Artificial intelligence-based digital cultural creative content generation method

Abstract

The application belongs to the technical field of artificial intelligence, and particularly relates to a digital cultural creative content generation method based on artificial intelligence. The method comprises the following steps: step 1, calculating global text embedding; step 2, obtaining latent variables after manifold transformation; step 3, taking the transformed latent variables as initial states, constructing a potential energy function in a latent space, which comprises text semantic constraints and a global regularization term, and realizing diffusion update of the latent variables through a discrete random differential equation; adopting an inverse Fourier transform method to generate a local texture map; and step 4, taking the generated local texture map as a local driving factor, constructing a partial differential equation based on diffusion and reaction mechanism, and realizing spatiotemporal evolution of image brightness distribution under a preset initial condition, so as to realize adaptive synthesis of image structure layout and generate image content. The application realizes intelligent generation from text description to high-quality image content.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a method for generating digital cultural and creative content based on artificial intelligence. Background Technology

[0002] With the continuous development of artificial intelligence, computer vision, and generative modeling technologies, methods for automatically generating digital cultural and creative content based on text descriptions have become a crucial direction for the cultural and creative industries. Currently, many digital art, film and television special effects, game design, and cultural heritage restoration tasks rely on artificial intelligence technology to automatically generate high-quality visual content. In particular, text-driven content generation technology enables creators to generate images that conform to specific artistic styles or cultural themes through natural language descriptions. This approach not only improves the efficiency of creative content production but also reduces reliance on professional graphic designers, providing new possibilities for the automated and intelligent development of the cultural and creative industries.

[0003] Currently, most text-driven image generation methods are based on GANs, VAEs, or diffusion models. The core idea of ​​these methods is to learn from large-scale image-text pairing data, enabling the model to find the mapping relationship between text descriptions and image content in a high-dimensional latent space. For example, StyleGAN performs excellently in high-resolution image generation, while CLIP-based generative models such as DALL·E can better understand text semantics and convert them into corresponding visual content. However, these methods generally suffer from the following problems: Although methods like CLIP enhance the correlation between text and images through cross-modal contrastive learning, semantic deviations may still occur during the generation process. For example, the user-input text description may contain multiple elements (such as "an ancient castle against a sunset background"), but the generative model may not accurately capture all the details, tending to generate an approximate scene rather than one that perfectly matches the text description. Traditional GANs or diffusion models often rely on end-to-end training processes, making it difficult to precisely adjust the texture, color, and lighting features of local areas when generating images. Especially in cultural and creative applications, such as digital painting and art style transfer, creators often want to maintain overall stylistic consistency while controlling the details of specific areas. However, most existing methods lack control over local features, resulting in generated images that fail to meet the demands of high-end artistic creation in terms of detail. Summary of the Invention

[0004] The main objective of this invention is to provide an artificial intelligence-based method for generating digital cultural and creative content. This method combines semantic parsing, latent variable optimization, stochastic diffusion modeling, frequency modulation, and partial differential equation control to achieve intelligent generation from text descriptions to high-quality image content. The method improves semantic consistency by constructing mathematical constraints, ensures the stability of latent variables through geometric optimization, and enhances the controllability of local textures by combining frequency domain modulation and Fourier transform. Furthermore, the introduction of a diffusion-reaction mechanism allows the generated content to exhibit richer variations across different instances while maintaining the stability of the artistic style.

[0005] To address the aforementioned technical problems, this invention provides a method for generating digital cultural and creative content based on artificial intelligence, the method comprising:

[0006] Step 1: Divide the input text description into several semantic segments according to predefined syntax and semantic rules. For each semantic segment, calculate the local semantic score using predefined word vector mapping and activation function. Calculate the global text embedding using the local semantic scores of each semantic segment and preset weights.

[0007] Step 2: Add the global text embedding to the noise generated using the standard normal distribution to construct the initial latent variable; regard the initial latent variable as being located on a hidden manifold, and construct a local loss function with the global text embedding as a reference. Calculate the deviation between the initial latent variable and the text embedding using a preset Riemann metric, and use the Riemann gradient descent method to perform multi-step iterative updates on the initial latent variable to obtain the latent variable after manifold transformation.

[0008] Step 3: Using the transformed latent variables as the initial state, construct a potential energy function in the latent space, including text semantic constraints and global regularization terms. The latent variables are updated through the discretized stochastic differential equations. The updated latent variables are used as inputs. The texture amplitude, center frequency, and frequency diffusion parameters are determined according to the preset frequency modulation function. The inverse Fourier transform method is used to generate a local texture map.

[0009] Step 4: Using the generated local texture map as a local driving factor, construct a partial differential equation based on diffusion and reaction mechanisms. This partial differential equation performs spatiotemporal evolution of the image brightness distribution under preset initial conditions to achieve adaptive synthesis of the image structure layout and generate image content.

[0010] Furthermore, in step 1, let the input text description be T = {w1, w2, ..., w...} j , ..., w M}; where w j This represents the j-th word, where j is an integer index ranging from 1 to M, and M is the total number of words; a predefined word vector mapping. The word wj Mapped to a real number vector; This represents the vocabulary, which is the set of all possible words. The dimension of the word vector indicates the word w j In word vector space there are Each word has a feature value; each word is mapped to a vector space to obtain the word vector φ(w). j Using predefined syntactic and semantic rules, the text is decomposed into N semantic segments, where the i-th semantic segment is... Let n be the j-th word in the i-th semantic segment; word index n i，j Indicates the position of the word within the entire text; L i is the number of words in the i-th semantic segment; N is the total number of semantic segments after the text is divided.

[0011] Furthermore, in step 1, a semantic scoring function is defined for each semantic segment:

[0012]

[0013] Where ||·||1 represents the L1 norm operation; ψ(T i The local semantic score is used; and the segment semantic vector is constructed using the Sigmoid activation function:

[0014] s i =σ(ψ(T) i )-θ)v0;

[0015] Among them, s i Let be the semantic vector of the i-th semantic segment; σ(·) is the Sigmoid activation function; θ is the activation threshold, a preset scalar used to shift ψ(T) i The output of ) controls the response range of the Sigmoid activation function; v0 is the basic semantic direction vector, which is a vector in the range of ) The preset vector in; d s The dimension of the semantic vector determines the size of the feature space for global text embedding; global text embedding is obtained by weighted averaging of the semantic vectors of each segment.

[0016]

[0017] in, represents global text embedding; ||·||2 represents the L2 norm.

[0018] Furthermore, in step 2, the initial latent variables are constructed through the following process: for k = 1, ..., d s ,definition:

[0019]

[0020] U k V k ~Uniform(0,1);

[0021] Among them, U k and V k For the k-th noise component δ k The noise is generated by independently sampling random variables from a uniform distribution (Uniform(0,1)); all noise components are used to construct a noise vector. d s A 3D identity matrix is ​​used to ensure that there is no correlation between the dimensions of the noise; the initial latent variable ξ0 is constructed using the noise vector:

[0022] ξ0=e+λδ;

[0023] λ adjusts the noise amplitude, with a value ranging from 0.1 to 0.5.

[0024] Furthermore, in step 2, assume that the initial latent variable ξ0 is distributed on the hidden manifold. The local geometry is described by a metric constructed based on constraints from global text embedding, defining a local loss function. for:

[0025] Furthermore, in step 2, the process of calculating the deviation between the latent variables and the text embedding using a preset Riemannian metric, and then iteratively updating the latent variables using the Riemann gradient descent method to obtain the latent variables after manifold transformation includes:

[0026] The default Riemann metric is g:

[0027]

[0028] Where ξm is the m-th component of ξ0; ξl is the l-th component of ξ0; e m e is the m-th component; l g is the l-th component of e; ml Let g be the element in the m-th row and l-th column of the Riemann metric g; The latent variables are updated using Riemann gradient descent using the following formula:

[0029]

[0030] Among them, g -1 The matrix representing the inverse of the Riemannian metric is used to obtain the latent variables after the manifold transformation. T1 represents the total number of iterations; Gradient operator; Let ξ be the latent variable obtained in the (t-1)th iteration. t-1 The local loss function; ξ t Let be the latent variables obtained in the t-th iteration.

[0031] Furthermore, in step 3, after obtaining... Building upon this foundation, a diffusion process is introduced to further balance textual semantic constraints and generative exploration, constructing the following potential function with a global regularization term.

