the loss value no longer decreases significantly, the training process ends.

[0012] After training, the validation set is used for verification, that is, the lightweight effect of the large model is evaluated based on the expected lightweight performance indicators in the annotation information corresponding to the validation set;

[0013] After the 3D model is lightweighted, the performance of the lightweight 3D model is evaluated through the result evaluation submodule. If the expected lightweight performance indicators in the annotation information do not meet the requirements, the verification results are fed back to the large model training submodule to adjust the large model hyperparameters. The 3D model lightweight training steps are then re-performed based on the large model framework, and iterative optimization is continuously carried out until the 3D model stably achieves the expected lightweight performance indicators on the validation set, resulting in a well-trained large model that can adaptively handle 3D model lightweighting.

[0014] After verification, the trained large model is finally tested using the test set to ensure that the large model can automatically control the 3D model to achieve stable lightweight performance on new data that has not been used in training and verification. Similarly, the lightweight performance of the 3D model is evaluated based on the annotation information corresponding to the test set. If the test results are unsatisfactory, it may be necessary to re-examine the large model architecture, training data, or lightweight algorithm selection, and make targeted improvements.

[0015] After passing the test, a large model that can perform adaptive lightweight processing on the 3D model is obtained, thus completing the automated optimization module;

[0016] Third, a large-model-driven adaptive lightweight 3D model system is constructed. The automated optimization module and the 3D model lightweight algorithm set are called through API interfaces or code modules embedded in the 3D geometry engine.

[0017] Fourth, 3D models and user lightweighting requirement data are imported. Users import 3D models into the 3D geometry engine and input application scenarios and lightweighting requirements in the form of text or natural language.

[0018] Fifth, the 3D model is adaptively processed and displayed in a lightweight manner. The automated optimization module automatically starts the data reading and preprocessing process, understands the user's lightweight needs through the trained large model, completes scene recognition, generates an algorithm combination strategy, selects various algorithms from the 3D model lightweight algorithm set in turn, and performs lightweight compression on the 3D model. Finally, the results are pushed to the 3D geometry engine for real-time display. Users can interactively view the effects and make fine adjustments as needed.

[0019] The technical problem to be solved by the present invention can also be achieved by the following technical solutions: the three-dimensional model lightweight algorithm set includes three-dimensional model lightweight algorithms under various dimensions, including but not limited to model geometry simplification methods, texture optimization algorithms, structure optimization methods, algorithm and hardware collaboration methods, and transmission and rendering optimization methods;

[0020] The model geometry simplification method includes but is not limited to quadratic error metric algorithm, edge collapse, vertex clustering, vertex downsampling, surface fitting, and octree;

[0021] The texture optimization algorithm includes but is not limited to format compression, texture thinning and mapping optimization, UV expansion, and ASTC algorithm;

[0022] The algorithm and hardware collaboration method include but are not limited to GPU parallel computing and video memory optimization management;

[0023] Transmission and rendering optimization methods include but are not limited to progressive transmission, viewpoint-dependent rendering, and dynamic loading of viewing distance.

[0024] The technical problem to be solved by the present invention can also be achieved through the following technical solutions, where the expected performance indicators of the lightweight target include but are not limited to the reduction ratio of the number of model faces, the texture compression ratio, and the rendering frame rate improvement ratio.

[0025] The technical problem to be solved by the present invention can also be achieved through the following technical solution. The method of feedback adjustment of large model parameters is that the training reaches the set number of iterations or the loss value no longer decreases significantly, and the model stably reaches the lightweight index on the verification set, thereby obtaining a well-trained model that can adaptively process three-dimensional models.

[0026] A system for implementing the above-mentioned three-dimensional model adaptive lightweighting method is characterized in that the system includes a three-dimensional geometry engine, an automated optimization module, and a three-dimensional model lightweighting algorithm set, wherein the three-dimensional geometry engine mainly performs model import, model display, and model scaling, translation, and rotation; the automated optimization module understands user needs, generates model lightweighting strategies, and automatically controls the lightweighting process based on large models; the three-dimensional model lightweighting algorithm set includes three-dimensional model lightweighting algorithms in various dimensions, including but not limited to model geometry simplification methods, model texture optimization algorithms, structural optimization methods, algorithm and hardware collaboration methods, and transmission and rendering optimization methods;

[0027] The automated optimization module includes a data management submodule, a large model training submodule, and a result evaluation submodule:

[0028] Data management submodule: responsible for the collection, preprocessing and annotation information management of 3D model data, providing data for the large model training submodule, and performing lightweight model and result storage.

