Performance-based 3D car model generation network, generation method and system

Abstract

This invention provides a 3D car model generation network, method, and system based on performance indicators, relating to the field of deep learning network technology. The generation network includes a multi-view diffusion model, a 3D generation model, and a drag prediction proxy model. Driven by size, angle, and performance parameters, the input data passes through an upscaling layer, a multilayer perceptron, and a multi-head attention block within the multi-view diffusion model to obtain multi-view features. The U-Net architecture generates four views of the car. These four views are then input into the 3D generation model, efficiently and accurately generating a 3D car model that meets engineering design requirements, improving the design iteration efficiency for creative designers. A drag prediction proxy model is added, comparing its predicted drag coefficient with the input theoretical drag coefficient. Backpropagation loss is applied to adjust the parameters of the multi-view diffusion model, ensuring that the four views of the car meet the performance parameter requirements, thereby ensuring that the 3D car model also meets the performance parameter requirements.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning network, in particular to a three-dimensional automobile model generation network based on performance indicators, a generation method and a system. BACKGROUND

[0002] In the current automobile R&D process, design and engineering verification are two relatively independent and serially connected stages. The designer team first creates an initial shape concept of the automobile according to market trends, brand language, aesthetic principles and ergonomics, etc., relying on their professional experience and design intuition. Then, the design scheme is handed over to the engineering team to perform high-fidelity simulation using computational fluid dynamics (CFD) to obtain key performance data such as drag coefficient. After the simulation results are obtained, the engineering team feeds back the performance analysis report and modification suggestions to the design team. The design team adjusts the model according to the feedback and then submits it again for a new round of simulation verification. The serial iteration process of “design-test-modification” may need to be repeated multiple times, resulting in a long cycle and high cost of the entire concept design phase, and a serious dependence on the prior knowledge of designers and engineers.

[0003] The existing technology presents a clear binary separation: on the one hand, the engineering field relies on highly time-consuming physical simulation tools (such as CFD) with high precision; on the other hand, the creative field applies generation models that can quickly generate visual content but have no knowledge of physical performance. Between “looking like a car” and “achieving specific engineering performance standards as a car”, there is a deep “semantic-performance gap”. Existing generation models are mainly based on a single picture or a piece of text to generate a corresponding three-dimensional automobile model, but the three-dimensional automobile model obtained is difficult to meet the performance indicators of engineering design. SUMMARY

[0004] The problem to be solved by the present application is that the existing technology generates a three-dimensional automobile model based on a single picture or a piece of text, which is difficult to meet the performance indicators of engineering design, and has low R&D efficiency.

[0005] To solve the above problems, in a first aspect, the present application provides a three-dimensional automobile model generation network based on performance indicators, comprising a multi-view diffusion model, a 3D generation model and a drag prediction proxy model.

[0006] The multi-view diffusion model comprises a dimension increasing layer, a multi-layer perception, a multi-head attention block and a U-Net architecture of the diffusion model connected in sequence; the residual module of the U-Net architecture comprises a residual block, a three-dimensional self-attention block and a cross-attention block connected in sequence.

[0007] The input data includes size parameters, angle parameters and performance parameters, which are input into a multi-head attention block after dimension lifting processing by a dimension lifting layer and projection by a corresponding multi-layer perception, to obtain multi-view view features; the multi-view view features, time steps and camera parameters are input into a U-Net architecture of a diffusion model, and are sequentially processed by a residual block, a three-dimensional self-attention block and a cross-attention block to generate a four-view image of a car;

[0008] The four-view image of the car and the multi-view view features are input into a 3D generation model to obtain a 3D car model;

[0009] The training process of the multi-view diffusion model includes:

[0010] The point cloud data of the 3D car model is input into a pre-trained wind resistance prediction proxy model to obtain a predicted wind resistance coefficient; a performance loss value is determined according to the predicted wind resistance coefficient and a theoretical wind resistance coefficient, and the performance loss value is back propagated to adjust the parameters of the multi-view diffusion model.

[0011] Optionally, before the time steps and the camera parameters are input into the U-Net architecture of the diffusion model, the camera parameters are mapped into embedding vectors of a fixed dimension by a multi-layer perception ; the time steps are encoded into time step embedding vectors ; and the embedding vectors and the time step embedding vectors are concatenated to form a condition vector e.

