Game model generation method and terminal
By using pre-trained GAN models and neural network models, the basic skeleton and material characteristics of the game model are automatically generated and replaced, and the problems of inconsistent styles and high labor costs in batch replacement of game models are solved, achieving rapid automatic replacement and unified art styles.
Patent Information
- Application Number
- CN202311542838.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-17
- Publication Date
- 2025-05-20
AI Technical Summary
In the batch replacement of game models in the prior art, the style cannot be unified and labor costs are huge.
The basic skeleton of the target game model is generated through the pre-trained GAN model, and the target material features are output using the pre-trained neural network model, and the target material features are set into the target game model to complete the final generation of the game model.
It realizes the rapid and automatic replacement of the game model without manual participation in design or unit replacement, which speeds up the replacement speed, ensures the unity of art style, and improves the player's gaming experience.
Smart Images
Figure CN120019837A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence, and particularly to a method for generating a game model and a terminal. Background Art
[0002] In the development of digital games, the design and replacement of game models are important components of game graphic art creation. Traditional game model replacement methods involve complex manual design, manual association, and cumbersome model testing processes. Currently, professional 3D modeling software is used on the market to create or modify game models, but this process is usually time-consuming and laborious, especially when a large number of models need to be batch replaced to update game scenes or new content, and the effect is more obvious.
[0003] In addition, texture mapping and material production in the prior art also require a large amount of human-computer interaction and the intervention of art professionals, which limits the iterative update speed of the model. In addition, the manual intervention design process may also lead to model inconsistencies, especially in large projects, and the style differences of different designers may have a negative impact on the final game product. Summary of the Invention
[0004] The technical problem to be solved by the present invention is: to provide a method for generating a game model and a terminal, which solve the problems that the batch replacement style of game models cannot be unified and the labor cost is huge.
[0005] To solve the above technical problem, the technical solution adopted by the present invention is:
[0006] A method for generating a game model, comprising the steps of:
[0007] S1. Pre-train a GAN model;
[0008] S2. Control the GAN model to output a target game model;
[0009] S3. Control the pre-trained neural network model to output a target material feature for the target game model;
[0010] S4. Set the target material feature to the target game model.
[0011] To solve the above technical problem, another technical solution adopted by the present invention is:
[0012] A game model generation terminal, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the computer program, the following steps are completed:
[0013] S1. Pre-train a GAN model;
[0014] S2. Control the GAN model to output a target game model;
[0015] S3. Control the pre-trained neural network model to output target material features for the target game model;
[0016] S4. Set the target material features to the target game model.
[0017] The beneficial effects of the present invention are as follows: A method and a terminal for generating a game model are provided. Through continuous adversarial training between the generator and the discriminator in the GAN model, it gradually approaches the target game model. When replacing game models in batches, there is no need for manual participation in designing the basic framework of each game model one by one. Subsequently, the target material features are output by the trained neural network model, and the target material features are combined with the target game model to complete the final generation of the game model. That is, the basic framework of the game model is generated using the GAN adversarial model, and the target material features of the framework are generated using the neural network model. After the two are combined, the game model is generated without manual participation in design or unit replacement, accelerating the replacement speed of the game model and realizing the rapid automatic replacement of batch game models. At the same time, the application of the GAN adversarial model and the neural network model ensures the unity of the art style and improves the game experience of players. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Figure 1 It is a flowchart of a method for generating a game model in an embodiment of the present invention;
[0019] Figure 2 It is a specific flowchart of a method for generating a game model in an embodiment of the present invention;
[0020] Figure 3 It is a schematic diagram of a game model generation terminal in an embodiment of the present invention;
[0021] Reference Numeral Description:
[0022] 1. A game model generation terminal; 2. A memory; 3. A processor. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0023] To describe the technical content, the achieved objectives, and the effects of the present invention in detail, the following is described in conjunction with the embodiments and the accompanying drawings.
[0024] Please refer to Figure 1 and Figure 2 , a method for generating a game model, including the steps:
[0025] S1. Pre-train a GAN model;
[0026] S2. Control the GAN model to output a target game model;
[0027] S3. Control the pre-trained neural network model to output target material features for the target game model;
[0028] S4. Set the target material features to the target game model.