[0029] Large model training submodule: It consists of a large model framework, training parameters, optimizer, attenuation strategy, etc. Based on the large model, it selects one or several algorithm combinations from the 3D model lightweight algorithm set, and calls the corresponding lightweight algorithms in sequence to perform lightweight processing on the 3D model to realize 3D model lightweight training. The training is terminated until the stop training signal is received from the result evaluation submodule.

[0030] Result evaluation submodule: When the set number of iterations is reached or the loss value no longer decreases significantly, a stop training learning message is sent to the large model training submodule; when verifying the large model, the performance of the lightweight 3D model is evaluated, such as calculating the rendering frame rate and the reduction ratio of the geometric data volume, and the evaluation results are fed back to the large model training submodule for further adjustment of the algorithm parameters or reselection of the lightweight algorithm. Among them, when the loss function is dynamically balanced or the key evaluation indicators on the test set reach the preset targets and there are no obvious signs of overfitting, a stop verification message is sent to the large model training submodule; when testing the adaptive lightweight effect of the 3D model of the large model, the 3D model lightweight performance evaluation is performed according to the annotation information corresponding to the test set, and the 3D model lightweight effect is fed back to the user.

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

[0032] (1) Intelligent adaptation: By learning a large amount of 3D model data features from a large model, the most suitable lightweight algorithm can be intelligently selected based on the characteristics of the input model without the need for human experience intervention. This greatly improves the adaptability and accuracy of lightweight processing and is suitable for 3D models of various types and complexities.

[0033] (2) High efficiency. The fast computing capability of large models and the intelligent scheduling of lightweight algorithms make the lightweighting process of three-dimensional models more efficient. It can complete the lightweight processing of complex three-dimensional models in a short time and meet the application scenarios with high processing speed requirements such as real-time rendering.

[0034] (3) Versatility: The method of the present invention is not limited to a specific type of 3D model or a specific lightweight algorithm. It can adapt to new 3D model types and lightweight requirements by continuously updating training data and optimizing large models. It has strong versatility and scalability.

[0035] The present invention automatically learns a large amount of 3D model data features through a large model, intelligently selects and calls appropriate lightweight algorithms, realizes efficient lightweighting of 3D models, meets users' 3D model lightweight performance requirements, and improves the versatility and adaptability of 3D model processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a system architecture diagram of the three-dimensional model adaptive lightweight system based on large model drive according to the present invention;

[0037] Figure 2 Provide a lightweight flow chart for the 3D model;

[0038] Figure 3 Flowchart for lightweight processing and training of 3D models based on large models. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0040] Reference Figure 1 The present invention provides a three-dimensional model adaptive lightweight method and system based on large model drive, which mainly includes a three-dimensional geometry engine, an automated optimization module, and a three-dimensional model lightweight algorithm set. The automated optimization module and the three-dimensional model lightweight algorithm set can be integrated into the three-dimensional geometry engine in the form of an API interface or encapsulated embedding. The three-dimensional geometry engine mainly performs operations such as three-dimensional model import, three-dimensional model display, and three-dimensional model scaling, translation, and rotation; the automated optimization module understands user needs, generates three-dimensional model lightweight strategies, and automatically controls the lightweighting process based on the large model; the three-dimensional model lightweight algorithm set includes three-dimensional model geometry simplification, texture optimization, structure optimization, algorithm and hardware collaboration, transmission and rendering optimization, and other three-dimensional model lightweight methods in various dimensions, which are automatically called by the automated optimization module according to the optimization strategy.