[0012] Optionally, in the cross-attention block, the time steps and the camera parameters jointly form a condition vector to guide the generation of the four-view image of the car; and a formalized representation of the cross-attention block is:

[0013]

[0014] wherein X represents a feature tensor of the multi-view view features, e represents a condition vector, represents an embedding form of the condition vector, Q(X) represents a query matrix of the feature tensor of the multi-view view features, represents a key matrix of the condition vector, represents a value matrix of the condition vector, and d represents a dimension of the key matrix.

[0015] Optionally, the 3D generation model includes a DINO encoder, a plurality of multi-layer perceptions, a diffusion Transformer block, a linear transformation layer, a shape latent encoding layer and a VAE-decoder;

[0016] The four-view image of the car is processed by the DINO encoder and the multi-layer perceptions to obtain an image encoding ; and the multi-view view features are processed by the multi-layer perceptions to obtain a condition vector encoding ; image encoding and conditional vector encoding are input into a diffusion Transformer block, multiple Transformer blocks in the diffusion Transformer block learn latent encoding, and a learning output result of the diffusion Transformer block is subjected to linear processing of a linear transformation layer and linear projection of a shape latent encoding layer to obtain a latent hidden shape feature; the latent hidden shape feature is subjected to refinement processing of multiple self-attention mechanism blocks stacked in a VAE-decoder to generate a 3D automobile model.

[0017] Optionally, the wind resistance prediction proxy model is pre-trained, including:

[0018] The wind resistance prediction proxy model is pre-trained offline using a data set of 3D automobile models paired with CFD wind resistance coefficients until convergence, and the parameters of the wind resistance prediction proxy model are fixed.

[0019] Optionally, the performance loss value is

[0020]

[0021] wherein, represents a performance loss value, represents a theoretical wind resistance coefficient, represents a predicted wind resistance coefficient.

[0022] In a second aspect, the present application further provides a three-dimensional automobile model generation method based on performance indicators, including:

[0023] Obtaining input data, the input data including size parameters, angle parameters and performance parameters;

[0024] Inputting the input data into the three-dimensional automobile model generation network based on performance indicators as described above to obtain a 3D automobile model output by the 3D generation model, wherein the 3D automobile model meets the performance requirements and wind resistance coefficient requirements of the input data.

[0025] In a third aspect, the present application further provides a three-dimensional automobile model generation system based on performance indicators, including:

[0026] A data acquisition module for acquiring input data, the input data including size parameters, angle parameters and performance parameters;

[0027] A 3D model generation module for inputting the input data into the three-dimensional automobile model generation network based on performance indicators as described above to obtain a 3D automobile model output by the 3D generation model, wherein the 3D automobile model meets the performance requirements and wind resistance coefficient requirements of the input data.

[0028] In a fourth aspect, the present application provides an electronic device comprising a memory and a processor;

[0029] The memory is configured to store a computer program;

[0030] The processor is configured to implement the performance index-based three-dimensional automobile model generation network or the performance index-based three-dimensional automobile model generation method when executing the computer program.

[0031] In a fifth aspect, the present application provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the performance index-based three-dimensional automobile model generation network or the performance index-based three-dimensional automobile model generation method is implemented.

[0032] The present application provides a performance index-based three-dimensional automobile model generation network, generation method and system. Compared with the prior art, the following beneficial effects are achieved:

[0033] A three-dimensional automobile model generation network is constructed, which can be driven by a set of size parameters and angle parameters and a performance parameter. The input data is processed and feature extracted via a dimensionality increasing layer, a multilayer perception and a multi-head attention block, to obtain multi-view view features, and then enhanced and denoised via residual blocks, three-dimensional self-attention blocks and cross-attention blocks in a U-Net architecture of a diffusion model, to generate four views of a car with consistent structure and appearance. The four views of the car are input to a 3D generation model, which directly generates a 3D automobile model meeting the requirements of engineering design with high efficiency and precision, greatly improving the iteration efficiency of creative designers in the early design process. Since the four views of the car directly determine the quality of the 3D automobile model generated by the 3D generation model, the three-dimensional automobile model generation network further adds a wind resistance prediction proxy model. In the training stage, the point cloud data of the generated 3D automobile model is input to the pre-trained wind resistance prediction proxy model, the predicted wind resistance coefficient is compared with the input theoretical wind resistance coefficient, the loss is back propagated, the multi-view diffusion model parameters are adjusted, so that the four views of the car output by the multi-view diffusion model meet the requirements of the performance parameter, so that the 3D automobile model output by the 3D generation model also meets the performance parameter requirements. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0035] Figure 1 A structural schematic diagram of a three-dimensional automobile model generation network based on performance indicators is provided for an embodiment of the present application.

