An artificial intelligence enabled game image generation system and method

Abstract

The application provides a game image generation system and method based on artificial intelligence empowerment, converts collected multi-dimensional data into an environment state description vector; performs convolutional feature extraction on a reference image in the environment state description vector to obtain a style representation tensor, fuses and reduces dimensions of the style representation tensor to obtain a target style identification vector; performs difference measurement on candidate image features of a candidate generated image and the target style identification vector to obtain a style difference matrix, determines a style consistency penalty term in image generation based on the style difference matrix, and constructs a game image generation model using the style consistency penalty term; and generates a game style image according to the game image generation model and the environment state description vector. The technical scheme provided by the application can extract a target style identification vector from multi-dimensional data flow, and construct a closed-loop control mechanism in which a style difference measurement is used as a penalty term and fed back to a generation model, so as to realize efficient and dynamic generation of game images.

Description

Technical Field

[0001] This application relates to the field of image generation technology, and more specifically, to a game image generation system and method based on artificial intelligence. Background Technology

[0002] In recent years, with the rapid development of artificial intelligence technology, especially the maturity of deep generative models such as generative adversarial networks and diffusion models, the application of artificial intelligence in the field of game image generation has become increasingly widespread. These technologies have been used to generate static assets such as concept art for game characters, scene textures, and item icons, significantly improving the efficiency of early-stage art design.

[0003] In existing technologies, artificial intelligence is deeply integrated into the game runtime environment to achieve real-time, dynamic, and adaptive image generation. However, when dynamically generating or adjusting images during game runtime, current methods often perform style matching or filtering after generation. This process is slow and computationally expensive, lacking a mechanism to directly apply strong style constraints during generation. In other words, it cannot start from a unified style identifier to actively guide and regulate the generation process, leading to a visual disconnect between dynamically generated content and existing assets, thus disrupting the unity of the game world. Therefore, the key challenge is how to extract target style identifier vectors from multi-dimensional data streams and construct a closed-loop control mechanism that uses style difference measurement as a penalty to feed back to the generation model, in order to achieve efficient dynamic generation of game images. Summary of the Invention

[0004] This application provides a game image generation system and method based on artificial intelligence, which can extract target style identifier vectors from multi-dimensional data streams and construct a closed-loop control mechanism that uses style difference measurement as a penalty to feed back to the generation model, so as to achieve efficient dynamic generation of game images.

[0005] In a first aspect, this application provides a game image generation method based on artificial intelligence, comprising the following steps:

[0006] The system collects multi-dimensional data streams during game runtime in real time and then transforms these multi-dimensional data streams into a unified semantic environment state description vector.

[0007] Convolutional feature extraction is performed on the reference image in the environmental state description vector to obtain a style representation tensor. The style representation tensor is then fused and dimensionality reduced to obtain the target style identifier vector.

[0008] Candidate generated images are obtained, and the difference is measured based on the candidate image features and the target style identifier vector to obtain a style difference matrix. Based on the style difference matrix, a style consistency penalty in image generation is determined, and a game image generation model is constructed using the style consistency penalty.

[0009] A game-style image is generated during game runtime based on the game image generation model and the environment state description vector.

[0010] In some embodiments, the multi-dimensional data stream includes character and NPC position and orientation, posture skeleton state, camera parameters, scene element visibility and spatial relationships, object material and texture state, light intensity and direction, shadow and reflection parameters, particle and special effects state, weather and environmental effect parameters, spatial distribution of important dynamic elements in the scene, and reference images.

[0011] In some embodiments, converting the multi-dimensional data stream into a unified semantic environment state description vector specifically includes:

[0012] The multi-dimensional data stream is divided into visual layer data, physical layer data, interaction layer data and system layer data according to its source and attributes. Then, time alignment, unit normalization and outlier removal are performed on the data of each layer respectively.

[0013] For each layer of data, a specific semantic mapping model is constructed, and then the semantic features of each layer are output.

[0014] A semantic space is obtained by learning semantic features from different layers through a semantic embedding network;

[0015] Within the semantic space, the multimodal semantic features at the current moment are aggregated using an aggregation function to generate a unified semantic environment state description vector.