[0041] The automated optimization module mainly includes a data management submodule, a large model training submodule, and a result evaluation submodule. The data management submodule is responsible for the collection, preprocessing, and annotation management of 3D model data, providing data for the large model training submodule, and performing lightweight model and result storage. The large model training submodule consists of a large model framework, training parameters, an optimizer, an attenuation strategy, etc. Based on the large model, it selects one or several algorithm combinations from the 3D model lightweight algorithm set, and sequentially calls the corresponding lightweight algorithms to perform lightweight processing on the model, realizing 3D model lightweight training. Training is terminated until the stop training signal is received from the result evaluation submodule. The result evaluation submodule is responsible for the performance evaluation of the large model. When the set number of iterations is reached or the loss value no longer decreases significantly, a stop training learning message is sent to the large model training submodule. When verifying the large model, the lightweight 3D model is evaluated for performance, such as calculating the rendering frame rate and the reduction ratio of the geometric data volume, and the evaluation results are fed back to the large model training submodule for further adjustment of the algorithm parameters or reselection of the lightweight algorithm. Among them, when the loss function is dynamically balanced or the key evaluation indicators on the test set reach the preset targets and there are no obvious signs of overfitting, a stop verification message is sent to the large model training submodule. When testing the adaptive lightweight effect of the 3D model of the large model, the 3D model lightweight performance evaluation is performed according to the annotation information corresponding to the test set, and the lightweight effect of the 3D model is fed back to the user.

[0042] Reference Figure 2 and Figure 3, Example 1, a three-dimensional model adaptive lightweight method and system based on large model drive mainly includes the following steps:

[0043] S100, main module design and development;

[0044] A 3D geometry engine is built based on open source frameworks such as three.js, which has interactive functions such as 3D model import, storage, display, translation, rotation, and scaling. A set of lightweight algorithms for 3D models is constructed and encapsulated, and lightweight algorithms are numbered, including model lightweight methods in various dimensions such as model geometry simplification, texture optimization, structure optimization, algorithm and hardware collaboration, transmission and rendering optimization, etc. A large model framework is built based on the Transformer architecture for lightweight training of 3D models to form an automated optimization module.

[0045] S200, 3D model processing and training based on large models, the detailed process is as follows;

[0046] S201 3D model acquisition and annotation: A large amount of 3D model data of different types and complexities is read into the 3D geometry engine to form a 3D model set M = {m1, m2, ..., m n}, covering model categories (such as industrial machinery, biological models, architectural scenes, etc.), application scenarios (real-time interaction, offline rendering, mobile display, etc.), expected lightweight degree indicators (number of faces reduced ratio r face , texture compression ratio r texture , storage volume reduction rate r volume etc.) to form a labeling information set L = {l1,l2,…,l n}, where each annotation l i Is a vector containing multiple attribute values. For each 3D model, one or several suitable algorithms are selected from the 3D model lightweight algorithm set to construct the available 3D model lightweight algorithm label set A = {a1, a2, ..., a n},a n is m n The algorithm combination number vector corresponding to the 3D model.

[0047] S202 Data Normalization: Perform standardization preprocessing on the 3D model data to make it have a uniform numerical range and distribution characteristics. For geometric coordinate data, let the model vertex coordinates be (x, y, z), and normalize according to the following formula:

[0048] x norm =(x-min(x all )) / (max(x all )-min(x all ))

[0049] ynorm =(y-min(y all )) / (max(y all )-min(y all ))

[0050] z norm =(z-min(z all )) / (max(z all )-min(z all ))

[0051] Where min(x all )、max(x all ) are the minimum and maximum values of the x coordinates of all model vertices, and the same applies to the y and z coordinates. For texture data, if an image is used to represent the texture, the pixel value is normalized to the [0,1] interval, and the pixel value is expressed as norm =(pixel-min(pixel all )) / (max(pixel all )-min(pixel all )) to ensure that data from different models are at the same level and improve the stability of subsequent model training.

[0052] S203 Generate feature vectors: Perform pre-training on the normalized 3D model data and generate a high-dimensional feature vector set F = {f1, f2, ..., f n}, each eigenvector f i The dimension is d, that is, f i =[v i1 ,v i2 ,…,v id ], where v ij Represents the jth eigenvalue of the i-th 3D model. These eigenvectors comprehensively reflect the geometry, texture, and topology of the model, and serve as the key information source for subsequent large model input.