[0036] Figure 2 A structural schematic diagram of a 3D generated model is provided for an embodiment of the present application.

[0037] Figure 3 A flowchart of a three-dimensional automobile model generation method based on performance indicators is provided for an embodiment of the present application.

[0038] Figure 4 A structural schematic diagram of a three-dimensional automobile model generation system based on performance indicators is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0039] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application are described clearly and completely. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0040] The technical solutions in the embodiments of the present application are to solve the above technical problems, and the general idea is as follows:

[0041] The "design-test-modification" cycle of traditional automobile design is a typical serial process, which fundamentally limits the research and development efficiency. Designers can only verify a few design schemes within the limited project time, and cannot extensively explore the potential optimal solution in the design space. In addition, due to the lack of real-time performance feedback, the scheme proposed by the designer in the early stage may have fundamental defects in aerodynamics, resulting in the need for substantial, costly modifications in the later stage, or even starting over, still failing to achieve "one-time" generation of the final design that meets all constraints.

[0042] Even now there are some models that can generate 3D models, but the condition inputs of these models are usually single modal, such as a piece of text or a picture, which is far from enough to describe a complex engineering design problem. None of the existing methods provides a unified framework that allows users to directly generate 3D automobile models that meet specified performance indicators through a comprehensive vector of geometric parameters. Existing generated models cannot parse and execute such complex constraints composed of multi-modal, quantitative parameters. Therefore, the present application proposes a three-dimensional automobile model generation network based on performance indicators to realize rapid iteration of design.

[0043] For better understanding of the above technical solutions, the above technical solutions will be described in detail below in combination with the drawings and specific embodiments of the specification.

[0044] As shown in Figure 1 The embodiment of the application provides a three-dimensional automobile model generation network based on a performance index, which comprises a multi-view diffusion model, a 3D generation model and a wind resistance prediction agent model.

[0045] The multi-view diffusion model comprises a dimension increasing layer, a multi-layer perception, a multi-head attention block and a U-Net architecture of the diffusion model connected in sequence; and the residual module of the U-Net architecture comprises a residual block, a three-dimensional self-attention block and a cross-attention block connected in sequence.

[0046] The input data comprises size parameters, angle parameters and performance parameters, the input data is subjected to dimension increasing processing of the dimension increasing layer and projection of the corresponding multi-layer perception, and then is input into the multi-head attention block to obtain multi-view view features; the multi-view view features, a time step and camera parameters are input into the U-Net architecture of the diffusion model, and are sequentially subjected to processing of the residual block, the three-dimensional self-attention block and the cross-attention block to generate automobile four views.

[0047] The automobile four views and the multi-view view features are input into the 3D generation model to obtain a 3D automobile model.

[0048] The training process of the multi-view diffusion model comprises:

[0049] The point cloud data of the 3D automobile model is input into the pre-trained wind resistance prediction agent model to obtain a predicted wind resistance coefficient; a performance loss value is determined according to the predicted wind resistance coefficient and a theoretical wind resistance coefficient, the performance loss value is back propagated, and the parameters of the multi-view diffusion model are adjusted.

[0050] Specifically, the size parameters include length, width and height dimensions, and more specifically underbody clearance, front overhang height, rear overhang height, roof height and vehicle height, etc.; the angle parameters include front windshield angle, rear windshield angle and roof end angle, etc.; the performance parameters include theoretical wind resistance coefficient. The multi-view diffusion model processes the input data, and processes the size parameters, angle parameters and performance parameters of different dimensions and different forms: first, the dimension of the input data is raised to 32 through the dimension raising layer; then the size parameters are projected and expanded to 256-dimensional embedding vectors, the angle parameters are projected and expanded to 128-dimensional embedding vectors, and the performance parameters are projected and expanded to 64-dimensional embedding vectors through the corresponding multi-layer perception; the embedding vectors output by the multi-layer perception are input together into the multi-head attention block for fusion, so as to capture the interaction between different attributes and form a unified feature representation. The size features, angle features and performance features in the input data are extracted through the multi-head attention mechanism, thereby forming multi-view view features of unified dimensions, such as front view features, left view features, rear view features and right view features. The multi-view diffusion model takes performance indicators as input to generate a set of multi-view two-dimensional images (front view, rear view, left view and right view) with consistent geometric dimensions. These images not only depict the appearance of the car, but more importantly, implicitly encode the geometric structure information required to meet specific performance constraints at the pixel level.