[0032] Where δ1>0 is a preset value, providing a global contraction effect; a stochastic differential equation for the diffusion process is introduced:

[0033]

[0034] Where D > 0 is the preset diffusion coefficient, and W(t) is the standard Wiener process; let , define the discretization step size Δt, and the iterative formula is:

[0035]

[0036] in, After T1 iterations, the final latent variables after diffusion update are obtained.

[0037] Furthermore, in step 3, the frequency modulation function The formula is:

[0038]

[0039] Where ω is the frequency variable, which belongs to the set of real numbers. a0 is the constant base value of amplitude modulation; a1 is the constant value of amplitude modulation. The relevant proportionality coefficient is a real constant; the center frequency ω base The reference center frequency is a preset constant; <·> indicates inner product operation; b is the proportionality coefficient; and the frequency standard deviation is... σ0 is the baseline standard deviation, a positive preset constant; the local texture map is obtained through inverse Fourier transform.

[0040] Where x is the coordinate of the texture map position; This indicates the operation of taking the real part; s is the imaginary unit; Ω x For texture graph space; x∈Ωx Phase function Where c0 and c1 are constants.

[0041] Furthermore, in step 4, the obtained local texture map is used. To generate the structural layout of a two-dimensional image; let the image be I(x, y, v), its evolution is driven by partial differential equations of diffusion and reaction mechanisms:

[0042]

[0043] Among them, the local diffusion coefficient Where κ0>0 is a preset value; τ>0 is a preset value to ensure that diffusion weakens at abrupt texture changes; v is the number of iterations; and the decay rate is... Where μ0>0 is a preset value; texture driving weight λ T >0 is the preset value; the initial condition is selected as I(x, y, 0) = I0(x, y), which is a zero-mean noise image. After T3 iterations, the image content I is obtained. init (x, y) = I(x, y, T3).

[0044] The AI-based digital cultural and creative content generation method of this invention has the following beneficial effects: Firstly, this invention effectively improves the accuracy of mapping text semantics to visual content. Existing text-to-image generation methods generally rely on large-scale training data for end-to-end learning. However, in complex semantic description scenarios, it is often difficult to ensure that the generated content accurately reflects every detail of the text description. This invention decomposes the input text into multiple semantic fragments by pre-setting grammatical and semantic rules, and constructs a semantic scoring function based on mathematical modeling to ensure that the text semantics maintain high integrity during the conversion into latent variables. This method not only improves the accuracy of text parsing but also enhances the model's adaptability to complex text structures, enabling the generated cultural and creative content to more accurately correspond to the description of the input text. Secondly, this invention achieves a balance between diversity and stability in the generated content by optimizing latent variables. In existing methods, the construction of latent variables usually relies on random noise or a standard Gaussian distribution, which to some extent limits the controllability of the generated content. This invention, by combining semantic embedding and geometric optimization methods, introduces specific mathematical constraints during the construction of latent variables, allowing them to be distributed across an implicit geometric structure. This optimization strategy not only ensures the stability of the generated results but also enhances the creativity of the generated content, allowing the same text input to exhibit subtle but reasonable variations across different instances, thereby enriching the cultural and creative content. Another significant benefit of this invention lies in its improved control over local details. In applications such as artistic creation and cultural heritage restoration, precise control over local areas of an image is often required to ensure that the texture, color, and structure of specific regions meet expectations. However, existing neural network-based methods have significant limitations in local control, primarily because the end-to-end black-box training model struggles to provide resolvable control variables. This invention introduces a frequency domain modulation method during the generation process, enabling local texture features to dynamically adjust based on the state of latent variables, and combines this with Fourier transform to achieve more precise local structural control. Compared to traditional methods, this approach can adjust the details of local areas without affecting the overall stylistic consistency, making the generated content more aligned with specific artistic needs. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0046] Figure 1A schematic diagram of the method flow for generating digital cultural and creative content based on artificial intelligence, provided in an embodiment of the present invention. Detailed Implementation

[0047] The method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0048] Example 1, Reference Figure 1 A method for generating digital cultural and creative content based on artificial intelligence, the method comprising:

[0049] Step 1: Divide the input text description into several semantic segments according to predefined syntax and semantic rules. For each semantic segment, calculate the local semantic score using predefined word vector mapping and activation function. Calculate the global text embedding using the local semantic scores of each semantic segment and preset weights.

[0050] This step employs an AI-based text semantic modeling method, dividing the text description into several independent but interconnected semantic segments through syntactic analysis and semantic parsing. This segmentation not only considers the logical hierarchy within the text but also preserves the contextual relationships between words, ensuring that the deep-seated features of cultural and creative content can be fully captured during subsequent semantic embedding calculations. After completing the text semantic decomposition, each semantic segment needs to be vectorized to facilitate high-dimensional semantic computation by the computer. To this end, this invention employs a pre-trained word vector mapping method, projecting each word into a high-dimensional space, enabling the quantification of its semantic information. In this way, each semantic segment of the text can form a corresponding vector representation, which not only captures the meaning of the word itself but also implicitly includes its contextual relationships. After obtaining the vector representations of the semantic segments, it is necessary to further evaluate the semantic importance of each segment to ensure that the contribution of different segments is fully considered when constructing the global text embedding. To this end, this invention employs a semantic score calculation method based on nonlinear transformation, introducing exponential mapping and activation functions to dynamically adjust the semantic score between different segments. This calculation method can effectively avoid the semantic dilution problem that may be caused by traditional weighted averaging methods, so that segments with higher semantic relevance can be given higher weights when constructing global text embeddings, thereby enhancing the semantic consistency of the generated images.

[0051] After obtaining the scores of all semantic segments, the next step is to construct a global text embedding. Unlike existing technologies, this invention does not simply average all segments, but introduces a global semantic fusion method based on weighted similarity. This method first calculates the mean of the semantic vectors of all segments, using this as a baseline representation of the overall text semantics. Then, it calculates the Euclidean distance between each segment and this baseline, and defines a weight function based on this distance to determine the segment's contribution to the global embedding. This weight allocation mechanism ensures that segments closer to the global semantic center play a dominant role in the final embedding, while segments that deviate significantly are assigned lower weights to avoid semantic drift during image generation. Furthermore, this invention normalizes the global text embedding to ensure numerical consistency with random noise and manifold transformations during the latent variable construction stage. This normalization method not only improves computational stability but also reduces the semantic scale unevenness caused by different text input lengths and complexities, enabling the final generated creative content to maintain stable semantic expressiveness. Through the above method, the text semantic decomposition and global embedding construction process described in this invention ensures that the semantic information of the input text is fully utilized and forms a stable and parsable semantic representation in a high-dimensional space. This semantic representation not only provides high-quality initialization conditions for subsequent latent variable construction but also plays a global constraint role throughout the generation process, enabling the final generated digital cultural and creative content to accurately reflect the core ideas of the text description. In existing technologies, text semantic processing typically relies on simple word vector averaging or sentence vector encoding methods. However, this invention significantly improves the accuracy and consistency of semantic expression by introducing mechanisms such as nonlinear activation, adaptive weight allocation, and global normalization.

[0052] Step 2: Add the global text embedding to the noise generated using the standard normal distribution to construct the initial latent variable; regard the initial latent variable as being located on a hidden manifold, and construct a local loss function with the global text embedding as a reference. Calculate the deviation between the initial latent variable and the text embedding using a preset Riemann metric, and use the Riemann gradient descent method to perform multi-step iterative updates on the initial latent variable to obtain the latent variable after manifold transformation.

[0053] In the first stage of this step, random noise is generated using a standard normal distribution and combined with global text embedding to construct initial latent variables. The essence of this process is to introduce a degree of randomness, allowing the same text input to exhibit different artistic styles in different generated instances, while maintaining the core semantic information of the text. Compared to traditional deterministic generation methods, this invention uses random perturbation to distribute latent variables across different semantic neighborhoods, thus forming a richer space for creative expression. The theoretical basis of this method lies in the variational autoencoding idea in probabilistic generation models, which expands the latent space of data through noise perturbation, enabling the generated results to cover a wider range of possibilities. The use of a standard normal distribution ensures the uniformity and controllability of the noise, allowing random perturbation to provide sufficient creativity without causing excessive deviation from the semantic expression of the text. Furthermore, an adjustable scaling factor is introduced during the noise injection process to control the degree of influence of noise on the latent variables. By adjusting this factor, the generated results can be flexibly adjusted between faithfully adhering to the text description and showcasing creative variations, making this method adaptable to the generation needs of different cultural and creative content.