[0053] S204 Dataset Division: First, divide the dataset into layers according to model category and application scenario, and then randomly divide the labeled 3D model dataset into training set M according to a certain ratio (such as the common 70%, 15%, and 15%) in each layer. train , validation set M val and the test set M test , the corresponding feature vector set is divided into F train 、F val 、F test , the annotation information set is divided into L train , L val , L test , ensuring balanced data distribution and avoiding overfitting.

[0054] S205: Lightweight processing training of 3D models based on large models: Set the embedding dimension, number of heads, number of layers, initialize model parameters, and learning rate of the large model. Input the training set data into the model in batches. Use the cross entropy loss function to measure the prediction error of the algorithm combination. Adjust the parameter θ using the AdamW optimizer. Train for multiple rounds (e.g., 100-500 rounds) to monitor the convergence of the validation set indicators. Input the training set feature vector F into the large model. train With annotation information L train Learn the mapping relationship between 3D model features and the optimal lightweight algorithm combination. The algorithm combination rules are based on the summary of each algorithm principle (such as geometric simplification based on vertex clustering, texture compression based on wavelet transform, etc.), applicable scenarios and past experimental results to form a knowledge base to assist large model decision-making. According to the algorithm combination number vector a i , call the corresponding algorithm from the 3D model lightweight algorithm set and execute it in sequence. For example, if a i =[3,7], then first call the lightweight algorithm numbered 3 to perform the three-dimensional model m i After processing, the results are input into the lightweight algorithm No. 7 for further optimization. Different algorithms share the model data structure during processing, reducing the overhead of intermediate data conversion.

[0055] S206 3D model lightweight training feedback: If the training reaches the set number of iterations or the loss value no longer decreases significantly, the training ends.

[0056] S207 Large Model Training Verification and Process Iteration: After the training phase, the validation set is used for verification. The multi-head attention mechanism in the Transformer large model can simultaneously focus on the correlation of features in different subspaces and mine complex patterns. Let the large model prediction function be P, the model parameter be θ, and for the feature vector f of the 3D model to be processed i , predict the output lightweight algorithm combination number vector a i =[a i1 ,a i2 ,…,a ik ], where k is the maximum length of possible algorithm combinations, predicted by:

[0057] a i =P(f i ,L train ;θ)

[0058] Here P is the prediction function of the large model, and θ is the large model parameter.

[0059] Different lightweight algorithms have different parameter configuration spaces. For example, algorithms based on face deletion may have parameters such as face deletion threshold and retained area weight. Set an adjustable parameter set for each lightweight algorithm called to further optimize the lightweight effect. Taking the face reduction algorithm as an example, let the face reduction threshold parameter p bej1 , through multiple tests of different thresholds on the validation set model to meet the reduction ratio r of the number of annotated faces face And try to maintain the model visual quality as the goal, find the parameter group that maximizes the comprehensive evaluation index I

[0060] I=w1·Quality-w2·(1-N light / N ori )

[0061] where N ori 、N light are the original and lightweight posterior numbers respectively, Quality is the visual quality assessment score (which can be calculated by image similarity algorithm), and w1 and w2 are weight coefficients.

[0062] According to the predicted lightweight algorithm number and algorithm parameter settings, the corresponding lightweight algorithm model is called in sequence to perform lightweight processing of the 3D model. val Evaluate whether the lightweight effect of large models is achieved. For example, calculate the actual reduction ratio of the number of faces. Expected ratio with annotation r face In contrast, if If the thresholds (∈) are met and indicators such as texture compression ratio and storage volume reduction rate are within the corresponding thresholds, the model is considered fully trained and the large model is obtained. Otherwise, the model is marked as unsatisfactory and hyperparameter tuning of the large model is performed. The 3D model lightweighting algorithm is re-executed based on the large model framework, and the selected training steps are continuously iterated and optimized until the 3D model stably achieves the desired lightweight performance indicators on the validation set, resulting in a fully trained large model that can adaptively handle 3D model lightweighting.

[0063] S208 testing verifies the large model that achieves adaptive lightweighting: After passing verification, the trained large model undergoes final testing using the test set to ensure stable lightweighting performance on new data that was not used in training and verification. Performance evaluation is also conducted based on the annotations corresponding to the test set. If the test results are unsatisfactory, it may be necessary to re-examine the large model architecture, training data, or algorithm selection, and make targeted improvements. After passing the test, a trained model capable of adaptive lightweighting of 3D models is obtained.