[0016] In some embodiments, performing convolutional feature extraction on the reference image in the environment state description vector to obtain a style representation tensor specifically includes:

[0017] The reference image in the environmental state description vector is input into the feature pyramid constructed by the multi-layer convolutional neural network for extraction, and the convolutional feature map corresponding to each layer is obtained.

[0018] The convolutional feature maps corresponding to each layer are fused according to preset weights to obtain fused convolutional feature maps.

[0019] The fused convolutional feature maps are mapped and integrated into a style representation tensor.

[0020] In some embodiments, fusing and reducing the dimensionality of the style representation tensor to obtain the target style identifier vector specifically includes:

[0021] The style representation tensor is weighted and converged by the attention module of the input channel to obtain the style feature vector;

[0022] The spatial dimension of the style feature vector is compressed to obtain the target style identifier vector.

[0023] In some embodiments, the process of measuring the difference between the candidate image features of the candidate generated image and the target style identifier vector to obtain a style difference matrix specifically includes:

[0024] The color difference is measured by analyzing the candidate image features of the candidate generated image and the target style identifier vector to obtain the color difference degree.

[0025] The brightness difference is measured by analyzing the candidate image features and the target style identifier vector of the candidate generated image to obtain the brightness difference degree.

[0026] The contrast difference is measured between the candidate image features of the candidate generated image and the target style identifier vector to obtain the contrast difference degree;

[0027] The candidate image features and the target style identifier vector of the candidate generated image are subjected to local texture difference measurement to obtain the local texture difference degree;

[0028] A style difference matrix is ​​obtained based on the color difference, the brightness difference, the contrast difference, and the local texture difference.

[0029] In some embodiments, generating a game-style image during game runtime based on the game image generation model and the environment state description vector involves inputting the environment state description vector as input data into the game image generation model, thereby generating a game-style image during game runtime.

[0030] Secondly, this application provides an AI-enabled game image generation system for executing an AI-enabled game image generation method, including:

[0031] The data acquisition module is used to collect multi-dimensional data streams during game runtime in real time, and then transform the multi-dimensional data streams into unified semantic environment state description vectors.

[0032] The style extraction module is used to perform convolutional feature extraction on the reference image in the environmental state description vector to obtain a style representation tensor, and to perform fusion and dimensionality reduction on the style representation tensor to obtain the target style identifier vector.

[0033] The model building module is used to acquire candidate generated images, measure the difference between the candidate image features and the target style identifier vector to obtain a style difference matrix, determine the style consistency penalty in image generation based on the style difference matrix, and use the style consistency penalty to build a game image generation model.

[0034] The image generation module is used to generate game-style images during game runtime based on the game image generation model and the environment state description vector.

[0035] Thirdly, this application provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described AI-enabled game image generation method.

[0036] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned AI-enabled game image generation method.

[0037] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects:

[0038] This application provides an AI-enabled game image generation system and method, which involves real-time acquisition of multi-dimensional data streams during game execution, transforming these data streams into a unified semantic environment state description vector; performing convolutional feature extraction on reference images within the environment state description vector to obtain a style representation tensor; fusing and reducing the dimensionality of the style representation tensor to obtain a target style identifier vector; acquiring candidate generated images; measuring the difference between the candidate image features and the target style identifier vector to obtain a style difference matrix; determining a style consistency penalty in image generation based on the style difference matrix; constructing a game image generation model using the style consistency penalty; and generating game style images during game execution based on the game image generation model and the environment state description vector.

[0039] Therefore, this application firstly transforms the multi-dimensional data stream during game execution into a semantic environment state description vector, enabling the generation system to understand the overall state of the game scene in a high-level, structured manner. Then, by performing convolutional feature extraction on the reference image in the environment state description vector, the multi-scale, global, and local style features of the reference image are encoded into a high-dimensional style representation tensor. This is then fused and reduced to obtain a compact target style identifier vector, allowing the generation model to acquire complete style information in a low-dimensional, operable form, removing redundant spatial information. Secondly, by combining candidate generated image features with the target style vector... The method measures the differences in quantity to form a style difference matrix, and calculates a style consistency penalty based on this matrix. This penalty is then fed back to the game image generation model in real time, enabling dynamic constraints on the style deviation of the generated images. Finally, the game image generation model directly generates stylized images that conform to the semantics of the current scene, thereby achieving synchronous control of content and style. This not only ensures that the generated images are consistent with the target style in multiple dimensions such as color, brightness, texture, and composition, but also significantly improves generation efficiency and real-time performance. This allows images in the game scene to maintain visual coherence and style unity under dynamic changing conditions, achieving efficient, stable, and controllable dynamic generation of game images.