[0029] As can be seen from the above description, the beneficial effects of the present invention are as follows: A method for generating a game model is provided. Through continuous confrontation between the generator and the discriminator in the GAN model, it gradually approaches the target game model. When replacing a batch of game models, there is no need for manual participation in designing the basic skeleton of each game model one by one. Subsequently, the target material features are output by the trained neural network model, and the target material features are combined with the target game model to complete the final generation of the game model, that is, the basic skeleton of the game model is generated using the GAN confrontation model, and the target material features of the skeleton are generated using the neural network model. After the two are combined, the game model is generated without manual participation in design or unit replacement, accelerating the replacement speed of the game model and realizing the rapid automatic replacement of a batch of game models; at the same time, the application of the GAN confrontation model and the neural network model ensures the unity of the art style and improves the game experience of players.
[0030] Specifically, the GAN model is trained based on a large amount of bone structure data. GAN stands for Generative Adversarial Network, which contains two important components: a generator and a discriminator. The generator is responsible for generating models with new bone structures, while the discriminator is responsible for judging whether the generated models are real or fake.
[0031] During the training process, the generator tries to generate as realistic models as possible, while the discriminator improves its judgment ability by comparing real data and the data generated by the generator. As the training progresses, the generator and the discriminator continuously conduct repeated confrontations. The generator tries to deceive the discriminator, while the discriminator tries to distinguish the data generated by the generator. Through this process of confrontation and learning, the generator gradually learns to generate realistic models.
[0032] An example of the application of the GAN model is as follows: Through the training data parameters, a specified model is generated, such as generating a batch of relatively strong male models. Subsequently, to meet the needs of the game, these male models can also be replaced with gorilla models through the replacement of existing models.
[0033] In some embodiments of the present invention, step S3 is specifically:
[0034] S31. Input the original image into the neural network model;
[0035] S32. Control the neural network model to extract approximate target features from the original image;
[0036] S33. Control the neural network model to stylize or reconstruct the original image using the approximate target feature to obtain the target material feature.
[0037] As can be seen from the above description, to solve the problem of batch replacement of target materials, in this example, a neural network model is used to analyze the input original image, extract the approximate target feature from the original image, and use the approximate target feature to reconstruct the original image in the neural network, thereby obtaining the target material feature.
[0038] Further, step S32 is specifically:
[0039] Control the neural network model to perform feature extraction on the original image and output texture feature maps at different levels and scales;
[0040] Use the neural network to screen out the approximate target feature from the texture feature maps at different levels and scales; the approximate target feature is the texture feature map with the highest correlation with the target material feature.
[0041] As can be seen from the above description, in order to extract the approximate target feature, the neural network model performs feature extraction on the original image and outputs texture feature maps at different levels and scales. Texture refers to the patterns or textures that repeat in an image, such as the leaves of trees, the lawns of grasslands, the grains of stones, etc. Specifically, the neural network model usually consists of multiple layers, and each layer processes the input image differently. Different levels are the levels on the neural network model, and different scales mean that the feature maps output by each corresponding level have different sizes. For example, a pre-trained convolutional neural network model is used for feature extraction, and an input image showing a stone texture is passed to the model. In the neural network model, there are usually multiple convolutional layers and pooling layers, and each layer can extract features at different levels and scales; in the bottom convolutional layer, the model will extract some low-level texture features, such as edges and colors. The size of these feature maps may be relatively large, such as 64x64 pixels. As the network level increases, the model will extract more advanced and abstract texture features, such as texture blocks, texture repetitions, and texture directions, etc. The size of these feature maps may gradually decrease, such as 16x16 pixels or smaller. Subsequently, the approximate target feature with the highest correlation with the target material feature is screened out from the output texture feature maps, and the original image is reconstructed using the approximate target feature to obtain the target material feature. Specifically, the reconstruction is performed using the approximate target feature extracted from the original image, rather than the texture features of other images, reducing the probability of style mismatch after reconstruction and improving the efficiency of the neural network in generating the target material feature.
[0042] In some embodiments of the present invention, step S1 is specifically:
[0043] S11. Generate a game model in the generator of the GAN model using adaptive normalization technology;
[0044] S12. Use a point cloud convolutional neural network to capture the shape and structural features of the game model and input them into the discriminator in the GAN model;
[0045] S13. Introduce a reconstruction loss function and a shape consistency loss function to guide the training of the GAN model.