[0054] In the second stage of this step, considering that the direct addition of text semantic embedding and random noise still belongs to a linear transformation of Euclidean space, a manifold transformation is further introduced to allow latent variables to be embedded more naturally into a high-dimensional latent space while maintaining reasonable semantic topological relationships in the local structure. The core idea of ​​manifold transformation is to define a metric space using Riemannian geometry, so that the update of latent variables can follow certain geometric rules, rather than simply moving along the gradient direction of Euclidean space. Specifically, this invention first defines a local loss function based on global text embedding, which measures the distance between latent variables and text embeddings, and defines the local geometric structure by constructing a Riemannian metric tensor based on this loss. The introduction of the metric tensor allows the curvature of the space to be dynamically adjusted, thereby guiding latent variables to evolve along a more suitable path during the optimization process. The advantage of this method is that it can ensure that latent variables maintain a close connection with text semantics during the change process, while allowing a certain degree of deformation within a semantically reasonable range to enhance the creativity of the generated content.

[0055] In the third stage of this step, a Riemann gradient descent-based method is used to optimize the latent variables, gradually converging them towards a better latent space location. In traditional gradient descent methods, the update direction of the optimized variables depends only on the gradient of the loss function. However, this method, through gradient calculation on the manifold, ensures that the update direction is influenced not only by the gradient but also by the local geometric structure. Specifically, during each iteration, the deviation between the current state of the latent variable and the text semantic embedding is first calculated, and the inverse matrix of the local metric tensor is calculated based on this deviation. Subsequently, the gradient on the manifold is calculated and normalized using Riemann metrics to obtain an optimization direction that better conforms to the topological structure of the latent space. This process ensures that the evolution of the latent variables is not merely a simple numerical optimization process, but a dynamic adjustment process based on geometric constraints, enabling the generated content to maintain a high degree of coherence and controllability across different styles and text descriptions.

[0056] Compared to existing text-to-image generation methods based on GANs or VAEs, this invention utilizes manifold transformations to construct latent variables and combines random perturbations and geometric optimization methods. This allows the latent variables to not only cover a richer creative space but also maintain high accuracy and stability in semantic expression. Traditional methods often employ fixed latent distributions, limiting the diversity of generated results. This method, by introducing manifold optimization under global textual semantic constraints, allows the generation process to have artistic freedom without completely deviating from the semantic intent of the text description. Furthermore, existing methods typically use Euclidean gradient descent to optimize latent variables. While computationally simple, this method cannot adapt to the complex variations in high-dimensional nonlinear spaces. This invention, by introducing Riemannian geometry, enables a smoother and more efficient search in higher-dimensional latent spaces, thereby improving the quality and diversity of generated results.

[0057] Step 3: Using the transformed latent variables as the initial state, construct a potential energy function in the latent space, including text semantic constraints and global regularization terms. The latent variables are updated through the discretized stochastic differential equations. The updated latent variables are used as inputs. The texture amplitude, center frequency, and frequency diffusion parameters are determined according to the preset frequency modulation function. The inverse Fourier transform method is used to generate a local texture map.

[0058] This step, based on a diffusion model of stochastic differential equations, dynamically evolves latent variables to make their distribution in the latent space more uniform and stable. In natural physical phenomena, diffusion processes typically describe the random motion of particles in fluids. In the generation system of this invention, the introduction of a diffusion process aims to enable latent variables to adaptively adjust within a semantically reasonable range, avoiding them from getting trapped in local extrema due to over-optimization, which would result in a lack of diversity in the generated results. The diffusion process combines textual semantic information with random perturbations by constructing a specific potential energy function, allowing latent variables to change along a path consistent with the logic of artistic expression. This potential energy function contains two key parts: first, a global textual semantic constraint term, which ensures that latent variables do not deviate from the original semantic information, ensuring that the generated image remains faithful to the text description; second, a regularization term, which controls the range of change of latent variables, preventing them from over-diverging or converging too quickly during the update process. By continuously adjusting the weight ratio of the potential energy function during the diffusion process, the evolution of latent variables has both strong stability and the ability to explore more possibilities to a certain extent, providing richer basic information for subsequent image synthesis.

[0059] After completing the diffusion update of latent variables, this step further employs a frequency domain modulation method to ensure the controllability of visual details in the generated content. Traditional pixel-space-based generation methods often struggle to precisely control the local texture distribution of an image, while frequency domain methods can directly adjust the style, level of detail, and local texture of an image by adjusting the weights of different frequency components, making the generated content more in line with specific artistic expression needs. In the method of this invention, the core idea of ​​frequency domain modulation is to construct a set of modulation functions based on Fourier transform. The amplitude and phase of these modulation functions are controlled by latent variables, allowing the final generated texture map to be naturally embedded into the overall cultural and creative image. First, by calculating the norm of the latent variables, the basic amplitude of the modulation function is determined. This amplitude determines the overall intensity of the texture, allowing the texture details to be adaptively adjusted under different semantic descriptions. Second, based on the relationship between the latent variables and the text semantic embedding, the center frequency of the texture is dynamically calculated, enabling different text inputs to correspond to different main texture structures, achieving stylistic consistency. Finally, by adjusting the frequency diffusion parameter, the texture details can undergo fine-grained changes according to the text description. For example, when the input text describes gentle cultural elements, the frequency diffusion range is smaller to generate a more uniform and smooth texture, while when describing content with a strong artistic impact, the frequency diffusion range is larger to enhance visual contrast and dynamic change effects.

[0060] After completing frequency domain modulation, this step uses inverse Fourier transform to convert the modulated frequency domain information back to the spatial domain, generating a local texture map. The role of inverse Fourier transform is to remap frequency components to pixel space, allowing the modulated features to be directly reflected in the final image generation process. This process also involves the introduction of a phase modulation function to ensure that the spatial distribution of the texture map matches the semantic information of the original text. The core idea of ​​phase modulation is to calculate the Euclidean distance between the latent variables and the text semantic embedding and map it to a phase offset, thereby fine-tuning the local features of the texture during reconstruction. This results in a final image that conforms to the overall semantic expression while possessing rich local detail variations. In this way, the method of this invention can enhance the artistic style of the image while maintaining overall visual aesthetics, making it more in line with the needs of the cultural and creative industries.

[0061] Step 4: Using the generated local texture map as a local driving factor, construct a partial differential equation based on diffusion and reaction mechanisms. This partial differential equation performs spatiotemporal evolution of the image brightness distribution under preset initial conditions to achieve adaptive synthesis of the image structure layout and generate image content.

[0062] Specifically, in this step, the local texture map generated in the previous stage is used as a driving factor to provide initial guidance for the structural layout of the entire image. Since the local texture map is generated after diffusion updates and frequency domain modulation of latent variables, it already contains a large number of visual features corresponding to textual semantics, such as color changes, boundary features, and texture density in different regions. Based on this texture map, a set of partial differential equations is constructed to describe the evolution of the image brightness distribution, enabling the local structure of the image to be dynamically adjusted in continuous time, thereby achieving the gradual construction from the initial state to the final complete image. The introduction of partial differential equations gives the entire generation process a physics-like simulation characteristic, ensuring that the features of different regions do not change abruptly but rather gradually change smoothly, thus forming a more natural composition. This method differs from traditional image generation networks, which typically generate the final image directly through end-to-end mapping of neural networks. This method introduces a physical constraint during the generation process, making the layout and detail evolution of the content more in line with natural laws. In constructing the partial differential equations, a combination of diffusion and reaction terms is used to simultaneously control the global smoothness and local detail changes of the image. The diffusion term, through the calculation of local gradients, enables interaction between adjacent pixels, ensuring a smooth image transition. The reaction term, based on driving information provided by the local texture map, enhances or suppresses specific structural features within a particular region. This combination ensures both global consistency and flexible adjustment within local regions, providing adaptability to different types of cultural and creative content. For example, when generating images in the style of traditional ink painting, the diffusion term controls the natural rendering of ink marks, while the reaction term adjusts the color intensity and brushstroke shape in different regions, making the final image more consistent with the visual characteristics of ink painting. When generating futuristic science fiction works, the reaction term enhances the high-contrast structure in local regions, making the final image more technological and futuristic. Therefore, by controlling different weight ratios of the diffusion and reaction terms, adaptive generation of images of different styles can be achieved, making this method more adaptable to the diversity of cultural and creative content generation. In the numerical solution of the partial differential equations, a discretization method is used to approximate the continuous time evolution process, ensuring that the generation process converges to a stable image structure within a finite time. The key to numerical solutions lies in designing reasonable boundary conditions to prevent discontinuities or abrupt transitions at the edges of the image. To this end, this method employs a Neumann boundary-based strategy, which forces the gradient to zero in the boundary regions of the image, thus ensuring consistency of content at the edges without noticeable truncation or jumps.Furthermore, to further enhance the level of detail in the generated results, this method introduces a dynamic adjustment mechanism based on local feature contrast in the reaction term. When calculating the reaction intensity, it considers not only the information provided by the local texture map but also the local contrast of the current image state. This ensures that high-contrast areas remain clear during evolution, while low-contrast areas transition more naturally. This dynamic adjustment approach ensures that the final generated image possesses strong visual hierarchy while maintaining an overall artistic style that aligns with the semantic intent of the text description.