[0051] The multi-head attention block takes the embedding vectors of the multi-view as batch input, and the embedding vectors (feature tensors) are denoted as: ; wherein, N represents the number of views; B represents the batch size, that is, the number of vectors processed simultaneously in one inference process; H represents the height of the feature tensor; W represents the width of the feature tensor; C represents the number of channels. The tensor is rearranged as: , and input into the standard self-attention mechanism:

[0052]

[0053] where Q, K and V come from linear transformations of , respectively representing the query matrix, key matrix and value matrix of the feature tensor, and d represents the dimension of the key matrix. In this way, the features of all views are placed in the same attention sequence, thereby realizing global information interaction across views.

[0054] The diffusion model receives multi-view view features, introduces a three-dimensional self-attention block and a cross-attention block in a residual module of an internal U-Net architecture, and the three-dimensional self-attention block enhances the cross-view modeling capability. In the cross-attention, a conditional vector is formed by the camera parameter embedding and the time step embedding, which guides the image generation process, so as to obtain four views of the car with unified dimensions. The 3D generation model can be selected from a NeRF (Neural Radiance Fields) or a Kiss3DGen model; the wind resistance prediction proxy model can be selected from a DNNFluid-Car, a DrivAer model, or can be replaced by ANSYS Fluent software, which is used to analyze the wind resistance coefficient of the 3D car model.

[0055] In the embodiment, a three-dimensional car model generation network is constructed, which can be driven by a set of size parameters and angle parameters and a performance parameter. After input data is processed and feature extracted through a dimension increasing layer, a multi-layer perception and a multi-head attention block, multi-view view features are obtained, and then the diffusion model is enhanced and denoised through a residual block, a three-dimensional self-attention block and a cross-attention block in the U-Net architecture, to generate four views of the car with consistent structure and appearance. The four views of the car are input into the 3D generation model, and the 3D car model meeting the engineering design requirements is directly generated with high efficiency and high precision, which greatly improves the iteration efficiency of the creative designer in the early design process. Since the four views of the car directly determine the quality of the 3D car model generated by the 3D generation model, in order to ensure that the generated 3D car model not only meets the basic size requirements, but also further meets the performance requirements of the wind resistance coefficient, the three-dimensional car model generation network further adds a wind resistance prediction proxy model. In the training stage, the point cloud data of the generated 3D car model is input into the pre-trained wind resistance prediction proxy model, the predicted wind resistance coefficient is compared with the input theoretical wind resistance coefficient, the loss is back propagated, the multi-view diffusion model is adjusted, so that the four views of the car output by the multi-view diffusion model meet the performance parameter requirements, so that the 3D car model output by the 3D generation model also meets the performance parameter requirements.

[0056] The various modules will be further described in detail below.

[0057] In the optional embodiment of the present application, as shown in Figure 1 , before the time step and the camera parameter are input into the U-Net architecture of the diffusion model, the camera parameter is mapped into a fixed-dimension embedding vector by a multi-layer perception ; the time step is encoded into a time step embedding vector ; the embedding vector and the time step embedding vector are spliced to form a conditional vector e.

[0058] Specifically, the camera parameters include position information (e.g., azimuth coordinates) and angle information (e.g., rotation angles), which are respectively mapped into fixed-dimension embedding vectors by multi-layer perceptron (MLP). Let the camera parameters be , and the embedding vectors thereof be represented as . The time steps of the diffusion process are encoded into time step embedding vectors , and the final conditioning vector is obtained by adding or concatenating the two, i.e., the conditioning vector .

[0059] In the cross-attention block, the time step and the camera parameters jointly constitute the conditioning vector, which guides the generation of the four views of the car; the formalized representation of the cross-attention block is as follows:

[0060]

[0061] where X represents the feature tensor of the multi-view features, e represents the conditioning vector, represents the embedding form of the conditioning vector, Q(X) represents the query matrix of the feature tensor of the multi-view features, represents the key matrix of the conditioning vector, represents the value matrix of the conditioning vector, and d represents the dimension of the key matrix.

[0062] In addition, in the process of locally training the U-Net architecture of the diffusion model, the optimization objective function used is as follows:

[0063]

[0064] where represents the noisy image at time step t, is the Gaussian noise, is the noise residual function predicted by the model, c refers to the conditioning parameter, t refers to the time step, and E refers to expectation. The objective function ensures that the U-Net architecture of the diffusion model can gradually restore the multi-view consistent image from the noise under the given text parameter and camera parameter conditions.