[0046] As can be seen from the above description, in order to improve the replacement speed of the game model, the generator in the GAN model uses adaptive normalization technology to improve the smoothness and authenticity of the 3D model; specifically, the adaptive normalization technology (Adaptive Normalization) usually refers to a method of dynamically normalizing the input data or the activation of the intermediate layer in the field of data processing and neural networks. Normalization is a technology for adjusting the scale and distribution of data, which is convenient for improving efficiency and stability in subsequent processing. The adaptive normalization technology allows the model to dynamically adjust the normalization parameters to adapt to the specific statistical attributes of the data.
[0047] At the same time, the discriminator in the GAN model applies a point cloud convolutional neural network to more efficiently capture the shape and structural information of the 3D model; specifically, in the discriminator model, the input point cloud data represents the shape and structural information of the model. The point cloud convolutional neural network is a convolutional neural network specifically designed for processing point cloud data. Through point cloud convolution operations, feature extraction and aggregation can be performed on the point cloud data, so as to learn more discriminative and distinguishable feature representations. Traditional convolution operations cannot be directly applied to point cloud data, while point cloud convolution can consider the disorder of the point cloud and the relationship between different points while retaining the point cloud features, and process and analyze the point cloud data. By using point cloud convolution, local and global features of the point cloud can be extracted, and both surface geometric features and geometric structure features can be well represented. The point cloud convolution model can perform tasks such as noise reduction, feature extraction, classification, and segmentation on the point cloud. Therefore, using point cloud convolution in the discriminator model can better capture the shape and structural information of the 3D model. In this example, the specific application logic of point cloud convolution in the GAN model is to obtain more discriminative model shape and structural information from the point cloud data through feature extraction and aggregation operations, so as to more efficiently capture the shape and structural information of the 3D model.
[0048] In addition, a reconstruction loss function and a shape consistency function are introduced to guide the training of the GAN model. The reconstruction loss function is used to measure the difference between the 3D model generated by the generation model and the real model. Taking the mean square error (Mean Square Error, MSE) and the symmetric Hausdorff distance as examples, the specific details are as follows:
[0049] Mean Square Error (MSE): MSE is widely used in reconstruction loss. It calculates the average of the squares of the Euclidean distances between corresponding points of the generated model and the real model. By minimizing the MSE, the generated model can be made as close as possible to the real model.
[0050] Hausdorff distance: The Hausdorff distance is a measure of the similarity between two point clouds. In the reconstruction task, the symmetric Hausdorff distance can be used to measure the degree of shape difference between the generated model and the real model. This distance calculates the larger value between the maximum distance from each point in the generated model to the nearest point in the real model and the maximum distance from each point in the real model to the nearest point in the generated model.
[0051] The shape consistency loss function is used to guide the generated model to maintain shape features similar to the real model. Taking the Chamfer distance and the Earth Mover's Distance (EMD) as examples, the details are as follows:
[0052] Chamfer distance: The Chamfer distance is a measure of the distance between two point sets and is commonly used in shape matching and corresponding point detection. In 3D model reconstruction, the Chamfer distance can be used to measure the shape consistency between the generated model and the real model. This distance calculates the sum of the average distance from each point in the generated model to the nearest point in the real model and the average distance from each point in the real model to the nearest point in the generated model.
[0053] Earth Mover's Distance (EMD): EMD is a measure used to measure the distance between two probability distributions and can be used to measure the shape consistency between the generated model and the real model. In 3D model generation, the generated model and the real model can be regarded as two point cloud distributions, and the EMD is used to measure the shape difference between them. The EMD calculates the minimum cost flow between the two distributions, representing the average distance from the points of the generated model point cloud to the points of the real model point cloud.
[0054] In some embodiments of the present invention, after step S2, step S30 is further included:
[0055] Perform a compatibility check on the target game model for the game engine. If the check passes, replace the existing model with the target game model; otherwise, convert the format of the target game model until the compatibility check passes.
[0056] As can be seen from the above description, adding a compatibility check link can avoid compatibility problems during batch replacement, resulting in the failure of all replaced models.
[0057] A generation terminal 1 of a game model, including a memory 2, a processor 3, and a computer program stored on the memory 2 and operable on the processor 3. When the processor 3 executes the computer program, the following steps are completed:
[0058] S1. Pre-train a GAN model;
[0059] S2. Control the GAN model to output a target game model;
[0060] S3. Control the pre-trained neural network model to output target material features for the target game model;
[0061] S4. Set the target material features to the target game model.