[0063] Example 2: In step 1, let the input text description be T = {w1, w2, ..., w...} j , ..., w M}; where w j This represents the j-th word, where j is an integer index ranging from 1 to M, and M is the total number of words; a predefined word vector mapping. The word w j Mapped to a real number vector; This represents the vocabulary, which is the set of all possible words. The dimension of the word vector indicates the word w j In word vector space there are Each word has a feature value; each word is mapped to a vector space to obtain the word vector φ(w). j Using predefined syntactic and semantic rules, the text is decomposed into N semantic segments, where the i-th semantic segment is... Let n be the j-th word in the i-th semantic segment; word index n i，j Indicates the position of the word within the entire text; L i is the number of words in the i-th semantic segment; N is the total number of semantic segments after the text is divided.

[0064] Specifically, the input text description is represented as a word sequence T = {w1, w2, ..., w...} M}, where w j Let represent the j-th word in the text, M be the total number of words, and j be the integer index indicating the word's order in the text. Since words in natural language text have different meanings and contexts, a predefined word vector mapping function φ(w) is needed. j ) each word w j This mapping is performed on a high-dimensional vector space, allowing the semantic information of words to be expressed numerically. This mapping function is typically trained by a deep learning model (such as the embedding layer of a Word2Vec, GloVe, or Transformer language model) to preserve the semantic relationships of words in the high-dimensional vector space. Specifically, φ(w jThe domain of a word is the vocabulary V, which is the set of all possible words, and the range is a... A vector space of dimension, where This represents the dimension of the word in the word vector space. Under this mapping, each word w... j Both can be represented as having The text is represented by real-valued vectors of features, allowing the semantic information of the text to be measured by calculating the similarity between vectors. After obtaining the vector representations of all words, the text needs further structuring to construct suitable semantic embeddings for computation in subsequent steps. Since the semantics of text are usually not determined by individual words, but by combinations of multiple words expressing more complex concepts, simply representing words independently is insufficient to support high-quality content generation. Therefore, in the second stage of this step, we split the text into N independent semantic segments T according to predefined syntactic and semantic rules. i Each semantic segment T i Each is a single phrase or sentence that can express complete semantic information. Each semantic segment T i From a group of words Composition, where n i，j This represents the position index of the word within the entire text, while L iThe number of words in the semantic segment represents the total number of semantic segments the text is divided into. The total number of semantic segments is denoted by N, which is determined by the text's length, complexity, and the semantic segmentation algorithm. During text segmentation, multiple factors need to be considered to ensure that the semantic segmentation accurately reflects the text's hierarchical structure and semantic logic. On one hand, syntactic segmentation rules ensure that the text is reasonably split into sentences or phrases, for example, using punctuation marks such as periods, commas, and semicolons as segmentation criteria. On the other hand, semantic segmentation rules can group semantically related words into the same segment based on semantic similarity, co-occurrence relationships, or dependency relationships between words. For example, in natural language processing, common methods include autoregressive segmentation based on attention mechanisms, clustering segmentation based on word embedding similarity, and tree-structure segmentation based on syntactic dependency relationships. These methods effectively capture the hierarchical relationships of the text, enabling the final semantic segments to more accurately reflect the core meaning of the text. Compared to traditional text processing methods, the method of this invention has stronger adaptability and intelligence in semantic decomposition. Traditional methods typically employ fixed window sizes or rule-based sentence segmentation. In contrast, the method of this invention incorporates artificial intelligence, dynamically adjusting the segmentation strategy based on the specific content of the text, resulting in more flexible semantic segmentation. For example, describing an abstract painting might require integrating descriptions of multiple visual elements into a longer semantic segment, while describing artworks with numerous details necessitates breaking the text down into smaller sub-units to more accurately express the features of each part. This dynamic adjustment capability allows the method to adapt to different expressive needs when processing various types of cultural and creative texts, thereby ensuring higher semantic consistency between the generated image content and the input text.

[0065] Example 3: In step 1, a semantic scoring function is defined for each semantic segment:

[0066]

[0067] Where ||·||1 represents the L1 norm operation; ψ(T i The local semantic score is used; and the segment semantic vector is constructed using the Sigmoid activation function:

[0068] s i =σ(ψ(T) i )-θ)v0;

[0069] Among them, s i Let be the semantic vector of the i-th semantic segment; σ(·) is the Sigmoid activation function; θ is the activation threshold, a preset scalar used to shift ψ(T) iThe output of ) controls the response range of the Sigmoid activation function; v0 is the basic semantic direction vector, which is a vector in the range of ) The preset vector in; d s The dimension of the semantic vector determines the size of the feature space for global text embedding; global text embedding is obtained by weighted averaging of the semantic vectors of each segment.

[0070]

[0071] in, represents global text embedding; ||·||2 represents the L2 norm.

[0072] Specifically, to measure the importance of different text fragments, this step defines a semantic scoring function ψ(T) based on nonlinear transformation. i This scoring function calculates the L1 norm of all word vectors in a segment and normalizes it using an exponential transformation, ensuring that the semantic score reflects the overall semantic activity of the segment. The use of the L1 norm ensures that the contribution of word vectors from all dimensions is considered, while the exponential transformation emphasizes words with high semantic intensity, avoiding unnecessary influence from words with low semantic intensity on the overall score. Furthermore, to ensure the numerical stability of the semantic score, a logarithmic transformation is introduced, allowing the scoring function to maintain good numerical distribution characteristics under different text input conditions. The essence of this scoring method is to perform a non-linear mapping of the features of different semantic segments in the word vector space, ensuring that the subsequent semantic vectors better reflect the hierarchical structure of the text. After calculating the semantic score of the text segment, this step further transforms the score using the Sigmoid activation function to obtain the semantic vector s of the segment. iThe introduction of the Sigmoid function normalizes the numerical range of the scores to (0, 1), and by controlling the activation threshold θ, different semantic scores of varying intensities can achieve different responses during the mapping process. The core function of this process is to weaken low-semantic-intensity segments and highlight high-semantic-intensity segments through nonlinear mapping, thus providing more discriminative input when constructing global text embeddings. Furthermore, a basic semantic direction vector v0 is introduced during semantic vector construction. This vector provides a reference direction throughout the embedding space, making the distribution of semantic representations of all segments more stable in the high-dimensional space and reducing instability caused by numerical differences between semantic segments. In this way, this step ensures that the representation of each semantic segment in the embedding space has strong discriminativeness and stability, laying the foundation for subsequent global embedding computation. After obtaining the semantic vectors of all segments, this step employs a weighted average-based global text embedding method to ensure that the global semantics can integrate information from different segments and has stronger robustness. Traditional text embedding methods typically employ simple averaging strategies. This invention introduces a Gaussian-weighted weighted averaging method. By calculating the Euclidean distance between each segment vector and the mean of all segment vectors, and using a Gaussian kernel function for weight normalization, segments closer to the global center receive higher weights, while segments farther from the center contribute less to the final embedding. The advantage of this method is its ability to adaptively adjust the influence of different segments, ensuring that the final global text embedding more accurately reflects the main semantic information of the text, while avoiding offsets caused by the presence of anomalous segments. Furthermore, the use of Gaussian weights effectively reduces noise, making the global embedding calculation smoother and ensuring the stability of the generated results.

