A method for optimizing the inference speed of image AIGC model

Abstract

The present invention discloses a method for optimizing the inference speed of an image AIGC model, and relates to the field of artificial intelligence technology. The method comprises: step 1: constructing a high-dimensional complex feature space of an input image; step 2: initializing a multimodal spectrum convolution kernel and defining a multimodal convolution operation; performing a multimodal spectrum convolution operation on the high-dimensional complex feature space of the input image, and calculating an initial convolution output; step 3: applying a complex activation function and a high-dimensional Laplace operator to the initial convolution output to obtain an optimized inference feature; step 4: the AIGC model performs model inference based on the optimized inference feature to generate an image; step 5: evaluating the quality of the generated image, obtaining a quality evaluation result, and optimizing the multimodal convolution kernel parameters based on the quality evaluation result. The present invention significantly improves the inference speed and generation quality of the image AIGC model.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method for optimizing the inference speed of an image AIGC model. Background Art

[0002] In recent years, with the rapid development of deep learning and computer vision technology, artificial intelligence generated content (AIGC) technology has made significant progress in image generation, image restoration, style transfer, etc. AIGC technology can generate high-quality images from scratch or repair and optimize existing images by training deep neural network models. These technologies have been widely used in many fields, including entertainment, medical imaging, and autonomous driving. However, although the existing AIGC technology has achieved remarkable results in generating image quality, there are still many problems with its inference speed and processing efficiency, which limits its widespread promotion in real-time and large-scale applications.

[0003] At present, generative adversarial networks (GAN) and variational autoencoders (VAE) are the most commonly used technologies in the field of AIGC. GAN can generate highly realistic images through adversarial training of generators and discriminators; VAE generates images through probabilistic models, which has a certain generation diversity while maintaining image quality. However, these technologies face the following problems in practical applications: GAN and VAE models usually contain a large number of parameters and hierarchical structures, and the training and inference processes require a lot of computing resources and time. Especially when generating high-resolution images, the computational complexity increases significantly, resulting in slow inference speed and difficulty in meeting the needs of real-time applications. GAN is prone to problems such as mode collapse and vanishing gradient during training, resulting in insufficient diversity and unstable quality of generated images. VAE also has a trade-off problem between maintaining the quality and diversity of generated images. Existing AIGC models still have limitations in feature extraction. Although convolutional neural networks (CNNs) perform well in feature extraction, when processing complex and highly diverse images, the features extracted by existing models may not be sufficient and accurate, affecting the quality of generated images. Existing AIGC technologies still need to be improved in terms of model parameter optimization. Traditional optimization methods are inefficient when dealing with large-scale and high-dimensional data, and are unable to quickly adapt to dynamically changing input data and generation requirements. Summary of the invention

[0004] In view of this, the present invention provides an optimization method for the inference speed of the image AIGC model, which significantly improves the inference speed and generation quality of the image AIGC model. At the same time, the dynamic parameter optimization mechanism based on the quality evaluation results makes the model highly robust and adaptable, and can perform well in various complex application scenarios.

[0005] The technical solution adopted by the present invention is as follows:

[0006] A method for optimizing the inference speed of an image AIGC model, the method comprising:

[0007] Step 1: Construct a high-dimensional complex feature space of the input image;

[0008] Step 2: Initialize the multimodal spectral convolution kernel and define the multimodal convolution operation; perform the multimodal spectral convolution operation on the high-dimensional complex feature space of the input image and calculate the initial convolution output;

[0009] Step 3: Apply complex activation function and high-dimensional Laplacian operator to the initial convolution output to obtain optimized inference features;

[0010] Step 4: The AIGC model performs model inference based on the optimized inference features and generates an image;

[0011] Step 5: Evaluate the quality of the generated image, obtain the quality assessment result, and optimize the multimodal convolution kernel parameters based on the quality assessment result.

[0012] Furthermore, step 1 specifically includes: calculating the complex matrix representation of the input image; using multi-scale decomposition in the complex domain to map the complex matrix to different frequencies and feature dimensions, while introducing fractal dimension parameters to expand and weight different frequencies and feature dimensions, and then smoothing through a complex Gauss-Legendre transform to obtain a high-dimensional complex feature space of the input image.

[0013] Furthermore, the complex feature representation of the input image is calculated using the following formula:

[0014]

[0015] Among them, I m，n is the complex matrix representation of the input image I; I m，n is the two-dimensional matrix representation of the input image; i is the imaginary unit, defined as i 2 = -1; To perform Hilbert transform on the input image, its instantaneous amplitude and phase information is obtained; m and n represent the pixel coordinates of the input image in the horizontal and vertical directions respectively; through the following formula, the multi-scale decomposition of the complex domain is used to map the complex matrix to different frequencies and feature dimensions, and the fractal dimension parameter is introduced to expand and weight different frequencies and feature dimensions to obtain the complex fractal representation F m，n，k :

[0016]

[0017] Among them, F m，n，k represents the complex fractal representation on the kth feature dimension; P and Q represent the frequency range in the horizontal and vertical directions respectively; f m and f n Represent the frequency components in the horizontal and vertical directions respectively; β is the fractal dimension parameter, which is used to adjust the weight in multi-scale decomposition; α k is the fractal dimension parameter, which controls the expansion degree of the feature dimension k; p and q are both subscript indices.

[0018] Furthermore, using the following formula, the complex fractal representation F m，n，k The high-dimensional complex feature space of the input image is obtained by smoothing the complex Gauss-Legendre transform:

[0019]

[0020] Where σ is the standard deviation of the Gaussian function; M and N are the number of pixels in the horizontal and vertical directions of the input image respectively; P m，n (z) is the Legendre polynomial; z is a high-dimensional complex variable.

[0021] Furthermore, the multimodal spectral convolution kernel initialized in step 2 is expressed using the following formula:

[0022]

[0023] Among them, W u，v，l represents the multimodal spectral convolution kernel at spatial position (u, v) and channel l; ω0 is the initial frequency parameter.

[0024] Furthermore, the following formula is used to perform a multimodal graph convolution operation on the high-dimensional complex feature space of the input image to calculate the initial convolution output:

[0025]

[0026] Among them, X(1) represents the initial convolution output; It means to sum all feature extraction paths, S represents the total number of feature extraction paths; s is the subscript index; represents the feature extraction path Γ s Perform integral operation; * indicates convolution operation; G m Ξ a graph structure representing a multimodal graph; s (θ) is the multimodal feature extraction function, defined as:

[0027]

[0028] Among them, θ s is the path angle center; θ is the angle variable; σ 2 represents the scale of the path angle; R is the number of time domain transformation functions; r is the subscript index; φ r (τ) is the time transform domain function, τ is the time variable; φ r (τ) = exp(-α r τ 2 )cos(β r τ); where α r is the time decay parameter; β r is the time frequency parameter.