[0062] As can be seen from the above description, an execution carrier of a method for generating a game model is provided. Through the continuous confrontation between the generator and the discriminator in the GAN model, it gradually approaches the target game model. When batch replacing game models, there is no need for manual participation in designing the basic skeleton of each game model one by one. Subsequently, the target material features are output by the trained neural network model, and the target material features are combined with the target game model to complete the final generation of the game model, that is, the basic skeleton of the game model is generated using the GAN confrontation model and the target material features of the skeleton are generated using the neural network model. After the two are combined, the game model is generated without manual participation in design or unit replacement, accelerating the replacement speed of game models and realizing the rapid automatic replacement of batch game models. At the same time, the application of the GAN confrontation model and the neural network model ensures the unity of the art style and improves the game experience of players.
[0063] In some embodiments of the present invention, step S3 is specifically:
[0064] S31. Input an original image into the neural network model;
[0065] S32. Control the neural network model to extract approximate target features from the original image;
[0066] S33. Control the neural network model to stylize or reconstruct the original image using the approximate target features to obtain target material features.
[0067] As can be seen from the above description, in order to solve the problem of batch replacement of target materials, in this example, the neural network model is used to analyze the input original image, extract approximate target features from the original image, and reconstruct the original image using the approximate target features in the neural network to obtain target material features.
[0068] In some embodiments of the present invention, step S32 is specifically:
[0069] Control the neural network model to extract features in the original image and output texture feature maps of different levels and scales;
[0070] Use the neural network to screen out approximate target features from the texture feature maps of different levels and scales; the approximate target features are the texture feature maps with the highest degree of correlation with the target material features.
[0071] As can be seen from the above description, in order to extract approximate target features, the neural network model extracts features from the original image and outputs texture feature maps of different levels and scales. Texture refers to the patterns or textures that repeatedly appear in the image, such as the leaves of trees, the lawns of grasslands, the grains of stones, etc. Specifically, different levels refer to XXX, and different scales refer to XXX. Subsequently, in the output texture feature maps, screen out the approximate target features with the highest degree of correlation with the target material features, and use the approximate target features to reconstruct the original image to obtain the target material features. Specifically, use the approximate target features extracted from the original image for reconstruction, rather than the texture features of other images, to reduce the probability of style mismatch after reconstruction and improve the efficiency of the neural network in generating target material features.
[0072] In some embodiments of the present invention, step S1 is specifically as follows:
[0073] S11. Use the adaptive normalization technique to generate a game model in the generator of the GAN model;
[0074] S12. Use the point cloud convolutional neural network to capture the shape and structural features of the game model and input them into the discriminator of the GAN model;
[0075] S13. Introduce a reconstruction loss function and a shape consistency loss function to guide the training of the GAN model.
[0076] In some embodiments of the present invention, after step S2, there is also step S30:
[0077] Perform a compatibility check on the target game model for the game engine. If the check passes, replace the existing model with the target game model; otherwise, convert the format of the target game model until the compatibility check passes.
[0078] As can be seen from the above description, adding a compatibility check link can avoid compatibility problems during batch replacement, resulting in the failure of all replaced models.
[0079] The present invention discloses a method and a terminal for generating a game model, which are mainly applied to the replacement of game models in the field of digital games. The following is a specific description in combination with embodiments.
[0080] Please refer toFigures 1 to 2 , Embodiment 1 of the present invention is as follows:
[0081] A method for generating a game model, comprising the steps of:
[0082] S1. Pre-train a GAN model;
[0083] Specifically, the GAN model is trained based on a large amount of bone structure data. GAN stands for Generative Adversarial Network, which contains two important components: a generator and a discriminator. The generator is responsible for generating models with new bone structures, while the discriminator is responsible for determining whether the generated models are real or fake.
[0084] During the training process, the generator tries to generate as realistic models as possible, while the discriminator improves its judgment ability by comparing real data and the data generated by the generator. As the training progresses, the generator and the discriminator continuously conduct repeated confrontations. The generator tries to deceive the discriminator, while the discriminator tries to distinguish the data generated by the generator. Through this process of confrontation and learning, the generator gradually learns to generate realistic models.
[0085] S2. Control the GAN model to output a target game model;
[0086] That is, use the GAN to generate a model according to the feature vector of the input bone structure. For example, when inputting the bone structure of an adult man, the GAN model may output one of the following models:
[0087] (1) Adult male model: The bone structure of this model is similar to the input bone structure, but there are detailed changes.