[0040] In summary, the technical solution adopted in this application can extract target style identifier vectors from multi-dimensional data streams and construct a closed-loop control mechanism that uses style difference measurement as a penalty to feed back to the generation model, so as to achieve efficient dynamic generation of game images. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this embodiment of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is an exemplary flowchart of an AI-enabled game image generation method according to some embodiments of this application;

[0043] Figure 2 This is an exemplary flowchart illustrating the determination of a style representation tensor according to some embodiments of this application;

[0044] Figure 3 This is a schematic diagram of the structure of an AI-enabled game image generation system according to some embodiments of this application;

[0045] Figure 4This is a schematic diagram of the structure of a computer device that implements an AI-enabled game image generation method according to some embodiments of this application. Detailed Implementation

[0046] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0047] This application provides an AI-enabled game image generation system and method. Its core is to collect multi-dimensional data streams during game execution in real time, and then transform these multi-dimensional data streams into a unified semantic environment state description vector. Convolutional feature extraction is performed on reference images in the environment state description vector to obtain a style representation tensor. This style representation tensor is then fused and dimensionality-reduced to obtain a target style identifier vector. Candidate generated images are acquired, and a difference measurement is performed between the candidate image features and the target style identifier vector to obtain a style difference matrix. A style consistency penalty is determined based on the style difference matrix, and a game image generation model is constructed using the style consistency penalty. Finally, a game style image is generated based on the game image generation model and the environment state description vector. This approach can extract the target style identifier vector from the multi-dimensional data stream and construct a closed-loop control mechanism that uses style difference measurement as a penalty to feed back to the generation model, thereby achieving efficient dynamic generation of game images.

[0048] To better understand the above technical solutions, a detailed description of the technical solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. (Refer to...) Figure 1 The figure is an exemplary flowchart of an AI-enabled game image generation method according to some embodiments of this application, which mainly includes the following steps:

[0049] In step S101, multi-dimensional data streams are collected in real time during game execution, and then the multi-dimensional data streams are transformed into unified semantic environment state description vectors.

[0050] It should be noted that, in this application, the multi-dimensional data stream includes the position and orientation of characters and NPCs, posture and skeleton status, camera parameters, visibility and spatial relationships of scene elements, material and texture status of objects, light intensity and direction, shadow and reflection parameters, particle and special effects status, weather and environmental effect parameters, spatial distribution of important dynamic elements in the scene, and reference images.

[0051] In practical implementation, firstly, lightweight acquisition interfaces are embedded in the game engine's rendering, scene management, character control, and special effects modules. A unified acquisition standard is defined for image generation-related events, including object position, orientation, pose, material state, lighting intensity, shadow parameters, particle effect state, camera parameters (viewpoint, focal length, frustum), scene element visibility, and texture loading information. Pre-set reference images are obtained, and each data point is tagged with a high-precision timestamp and context identifier. Then, all acquired data is stored in a local buffer queue and aggregated by frame or time window to ensure continuous sampling of frequently changing rendering states, character actions, and lighting changes. Simultaneously, timestamps are uniformly synchronized to ensure strict temporal alignment of multi-dimensional data. Secondly, the acquired raw rendering data, object states, and lighting parameters are deduplicated, normalized, and compressed to reduce transmission load and maintain data accuracy, while removing metrics irrelevant to image generation. Furthermore, key rendering states, object attributes, and lighting information are uploaded to the generation module in real time via a low-latency channel, while secondary or high-frequency data is uploaded in batches to ensure real-time performance and integrity. Finally, the server decodes, verifies, and reconstructs the timing of the received data, organizing the multi-source object states, lighting information, and camera parameters into a unified data structure, thereby obtaining a multi-dimensional data stream during game runtime.

[0052] In some embodiments, the transformation of the multi-dimensional data stream into a unified semantic environment state description vector can be achieved in the following ways:

[0053] The multi-dimensional data stream is divided into visual layer data, physical layer data, interaction layer data and system layer data according to its source and attributes. Then, time alignment, unit normalization and outlier removal are performed on the data of each layer respectively.