[0065] The U-Net architecture of the diffusion model gradually reduces the noise based on the specified performance indicator (i.e., the theoretical drag coefficient), and finally outputs the images of multiple views such as the front view, the rear view, the left view, and the right view, while ensuring the consistency of the structure and the appearance, thereby providing reliable two-dimensional images for subsequent stable three-dimensional reconstruction.

[0066] In an optional embodiment of the present application, as shown in Figure 2 , the 3D generation model includes a DINO encoder, a plurality of multi-layer perceptrons, a diffusion Transformer block, a linear transformation layer, a shape latent encoding layer, and a VAE-decoder.

[0067] The four views of the car are processed by a DINO encoder and a multi-layer perception machine to obtain image encoding ; the multi-view view features are processed by a multi-layer perception machine to obtain conditional vector encoding ; the image encoding and the conditional vector encoding are jointly input into a diffusion Transformer block, multiple Transformer blocks in the diffusion Transformer block learn latent encoding; the learning output result of the diffusion Transformer block is linearly processed by a linear transformation layer and linearly projected by a shape latent encoding layer to obtain latent hidden shape features; the latent hidden shape features are refined by multiple self-attention mechanism blocks stacked in a VAE-decoder to generate a 3D car model.

[0068] Specifically, as shown in Figure 2 , the diffusion Transformer block includes multiple Transformer blocks, each of which includes, in sequence, a latent encoding layer, a multi-layer perception machine, a normalization layer, a self-attention mechanism, a cross-attention mechanism, a normalization layer, and an expert hybrid model (MoE); the linear transformation layer includes, in sequence, layer normalization, a scaling offset linear transformation layer, and a linear layer; the VAE-decoder includes multiple self-attention mechanism blocks connected in sequence. The 3D generation model also inputs a time step, which cooperates with the time step of the multi-view diffusion model to control the multi-view diffusion model and the 3D generation model to perform operations in sequence on a time line.

[0069] The four views of the car are processed by a DINO encoder to extract high-level geometric and semantic features, and then mapped to a latent space by an MLP (multi-layer perception machine) to obtain a unified visual encoding:

[0070]

[0071] wherein, represents the input set of four views of the car.

[0072] In an optional embodiment of the present application, the wind resistance prediction proxy model is pre-trained, comprising:

[0073] The wind resistance prediction proxy model is pre-trained offline using a dataset of 3D car models paired with CFD wind resistance coefficients until convergence, and the parameters of the wind resistance prediction proxy model are fixed. At the same time, since the 3D generation model itself has no performance constraint condition, the performance constraint index (i.e., the multi-view view feature) of the first stage processing is embedded as conditional vector encoding after being processed by an MLP. The image encoding and the conditional vector encoding The consistency multi-view view features are input into the diffusion Transformer block. The diffusion Transformer block includes multiple Transformer blocks, and the self-attention mechanism in each Transformer block is used to capture the global consistency within the latent representation, the cross-attention mechanism realizes the cross-modal interaction between the image features and the latent shape representation, and the MoE enhances the modeling capability of the model for complex geometric patterns. The latent hidden shape features are sent to the VAE-based decoder, which gradually recovers the three-dimensional geometric structure through the layer-stacked self-attention mechanism, and finally generates a complete 3D car model. Under the constraint of consistent multi-view view features, the performance condition can be fine-tuned, which not only ensures the geometric and semantic consistency across views, but also realizes the collaborative satisfaction of geometric precision and performance constraints in the generated results, thereby obtaining a high-fidelity 3D car model.

[0074] In an optional embodiment of the present application, the wind resistance prediction proxy model is pre-trained, comprising:

[0075] The wind resistance prediction proxy model is pre-trained offline using a dataset of 3D car models paired with CFD wind resistance coefficients until convergence, and the parameters of the wind resistance prediction proxy model are fixed.