[0088] (2) Young male model: The GAN model may have learned the distribution of the bone structure of young men during pre-training, so it can generate a young male model.
[0089] (3) Female model: The GAN model may have learned the differences between male and female bone structures during pre-training, so it can generate a female model corresponding to the input bone structure.
[0090] (4) Unrelated models: The GAN model has creativity and randomness, so it may generate unrelated models, such as bird models, tree models, etc.
[0091] According to the evaluation index in the discriminator of the GAN model, finally generate and select the target game model, that is, the adult male model in the above (1).
[0092] S3. Control the pre-trained neural network model to output target material features for the target game model;
[0093] S4. Set the target material feature to the target game model.
[0094] That is, the target material feature is output through the trained neural network model, and the target material feature is combined with the target game model to complete the final generation of the game model. That is, the basic skeleton of the game model is generated using the GAN adversarial model, and the target material feature of the skeleton is generated using the neural network model. After the two are combined, the game model is generated without manual design or unit replacement, which speeds up the replacement speed of the game model and realizes the rapid automatic replacement of batch game models. At the same time, the application of the GAN adversarial model and the neural network model ensures the unity of the art style and improves the player's game experience.
[0095] Please refer to Figures 1 to 2 , the second embodiment of the present invention is:
[0096] Based on the first embodiment, step S3 is specifically:
[0097] S31. Input the original image into the neural network model;
[0098] S32. Control the neural network model to extract approximate target features from the original image;
[0099] S33. Control the neural network model to stylize or reconstruct the original image using the approximate target features to obtain the target material feature.
[0100] That is, to solve the problem of batch replacement of target materials, in this example, the neural network model is used to analyze the input original image, extract approximate target features from the original image, and reconstruct the original image using the approximate target features in the neural network to obtain the target material feature.
[0101] Specifically, step S32 is specifically:
[0102] Control the neural network model to perform feature extraction in the original image and output texture feature maps of different levels and scales;
[0103] Use the neural network to screen out approximate target features from the texture feature maps of different levels and scales; the approximate target feature is the texture feature map with the highest correlation with the target material feature.
[0104] The application example is as follows: There is an original image showing a stone texture, and the target material feature is the material feature of wood. The specific execution steps in the neural network are as follows:
[0105] (1) Load the pre-trained neural network model and pass the original image to the model for feature extraction. The model will output a series of feature maps, where each feature map represents the texture features of different levels and scales in the image.
[0106] (2) Select a feature map related to the target material characteristics according to the target material characteristics. In this example, select the feature map with wood texture characteristics as the approximate target feature.
[0107] (3) Then use the approximate target feature to stylize or reconstruct the original image. That is, fuse the texture feature of the original image with the approximate target feature to generate an image with the style of wood material.
[0108] Please refer to Figures 1 to 2 , Embodiment 3 of the present invention is:
[0109] Based on Embodiment 1, step S1 is specifically:
[0110] S11. Generate a game model in the generator of the GAN model using the adaptive normalization technique;
[0111] S12. Capture the shape and structural features of the game model using a point cloud convolutional neural network and input them into the discriminator in the GAN model;
[0112] S13. Introduce a reconstruction loss function and a shape consistency loss function to guide the training of the GAN model.
[0113] As can be seen from the above description, in order to improve the replacement speed of the game model, the generator in the GAN model uses the adaptive normalization technique to improve the smoothness and authenticity of the 3D model; specifically, the adaptive normalization technique (Adaptive Normalization) usually refers to a method of dynamically normalizing the input data or the activation of the intermediate layer in the fields of data processing and neural networks. Normalization is a technique for adjusting the scale and distribution of data to facilitate improving efficiency and stability in subsequent processing. The adaptive normalization technique allows the model to dynamically adjust the normalization parameters to adapt to the specific statistical attributes of the data. At the same time, the discriminator in the GAN model applies a point cloud convolutional neural network to more efficiently capture the shape and structural information of the 3D model;
[0114] Please refer to Figures 1 to 2 , Embodiment 4 of the present invention is:
[0115] Based on Embodiment 1, after step S2, step S30 is further included:
[0116] Check the compatibility of the target game model with the game engine. If the check passes, replace the existing model with the target game model; otherwise, convert the format of the target game model until the compatibility check passes. That is, add a compatibility check link to avoid compatibility problems during batch replacement, resulting in the failure of all replaced models.