[0054] For each layer of data, a specific semantic mapping model is constructed, and then the semantic features of each layer are output.

[0055] A semantic space is obtained by learning semantic features from different layers through a semantic embedding network;

[0056] Within the semantic space, the multimodal semantic features at the current moment are aggregated using an aggregation function to generate a unified semantic environment state description vector.

[0057] In practical implementation, firstly, the real-time acquired multi-dimensional data stream is divided into visual layer data (rendering and scene object information), physical layer data (position, velocity, collision, force feedback), interaction layer data (player input, UI events, interaction targets), and system layer data (frame rate, performance, network status) according to its source and attributes. Time alignment, unit normalization, and outlier removal are performed on the data of each layer to ensure that each dimension can be expressed on a unified timeline. Then, a dedicated semantic mapping model can be designed for each layer—for example, the visual layer uses a lightweight object detection / semantic segmentation and feature pyramid encoder to extract object category, visibility, and saliency vectors; the physical layer uses a dynamic encoder to map position, velocity, and contact relationships into entity state vectors; the interaction layer converts input sequences and intent labels into operation intent vectors using a sequence encoder or behavior embedder; and the system layer maps frame rate, rendering load, and texture status into resource quality vectors. Through these methods, the semantic features of each layer can be obtained.

[0058] In addition, in practical implementation, semantic embedding networks can be used to learn semantic features from different layers. That is, multimodal semantic features of a unified dimension are fed into the semantic embedding network (e.g., a fusion network based on multi-head attention or a graph neural network). Within a time window, the network uses attention weights or edge weights to model the spatial-temporal relationships between entities and regions, learning a coherent semantic space representation, thus obtaining the semantic space. Then, within the semantic space, the multimodal semantic features at the current moment can be aggregated using an aggregation function to generate a unified semantic environment state description vector. That is, within this semantic space, an aggregation function (e.g., weighted attention convergence, combined with sliding window time smoothing or recursive update mechanism) is applied to the multimodal features at the current moment or within the time window to generate a fixed-dimensional environment state description vector, with attached meta-information indexes for reverse reconstruction to local semantic constraints when necessary. Here, the environment state description vector is a feature vector that comprehensively reflects the global semantic state of the game scene.

[0059] It should be noted that by unifying the multi-dimensional data streams during game runtime into semantic environment state description vectors, the generation system can understand the overall state of the game scene in a high-level, structured way, thereby achieving intelligent perception and adaptive control of image generation.

[0060] In step S102, convolutional feature extraction is performed on the reference image in the environmental state description vector to obtain a style representation tensor. The style representation tensor is then fused and dimensionality reduced to obtain the target style identifier vector.

[0061] Preferably, in some embodiments, reference is made to Figure 2As shown, this figure is an exemplary flowchart of determining a style representation tensor according to some embodiments of this application. In this embodiment, the style representation tensor is obtained by performing convolutional feature extraction on the reference image in the environment state description vector using the following steps:

[0062] In step S1021, the reference image in the environmental state description vector is input into the feature pyramid constructed by the multi-layer convolutional neural network for extraction, so as to obtain the convolutional feature map corresponding to each layer.

[0063] In step S1022, the convolutional feature maps corresponding to each layer are fused according to preset weights to obtain fused convolutional feature maps;

[0064] In step S1023, the fused convolutional feature map is mapped and integrated into a style representation tensor.

[0065] In practical implementation, firstly, the reference image in the environment state description vector can be input into the feature pyramid constructed by a multi-layer convolutional neural network for extraction. That is, the reference image corresponding to the environment state description vector is used as input and fed into the feature pyramid constructed by the multi-layer convolutional neural network. Different layers of the multi-layer convolutional neural network extract features at different scales: shallow layers capture color distribution, brightness changes, and local texture; middle layers capture brush stroke direction, texture patterns, and edge information; and deep layers extract overall composition, spatial layout, and lighting relationships. The convolutional feature maps output by each layer retain spatial structure information for subsequent fusion, thus obtaining the convolutional feature map corresponding to each layer. Then, the convolutional feature maps corresponding to each layer can be fused according to preset weights, i.e., using channel attention or layer-by-layer weighted summation, to obtain a fused convolutional feature map. This ensures that the fused feature map retains both local texture details and highlights global composition and lighting information, thereby strengthening key style features. Finally, the fused convolutional feature maps can be integrated into a style representation tensor, which is a high-dimensional, multi-channel feature data structure used to comprehensively represent the visual style features of the reference image. This style representation tensor not only contains global information such as color, brightness, and contrast, but also multi-level local features such as brushstrokes, texture, local details, composition, and lighting. Therefore, it can describe the overall style of the image in a multi-dimensional and continuous manner. In actual implementation, the fused convolutional feature maps can be integrated through a fully connected mapping layer to form a high-dimensional style representation tensor.