[0076] Specifically, the wind resistance prediction proxy model is pre-trained offline using a large dataset (DrivAerNet++) containing a large number of "3D model-CFD wind resistance coefficient" paired data until convergence. During the online fine-tuning phase (training phase) of the entire three-dimensional car model generation network, the parameters of the wind resistance prediction proxy model will be frozen. The wind resistance prediction proxy model receives the generated 3D car model (usually input in the form of point cloud or mesh) as input and outputs the predicted wind resistance coefficient . As Figure 1 shown, a closed-loop fine-tuning chain is formed between the three models, and the closed-loop fine-tuning is driven by a performance loss function that quantifies the gap between the theoretical wind resistance coefficient and the predicted wind resistance coefficient output by the wind resistance prediction proxy model:

[0077]

[0078] wherein, represents the performance loss value, represents the theoretical wind resistance coefficient, represents the predicted wind resistance coefficient. The performance loss value is back-propagated to update part of the parameters of the first-stage multi-view diffusion model. This feedback loop is iterated to continuously adjust the generation strategy of the first-stage multi-view diffusion model, so that the multi-view images generated by the first-stage multi-view diffusion model can reconstruct a three-dimensional model closer to the target performance, and finally realize accurate control of the performance indicators.

[0079] The three-dimensional vehicle model generation network is constructed and trained using the data set, and a specific operation example is as follows.

[0080] Taking generation of a 3D vehicle model using the DrivAerNet++ data set as an example. The data set is a data set that meets the requirements of training and verification, and it provides all the necessary data modalities. The data set contains 8000 three-dimensional vehicle grid models of different configurations, table data of 26 detailed geometric parameters corresponding to each model, and aerodynamic performance data including wind resistance coefficients obtained through high-fidelity CFD simulation.

[0081] 1. Wind resistance prediction proxy model pre-training. Train the wind resistance prediction proxy model on the DrivAerNet++ data set. Use the 3D point cloud data of the three-dimensional vehicle model in the data set as input, and the theoretical wind resistance coefficient calculated by CFD as label, train the model until high precision is achieved on the validation set.

[0082] 2. Training of multi-view diffusion model and 3D generation model. Sample from the DrivAerNet++ data set, use geometric parameters and known theoretical wind resistance coefficients as input data, train the multi-view diffusion model of the first stage to generate corresponding vehicle four views, and use the 3D generation model of the second stage to reconstruct the 3D vehicle model. The training goal of this stage is to enable the model to learn to generate correct vehicle shapes according to the parameters.

[0083] 3. Online fine-tuning. In actual use, the user provides a set of customized target parameters (vehicle length, vehicle width, vehicle height, ground clearance, front suspension length, rear suspension length, end of hood height, front windshield angle, rear windshield angle, end of hood angle, vehicle head height, and wind resistance coefficient).

[0084] (1) The trained multi-view diffusion model and 3D generation model generate an initial 3D vehicle model according to these target parameters.

[0085] (2) Input the 3D vehicle model into the trained and frozen wind resistance prediction proxy model to obtain the predicted wind resistance coefficient.

[0086] (3) Calculate the performance loss value, and use the loss value to adjust the multi-view diffusion model of the first stage for a small number of steps, and then regenerate.

[0087] (4) Repeat the process until the difference between the predicted wind resistance coefficient and the theoretical wind resistance coefficient is less than a preset threshold, and finally output a 3D vehicle model that meets the performance requirements.

[0088] As shown in Figure 3 , the three-dimensional vehicle model generation method based on performance indicators provided by the embodiments of the present application comprises:

[0089] S1: obtaining input data, the input data comprising size parameters, angle parameters and performance parameters;

[0090] S2: inputting the input data into the performance index-based three-dimensional automobile model generation network as described above to obtain a 3D automobile model output by a 3D generation model, wherein the 3D automobile model meets the performance requirements and the wind resistance coefficient requirements of the input data.

[0091] Compared with the prior art, the performance index-based three-dimensional automobile model generation method has the following beneficial effects:

[0092] 1. The research and development efficiency is improved. The "design-simulation" verification period in traditional automobile design in units of days or weeks is shortened to several minutes. This leap in efficiency enables designers and engineers to explore thousands of design possibilities in the early concept stage, greatly broadening the design space and increasing the opportunity to discover breakthrough innovative designs.

[0093] 2. The design paradigm is changed from "exploratory" to "targeted". Unlike existing generation models that are exploratory (post-generation evaluation), the present application is a target-oriented generation method. Users can directly specify performance targets, and the system directly "solves" the design scheme that meets the target. This represents a fundamental change in design paradigm from "what-if" analysis to direct problem-solving.

[0094] 3. Performance-aware design optimization. By implicitly encoding complex aerodynamic knowledge into the generation model, designers without CFD expertise can create three-dimensional models that meet aerodynamic requirements at the design stage, reducing the technical threshold for high-performance design.

[0095] As shown in Figure 4 The performance index-based three-dimensional automobile model generation system provided by the embodiment of the present application comprises:

[0096] The data acquisition module 10 is configured to obtain input data, the input data comprising size parameters, angle parameters and performance parameters.