[0029] Furthermore, in step 3, a complex activation function is applied to the initial convolution output by the following formula:

[0030]

[0031] Among them, φ(X (1) ) represents a complex activation function.

[0032] Furthermore, by using the following formula, after applying the complex activation function to the initial convolution output, it is multiplied by the high-dimensional Laplace operator to obtain the optimized inference feature.

[0033] Furthermore, in step 5, the quality assessment result is obtained by evaluating the difference between the generated image and the expected generated image, specifically including: calculating the pixel difference between the generated image and the expected generated image at the corresponding position, obtaining the sum of the pixel differences at all positions, and dividing it by the total number of pixels as the quality assessment result; dividing the initial frequency parameter in the multimodal convolution kernel by the quality assessment result to complete the optimization.

[0034] By adopting the above technical scheme, the present invention produces the following beneficial effects: the present invention effectively improves the representation ability of image features by constructing a high-dimensional complex feature space. The complex feature representation not only contains amplitude information, but also introduces phase information, making the feature representation richer and more comprehensive. In this way, the efficiency of feature extraction can be significantly improved without increasing the computational complexity, thereby accelerating the inference speed of the model. The design of the multimodal spectral convolution kernel combines Gaussian function and plane wave function, so that the convolution operation can capture local features while maintaining global phase information and frequency information. The Gaussian function part ensures the smooth transition and noise suppression of the convolution kernel at different spatial positions through the reasonable distribution of weights, while the complex plane wave function part enhances the responsiveness of the convolution kernel to complex image features through directionality and frequency selectivity. The use of this multimodal spectral convolution kernel greatly improves the efficiency of the convolution operation, thereby accelerating the speed of feature extraction and processing. The complex activation function enhances the complexity and richness of the feature representation by performing a nonlinear transformation on the initial convolution output. The introduction of the high-dimensional Laplace operator, by calculating the second-order derivative of the feature representation, highlights the local changes in the feature and effectively enhances the edge and detail information in the feature. The combination of the two makes the feature representation not only more nonlinear and complex, but also better reflects the local and global information in the image, thereby optimizing the inference process and improving the overall computational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flowchart of a method for optimizing the image AIGC model inference speed in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] All features disclosed in this specification, or steps in all methods or processes disclosed, except mutually exclusive features and / or steps, can be combined in any manner.

[0037] Any feature disclosed in this specification (including any additional claims and abstract), unless otherwise stated, may be replaced by other equivalent or alternative features having similar purposes. That is, unless otherwise stated, each feature is only an example of a series of equivalent or similar features.

[0038] Example 1: Reference Figure 1 , a method for optimizing the inference speed of an image AI GC model, the method comprising:

[0039] Step 1: Construct a high-dimensional complex feature space of the input image;

[0040] The construction of a high-dimensional complex feature space is a key step because it can add additional dimensions to the original image to represent complex image features. The introduction of complex features is not just a simple extension of the image pixel values, but rather captures the amplitude and phase information of the image by representing the image pixels as complex numbers. The amplitude information corresponds to the intensity changes of the image, while the phase information reflects the edge, texture, and structural changes of the image. This complex feature representation can better capture the global and local information in the image, thereby improving the richness and accuracy of feature expression.

[0041] Specifically, in traditional image processing methods, images are usually represented as two-dimensional or three-dimensional real number matrices containing the grayscale values ​​or color values ​​of pixels. However, this representation may be insufficient in processing complex image features, especially in capturing subtle structural and texture changes. By constructing a high-dimensional complex feature space, each pixel of the image can be represented as a complex number, where the real and imaginary parts represent different feature dimensions. This representation not only increases the dimension of the feature, but also introduces more feature interactions and combinations through complex operations, thereby enhancing the expressiveness of image features. When constructing a high-dimensional complex feature space, we first need to convert the data of the original image into complex form. Usually, this step can be achieved through Fourier transform or other transformation methods to convert the image from the spatial domain to the frequency domain, thereby obtaining a representation containing complex features. Through this transformation, the spatial information of the image is converted into frequency information, and the changes in amplitude and phase reflect the characteristics of the image at different scales and directions. Then, by selecting the appropriate high-dimensional complex feature dimension, the details and structures in the image can be better captured, thereby providing richer input for subsequent feature extraction and convolution operations.

[0042] Another advantage of high-dimensional complex feature space is that it can effectively handle noise and interference. In practical applications, image data is often affected by various noises, such as sensor noise, environmental interference, etc. These noises affect the quality of the image and the accuracy of feature extraction. However, through complex feature representation, the noise in the image can be better separated and processed, because the complex representation can provide additional dimensions to distinguish between signals and noise. In addition, the complex feature space can also identify and filter out interference signals through phase information, thereby improving the stability and robustness of feature extraction. By constructing a high-dimensional complex feature space, the image AI GC model can use richer and more accurate features during the inference process, thereby improving the inference speed of the model and the quality of the generated image. This method is not just a simple extension of image data, but enhances the diversity and expression of feature representation by introducing complex features. In the subsequent convolution operation and feature extraction process, the introduction of complex features can provide higher feature extraction accuracy and efficiency, thereby effectively improving the overall performance of the model. Finally, by optimizing the construction method of the high-dimensional complex feature space, the inference speed and quality of the image AI GC model can be significantly improved, providing a more efficient and stable solution for practical applications.

[0043] Step 2: Initialize the multimodal spectral convolution kernel and define the multimodal convolution operation; perform the multimodal spectral convolution operation on the high-dimensional complex feature space of the input image and calculate the initial convolution output;

[0044] The multimodal spectral convolution kernel is designed to fuse information from different modalities to capture the features in the image more comprehensively. In traditional convolution operations, convolution kernels usually only process the features of a single modality, while multimodal spectral convolution kernels can simultaneously process feature information of multiple modalities, such as spatial features, frequency features, and phase features. This convolution kernel that fuses features of different modalities can not only capture the diversity of information in the image, but also improve the depth and breadth of feature extraction, thereby providing richer feature inputs for subsequent model inference. The definition of multimodal convolution operations is also crucial. In multimodal convolution operations, features of different modalities are fused through specific convolution operations to produce initial convolution outputs. Traditional convolution operations are usually performed in the spatial domain or frequency domain, while multimodal convolution operations require processing feature information of multiple modalities simultaneously in a high-dimensional complex feature space. This operation not only needs to consider the complementarity of different modal features, but also needs to ensure the computational efficiency and accuracy of the convolution process. By rationally designing multimodal convolution operations, efficient feature fusion can be achieved at a lower computational complexity, thereby significantly improving the inference speed of the model.