[0066] In some embodiments, the style representation tensor is fused and dimensionality reduced to obtain the target style identifier vector, which can be achieved in the following manner:

[0067] The style representation tensor is weighted and converged by the attention module of the input channel to obtain the style feature vector;

[0068] The spatial dimension of the style feature vector is compressed to obtain the target style identifier vector.

[0069] In practical implementation, firstly, the style representation tensor input channel attention module can be weighted and converged. This involves assigning attention weights to each channel of the style representation tensor, enabling the network to automatically identify and enhance the most important feature channels for overall style expression, such as color, brushstrokes, texture, or lighting features, while suppressing redundant or noisy information. Through weighted summation and channel attention mechanisms, multi-channel features are integrated into a preliminary style feature vector, preserving various style components while highlighting key visual features. Next, the spatial dimension of the style feature vector can be compressed. This compression can be achieved using methods such as adaptive average pooling, global average pooling, or principal component analysis (PCA), without limitation. The goal is to aggregate high-dimensional spatial information into a single value, thereby removing unnecessary spatial redundancy while maintaining the complete expression of global style features. The resulting vector can then be used as the target style identifier vector, which is a low-dimensional vector obtained from the high-dimensional style representation tensor through fusion, weighting, and dimensionality reduction. This target style identifier vector not only includes key visual style features such as color, brushstrokes, texture, lighting, and composition, but also removes redundant spatial information and unimportant channel information, making it both compact and able to fully express the overall style of the image.

[0070] It should be noted that by performing convolutional feature extraction on the reference image in the environmental state description vector, the multi-scale, global and local style features of the reference image are encoded into a high-dimensional style representation tensor. Then, through fusion and dimensionality reduction, a compact target style identifier vector is obtained, enabling the generative model to obtain complete style information in a low-dimensional and operable form. This not only preserves core features such as color, brushstrokes, texture, lighting and composition, but also removes redundant spatial information, allowing the model to quickly and accurately match the reference style during the generation process.

[0071] In step S103, candidate generated images are obtained, and a difference measurement is performed based on the candidate image features of the candidate generated images and the target style identifier vector to obtain a style difference matrix. Based on the style difference matrix, a style consistency penalty in image generation is determined, and a game image generation model is constructed using the style consistency penalty.

[0072] In practice, image candidates can be initially generated using a generator based on a generative adversarial network, i.e., candidate generated images.

[0073] In some embodiments, the style difference matrix is ​​obtained by performing a difference measurement based on the candidate image features of the candidate generated image and the target style identifier vector. Specifically, this can be achieved in the following manner:

[0074] The color difference is measured by analyzing the candidate image features of the candidate generated image and the target style identifier vector to obtain the color difference degree.

[0075] The brightness difference is measured by analyzing the candidate image features and the target style identifier vector of the candidate generated image to obtain the brightness difference degree.

[0076] The contrast difference is measured between the candidate image features of the candidate generated image and the target style identifier vector to obtain the contrast difference degree;

[0077] The candidate image features and the target style identifier vector of the candidate generated image are subjected to local texture difference measurement to obtain the local texture difference degree;

[0078] A style difference matrix is ​​obtained based on the color difference, the brightness difference, the contrast difference, and the local texture difference.

[0079] In specific implementation, firstly, feature extraction can be performed on the candidate generated image using the same feature extraction method as the reference image, thus obtaining candidate image features. Then, the candidate image features are compared with the target style identifier vector in the color correlation channel; that is, the cross-similarity of the color histogram is calculated, and this cross-similarity is used as the color difference, which quantifies the degree of deviation between the candidate image and the target style in the overall color distribution. Secondly, the difference between the variances of the brightness channel of the candidate image features and the corresponding dimension of the target style vector is calculated, and the result is used as the brightness difference, which is used to evaluate the deviation of the candidate generated image in terms of brightness levels and illumination intensity. Furthermore, the difference between the feature contrast in the candidate image features and the corresponding feature in the target style identifier vector is used as the contrast difference, which quantifies the degree of difference in the overall visual impact of the candidate generated image.