[0117] Please refer toFigure 3 , Embodiment 5 of the present invention is as follows: A generation terminal 1 of a game model, including a memory 2, a processor 3, and a computer program stored on the memory 2 and operable on the processor 3. When the processor 3 executes the computer program, it completes the steps in the generation method of any one of the above-mentioned Embodiments 1 to 4 of the game model.
[0118] In summary, the generation method and terminal of a game model provided by the present invention use the generator and discriminator in the GAN model to continuously conduct adversarial training to gradually approach the target game model. When replacing game models in batches, there is no need for manual participation in designing the basic skeleton of each game model one by one. Subsequently, the target material features are output through the trained neural network model, and the target material features are combined with the target game model to complete the final generation of the game model, that is, the basic skeleton of the game model is generated using the GAN adversarial model and the target material features of the skeleton are generated using the neural network model. After the two are combined, the game model is generated without manual participation in design or unit replacement, accelerating the replacement speed of game models and realizing the rapid automatic replacement of game models in batches. At the same time, the application of the GAN adversarial model and the neural network model ensures the unity of the art style and improves the game experience of players.
[0119] The above are only the embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent transformation made using the content of the specification and drawings of the present invention, or directly or indirectly applied in the relevant technical fields, shall be equally included in the patent protection scope of the present invention.
Claims
1. A method for generating a game model, characterized in that: Includes steps: S1, pre-training GAN model; S2, control the GAN model to output the target game model; S3, controlling the pre-trained neural network model to output target material features for the target game model; S4. Setting the target material characteristics to the target game model.
2. A method for generating a game model according to claim 1, characterized in that: The step S3 is specifically as follows: S31, inputting the original image into the neural network model; S32, controlling the neural network model to extract approximate target features in the original image; S33, controlling the neural network model to stylize or reconstruct the original image using the approximate target features to obtain target material features.
3. A method for generating a game model according to claim 2, characterized in that: The step S32 is specifically as follows: Controlling the neural network model to extract features from the original image and output texture feature maps of different levels and scales; The neural network is used to screen out approximate target features from texture feature maps at different levels and scales; the approximate target features are the texture feature maps with the greatest correlation with the target material features.
4. The method for generating a game model according to claim 1, characterized in that: The step S1 is specifically as follows: S11. Generate a game model in the generator of the GAN model using adaptive normalization technology; S12, using point cloud convolutional neural network to capture the shape and structural features of the game model and input them into the discriminator in the GAN model; S13. Reconstruction loss function and shape consistency loss function are introduced to guide the training of the GAN model.
5. The method for generating a game model according to claim 1, characterized in that: The step S2 further includes a step S30: The target game model is subjected to a compatibility check for the game engine. If the check passes, the target game model replaces the existing model; otherwise, the format of the target game model is converted until the compatibility check passes.
6. A game model generation terminal, characterized in that: The invention comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the following steps are performed: S1, pre-training GAN model; S2, control the GAN model to output the target game model; S3, controlling the pre-trained neural network model to output target material features for the target game model; S4. Setting the target material characteristics to the target game model.
7. A game model generation terminal according to claim 6, characterized in that: The step S3 is specifically as follows: S31, inputting the original image into the neural network model; S32, controlling the neural network model to extract approximate target features in the original image; S33, controlling the neural network model to stylize or reconstruct the original image using the approximate target features to obtain target material features.
8. A game model generation terminal according to claim 7, characterized in that: The step S32 is specifically as follows: Controlling the neural network model to extract features from the original image and output texture feature maps of different levels and scales; The neural network is used to screen out approximate target features from texture feature maps at different levels and scales; the approximate target features are the texture feature maps with the greatest correlation with the target material features.
9. A game model generation terminal according to claim 6, characterized in that: The step S1 is specifically as follows: S11. Generate a game model in the generator of the GAN model using adaptive normalization technology; S12, using point cloud convolutional neural network to capture the shape and structural features of the game model and input them into the discriminator in the GAN model; S13. Reconstruction loss function and shape consistency loss function are introduced to guide the training of the GAN model.
10. A game model generation terminal according to claim 6, characterized in that: The step S2 further includes a step S30: The target game model is subjected to a compatibility check for the game engine. If the check passes, the target game model replaces the existing model; otherwise, the format of the target game model is converted until the compatibility check passes.