[0045] When performing multimodal graph convolution operations, the high-dimensional complex feature space of the input image is used as the input of the convolution operation, and the convolution operation is performed through the multimodal spectral convolution kernel to obtain the initial convolution output. The key to this process is how to efficiently fuse the feature information of different modalities so that the initial convolution output can fully represent the complex features of the input image. In the multimodal convolution operation, the weights and combination methods of different modal features need to be optimized according to the specific application scenario to ensure the effectiveness and robustness of feature fusion. The introduction of multimodal spectral convolution kernels and multimodal convolution operations provides richer and more diverse feature information for the feature extraction process, thereby improving the expressiveness and accuracy of feature representation. This method can not only capture the subtle structure and global information in the image, but also improve the stability and robustness of feature extraction through the complementarity of multimodal features. In the model inference process, the initial convolution output serves as the basis for subsequent feature extraction and inference, which directly affects the overall performance and inference speed of the model. Therefore, by optimizing the multimodal spectral convolution kernel and the multimodal convolution operation, the feature extraction process can be significantly optimized without significantly increasing the computational burden, thereby improving the inference speed of the image AI GC model. In addition, the design of the multimodal spectral convolution kernel also needs to consider the optimization of computing resources. In practical applications, the limitation of computing resources is often one of the key factors affecting the inference speed of the model. By rationally designing the multimodal spectral convolution kernel, the computational complexity of the convolution operation can be reduced while ensuring the feature extraction effect, thereby achieving efficient utilization of computing resources. The optimization of the multimodal convolution operation also needs to find a balance between computational efficiency and feature extraction quality. By adjusting the parameters and operation mode of the convolution kernel, the computational efficiency of the convolution operation can be significantly improved while ensuring the feature extraction effect.

[0046] Step 3: Apply complex activation function and high-dimensional Laplacian operator to the initial convolution output to obtain optimized inference features;

[0047] Traditional activation functions, such as ReLU and Sigmoid, are mainly used for nonlinear transformations in the real domain, while complex activation functions extend this concept to enable it to process data in the complex domain. By applying a complex activation function to the initial convolution output, nonlinear features can be effectively introduced, thereby improving the richness and discriminability of feature representation. Complex activation functions can capture the amplitude and phase changes in complex features, which is particularly important for image feature extraction. Since complex features in images often contain amplitude and phase information, the introduction of complex activation functions can capture this information more comprehensively, thereby enhancing the expressive power of feature representation. The Laplacian operator is widely used in edge detection and image enhancement in image processing and computer vision. Its role is to highlight the edge and texture information of the image by calculating the second-order derivative of the pixel value in the image. The high-dimensional Laplacian operator extends this concept to the high-dimensional complex feature space. By calculating the second-order derivative of the feature in the high-dimensional space, the local structure and global information of the feature can be extracted more accurately. Applying a high-dimensional Laplacian operator can further enhance the edge and texture information of the features, making the optimized inferred features clearer and more stable, thereby providing higher quality input for subsequent model inference.

[0048] The quality and stability of the optimized inferred features have been significantly improved by the combined application of complex activation functions and high-dimensional Laplace operators. The complex activation function makes the feature representation more discriminative and rich by introducing nonlinear transformations, while the high-dimensional Laplace operator makes the feature representation clearer and more stable by enhancing the edge and texture information of the features. The combination of the two can not only improve the quality of feature extraction, but also improve the efficiency of feature extraction, thereby significantly improving the inference speed of the model while ensuring the inference quality. The advantage of this method is that it can improve the quality and efficiency of feature extraction without significantly increasing the computational complexity. The computational complexity of complex activation functions and high-dimensional Laplace operators is relatively low, and efficient calculations can be achieved through technical means such as parallel computing and hardware acceleration. Therefore, by introducing these two technologies in the feature extraction process, the computational efficiency can be significantly improved while ensuring the feature extraction effect, thereby accelerating the inference process of the model. In addition, the quality of the optimized inferred features also has a direct impact on the quality of the generated images. High-quality feature input allows the AI ​​GC model to more accurately capture the details and structure of the image during the inference process, thereby generating higher quality images. The combined application of complex activation functions and high-dimensional Laplacian operators can significantly improve the quality of generated images by enhancing the richness and stability of feature representation, thereby meeting the high requirements in practical applications. Step 3 significantly improves the quality and efficiency of feature extraction by applying complex activation functions and high-dimensional Laplacian operators to the initial convolution output, thereby providing higher quality and more stable feature inputs for subsequent model inference. The complex activation function makes the feature representation more discriminative and rich by introducing nonlinear transformations, while the high-dimensional Laplacian operator makes the feature representation clearer and more stable by enhancing the edge and texture information of the features. The combined application of the two can not only improve the quality of feature extraction, but also improve the efficiency of feature extraction, thereby significantly improving the inference speed of the model while ensuring the quality of the generated image. This method is of great significance in practical applications and provides new ideas and solutions for efficient image generation and processing.

[0049] Step 4: The AI ​​GC model performs model inference based on the optimized inference features and generates an image;

[0050] After being processed by complex activation functions and high-dimensional Laplacian operators, these features have been refined and enhanced, with rich nonlinear information and clear edge and texture details. Therefore, when these optimized features are used as input for model inference, the efficiency and accuracy of inference can be significantly improved. By parsing and processing these high-quality features, the AI ​​GC model can perform data operations and reasoning more quickly, thereby reducing computing time and resource consumption. The optimization of the model inference process also depends on the internal architecture design of the AI ​​GC model. Modern AI GC models usually use deep learning techniques such as convolutional neural networks (CNNs) and generative adversarial networks (GANs) to achieve efficient feature extraction and image generation. The design of these network architectures is optimized to maintain efficient operation when processing complex data. Through the rapid parsing of optimized features, the model can complete multi-layer convolution and nonlinear activation operations in a shorter time, thereby quickly generating high-quality images.

[0051] In addition, the improvement in model inference speed also benefits from the application of parallel computing and hardware acceleration technology. The use of hardware accelerators such as graphics processing units (GPUs) and tensor processing units (TPUs) can greatly improve the computing speed of deep learning models. The optimized feature input can be calculated simultaneously in multiple computing units through parallel processing technology, thereby significantly shortening the model inference time. This technology is particularly important when processing large-scale image data, and can achieve efficient real-time processing while ensuring the quality of generation. In practical applications, the quality of generated images is one of the key indicators for evaluating the performance of AI GC models. High-quality images not only need to have clear details and accurate color representation, but also good texture and structural consistency. By optimizing the input of inferred features, the AI ​​GC model can more accurately capture the details and global structure of the image during the inference process, thereby generating high-quality images. The optimized features contain rich nonlinear information and enhanced edge and texture features, which are fully utilized in the model inference process to ensure the delicacy and realism of the generated images. It is worth noting that the optimization of inference speed does not come at the expense of image quality. On the contrary, by reasonably designing and optimizing the feature processing process, the AI ​​GC model can maintain or even improve the quality of the generated image while improving computational efficiency. Optimizing inference features enhances information density and extraction accuracy, enabling the model to generate higher quality images in a shorter time, thereby achieving dual optimization of speed and quality.