[0073] Example 4: In step 2, the initial latent variables are constructed through the following process: for k = 1, ..., d s ,definition:

[0074]

[0075] U k V k ~Uniform(0,1);

[0076] Among them, U k and V k For the k-th noise component δ k The noise is generated by independently sampling random variables from a uniform distribution (Uniform(0,1)); all noise components are used to construct a noise vector. d sA 3D identity matrix is ​​used to ensure that there is no correlation between the dimensions of the noise; the initial latent variable ξ0 is constructed using the noise vector:

[0077] ξ0=e+λδ;

[0078] λ adjusts the noise amplitude, with a value ranging from 0.1 to 0.5.

[0079] Specifically, the core idea of ​​the Box-Muller transform is to construct a standard normally distributed random variable from a uniformly distributed random variable, ensuring that the noise component in each dimension of the high-dimensional space conforms to the property of a normal distribution. Specifically, for each dimension k of the latent variable, two independent random variables U are first sampled from a uniform distribution in the interval [0, 1]. k and V k Then, the noise component δ, which follows a standard normal distribution, is calculated using the transformation relationship. k During this process, lnU k The operation is used to transform a uniformly distributed variable into one that conforms to an exponential distribution, while cos(2πV) kThis value is then mapped to the angular space of a standard normal distribution, ensuring that the final noise follows a normal distribution with a mean of zero and a variance of one. The advantage of this method is that it can efficiently generate standard normal distribution variables from independent uniformly distributed variables through simple mathematical transformations, and the generated noise is independent across different dimensions, ensuring that the final constructed latent variables maintain good distributional properties. This approach ensures the uniformity of the noise throughout the latent space, allowing the generated cultural and creative content to exhibit rich diversity across different instances, rather than being limited to specific patterns or styles. After generating the high-dimensional standard normal noise vector, the next step is to combine this noise vector with the text semantic embedding vector to construct the initial latent variables. The text semantic embedding vector *e*, calculated in the previous steps, contains the core semantic information of the input text and is the foundation of the entire generation system. However, if only the semantic embedding vector is used for generation, the final cultural and creative images will lack diversity because the same text input will always map to the same embedding space point, leading to a fixed pattern of generated content. Therefore, in this step, we introduce a noise vector and control its influence with a scaling factor λ to construct a latent variable ξ0 that contains both semantic information and a certain degree of randomness. The role of λ is particularly crucial, as it determines the trade-off between the semantic embedding vector and random noise. If λ is small, the initial latent variable is almost entirely dominated by the textual semantic embedding, and the generated cultural and creative content will mainly rely on the input text, resulting in lower randomness and more stable generated content for the same text input. If λ is large, the influence of noise increases, leading to greater diversity in generated content, but potentially causing deviations from the textual semantics. Therefore, λ needs to be adjusted within a reasonable range to ensure that the generated content is both faithful to the textual description and possesses sufficient artistic innovation. In this invention, the value of λ is set to a range of 0.1 to 0.5 to ensure that the generation style can be flexibly adjusted in different types of cultural and creative content generation tasks, making it suitable for different application scenarios, such as virtual art creation, stylized painting generation, and cultural visual design. Compared to traditional text-to-image generation methods, this invention introduces more rigorous mathematical methods in the initialization of latent variables to ensure that the generated noise conforms to a normal distribution and provides reasonable perturbation in the latent space. This allows the generated content to exhibit stronger creative variations across different instances. Traditional latent variable initialization methods typically employ simple Gaussian noise superposition or random sampling. While these methods provide a degree of randomness, they often result in uneven noise distribution in high-dimensional space, thus affecting the quality of the final generated content. This invention generates standard normal noise through Box-Muller transform and ensures the independence of each dimension of the noise, thereby ensuring a more reasonable distribution of the initial latent variables throughout the latent space and making the subsequent generation process more stable.Meanwhile, the adaptive adjustment mechanism of λ ensures the organic combination of textual semantic embedding and random noise, enabling the generated content to maintain semantic consistency while showcasing more stylistic variations at the detail level. The advantage of this method lies in its ability to provide more flexible control in artistic image generation tasks, allowing different text inputs to exhibit varying degrees of variation in the generated content by adjusting the randomness weights, thereby meeting diverse creative needs.

[0080] Example 5: In step 2, assume that the initial latent variable ξ0 is distributed in the hidden manifold The local geometry is described by a metric constructed based on constraints from global text embedding, defining a local loss function. for:

[0081] Specifically, in traditional text-to-image generation methods, latent variables are typically initialized using random distribution or direct sampling from Gaussian noise. While this provides a degree of randomness, it often lacks fine-grained control over text semantics, easily leading to instability in style and structure of the generated results. In the method of this invention, we assume that the latent variables are not uniformly distributed throughout the high-dimensional space, but are constrained by text semantics and distributed on a specific manifold M. Therefore, in optimizing the latent variables, a reasonable loss function needs to be defined to adjust within the local geometry of the manifold. This loss function minimizes the distance between the initial latent variable ξ0 and the global text semantic embedding e, ensuring stable convergence of the latent variables on the manifold without excessive deviation or random drift. Simultaneously, cos(2πV) k As a periodic modulation term, its main function is to introduce controlled fluctuations into the latent variables, allowing them to exhibit stylistic variations across different instances without falling into a fixed pattern. The core mathematical idea behind this design originates from modulation methods in signal processing, where a periodic factor is applied to a variable to cause it to change regularly within a specific range, thereby enhancing the stylistic diversity of the generated content. In this step, the introduction of this periodic term prevents the latent variables from falling into local extrema during optimization. Instead, it allows for reasonable exploration of the latent manifold while adhering to the semantics of the text, resulting in cultural and creative content that possesses both stable semantic expression and controllable stylistic variations.

[0082] Compared to the traditional L2 loss function, the loss function of this invention, by incorporating periodic modulation, makes the optimization process of latent variables more consistent with the geometric characteristics of high-dimensional manifolds. During the training of the generative model, if the optimization process of latent variables relies too heavily on a fixed L2 loss, it can easily lead to all latent variables converging to a static mean point, limiting the diversity of the generated results. However, in this method, cos(2πV) k This allows the loss term to have a certain degree of variability in different optimization iterations, thereby avoiding the latent variables from getting trapped in local minima and promoting natural variations in the generated content across different styles. For example, in a cultural and creative content generation task, if the input text describes a visual art work with an Eastern ink painting style, the periodic modulation term can make the generated brushstroke details vary between different instances during the optimization of the latent variables, making the texture of some parts exhibit a stronger brushstroke feel, while other parts may be more delicate and soft. If the input text describes an art work with a digital fantasy style, this modulation term can be used to enhance the modulation of different frequencies in local areas of the latent variables, so that the final generated work can exhibit a more dreamlike and futuristic light and shadow effect. The introduction of this controllability allows the generation method of this invention to not only maintain consistency with the text description in overall style, but also to achieve highly flexible style adjustments at the detail level, in order to meet the needs of different application scenarios in the digital cultural and creative industries. From the perspective of geometric optimization, the loss function in this step can be understood as a constrained optimization problem, that is, on a high-dimensional manifold M, the latent variables need to maintain the smoothness and coherence of the manifold as much as possible while satisfying certain textual semantic constraints. Since a manifold is a nonlinear subspace, the gradient descent direction of the latent variables during optimization is constrained not only by the L2 norm but also by local geometry. The introduction of a periodic modulation term allows the update direction of the latent variables to exhibit variations across different instances, enhancing the dynamism of the final generated content. This method is particularly important in artistic creation tasks, as artworks often possess stylistic characteristics while also needing to showcase unique nuances across different instances. This method achieves this goal through geometric optimization of latent variables, resulting in generated works that possess both a unified stylistic character and rich variations in specific details. For example, in the task of digitally restoring cultural heritage, different input texts may describe different parts of the same ancient mural. This method ensures that, while maintaining overall stylistic consistency, local adjustments to the artistic style of different areas are made to more accurately restore the detailed features of the cultural heritage. Furthermore, in immersive art generation tasks, the optimization process ensures that the generated content still conforms to specific visual style requirements under different angles and lighting conditions, thereby improving the stability and adaptability of the generation system.