[0097] The 3D model generation module 20 is configured to input the input data into the performance index-based three-dimensional automobile model generation network as described above to obtain a 3D automobile model output by a 3D generation model, wherein the 3D automobile model meets the performance requirements and the wind resistance coefficient requirements of the input data.

[0098] In this embodiment, the performance index-based three-dimensional automobile model generation system has similar beneficial effects to the performance index-based three-dimensional automobile model generation method described above, and will not be described again here.

[0099] The electronic device provided in the embodiments of the present application includes a memory and a processor; the memory is configured to store a computer program; and the processor is configured to implement the performance index based three-dimensional automobile model generation network or the performance index based three-dimensional automobile model generation method when executing the computer program.

[0100] The computer readable storage medium provided in the embodiments of the present application has a computer program stored thereon, and when the computer program is executed by a processor, the performance index based three-dimensional automobile model generation network or the performance index based three-dimensional automobile model generation method is implemented.

[0101] In the embodiments, the electronic device and the computer readable storage medium have similar advantages to the performance index based three-dimensional automobile model generation network or the performance index based three-dimensional automobile model generation method, and will not be described here again.

[0102] Now, an electronic device that can be a server or a client of the present application will be described, which is an example of a hardware device that can be applied to various aspects of the present application. The electronic device is intended to represent a variety of forms of digital electronic computing devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistant devices, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown in the figures, their connections, and relationships, and their functions, are meant only as examples, and are not meant to limit implementations of the present application described and / or claimed in this document.

[0103] The electronic device includes a computing unit that can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) or a computer program loaded from a storage unit into a random access memory (RAM). In the RAM, various programs and data required for device operation can also be stored. The computing unit, the ROM, and the RAM are connected to each other through a bus. An input / output (I / O) interface is also connected to the bus.

[0104] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing relevant hardware through a computer program. The program can be stored in a computer readable storage medium. When the program is executed, the processes of the above-mentioned embodiment methods can be included. The storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM) or a random access memory (RAM), etc. In the present application, the modules described separately can or can not be physically separated. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiments of the present application. In addition, the functional modules in each embodiment of the present application can be integrated in one processing unit, or each module can be physically present separately, or two or more modules can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0105] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element preceded by "comprises a" does not, without more constraints, foreclose the existence of additional identical elements in the process, method, article, or apparatus that comprises the recited element.

[0106] The above embodiments are only used to illustrate the technical solutions of the present application, but not limit the technical solutions of the present application; although the technical solutions of the present application are described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced by equivalent features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A performance index-based three-dimensional automobile model generation method, characterized by, The method comprises: obtaining input data, the input data comprising size parameters, angle parameters and performance parameters; inputting the input data into a performance index-based three-dimensional automobile model generation network to obtain a 3D automobile model output by a 3D generation model, wherein the 3D automobile model meets the performance requirements and the wind resistance coefficient requirements of the input data; the performance index-based three-dimensional automobile model generation network comprises a multi-view diffusion model, a 3D generation model and a wind resistance prediction agent model; the multi-view diffusion model comprises a dimension increasing layer, a multi-layer perception, a multi-head attention block and a U-Net architecture of the diffusion model connected in sequence; the residual module of the U-Net architecture comprises a residual block, a three-dimensional self-attention block and a cross-attention block connected in sequence; the input data comprises size parameters, angle parameters and performance parameters, and is subjected to dimension increasing processing by the dimension increasing layer and projection by the corresponding multi-layer perception, and then is inputted into the multi-head attention block to obtain multi-view view features; the multi-view view features, time steps and camera parameters are inputted into the U-Net architecture of the diffusion model, and are subjected to processing by the residual block, the three-dimensional self-attention block and the cross-attention block in sequence to generate automobile four views; the automobile four views and the multi-view view features are inputted into the 3D generation model to obtain the 3D automobile model; the training process of the multi-view diffusion model comprises: point cloud data of the 3D automobile model is inputted into the pre-trained wind resistance prediction agent model to obtain a predicted wind resistance coefficient; a performance loss value is determined according to the predicted wind resistance coefficient and a theoretical wind resistance coefficient, the performance loss value is back propagated, and parameters of the multi-view diffusion model are adjusted.