[0052] Step 5: Evaluate the quality of the generated image, obtain the quality assessment result, and optimize the multimodal convolution kernel parameters based on the quality assessment result.

[0053] The quality of generated images is usually evaluated through a variety of indicators, including but not limited to image clarity, color accuracy, texture consistency, and structural integrity. Common quality assessment methods include human visual evaluation, structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), etc. These evaluation methods can quantify the quality of generated images and provide objective evaluation criteria. In practical applications, automated quality assessment methods are particularly important because they can quickly and accurately evaluate a large number of generated images without human intervention. After obtaining the quality assessment results, optimizing the parameters of the multimodal convolution kernel based on these results is the key to improving the model inference speed. The parameter optimization process usually uses iterative optimization algorithms such as gradient descent, stochastic gradient descent (SGD), and Adam optimizer. These algorithms gradually improve the quality of generated images by continuously adjusting the parameters of the convolution kernel. During the optimization process, the parameters of the convolution kernel are slightly adjusted according to the feedback of the quality assessment results, so that the model can extract image features faster and more accurately when generating images. The feedback mechanism plays a vital role in this process. By feeding the quality assessment results back to the model, the parameters of the convolution kernel can be adjusted in real time, thereby achieving adaptive optimization. This feedback mechanism can ensure that the model always maintains efficient feature extraction and inference capabilities under different input images and feature distributions. The introduction of the feedback mechanism can not only improve the adaptability of the model, but also improve the overall performance and inference speed of the model through continuous optimization of parameter settings.

[0054] In addition, the efficiency and effect of parameter optimization are also affected by the selection of optimization algorithms. Different optimization algorithms have different convergence speeds and stability. Choosing a suitable optimization algorithm can achieve better parameter settings in a shorter time. The gradient descent algorithm is widely used because of its simple calculation and fast convergence speed, while the Adam optimizer improves the stability and efficiency of the optimization process by adaptively adjusting the learning rate. In practical applications, by combining multiple optimization algorithms, the parameter optimization strategy can be flexibly adjusted in different stages and scenarios to achieve the optimal inference speed and generation quality. By continuously optimizing the parameters of the multimodal convolution kernel, the model can continuously improve the inference speed and the quality of the generated image. This optimization method can not only significantly improve the model performance in the short term, but also achieve gradual improvement of the model through long-term continuous optimization. For the image AI GC model, the efficiency and accuracy of the parameter optimization process directly determine the inference speed and generation effect of the model. Through a reasonable parameter optimization strategy, the model can always maintain efficient feature extraction and fast inference capabilities in the process of generating images.

[0055] Embodiment 2: Step 1 specifically includes: calculating the complex matrix representation of the input image; using multi-scale decomposition in the complex domain to map the complex matrix to different frequencies and feature dimensions, while introducing fractal dimension parameters to expand and weight different frequencies and feature dimensions, and then smoothing through a complex Gauss-Legendre transform to obtain a high-dimensional complex feature space of the input image.

[0056] Specifically, the input image is usually a pixel matrix in the real domain. By converting it into a complex matrix representation, the information in the image can be captured more comprehensively. The complex matrix representation expands each pixel value of the image into a complex form, including real and imaginary parts. This representation can capture the amplitude and phase information of the image at the same time. The amplitude information reflects the brightness change of the image, while the phase information reveals the edge and texture characteristics of the image. In this way, the complex matrix representation provides richer basic data for subsequent feature extraction and processing. Next, using multi-scale decomposition in the complex domain to map the complex matrix to different frequencies and feature dimensions is one of the key steps. Multi-scale decomposition is an important signal processing technique that can decompose an image into different scales and frequency components to capture the global and local features of the image. In the complex domain, multi-scale decomposition can more effectively process complex features in the image, especially those detailed information containing phase changes. By mapping the complex matrix to different frequencies and feature dimensions, the diverse features in the image can be more comprehensively captured, thereby providing richer input for subsequent feature fusion and processing. In this process, the introduction of fractal dimension parameters to expand and weight different frequencies and feature dimensions is an innovation. Fractal dimension parameters are a geometric dimension used to describe complex systems. By introducing fractal dimension parameters, complex features in images can be described and processed more accurately. The introduction of fractal dimension parameters can expand and weight different frequencies and feature dimensions, making feature representation more flexible and diverse. In this way, the details and structural changes in the image can be better captured, thereby improving the accuracy and richness of feature extraction. Finally, smoothing through complex Gauss-Legendre transform is one of the key steps in feature space construction. Complex Gauss-Legendre transform is a mathematical transformation used for smoothing, which can eliminate noise and interference while retaining feature information. Through complex Gauss-Legendre transform, the expanded and weighted features can be smoothed to obtain a more stable and clear feature representation. This smoothing process can effectively improve the stability and robustness of feature representation, thereby providing higher quality input for subsequent model inference.

[0057] Embodiment 3: Use the following formula to calculate the complex feature representation of the input image:

[0058]

[0059] Among them, I m,n is the complex matrix representation of the input image I; I m,n is the two-dimensional matrix representation of the input image; i is the imaginary unit, defined as i 2 = -1; To perform Hilbert transform on the input image, its instantaneous amplitude and phase information is obtained; m and n represent the pixel coordinates of the input image in the horizontal and vertical directions respectively; through the following formula, the multi-scale decomposition of the complex domain is used to map the complex matrix to different frequencies and feature dimensions, and the fractal dimension parameter is introduced to expand and weight different frequencies and feature dimensions to obtain the complex fractal representation F m,n,k :

[0060]

[0061] Among them, F m,n,k represents the complex fractal representation on the kth feature dimension; P, Q represent the frequency range in the horizontal and vertical directions respectively; f m and f n Represent the frequency components in the horizontal and vertical directions respectively; β is the fractal dimension parameter, which is used to adjust the weight in multi-scale decomposition; α k is the fractal dimension parameter, which controls the expansion degree of the feature dimension k; p and q are both subscript indices.