[0080] In addition, in practical implementation, the high-frequency texture features in the candidate image features (such as those represented by convolutional features, Fourier spectra, or Gram matrices) can be compared with the corresponding texture features in the target style vector. For example, the Gram matrix Frobenius norm or local autocorrelation difference can be used to calculate the similarity. The result can then be used to obtain the local texture difference, which represents the degree of difference in local texture contrast between the candidate generated images. Finally, the data matrix composed of color difference, brightness difference, contrast difference, and local texture difference can be used as the style difference matrix. This involves arranging the four types of difference measures according to feature type, with each row corresponding to different semantic regions of the candidate generated image, and each column corresponding to color, brightness, contrast, and local texture differences. Each difference value is then filled into a matrix cell to form the complete style difference matrix.

[0081] In some embodiments, a style consistency penalty in image generation is determined based on the style difference matrix. It should be noted that, in this application, the style consistency penalty is a loss or penalty signal used to constrain the image generation model, making the generated image as visually consistent as possible with the reference style. This style consistency penalty quantifies the difference between the candidate generated image and the target style identifier vector, converting this deviation into a calculable value, and adding it as a constraint to the loss function during the training or inference process of the generation model, guiding the model to adaptively adjust its output. Specifically, all differences in the style difference matrix can be normalized, and then the normalized differences can be weighted and summed. The corresponding weights can be set according to historical experiments, and the calculation result is used as the style consistency penalty in image generation.

[0082] In some embodiments, the style consistency penalty is used to construct a game image generation model. Specifically, firstly, the previously calculated style consistency penalty can be weighted and fused with the original loss function to form a total loss function. This allows the generator to optimize the accuracy of image content during training or inference while constraining the image to maintain consistency with the target style in dimensions such as color, brightness, contrast, and texture. Then, during training, the total loss function is backpropagated to the parameters of the generation model, iteratively updating the network weights so that the generator gradually learns to generate style-consistent images under different scenes and semantic conditions. During inference, incremental updates or iterative iteration strategies can be used to allow the style consistency penalty to guide the generation process in real time. The game image generation model can then be constructed in this way.

[0083] It should be noted that a style difference matrix is ​​formed by measuring the difference between the features of the candidate generated images and the target style vector. A style consistency penalty is then calculated based on this style difference matrix and fed back to the game image generation model in real time. This achieves dynamic constraints on the style deviation of the generated images, ensuring that the generated images are consistent with the target style in multiple dimensions such as color, brightness, contrast, and texture. It also allows the game image generation model to adaptively adjust its output according to scene changes, thereby significantly improving the efficiency, real-time performance, and style consistency of image generation. This makes the dynamically generated images in the game scene both visually coherent and in line with design requirements.

[0084] In step S104, a game style image is generated based on the game image generation model and the environment state description vector during game runtime.

[0085] In some embodiments, generating a game-style image during game runtime based on the game image generation model and the environment state description vector involves inputting the environment state description vector as input data into the game image generation model, thereby generating a game-style image during game runtime.

[0086] In practice, firstly, the environmental state description vector, acquired in real time and processed, is provided as input data to the game image generation model. Then, upon receiving this input, the game image generation model maps the environmental semantic information to the image space through forward propagation calculations using a neural network. Simultaneously, during the generation process, it is constrained by a style consistency penalty term, dynamically adjusting the output features to match the target style. Finally, the game image generation model outputs a game image that conforms to the current environmental state and has a unified style—that is, a game-style image during game runtime—which can be used for real-time rendering or subsequent image processing and display, achieving dynamic unification of scene content and visual style.

[0087] It should be noted that by directly generating stylized images that conform to the semantics of the current scene through the game image generation model, the synchronous control of content and style can be achieved. This not only ensures that the generated images are consistent with the target style in multiple dimensions such as color, brightness, texture and composition, but also significantly improves the generation efficiency and real-time performance. This enables the images in the game scene to maintain visual coherence and stylistic unity under dynamic changing conditions, achieving efficient, stable and controllable dynamic generation of game images.