2. The performance index based three-dimensional automobile model generating method of claim 1, wherein, The time steps and camera parameters are mapped by a multi-layer perceptron into fixed-dimension embedding vectors before being input to a U-Net architecture of the diffusion model ; the time steps are encoded into time step embedding vectors ; the embedding vectors and the time step embedding vectors are concatenated to form a conditioning vector e.

3. The performance index based three-dimensional automobile model generating method of claim 2, wherein, in the cross-attention block, the time steps and the camera parameters jointly constitute a condition vector to guide the generation of the automobile four views; and a formalized representation of the cross-attention block is: where X denotes a feature tensor of multi-view features, e denotes a condition vector, denotes an embedding form of the condition vector, Q(X) denotes a query matrix of the feature tensor of multi-view features, denotes a key matrix of the condition vector, denotes a value matrix of the condition vector, d denotes a dimension of the key matrix.

4. The performance index based three-dimensional automobile model generating method of claim 1, wherein, the 3D generation model comprises a DINO encoder, a plurality of multi-layer perceptions, a diffusion Transformer block, a linear transformation layer, a shape latent encoding layer and a VAE-decoder; The automobile four-view image is processed by a DINO encoder and a multi-layer perception machine to obtain image encoding The multi-view feature is processed by a multi-layer perception machine to obtain conditional vector encoding The image encoding and the conditional vector encoding are jointly input into a diffusion Transformer block, and multiple Transformer blocks in the diffusion Transformer block learn latent encoding the learning output result of the diffusion Transformer block is subjected to linear processing by the linear transformation layer and linear projection by the shape latent encoding layer to obtain a latent hidden shape feature; the latent hidden shape feature is subjected to refinement processing by a plurality of self-attention mechanism blocks in the VAE-decoder to generate the 3D automobile model.

5. The performance index based three-dimensional automobile model generating method as claimed in claim 1, wherein, the wind resistance prediction agent model is pre-trained, comprising: using a data set of 3D automobile models paired with CFD wind resistance coefficients to pre-train the wind resistance prediction agent model offline until convergence, and fixing parameters of the wind resistance prediction agent model.

6. The performance index based three-dimensional automobile model generating method of claim 1, wherein, the performance loss value is wherein, represents a performance loss value, represents a theoretical wind resistance coefficient, represents a predicted wind resistance coefficient.

7. A performance index-based three-dimensional automobile model generation system characterized by comprising: the system for implementing the performance index-based three-dimensional automobile model generation method according to any one of claims 1-6, the system comprising: a data acquisition module configured to obtain input data, the input data comprising size parameters, angle parameters and performance parameters; a 3D model generation module configured to input the input data into a performance index-based three-dimensional automobile model generation network to obtain a 3D automobile model output by a 3D generation model, wherein the 3D automobile model meets the performance requirements and the wind resistance coefficient requirements of the input data; and The performance index-based three-dimensional automobile model generation network comprises a multi-view diffusion model, a 3D generation model and a wind resistance prediction agent model; The multi-view diffusion model comprises a dimension increasing layer, a multi-layer perception, a multi-head attention block and a U-Net architecture of the diffusion model connected in sequence; the residual module of the U-Net architecture comprises a residual block, a three-dimensional self-attention block and a cross-attention block connected in sequence; The input data comprises size parameters, angle parameters and performance parameters, the input data is subjected to dimension increasing processing by the dimension increasing layer and projection by the corresponding multi-layer perception, and then is input into the multi-head attention block to obtain multi-view view features; the multi-view view features, a time step and camera parameters are input into the U-Net architecture of the diffusion model, and are sequentially subjected to processing by the residual block, the three-dimensional self-attention block and the cross-attention block to generate automobile four views; The automobile four views and the multi-view view features are input into the 3D generation model to obtain a 3D automobile model; The training process of the multi-view diffusion model comprises: Point cloud data of the 3D automobile model is input into the pre-trained wind resistance prediction agent model to obtain a predicted wind resistance coefficient; a performance loss value is determined according to the predicted wind resistance coefficient and a theoretical wind resistance coefficient, the performance loss value is back propagated, and parameters of the multi-view diffusion model are adjusted.

8. An electronic device, comprising: comprise a memory and a processor; The memory is configured to store a computer program; The processor is configured to implement the performance index-based three-dimensional automobile model generation method according to any one of claims 1 to 6 when executing the computer program.

9. A computer-readable storage medium, characterized in that, The storage medium has a computer program stored thereon, and the computer program, when executed by a processor, implements the performance index-based three-dimensional automobile model generation method according to any one of claims 1 to 6.

Images