[0062] Specifically, the complex matrix mentioned in the formula represents I m,n As input, it is mapped to different frequencies and feature dimensions through multi-scale decomposition in the complex domain. The basic idea of ​​multi-scale decomposition is to decompose the complex matrix representation of the input image into components of different scales and frequencies in order to more comprehensively capture the details of the image. In the complex domain, this decomposition process is particularly important because it can simultaneously process the amplitude and phase information of the image, thereby providing a richer feature representation. In the multi-scale decomposition process, the introduction of frequency components is achieved through the complex exponential function. The function of this part is to transform the input features in the frequency domain to capture the feature changes at different frequencies. The phase part of the complex exponential function (i.e. -i2π(pf m +qf n )) can reflect the phase change in the image, so that the feature representation contains not only intensity information but also phase information. Phase information is crucial for capturing detailed features such as edges and textures of images, and can significantly improve the accuracy of feature extraction. In order to further improve the diversity and accuracy of feature representation, the fractal dimension parameters β and α are introduced k The fractal dimension parameters are used as part of the denominator in the formula to adjust the weights and feature dimensions in the multi-scale decomposition. Specifically, β is used to adjust the weights of different frequency components by affecting This enables the feature representation to better balance the influence of different frequency components. k It is used to control the expansion degree of feature dimension k by affecting This allows feature representation to have different expansion and weighting effects on different feature dimensions. This approach allows feature representation to more flexibly adapt to different frequency components and feature dimensions, thereby improving the richness and accuracy of feature extraction.

[0063] The combination of multi-scale decomposition and fractal dimension parameters can effectively enhance the stability and robustness of the feature extraction process. In practical applications, image data often has complex multi-scale features, and the feature representation at a single scale or frequency may not be able to fully capture the detailed information in the image. Through multi-scale decomposition, the complex feature representation of the image can be refined at different scales and frequencies, thereby more comprehensively reflecting the complex features in the image. The introduction of fractal dimension parameters further refines this decomposition process, so that the feature representation can be reasonably expanded and weighted at each scale and frequency, thereby improving the effect of feature extraction. Finally, the complex fractal representation F obtained through these steps is m,n,k , is the optimized high-dimensional complex feature space. These feature representations not only contain rich multi-scale and multi-frequency information, but also enhance the flexibility and diversity of feature representation through the adjustment of fractal dimension parameters. In this way, the image AIGC model can use these optimized features to perform data operations and image generation with higher efficiency and accuracy when inferring. The diversity and accuracy of feature extraction are significantly improved through the multi-scale decomposition of the complex domain and the introduction of fractal dimension parameters. In this way, the image AIGC model can significantly improve the inference speed and generation quality without significantly increasing the computational complexity. This optimization process not only enhances the richness and accuracy of feature representation, but also makes the feature extraction process more flexible and efficient through reasonable parameter adjustment, thereby providing strong technical support for image generation and processing.

[0064] Example 4: Using the following formula, the complex fractal representation F m,n,k The high-dimensional complex feature space of the input image is obtained by smoothing the complex Gauss-Legendre transform:

[0065]

[0066] Where σ is the standard deviation of the Gaussian function; M and N are the number of pixels in the horizontal and vertical directions of the input image respectively; P m,n (z) is the Legendre polynomial; z is a high-dimensional complex variable.

[0067] Specifically, the Gaussian function is a bell-shaped curve with good smoothness. By performing convolution operations on image data, it can effectively reduce noise and interference and highlight the main features of the image. The Gaussian function part in the formula is expressed in the form of exponential decay. The standard deviation σ is used as a parameter to weight the pixels in the horizontal and vertical directions of the image. The standard deviation σ determines the degree of smoothing: a larger σ value leads to a stronger smoothing effect, thereby eliminating more detail noise; while a smaller σ value retains more detail information. This weighted processing enables the complex feature representation of the input image to maintain the integrity of key features when smoothing, while reducing the impact of high-frequency noise. The role of Gaussian smoothing is not only to remove noise, but more importantly, it provides a smoother and more continuous basis for feature extraction. For the image AIGC model, the smoothness and continuity of the input features directly affect the computational efficiency and stability of the model. Through Gaussian smoothing, the mutations and discontinuities in the feature representation are reduced, which makes the model more efficient when performing convolution operations and feature extraction, thereby improving the overall inference speed.

[0068] Secondly, the introduction of Legendre polynomials provides a powerful mathematical tool for feature optimization. Legendre polynomials P m,n(z) is a set of orthogonal polynomials with good feature separation and representation capabilities. In the complex Gauss-Legendre transform, the Legendre polynomials further optimize and normalize the feature representation through their orthogonality and normalization. The integral part contained in the formula is used to normalize the Legendre polynomials, which ensures the consistency and stability of the feature representation in different dimensions. The characteristics of orthogonal polynomials reduce the mutual interference between different feature dimensions, thereby improving the clarity and separation of feature representation. The role of Legendre polynomials in optimizing feature representation is reflected in two aspects: first, through the orthogonal property, the Legendre polynomials can effectively separate different components in the feature space and reduce the overlap and interference between features; second, through normalization, the Legendre polynomials ensure the balance and stability of feature representation in different dimensions. This optimization process enables the feature representation to match and identify features with higher accuracy and clarity when performing model inference, thereby improving the inference speed and accuracy. The combination of complex Gauss-Legendre transform, through the smoothing of Gaussian function and the optimization of Legendre polynomials, makes the complex fractal representation significantly improved in terms of smoothness and accuracy. This process not only improves the stability and robustness of feature representation, but also enhances the separation and clarity of features. For the image AIGC model, this means that the model can perform data operations and image generation more quickly and accurately when processing these optimized features, thereby significantly improving the inference speed. In addition, this optimization process also has a positive impact on the overall performance of the model. By reducing noise and improving the clarity of feature representation, the model needs to process less data during inference, and the computational complexity is reduced, thereby improving computational efficiency. Smooth and accurate feature representation not only improves the inference speed of the model, but also improves the quality of the generated image, making the generated results more realistic and delicate. The complex fractal representation is optimized through the complex Gauss-Legendre transform, combined with the smoothing effect of the Gaussian function and the orthogonal characteristics of the Legendre polynomials. This process significantly improves the quality and stability of feature representation, providing a solid foundation for efficient inference of image AI GC models. Through this optimization method, the model can achieve fast and efficient image generation while ensuring the high quality of the generated images, thereby improving the overall performance and practicality.

[0069] Example 5: The initialized multimodal spectral convolution kernel in step 2 is expressed using the following formula:

[0070]

[0071] Among them, W u,v,l represents the multimodal spectral convolution kernel at spatial position (u, v) and channel l; ω0 is the initial frequency parameter.