[0083] Example 6: In step 2, the process of calculating the deviation between the latent variables and the text embedding using a preset Riemann metric, and then using the Riemann gradient descent method to perform multi-step iterative updates on the latent variables to obtain the latent variables after manifold transformation includes:

[0084] The default Riemann metric is g:

[0085]

[0086] Where, ξ m Let ξ be the m-th component of ξ0; let ξl be the l-th component of ξ0; e m e is the m-th component; l g is the l-th component of e; ml Let g be the element in the m-th row and l-th column of the Riemann metric g; The latent variables are updated using Riemann gradient descent using the following formula:

[0087]

[0088] Among them, g -1 The matrix representing the inverse of the Riemannian metric is used to obtain the latent variables after the manifold transformation. T1 represents the total number of iterations; Gradient operator; Let ξ be the latent variable obtained in the (t-1)th iteration. t-1 The local loss function; ξ t Let be the latent variables obtained in the t-th iteration.

[0089] Specifically, in this optimization step, the first step is to define the geometric structure of the space containing the latent variables. Assume that the manifold M containing the latent variable ξ0 is not a uniform Euclidean space, but rather a space with a certain curvature influenced by the text semantic embedding e. To describe the local geometric properties of this space, a Riemannian metric g is introduced. Its definition ensures that the influence of the metric is smaller when the latent variable is close to the text embedding, making the optimization process closer to gradient descent in standard Euclidean space. Conversely, the influence of the metric gradually increases as the latent variable moves away from the text embedding, thus guiding the latent variable back towards the semantic center. The core idea of ​​this metric is to adjust the local geometric structure of the space so that the latent variable can change along a more reasonable path during the optimization process, rather than simply updating along the gradient direction, thereby avoiding semantic drift caused by over-optimization. The construction of the Riemannian metric fully considers the relative relationship between the latent variables and the text embedding, with the key correction term (ξ) being... m -e m (ξ) l -el This demonstrates how the variations of latent variables across different dimensions align with the semantic embedding of the text, while also using normalization factors. To ensure the numerical stability of the optimization process, the optimization can still converge smoothly even when the latent variables are very close to the text embedding. This approach is actually a weighted method based on local structure. When the latent variables are close to the text embedding, the optimization process is subject to less geometric constraint, allowing for more free exploration. Conversely, when the latent variables are far from the text embedding, the geometric constraint is strengthened, making the optimization direction more controlled, thus ensuring that the final generated content still conforms to the semantic description of the text.

[0090] After defining the Riemann metric, the core of the optimization process lies in how to update the gradient based on this metric. In traditional gradient descent methods, variable updates rely solely on the calculation of local gradients, i.e., stepping along the direction of gradient descent. However, in Riemann space, due to the influence of the metric, the gradient needs to be corrected through the inverse transformation of the metric tensor, making the update direction more consistent with the geometry of the manifold. This means that in this method, the gradient descent update formula not only involves the gradient of the loss function but also requires transformation of the gradient using the inverse matrix of the metric g to ensure that the optimization path is adjusted along the optimal manifold direction. The advantage of this method is that it makes the optimization process of latent variables more consistent with the structure of the data manifold, thereby ensuring the stability and semantic consistency of the generated content. For example, in stylized art generation tasks, traditional methods often struggle to achieve smooth transitions between different styles, while this method, by applying geometric constraints to the optimization path of latent variables, makes the changes between different styles more natural and maintains a certain degree of coherence between different instances. Furthermore, the learning rate setting in this method has also been finely adjusted during the optimization process. Since the step size directly affects the convergence speed and stability of the optimization during gradient descent, a fixed step size of 0.05 is used for updates to ensure that the optimization process can converge to a stable state within a reasonable time. Compared to adaptive learning rate methods, this method uses a fixed step size, making the update process of latent variables more stable and avoiding optimization oscillations caused by excessive changes in the learning rate. Furthermore, since the optimization process is based on manifold geometry, the step size setting needs to match the numerical range of the metric to ensure that the optimization process maintains smooth updates throughout the latent space. Through this optimization method, the final latent variables are obtained... After multiple iterative convergences to a stable point on the manifold, the generated content achieves a balance between artistic style and semantic consistency. Compared to traditional text-to-image generation methods, this invention introduces the concept of Riemannian geometry into the latent variable optimization process, making the generated content not only more stable in overall style but also exhibiting richer variations in local details. For example, in digital art creation, certain styles of works may require different brushstroke features in different local areas. This method, through geometric constraints in the optimization process, ensures that these stylistic features remain consistent across different instances while still showcasing local diversity. Furthermore, in the task of personalized customization of cultural and creative content, this method can adaptively adjust the optimization path in the latent space based on the features of different text inputs to ensure that the generated results meet the semantic requirements of the specific text description while also exhibiting rich variations in local details.

[0091] Example 7: In step 3, after obtaining Building upon this foundation, a diffusion process is introduced to further balance textual semantic constraints and generative exploration, constructing the following potential function with a global regularization term.

[0092] Where δ1>0 is a preset value, providing a global contraction effect; a stochastic differential equation for the diffusion process is introduced:

[0093]

[0094] Where D > 0 is the preset diffusion coefficient, and W(t) is the standard Wiener process; let , define the discretization step size Δt, and the iterative formula is:

[0095]

[0096] in, After T1 iterations, the final latent variables after diffusion update are obtained.

[0097] Specifically, this step first defines a potential energy function. It consists of two main parts. The first is to ensure that the latent variables do not deviate too far from the constraints of the textual semantic embedding, i.e. This part ensures that the latent variables are always adjusted around the textual semantic embedding *e*, so that the final generated cultural and creative content always maintains a close connection with the original text description. This constraint avoids semantic drift in the generated content due to increased randomness during the optimization process, ensuring that the final generated image not only possesses high artistic value but also accurately reflects the core meaning of the text. Secondly, This regularization term introduces a global contraction effect, ensuring that latent variables do not expand indefinitely but maintain a certain convergence trend overall. This global regularization mechanism guarantees that the evolution of latent variables will not diverge excessively due to the complexity of high-dimensional space, thereby improving the stability of the optimization process. Furthermore, the parameter δ1, as an adjustable global contraction parameter, directly affects the stability and exploratory capability of the final generated content. When δ1 is small, the latent variables have higher degrees of freedom, enabling a wider range of exploration within the semantic space, resulting in richer variations in the generated content; while when δ1 is large, the distribution of latent variables becomes more concentrated, leading to a more consistent style in the generated content. Therefore, the choice of δ1 needs to be adapted to the specific cultural and creative application scenario. For example, in digital art drawing tasks, higher variability in the generated results may be desired, while in cultural heritage digital restoration tasks, stable style consistency may be more necessary. After constructing the potential function, this method further introduces a diffusion process to iteratively update the latent variables through stochastic differential equations. The basic idea of ​​the diffusion process is to decompose the changes in latent variables into two parts: one part is controlled by the gradient of the potential function, whose main function is to guide the latent variables to converge along the semantic embedding direction; the other part is controlled by a random noise term driven by the standard Wiener process, allowing the latent variables to explore moderately in high-dimensional space, so as to ensure that the generated content still has a certain degree of style variation among different instances. The introduction of this diffusion mechanism makes the optimization of latent variables no longer a simple gradient descent process, but a controlled evolution in a dynamic potential field, thus allowing a certain degree of style shift while ensuring the semantic consistency of the generated content.