[0088] Therefore, this application firstly transforms the multi-dimensional data stream during game execution into a semantic environment state description vector, enabling the generation system to understand the overall state of the game scene in a high-level, structured manner. Then, by performing convolutional feature extraction on the reference image in the environment state description vector, the multi-scale, global, and local style features of the reference image are encoded into a high-dimensional style representation tensor. This is then fused and reduced to obtain a compact target style identifier vector, allowing the generation model to acquire complete style information in a low-dimensional, operable form, removing redundant spatial information. Secondly, by combining candidate generated image features with the target style vector... The method measures the differences in quantity to form a style difference matrix, and calculates a style consistency penalty based on this matrix. This penalty is then fed back to the game image generation model in real time, enabling dynamic constraints on the style deviation of the generated images. Finally, the game image generation model directly generates stylized images that conform to the semantics of the current scene, thereby achieving synchronous control of content and style. This not only ensures that the generated images are consistent with the target style in multiple dimensions such as color, brightness, texture, and composition, but also significantly improves generation efficiency and real-time performance. This allows images in the game scene to maintain visual coherence and style unity under dynamic changing conditions, achieving efficient, stable, and controllable dynamic generation of game images.

[0089] In summary, the technical solution adopted in this application can extract target style identifier vectors from multi-dimensional data streams and construct a closed-loop control mechanism that uses style difference measurement as a penalty to feed back to the generation model, so as to achieve efficient dynamic generation of game images.

[0090] Furthermore, in another aspect of this application, in some embodiments, this application provides a game image generation system based on artificial intelligence, referencing... Figure 3 The figure is a schematic diagram of the structure of an AI-enabled game image generation system according to some embodiments of this application. The AI-enabled game image generation system includes:

[0091] Data acquisition module 201 is used to acquire multi-dimensional data streams during game runtime in real time, and then convert the multi-dimensional data streams into unified semantic environment state description vectors.

[0092] The style extraction module 202 is used to perform convolutional feature extraction on the reference image in the environmental state description vector to obtain a style representation tensor, and to perform fusion and dimensionality reduction on the style representation tensor to obtain the target style identifier vector.

[0093] The model building module 203 is used to acquire candidate generated images, perform difference measurement based on the candidate image features of the candidate generated images and the target style identifier vector to obtain a style difference matrix, determine the style consistency penalty in image generation based on the style difference matrix, and use the style consistency penalty to build a game image generation model.

[0094] The image generation module 204 is used to generate a game-style image during game runtime based on the game image generation model and the environment state description vector.

[0095] In addition, this application also provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described AI-enabled game image generation method.

[0096] In some embodiments, reference Figure 4 The figure is a schematic diagram of the structure of a computer device implementing an AI-enabled game image generation method according to some embodiments of this application. The AI-enabled game image generation method in the above embodiments can... Figure 4 The computer device shown is used to implement this, and the computer device includes at least one processor 301, a communication bus 302, a memory 303, and at least one communication interface 304.

[0097] The processor 301 may be a general-purpose central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more devices used to control the execution of the AI-enabled game image generation method described in this application.

[0098] The communication bus 302 can be used to transmit information between the aforementioned components.

[0099] The memory 303 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 303 may exist independently and be connected to the processor 301 via the communication bus 302. The memory 303 may also be integrated with the processor 301.

[0100] The memory 303 stores program code for executing the scheme of this application, and its execution is controlled by the processor 301. The processor 301 executes the program code stored in the memory 303. The program code may include one or more software modules. In the above embodiments, the determination of the AI-enabled game image generation method can be implemented by the processor 301 and one or more software modules in the program code in the memory 303.

[0101] Communication interface 304 uses any transceiver-like device for communicating with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0102] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single-CPU) processor or a multi-core (multi-CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0103] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.

[0104] In addition, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described AI-enabled game image generation method.