[0072] Specifically, the design of the multimodal spectral convolution kernel combines the characteristics of Gaussian smoothing and complex plane wave functions. First, the Gaussian function part ensures the smooth transition of the convolution kernel at different spatial positions by smoothing the distribution of weights. In the formula of this part, the standard deviation σ determines the degree of diffusion of the Gaussian function. A larger σ value will lead to a wider distribution of weights, resulting in a larger receptive field in the convolution operation, allowing the convolution kernel to capture a wider range of features. A smaller σ value will concentrate the weights in the central area, improving the ability to capture local features. This smoothing property of the Gaussian function helps to reduce the interference caused by noise and discontinuity during the convolution process, thereby improving the stability and accuracy of feature extraction. Secondly, the introduction of the complex plane wave function makes the convolution kernel directional and frequency selective. By combining cosine and sine functions, the complex plane wave function can capture the periodic features and phase information in the image. The form of the complex plane wave function cos(ω0u)+isin(ω0v) provides the flexibility and expressiveness of the convolution kernel in processing complex image features. The cosine function and the sine function capture frequency information in different directions respectively, so that the convolution kernel can more accurately identify detail features such as edges and textures in the image. In particular, the combination of sine and cosine functions enhances the responsiveness of the convolution kernel in different directions and frequencies, thereby improving the diversity and accuracy of feature extraction. In addition, the complex plane wave function part in the formula, through its complex domain characteristics, can capture the image amplitude information while retaining the phase information. Phase information is crucial for the precise positioning and detailed description of image features. Traditional convolution kernels can usually only process amplitude information and ignore phase information, which may lead to incomplete feature extraction. The introduction of the complex plane wave function, by capturing and retaining phase information, enables the convolution kernel to more comprehensively reflect the characteristics of the image, enhancing the completeness and accuracy of feature representation.

[0073] The initialization of the convolution kernel significantly improves the efficiency of the convolution operation and the accuracy of feature extraction through the combination of Gaussian smoothing and complex plane wave function. This design can not only effectively reduce the noise and interference in the calculation process, but also enhance the response ability of the convolution kernel to different directions and frequency features, thereby improving the diversity and accuracy of feature extraction. In practical applications, the design of this multimodal spectrum convolution kernel enables the image AIGC model to extract and process features more efficiently and accurately when processing complex image data, thereby significantly improving the inference speed of the model. Through this optimized design, the model can complete feature extraction and data processing more quickly when performing convolution operations, thereby reducing computing time and resource consumption. The Gaussian smoothing part ensures the smooth transition and noise suppression of the convolution kernel at different spatial positions through the reasonable distribution of weights. The complex plane wave function part enhances the response ability of the convolution kernel to complex image features through directionality and frequency selectivity. The combination of the two enables the convolution kernel to retain more image details and feature information while ensuring efficient calculation, thereby achieving accuracy and diversity in feature extraction. By combining Gaussian functions and plane wave functions, the design of the multimodal spectral convolution kernel is significantly optimized. This optimized design not only improves the efficiency of the convolution operation, but also enhances the accuracy and diversity of feature extraction, thereby significantly improving the inference speed of the image AI GC model. This method is of great significance in practical applications and provides a new technical path and solution for efficient image generation and processing. By reasonably designing and optimizing the convolution kernel, the model can achieve fast and efficient image generation without significantly increasing the computational complexity, thereby improving overall performance and practicality.

[0074] Example 6: Use the following formula to perform a multimodal graph convolution operation on the high-dimensional complex feature space of the input image to calculate the initial convolution output:

[0075]

[0076] Among them, X (1) represents the initial convolution output; It means to sum all feature extraction paths, S represents the total number of feature extraction paths; s is the subscript index; represents the feature extraction path Γ s Perform integral operation; * indicates convolution operation; G m Ξ a graph structure representing a multimodal graph; s (θ) is the multimodal feature extraction function, defined as:

[0077]

[0078] Among them, θ s is the path angle center; θ is the angle variable; σ2 represents the scale of the path angle; R is the number of time domain transformation functions; r is the subscript index; φ r (τ) is the time transform domain function, τ is the time variable; φ r (τ) = exp(-α r τ 2 )cos(β r τ); where α r is the time decay parameter; β r is the time frequency parameter.

[0079] Specifically, the initial convolution output X described in the formula (1) It is obtained by convolution operation and integral calculation of multiple feature extraction paths. The basic idea of ​​multimodal graph convolution operation is to transform the high-dimensional complex feature space of the input image into With the multimodal spectral convolution kernel W u,v,l Convolution processing is performed, and the multimodal feature extraction function and the time domain transformation function are combined to extract the complex features in the image. represents the feature extraction path Γ s The introduction of feature extraction paths enables convolution operations to extract features on different paths, increasing the diversity and depth of feature extraction. The design of the integral path takes into account a variety of possible feature extraction methods, making feature extraction more flexible and comprehensive. By summing all feature extraction paths It is possible to integrate feature information on multiple paths to obtain a richer and more comprehensive initial convolution output. This multi-path feature extraction method effectively increases the robustness of the model and the richness of feature extraction. Secondly, the convolution operator * represents the interaction of local features in the spatial position. Through the convolution operation, the local features of the input image can be dot-producted with the convolution kernel to extract feature information with local similarity. Multimodal spectral convolution kernel W u,v,l It plays a key role here. It combines the characteristics of Gaussian smoothing and complex plane wave functions, so that the convolution operation can capture local features while maintaining global phase and frequency information. The Gaussian function part ensures the smooth transition and noise suppression of the convolution kernel at different spatial positions through the reasonable distribution of weights, while the complex plane wave function part enhances the responsiveness of the convolution kernel to complex image features through directionality and frequency selectivity.

[0080] Multimodal feature extraction function Ξ s The introduction of (θ) enables the convolution operation to combine features of different angles and scales for processing. The feature extraction function is defined as where θ s is the path angle center, θ is the angle variable, Represents the scale of the path angle. This function is similar to the Gaussian function. By weighting the angle variable, feature extraction can be performed within different angle ranges, thereby enhancing the directionality and flexibility of feature extraction. The introduction of angles enables the feature extraction function to more accurately capture feature information in different directions, especially when processing images with complex textures and edge information. This method can significantly improve the effect of feature extraction. Time domain transformation function φ r (τ) further enriches the dimension of feature extraction. The time domain transformation function is defined as φ r (τ) = exp(-α r τ 2 )cos(β r τ); where α r is the time decay parameter, β r is the time frequency parameter. Through time domain transformation, image features can be mapped to the time domain, thereby capturing the dynamic changes and time correlation of image features. The Gaussian attenuation part exp(-α r τ 2 ) controls the range of time transformation, and the larger α r The value makes the time transformation more concentrated, while the smaller α r The cosine part cos(β r τ) introduces time-frequency information, so that the time transformation can capture the feature changes at different frequencies.