[0064] S300: 3D model reading adaptive lightweight;

[0065] The background thread starts the moment the user-side software (i.e., the 3D geometry engine) receives the user's 3D model. It uses the trained large model to quickly extract features, predict algorithm combinations and parameters, and adaptively calls the lightweight algorithm in sequence to complete the lightweight compression operation of the 3D model. The 3D model is pushed to the 3D engine scene through the engine interface. The user operates the sliders and buttons in the engine interaction interface to fine-tune the lightweight parameters (such as adjusting the slider for the degree of face reduction). The system responds by re-executing the lightweight optimization and updating the display in real time.

[0066] The three-dimensional model adaptive lightweighting method and system described in the present invention can automatically read various three-dimensional models, use large models to intelligently select and execute the optimal lightweighting algorithm, quickly obtain a model that meets lightweight performance indicators, and facilitate users to operate in the three-dimensional engine.

[0067] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A three-dimensional model adaptive lightweight method based on large model drive, characterized in that: The method comprises the following steps, First, develop a 3D geometry engine and a 3D model lightweight algorithm set. Select a set of suitable algorithm numbers for lightweighting the model from the 3D model lightweight algorithm set to form a lightweight algorithm label. Build an automated optimization module framework based on the Transformer open source library, and initialize the learning rate, batch size, number of attention heads, and number of iterations of the automated optimization module; Second, data processing and lightweight training of 3D models based on large models are performed to form a fully functional automated optimization module, which includes a data management submodule, a large model training submodule, and a result evaluation submodule; Import a large amount of 3D model data of different types and complexities through the 3D geometry engine. Use the data management submodule to annotate each 3D model, record the type of 3D model, application scenario, original performance indicators, and expected lightweight performance indicators. Convert the 3D model data into a unified data format and extract the model's geometric and topological features to generate feature vectors. The preprocessed 3D model data is divided into training set, validation set and test set. Each sample in the training set consists of a feature vector of a 3D model and a corresponding lightweight algorithm label set. For each 3D model in the training set, the feature vector is input into the large model training submodule. The corresponding algorithm is called from the 3D model lightweight algorithm set in sequence according to the algorithm number set to perform 3D model lightweight processing, and the lightweight optimization capability of the large model is trained. The performance of the lightweight 3D model is evaluated through the result evaluation submodule, such as calculating the rendering frame rate and the reduction ratio of the 3D model data volume. The evaluation results are fed back to the large model training submodule. If the expected performance indicators do not meet the expected training requirements, the training reaches the set number of iterations, or the loss value no longer decreases significantly, the training process ends. After training, the validation set is used for verification, that is, the lightweight effect of the large model is evaluated based on the expected lightweight performance indicators in the annotation information corresponding to the validation set; After the 3D model is lightweighted, the performance of the lightweight 3D model is evaluated through the result evaluation submodule. If the expected lightweight performance indicators in the annotation information do not meet the requirements, the verification results are fed back to the large model training submodule to adjust the large model hyperparameters. The 3D model lightweight training steps are then re-performed based on the large model framework, and iterative optimization is continuously carried out until the 3D model stably achieves the expected lightweight performance indicators on the validation set, resulting in a well-trained large model that can adaptively handle 3D model lightweighting. After verification, the trained large model is finally tested using the test set to ensure that the large model can automatically control the 3D model to achieve stable lightweight performance on new data that has not been used in training and verification. Similarly, the lightweight performance of the 3D model is evaluated based on the annotation information corresponding to the test set. If the test results are unsatisfactory, it may be necessary to re-examine the large model architecture, training data, or lightweight algorithm selection, and make targeted improvements. After passing the test, a large model that can perform adaptive lightweight processing on the 3D model is obtained, thus completing the automated optimization module; Third, a large-model-driven 3D model adaptive lightweight system is constructed. The automated optimization module and 3D model lightweight algorithm set are called through API interfaces or code modules embedded in the 3D geometry engine. Fourth, 3D models and user lightweighting requirement data are imported. Users import 3D models into the 3D geometry engine and input application scenarios and lightweighting requirements in the form of text or natural language. Fifth, the 3D model is adaptively processed and displayed in a lightweight manner. The automated optimization module automatically starts the data reading and preprocessing process, understands the user's lightweight needs through the trained large model, completes scene recognition, generates an algorithm combination strategy, selects various algorithms from the 3D model lightweight algorithm set in turn, and performs lightweight compression on the 3D model. Finally, the results are pushed to the 3D geometry engine for real-time display. Users can interactively view the effects and make fine adjustments as needed.