[0105] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0106] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A game image generation method based on artificial intelligence, characterized in that, Includes the following steps: The system collects multi-dimensional data streams during game runtime in real time and then transforms these multi-dimensional data streams into a unified semantic environment state description vector. Convolutional feature extraction is performed on the reference image in the environmental state description vector to obtain a style representation tensor. The style representation tensor is then fused and dimensionality reduced to obtain the target style identifier vector. Candidate generated images are acquired, and color difference, brightness difference, contrast difference, and local texture difference are measured on the candidate image features and the target style identifier vector of the candidate generated images, respectively, to obtain color difference degree, brightness difference degree, contrast difference degree, and local texture difference degree. Based on the color difference degree, brightness difference degree, contrast difference degree, and local texture difference degree, a style difference matrix is ​​obtained. Based on the style difference matrix, a style consistency penalty term in image generation is determined, and a game image generation model is constructed using the style consistency penalty term. Among them, color difference degree is used to quantify the degree of deviation between the candidate image and the target style in the overall color distribution; brightness difference degree is used to evaluate the deviation of the candidate generated image in terms of brightness and darkness levels and illumination intensity; contrast difference degree is used to quantify the degree of difference in the overall visual impact of the candidate generated image; and local texture difference degree is used to represent the degree of difference in the local texture contrast of the candidate generated image. A game-style image is generated during game runtime based on the game image generation model and the environment state description vector.

2. The game image generation method based on artificial intelligence as described in claim 1, characterized in that, The multi-dimensional data stream includes the position and orientation of characters and NPCs, posture and skeleton status, camera parameters, visibility and spatial relationships of scene elements, material and texture status of objects, light intensity and direction, shadow and reflection parameters, particle and special effects status, weather and environmental effect parameters, spatial distribution of important dynamic elements in the scene, and reference images.

3. The game image generation method based on artificial intelligence as described in claim 1, characterized in that, Transforming the multi-dimensional data stream into a unified semantic environment state description vector specifically includes: The multi-dimensional data stream is divided into visual layer data, physical layer data, interaction layer data and system layer data according to its source and attributes. Then, time alignment, unit normalization and outlier removal are performed on the data of each layer respectively. For each layer of data, a specific semantic mapping model is constructed, and then the semantic features of each layer are output. A semantic space is obtained by learning semantic features from different layers through a semantic embedding network; Within the semantic space, the multimodal semantic features at the current moment are aggregated using an aggregation function to generate a unified semantic environment state description vector.

4. The game image generation method based on artificial intelligence as described in claim 1, characterized in that, Convolutional feature extraction is performed on the reference image in the environmental state description vector to obtain the style representation tensor, specifically including: The reference image in the environmental state description vector is input into the feature pyramid constructed by the multi-layer convolutional neural network for extraction, and the convolutional feature map corresponding to each layer is obtained. The convolutional feature maps corresponding to each layer are fused according to preset weights to obtain fused convolutional feature maps. The fused convolutional feature maps are mapped and integrated into a style representation tensor.

5. The game image generation method based on artificial intelligence as described in claim 1, characterized in that, The fusion and dimensionality reduction of the style representation tensor to obtain the target style identifier vector specifically includes: The style representation tensor is weighted and converged by the attention module of the input channel to obtain the style feature vector; The spatial dimension of the style feature vector is compressed to obtain the target style identifier vector.

6. The game image generation method based on artificial intelligence as described in claim 1, characterized in that, The process of generating a game-style image during game runtime based on the game image generation model and the environment state description vector involves inputting the environment state description vector as input data into the game image generation model, thereby generating the game-style image during game runtime.

7. A game image generation system based on artificial intelligence, used to execute the game image generation method based on artificial intelligence as described in any one of claims 1 to 6, characterized in that, include: The data acquisition module is used to collect multi-dimensional data streams during game runtime in real time, and then transform the multi-dimensional data streams into unified semantic environment state description vectors. The style extraction module is used to perform convolutional feature extraction on the reference image in the environmental state description vector to obtain a style representation tensor, and to perform fusion and dimensionality reduction on the style representation tensor to obtain the target style identifier vector. The model building module is used to acquire candidate generated images, measure the difference between the candidate image features and the target style identifier vector to obtain a style difference matrix, determine the style consistency penalty in image generation based on the style difference matrix, and use the style consistency penalty to build a game image generation model. The image generation module is used to generate game-style images during game runtime based on the game image generation model and the environment state description vector.

8. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing code, and the processor being configured to retrieve the code and execute the AI-enabled game image generation method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the AI-enabled game image generation method as described in any one of claims 1 to 6.

Images