[0081] In the formula, the multimodal feature extraction function and the time domain transformation function work together to map the high-dimensional complex feature space of the input image to the initial convolution output X through convolution and integration operations. (1) . This multimodal graph convolution operation can effectively combine information in spatial, angular and temporal dimensions, enhancing the comprehensiveness and accuracy of feature extraction. Through multi-dimensional information fusion, the feature extraction process can not only capture the static features of the image, but also reflect the dynamic changes and multi-scale features of the image. A significant advantage of this method is that it can significantly improve the quality and efficiency of feature extraction without significantly increasing the computational complexity. Through reasonably designed convolution kernels and feature extraction functions, the model can complete feature extraction and data processing more quickly when performing convolution operations, thereby reducing computing time and resource consumption. The combination of multipath feature extraction and time domain transformation enables the model to make full use of multimodal information when processing complex image data, thereby improving the diversity and accuracy of feature extraction. The formula in Example 6 processes the high-dimensional complex feature space of the input image through a multimodal graph convolution operation, combined with a multimodal feature extraction function and a time domain transformation function, and calculates the initial convolution output X (1). This process significantly improves the quality and efficiency of feature extraction, thereby optimizing the inference speed of the model. By combining information in spatial, angular, and temporal dimensions, the feature extraction process can not only capture the static features of the image, but also reflect the dynamic changes and multi-scale features of the image. This method is of great significance in practical applications and provides a new technical path and solution for efficient image generation and processing. By reasonably designing and optimizing the convolution kernel, the model can achieve fast and efficient image generation without significantly increasing the computational complexity, thereby improving the overall performance and practicality.

[0082] To further understand how this formula works, we can conduct a deeper analysis from several perspectives. The first is the feature representation in the complex domain. This representation captures the amplitude and phase information of the image through the real and imaginary parts of the complex number, respectively, making the feature representation richer and more comprehensive. The feature representation in the complex domain can provide higher resolution and accuracy when processing complex textures and details in the image. The second is the design of the multimodal atlas convolution kernel. The multimodal atlas convolution kernel combines Gaussian smoothing and complex plane wave functions, so that the convolution operation can capture local features while maintaining global phase and frequency information. The Gaussian function part ensures the smooth transition and noise suppression of the convolution kernel at different spatial positions through the reasonable distribution of weights, while the complex plane wave function part enhances the responsiveness of the convolution kernel to complex image features through directionality and frequency selectivity. The introduction of multimodal feature extraction functions and time domain transformation functions further enriches the dimension of feature extraction. The feature extraction function weights the angle variables so that feature extraction can be performed within different angle ranges, thereby enhancing the directionality and flexibility of feature extraction. The time domain transformation function introduces time attenuation and frequency information, so that feature extraction can capture the dynamic changes and time correlation of the image. This multi-dimensional information fusion enables the feature extraction process to not only capture the static features of the image, but also reflect the dynamic changes and multi-scale features of the image. In general, the formula in Example 6 processes the high-dimensional complex feature space of the input image through a multimodal graph convolution operation, combined with a multimodal feature extraction function and a time domain transformation function, and calculates the initial convolution output X (1) . This process significantly improves the quality and efficiency of feature extraction, thereby optimizing the inference speed of the model. By properly designing and optimizing the convolution kernel, the model can achieve fast and efficient image generation without significantly increasing the computational complexity, thereby improving overall performance and practicality. This method is of great significance in practical applications and provides a new technical path and solution for efficient image generation and processing.

[0083] Example 7: In step 3, a complex activation function is applied to the initial convolution output using the following formula:

[0084]

[0085] Among them, φ(X (1) ) represents a complex activation function.

[0086] Specifically, the definition of the complex activation function fully considers the characteristics of complex data and is implemented through two parts: amplitude normalization and phase adjustment. These two parts are responsible for processing the modulus and phase of the complex number, respectively, so that the activation function can unify the amplitude of the input data and retain and utilize its phase information. Amplitude normalization eliminates the difference in amplitude between different complex inputs by normalizing the modulus of the complex input to the unit circle. This normalization operation ensures that the amplitudes of all input complex numbers are consistent, thereby simplifying the subsequent calculation process and improving the consistency and stability of feature representation. Especially when processing complex and diverse image features, this unified scale operation can effectively reduce the computational instability caused by input amplitude differences, thereby enhancing the robustness of the model. On the other hand, phase adjustment enables the activation function to more accurately capture the detailed features in the image by maintaining and utilizing the phase information of the complex number. Phase information is of great significance in complex domain feature representation, especially when describing the edge and texture features of the image, the phase information can provide more directionality and details. By introducing the exponential function, the phase adjustment part effectively enhances the nonlinearity of the complex activation function, making the feature representation more complex and rich. This phase adjustment not only retains the key phase information of the input complex number, but also optimizes the feature representation, so that the model can more accurately capture and express complex features when processing complex domain data. The application of complex activation function in model inference significantly improves the nonlinearity and complexity of feature representation, thereby improving the inference ability of the model. By processing the initial convolution output, the complex activation function optimizes and enhances the complex feature information, so that the model can extract features and process data more quickly and accurately during the inference process. The enhancement of feature representation not only improves the accuracy and reliability of inference, but also enables the model to complete complex calculation tasks in a shorter time, thereby improving the overall inference speed. In addition, the design of the complex activation function also considers the issue of computational efficiency. Through the combination of amplitude normalization and phase adjustment, the complex activation function effectively reduces unnecessary complex operations, making the calculation process more efficient. Amplitude normalization simplifies the processing of complex moduli during the calculation process, while phase adjustment reduces the processing burden of phase information in subsequent calculations by optimizing feature representation. This optimized design enables the model to significantly improve computational efficiency while ensuring high accuracy when processing large-scale data.

[0087] Embodiment 8: Through the following formula, after applying a complex activation function to the initial convolution output, multiplying it by a high-dimensional Laplace operator, an optimized inference feature is obtained.

[0088] The high-dimensional Laplacian operator is a mathematical tool commonly used in image processing. It highlights local changes in features by calculating the second-order derivative of feature representation. This operation can effectively enhance the edge and detail information in the feature, allowing the model to capture these important features more accurately when processing complex images. The high-dimensional Laplacian operator reduces the noise and instability in the feature representation by further optimizing the feature representation, thereby improving the overall accuracy and stability of the feature representation. The combination of complex activation function and high-dimensional Laplacian operator makes the feature representation not only more nonlinear and complex, but also better reflects the local and global information in the image. This combination has significant advantages in optimizing the inference speed of the image AI GC model. First, the complex activation function enables the feature representation to more accurately capture the complex features in the image through nonlinear transformation, thereby improving the accuracy and reliability of inference. Secondly, the introduction of the high-dimensional Laplacian operator further enhances the clarity and stability of the feature representation by highlighting the edge and detail information in the feature. The result of this double optimization is that the feature representation is not only richer and more diverse, but also can complete the feature extraction and data processing tasks more quickly and accurately in the subsequent inference process. In addition, this optimization method also has significant advantages in computational efficiency. Through amplitude normalization and phase adjustment of the complex activation function, the feature representation is simplified and optimized, making the calculation process more efficient. The introduction of the high-dimensional Laplace operator further reduces the noise and instability factors in the calculation by calculating the second-order derivative of the feature representation, thereby improving the overall computational efficiency. This efficient calculation method enables the model to significantly improve the calculation speed while ensuring high accuracy when processing large-scale data.