[0098] In the implementation, a discretized iterative formula is used to simulate the diffusion process, allowing the latent variables to gradually evolve to a stable state over multiple time steps. The iterative update method ensures that each update is adjusted based on the current potential gradient. Furthermore, by introducing an additional random noise term, the optimization process avoids getting trapped in local optima and instead adaptively adjusts within a reasonable range. (Gradient term) Depend on This approach ensures that the optimization direction is always constrained by the semantic embedding of the text, while maintaining an overall contraction effect, making the optimization path more aligned with the generation needs of cultural and creative content. The random noise term... This method provides additional exploratory capabilities during the optimization process, allowing latent variables to evolve freely within a certain range rather than being confined to a fixed path. This results in richer stylistic variations in the generated cultural and creative content. The parameter D, acting as a diffusion coefficient, controls the influence of random noise on the optimization process. A smaller D value means the evolution of latent variables is primarily constrained by the potential gradient, leading to more stable generated content. Conversely, a larger D value amplifies the influence of noise, increasing the range of variation in the generated content. Therefore, the D value needs to be adjusted based on specific application requirements to achieve a balance between stability and diversity. Compared to traditional text-to-image generation methods, this invention's optimization method not only relies on deterministic gradient optimization but also incorporates a random diffusion mechanism, enabling the generated cultural and creative content to exhibit richer variations across different instances. Traditional optimization methods typically rely on fixed gradient directions. While this ensures consistency between the generated content and the text description, it often results in insufficient variation across different instances. This method, by introducing controlled random diffusion during optimization, allows generated content with the same text description to exhibit different visual features to a certain extent, thereby enhancing the expressive power of the cultural and creative content. For example, in artistic stylization tasks, this method ensures that, with the same text input, the generated images maintain basic stylistic consistency while exhibiting different artistic expressions at the level of detail, thus avoiding overly mechanical repetition of generated content. Furthermore, in digital cultural heritage restoration tasks, this method ensures that the generated cultural elements maintain basic historical consistency across different instances, while showcasing different artistic variations in local details to better match diverse visual expression needs.

[0099] Example 8: In step 3, the frequency modulation function The formula is:

[0100]

[0101] Where ω is the frequency variable, which belongs to the set of real numbers. a0 is the constant base value of amplitude modulation; a1 is the constant value of amplitude modulation. The relevant proportionality coefficient is a real constant; the center frequency ω base The reference center frequency is a preset constant; <·> indicates inner product operation; b is the proportionality coefficient; and the frequency standard deviation is... σ0 is the baseline standard deviation, a positive preset constant; the local texture map is obtained through inverse Fourier transform.

[0102] Where x is the coordinate of the texture map position; This indicates the operation of taking the real part; s is the imaginary unit; Ω x For texture graph space; x∈Ω x Phase function Where c0 and c1 are constants.

[0103] Specifically, frequency modulation function This is the core part of this step, determining the final texture's frequency distribution and the detail variations in different regions. This function consists of an amplitude modulation term and a center frequency control term, allowing the frequency characteristics to adaptively adjust according to the state of the latent variables. In the amplitude modulation part, The overall intensity of the frequency amplitude is controlled, where a0 is the reference amplitude, and... By introducing the norm of the latent variables, the amplitude can be dynamically adjusted according to the state of the latent variables. This mechanism ensures that the overall amplitude of the texture is more stable when the latent variables are close to the text semantic embedding, while the intensity of the texture can change when the latent variables are far from the text embedding, thus revealing richer texture features in local details. Compared with traditional fixed amplitude modulation methods, this method, through adaptive control of latent variables, enables the generated images to maintain semantic consistency across different instances while exhibiting richer stylistic variations. The setting of the center frequency is also crucial, as it determines the main frequency components of the texture and affects the detailed features of the final generated image. In this method, the center frequency is determined by... Given, where ω base It is the reference center frequency, and By calculating the inner product of latent variables and text semantic embeddings, the center frequency can adaptively adjust according to the state of the latent variables. The mathematical significance of this mechanism lies in... The inner product with e reflects the similarity between the latent variable and the text semantics. Therefore, when the latent variable is close to the text semantic embedding, the center frequency is close to the preset baseline value. When the latent variable changes, the center frequency can adjust accordingly, ensuring that the texture features of the generated content can adaptively adjust with changes in text semantics. Compared to traditional fixed center frequency methods, this method, through center frequency control based on latent variables, enables the generated image to better match the style requirements of different text descriptions, while also showcasing more refined texture variations in local details. Furthermore, to further control the frequency diffusion range, this method introduces the frequency standard deviation. Where σ0 is the baseline standard deviation, and the denominator contains... By measuring the distance between latent variables and textual semantic embeddings, the standard deviation can be dynamically adjusted according to the state of the latent variables. When the latent variables are close to the textual semantic embeddings, A larger value indicates a wider frequency diffusion, resulting in a softer texture; while when latent variables are far from the text semantic embedding, The smaller size of the frequency distribution makes it more concentrated, thereby enhancing local detail. This mechanism ensures that the generated cultural and creative content maintains overall consistency across different instances while exhibiting rich variations in local details, thus improving the artistic expressiveness of the generated results.

[0104] After constructing the frequency modulation function, we convert it into a local texture map T(x) using inverse Fourier transform. The purpose of inverse Fourier transform is to convert the modulation information in the frequency domain back to the spatial domain, so that the final generated image can reflect the characteristics of frequency modulation at the pixel level. Among these, the phase function... By controlling phase changes at different locations, the texture can exhibit different local features in different regions. This phase control mechanism ensures that the generated cultural and creative content displays different texture features in different regions, thereby enhancing the layering and dynamic variation capabilities of the generated image. Compared to traditional direct texture mapping methods, this method uses inverse Fourier transform for modulation in the frequency domain, resulting in richer local details and greater controllability in the generated image. Compared to traditional pixel-space-based generation methods, the core innovation of this method lies in utilizing frequency domain modulation technology, allowing changes in latent variables to be directly mapped to frequency characteristics, thus enabling the generated cultural and creative content to exhibit richer artistic expression in local details. Traditional methods typically rely on end-to-end training of neural networks, while this method uses mathematical analysis to directly control the frequency domain, resulting in generated content that is not only more natural in style but also has higher precision in local details. For example, in digital art stylization tasks, this method can ensure that the generated content maintains a consistent overall style under different text inputs, while exhibiting different texture variations in local details. Furthermore, in the task of digital restoration of cultural heritage, this method can ensure that different artistic features are displayed in local areas while maintaining a consistent overall visual style, thereby better matching the visual expression needs of historical culture.

[0105] Example 9: In step 4, the obtained local texture map is used. To generate the structural layout of a two-dimensional image; let the image be I(x, y, v), its evolution is driven by partial differential equations of diffusion and reaction mechanisms:

[0106]

[0107] Among them, the local diffusion coefficient Where κ0>0 is a preset value; τ>0 is a preset value to ensure that diffusion weakens at abrupt texture changes; v is the number of iterations; and the decay rate is... Where μ0>0 is a preset value; texture driving weight λ T >0 is the preset value; the initial condition is selected as I(x, y, 0) = I0(x, y), which is a zero-mean noise image. After T3 iterations, the image content I is obtained. init (x, y) = I(x, y, T3).

[0108] Specifically, the key to this step lies in defining a partial differential equation based on diffusion and reaction mechanisms to describe the image's evolutionary behavior. The image evolution is controlled by the time variable t, and the image's state at any given time is defined as I(x, y, v), where (x, y) are the image's spatial coordinates, and v represents the iteration number. The core driving force of the entire evolutionary process comes from local diffusion, attenuation mechanisms, and texture-driven generation terms, enabling the image to gradually form a final form that conforms to the semantic description of the text during adaptive adjustment. (Diffusion term) The smoothing term is responsible for smoothing image content, allowing local pixel values ​​to propagate within their neighborhoods, thus ensuring the continuity of the overall structure. The reactive term -μ(x, y, T)I(x, y, v) is responsible for selectively suppressing certain features in the image, ensuring that the final image content displays more distinct local features. Furthermore, the external driving term λ... T T(x) is directly controlled by the local texture map T(x), and its main function is to ensure that the generated image content can reflect a specific visual style and structural layout in local regions, so that the final generated image is not just a smooth evolutionary result, but can exhibit visual features with a specific style. The diffusion coefficient κ(x, y, T) ensures that the diffusion process has different intensities in different regions, thus allowing the image content to expand more strongly in some areas while maintaining higher local detail in others. Specifically, the diffusion coefficient is defined as follows: In this method, κ0 is used as a preset value to control the global diffusion intensity, while the exponential term ensures that the diffusion intensity is low in regions with large texture gradients (i.e., regions with abrupt local texture changes), thus preserving the detailed information in these regions. The core function of this mechanism is to prevent the image from becoming overly smooth during evolution, ensuring that the generated cultural and creative content can exhibit clear local features while maintaining good coherence in smooth regions. Compared to traditional global diffusion methods, the innovation of this method lies in its adaptive adjustment of the diffusion coefficient through texture gradients, allowing different regions of the image to adjust according to local features, thereby exhibiting more flexible adaptability in different style generation tasks.