2. The method for adaptive lightweighting of a three-dimensional model based on large model driving according to claim 1, characterized in that: The three-dimensional model lightweight algorithm set includes three-dimensional model lightweight algorithms in various dimensions, including but not limited to model geometry simplification methods, model texture optimization algorithms, structure optimization methods, algorithm and hardware collaboration methods, and transmission and rendering optimization methods; The model geometry simplification method includes but is not limited to quadratic error metric algorithm, edge collapse, vertex clustering, vertex downsampling, surface fitting, and octree; The texture optimization algorithm includes but is not limited to format compression, texture thinning and mapping optimization, UV expansion, and ASTC algorithm; The algorithm and hardware collaboration method include but are not limited to GPU parallel computing and video memory optimization management; Transmission and rendering optimization methods include but are not limited to progressive transmission, viewpoint-dependent rendering, and dynamic loading of viewing distance.

3. The method for adaptive lightweighting of a three-dimensional model based on large model driving according to claim 1, characterized in that: The expected performance indicators of the lightweight target include but are not limited to the reduction ratio of the number of model faces, the texture compression ratio, and the rendering frame rate improvement ratio.

4. The method for adaptive lightweighting of a three-dimensional model based on large model driving according to claim 1, characterized in that: The method of feedback adjustment of large model parameters is as follows: when the training reaches a set number of iterations or the loss value no longer decreases significantly, the model stably reaches the lightweight index on the validation set, and a well-trained model capable of adaptively processing three-dimensional models is obtained.

5. A system for implementing the three-dimensional model adaptive lightweighting method according to any one of claims 1 to 4, characterized in that: The system includes a 3D geometry engine, an automated optimization module, and a 3D model lightweight algorithm set. The 3D geometry engine mainly performs model import, model display, and model scaling, translation, and rotation. The automated optimization module understands user needs, generates model lightweight strategies, and automatically controls the lightweighting process based on large models. The 3D model lightweight algorithm set includes 3D model lightweight algorithms in various dimensions, including but not limited to model geometry simplification methods, mold texture optimization algorithms, structural optimization methods, algorithm and hardware collaboration methods, and transmission and rendering optimization methods. The automated optimization module includes a data management submodule, a large model training submodule, and a result evaluation submodule; Data management submodule: responsible for the collection, preprocessing and annotation information management of 3D model data, providing data for the large model training submodule, and performing lightweight model and result storage; The large model training submodule consists of a large model framework, training parameters, an optimizer, and an attenuation strategy. Based on the large model, it selects one or several algorithms from the 3D model lightweight algorithm set, and sequentially calls the corresponding lightweight algorithms to perform lightweight processing on the 3D model, thus implementing lightweight training of the 3D model. Training is terminated when a stop training signal is received from the result evaluation submodule. The result evaluation submodule sends a stop training learning message to the large model training submodule when the set number of iterations is reached or the loss value no longer decreases significantly; when verifying the large model, the performance of the lightweight 3D model is evaluated, such as calculating the rendering frame rate and the reduction ratio of the geometric data volume, and the evaluation results are fed back to the large model training submodule for further adjustment of the algorithm parameters or reselection of the lightweight algorithm. Among them, when the loss function is dynamically balanced or the key evaluation indicators on the test set reach the preset targets and there are no obvious signs of overfitting, a stop verification message is sent to the large model training submodule; when testing the adaptive lightweight effect of the 3D model of the large model, the 3D model lightweight performance evaluation is performed according to the annotation information corresponding to the test set, and the 3D model lightweight effect is fed back to the user.

Images