[0089] Embodiment 9: In step 5, the quality assessment result is obtained by evaluating the difference between the generated image and the expected generated image, specifically including: calculating the pixel difference between the generated image and the expected generated image at the corresponding position, obtaining the sum of the pixel differences at all positions, and dividing it by the total number of pixels as the quality assessment result; dividing the initial frequency parameter in the multimodal convolution kernel by the quality assessment result to complete the optimization.

[0090] Specifically, quality assessment first needs to accurately measure the difference between the generated image and the expected generated image. This is usually done by calculating the pixel difference between the two images at corresponding positions. Specifically, by comparing each pixel of the generated image and the expected generated image, calculating the difference between them, and then accumulating all pixel differences to obtain the total pixel difference. This method can effectively capture the difference in local details of the image, thereby providing an objective evaluation criterion. After accumulating all pixel differences, divide them by the total number of pixels of the image to obtain an average difference as the quality assessment result. The smaller the average difference, the closer the generated image is to the expected image, and the higher the quality. Once the quality assessment result is obtained, this result can be used to adjust the parameters of the multimodal convolution kernel. In Example 9, the specific optimization strategy is to divide the initial frequency parameter in the multimodal convolution kernel by the quality assessment result. The logic behind this adjustment mechanism is that the quality assessment results reflect the accuracy and quality of the images generated by the current model. If the assessment result is poor (large difference), it means that the gap between the generated image and the expected image is large, and a greater degree of parameter adjustment is required; if the assessment result is good (small difference), it means that the generated image is close to the expected image and only fine-tuning of the parameters is required. Through this adaptive parameter adjustment mechanism, the model can dynamically optimize the generation process, thereby gradually improving image quality.

[0091] This quality assessment-based optimization strategy has multiple advantages. First, it implements a feedback mechanism that monitors the quality of generated images in real time and makes corresponding adjustments based on the quality assessment results. This real-time feedback and adjustment mechanism enables the model to respond quickly to changes in generation quality and perform adaptive optimization, thereby improving the overall inference speed and generation quality. Second, this approach allows for high flexibility and accuracy in parameter adjustment. Through parameter adjustment driven by quality assessment results, it is possible to accurately optimize problems that arise during the generation process, rather than relying on a pre-set fixed parameter adjustment strategy. This flexibility enables the model to maintain efficient operation in a variety of complex application scenarios. In addition, this optimization strategy also contributes significantly to improving the robustness and adaptability of the model. In practical applications, the quality of generated images may be affected by a variety of factors, such as the diversity of input data, noise interference, etc. Through dynamic parameter adjustment based on quality assessment results, the model can better adapt to these changes, maintain stable generation quality and efficient inference speed. Especially when dealing with complex and high-dimensional data, this optimization method can significantly improve the overall performance of the model, enabling it to perform well in various complex environments.

[0092] The present invention is not limited to the above-mentioned specific embodiments, but extends to any new features or any new combination disclosed in this specification, as well as any new method or process steps or any new combination disclosed.

Claims

1. A method for optimizing the inference speed of an image AIGC model, characterized in that: The method comprises: Step 1: Construct a high-dimensional complex feature space of the input image; Step 2: Initialize the multimodal spectral convolution kernel and define the multimodal convolution operation; perform the multimodal spectral convolution operation on the high-dimensional complex feature space of the input image and calculate the initial convolution output; Step 3: Apply complex activation function and high-dimensional Laplacian operator to the initial convolution output to obtain optimized inference features; Step 4: The AIGC model performs model inference based on the optimized inference features and generates an image; Step 5: Evaluate the quality of the generated image, obtain the quality evaluation result, and optimize the multimodal convolution kernel parameters based on the quality evaluation result; Calculate the complex matrix representation of the input image; use multi-scale decomposition in the complex domain to map the complex matrix to different frequencies and feature dimensions, introduce fractal dimension parameters to expand and weight different frequencies and feature dimensions, and then smooth it through complex Gauss-Legendre transform to obtain the high-dimensional complex feature space of the input image; Use the following formula to calculate the complex feature representation of the input image: ； in, For the input image The complex matrix representation of ; is the two-dimensional matrix representation of the input image; is an imaginary unit, defined as ; To perform Hilbert transform on the input image to obtain its instantaneous amplitude and phase information; and Respectively represent the pixel coordinates of the input image in the horizontal and vertical directions; through the following formula, the complex matrix is ​​mapped to different frequencies and feature dimensions using multi-scale decomposition in the complex domain, and the fractal dimension parameter is introduced to expand and weight different frequencies and feature dimensions to obtain the complex fractal representation. : ； in, Indicates Complex fractal representation on feature dimensions; Respectively represent the frequency range in the horizontal direction and the vertical direction; and Represent the frequency components in the horizontal and vertical directions respectively; is the fractal dimension parameter, which is used to adjust the weight in multi-scale decomposition; is the fractal dimension parameter, controlling the feature dimension the extent of expansion; and All are subscript indexes; Complex fractal representation The high-dimensional complex feature space of the input image is obtained by smoothing the complex Gauss-Legendre transform: ； in, is the standard deviation of the Gaussian function; and are the number of pixels of the input image in the horizontal and vertical directions respectively; is the Legendre polynomial; is a high-dimensional complex variable; The initialized multimodal spectral convolution kernel in step 2 is expressed using the following formula: ； in, Indicates the position in space and Channel Multimodal spectral convolution kernel on ; is the initial frequency parameter; Perform a multimodal graph convolution operation on the high-dimensional complex feature space of the input image and calculate the initial convolution output: ； in, represents the initial convolution output; represents the sum of all feature extraction paths. Represents the total number of feature extraction paths; is the subscript index; Represents along the feature extraction path Perform integral operations; Represents the convolution operation; Graph structures representing multimodal graphs; is a multimodal feature extraction function, defined as: ； in, is the path angle center; is the angle variable; A scale that represents the angle of the path; is the number of time domain transformation functions; is the subscript index; is the time transform domain function, is the time variable; ;in, is the time decay parameter; is the time frequency parameter; In step 3, a complex activation function is applied to the initial convolution output using the following formula: ； in, represents a complex activation function; The following formula is used to apply the complex activation function to the initial convolution output and then multiply it by the high-dimensional Laplacian operator to obtain the optimized inference features; In step 5, the quality assessment result is obtained by evaluating the difference between the generated image and the expected generated image, specifically including: calculating the pixel difference between the generated image and the expected generated image at the corresponding position, obtaining the sum of the pixel differences at all positions, and dividing it by the total number of pixels as the quality assessment result; the initial frequency parameter in the multimodal convolution kernel is divided by the quality assessment result to complete the optimization.

Images