[0109] Furthermore, the attenuation rate μ(x, y, T) is also a key factor affecting the final image generation quality. This parameter controls the attenuation intensity of the image during its evolution, ensuring that key features in the image are preserved to a certain extent. Specifically, it is defined as μ0(1-exp(-||T(x, y)||2)), where μ0 is a global adjustment parameter, and the exponential term ensures that when the value of the local texture map is small, the attenuation rate is close to zero, allowing sufficient diffusion in that region. In regions with strong local texture, the attenuation rate increases, suppressing structural changes in the image to a certain extent, thus preserving key visual details. This attenuation control mechanism ensures that visual features in specific regions are not lost due to excessive diffusion during image generation, allowing the final generated image to exhibit clear structural features in local details. Texture-driven weight λ T This controls the degree to which the local texture map influences the overall image evolution. When λ T When λ is large, the local texture map has a stronger influence on the final generated image, allowing the image content to more closely follow the structure of the local texture map. Conversely, when λ is large... T When the value is small, the image generation process is more affected by the diffusion and attenuation terms, resulting in a softer overall style. Therefore, by adjusting λ... T The value of can flexibly control the overall style of the image content, enabling the final cultural and creative image to showcase stylized content with clear details while maintaining high stability in its overall structure. In the implementation process, the initial image I(x, y, 0) is set as a zero-mean noise image. After T3 iterations, the final image content I is obtained. init (x, y) = I(x, y, T3). This iterative process ensures that the image can be optimized according to the current state at each time step, eventually converging to a stable style structure. Compared with traditional neural network direct mapping methods, the advantage of this method lies in its physical simulation-based partial differential equation framework, which makes the image generation process not only mathematically interpretable but also provides higher controllability and adaptive adjustment capabilities. Traditional image generation methods usually rely on data-driven end-to-end training, while this method, through an explicit diffusion-reaction equation, allows the image generation process to evolve under specific rules, thereby ensuring a good balance between style consistency and detail representation in the generated results.

[0110] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result according to substantially the same method falls within the scope of the present invention. Therefore, the scope of the present invention is defined only by the appended claims.

Claims

1. A method for generating digital cultural and creative content based on artificial intelligence, characterized in that, The method includes: Step 1: Divide the input text description into several semantic segments according to predefined syntax and semantic rules. For each semantic segment, calculate the local semantic score using predefined word vector mapping and activation function. Calculate the global text embedding using the local semantic scores of each semantic segment and preset weights. Step 2: Add the global text embedding to the noise generated using the standard normal distribution to construct the initial latent variable; regard the initial latent variable as being located on a hidden manifold, and construct a local loss function with the global text embedding as a reference. Calculate the deviation between the initial latent variable and the text embedding using a preset Riemann metric, and use the Riemann gradient descent method to perform multi-step iterative updates on the initial latent variable to obtain the latent variable after manifold transformation. Step 3: Using the transformed latent variables as the initial state, construct a potential energy function in the latent space, including text semantic constraints and global regularization terms. The latent variables are updated through the discretized stochastic differential equations. The updated latent variables are used as inputs. The texture amplitude, center frequency, and frequency diffusion parameters are determined according to the preset frequency modulation function. The inverse Fourier transform method is used to generate a local texture map. Step 4: Using the generated local texture map as a local driving factor, construct a partial differential equation based on diffusion and reaction mechanisms. This partial differential equation performs spatiotemporal evolution of the image brightness distribution under preset initial conditions to achieve adaptive synthesis of the image structure layout and generate image content.

2. The method for generating digital cultural and creative content based on artificial intelligence as described in claim 1, characterized in that, In step 1, let the input text be described as follows: ;in Indicates the first One word, The index is an integer subscript, with a value ranging from 1 to... , Total number of words; predefined word vector mapping , will words Mapped to a real number vector; This represents the vocabulary, which is the set of all possible words. The dimension of the word vector indicates the word's... In word vector space there are Each word has a feature value; each word is mapped to a vector space to obtain word vectors. Using predefined syntactic and semantic rules, the text is decomposed into... The semantic fragment, the first The semantic fragments are ; For the first The first semantic segment in the Words; Word Index Indicates the position of the word within the entire text; For the first The number of words in a semantic segment; This represents the total number of semantic segments after the text has been divided.

3. The method for generating digital cultural and creative content based on artificial intelligence as described in claim 2, characterized in that, In step 1, a semantic scoring function is defined for each semantic segment: ； in, express Norm operations; Local semantic scores are calculated; and fragment semantic vectors are constructed using the Sigmoid activation function: ； in, For the first The segment semantic vector of a semantic segment; Use the Sigmoid activation function; The activation threshold is a preset scalar value used for translation. The output of controls the response range of the Sigmoid activation function; The basic semantic direction vector is a... The preset vector in; The dimension of the semantic vector determines the size of the feature space for global text embedding; global text embedding is obtained by weighted averaging of the semantic vectors of each segment. ； in, Indicates global text embedding; This represents the L2 norm.

4. The method for generating digital cultural and creative content based on artificial intelligence as described in claim 3, characterized in that, In step 2, the initial latent variables are constructed through the following process: For ,definition: ； ； in, and For the first noise component Independent sampling self-uniform distribution used during generation The random variable; constructing a noise vector from all noise components. ; express A 3D identity matrix is ​​used to ensure that there is no correlation between the dimensions of the noise; initial latent variables are constructed using the noise vector. : ； in, Adjust the noise amplitude, with a value ranging from 0.1 to 0.

5.

5. The method for generating digital cultural and creative content based on artificial intelligence as described in claim 4, characterized in that, In step 2, let the initial latent variables be set. Distributed in hidden manifolds The local geometry is described by a metric constructed based on constraints from global text embedding, defining a local loss function. for: .

6. The method for generating digital cultural and creative content based on artificial intelligence as described in claim 5, characterized in that, In step 2, the process of calculating the deviation between the latent variables and the text embedding using a preset Riemannian metric, and then iteratively updating the latent variables using the Riemann gradient descent method to obtain the latent variables after manifold transformation includes: The default Riemann metric is : ； in, for The One component; for The One component; for The One component; for The One component; Riemannian measurement The first in Okay, number Column elements; The latent variables are updated using Riemann gradient descent according to the following formula: ； in, The matrix representing the inverse of the Riemannian metric is used to obtain the latent variables after the manifold transformation. ; This represents the total number of iterations. Gradient operator; For the first Latent variables obtained in the second iteration The local loss function; For the first The latent variables obtained in the next iteration.

7. The method for generating digital cultural and creative content based on artificial intelligence as described in claim 6, characterized in that, In step 3, after obtaining Building upon this foundation, a diffusion process is introduced to further balance textual semantic constraints and generative exploration, constructing the following potential function with a global regularization term. : ； in, , as a preset value, provides a global contraction effect; introduces a stochastic differential equation for the diffusion process: ； in, , where is the preset diffusion coefficient. For the standard Wiener process; define the discretization step size. The iterative formula is: ； in, After passing through After several iterations, the latent variables updated by diffusion are finally obtained. .

8. The method for generating digital cultural and creative content based on artificial intelligence as described in claim 7, characterized in that, In step 3, the frequency modulation function The formula is: ； in, It is a frequency variable and belongs to the set of real numbers. ; This is the constant base value for amplitude modulation; For amplitude modulation and The relevant proportionality coefficient is a real constant; the center frequency ; The reference center frequency is a preset constant. Indicates inner product operation; ; is the proportionality coefficient; frequency standard deviation ; The baseline standard deviation is a positive preset constant; the local texture map is obtained through inverse Fourier transform. : ； in, These are the coordinates of the texture map position. This indicates the operation of taking the real part; The imaginary unit; For texture graph space domain; Phase function ;in, and It is a constant.

9. The method for generating digital cultural and creative content based on artificial intelligence as described in claim 8, characterized in that, In step 4, the obtained local texture map is used. To generate the structural layout of a two-dimensional image; let the image be Its evolution is driven by partial differential equations of diffusion and reaction mechanisms: ； Among them, the local diffusion coefficient ;in, This is the preset value; , is a preset value to ensure that diffusion weakens at abrupt texture changes; is the number of iterations; decay rate. ;in, , which is the preset value; texture-driven weight , is the preset value; the initial condition is selected as The image is a zero-mean noise image, after... Iteration yields the image content. .

Images