Intelligent image fast rendering processing method and system for cultural and creative industry

Abstract

The application provides an intelligent image fast rendering processing method and system for a text and creative industry, relates to the technical field of image processing, and comprises the following steps: constructing a dual-domain decoupling representation model to extract hierarchical decoupling features of an image, combining a text and creative idea to perform cross-modal alignment and joint optimization, generating an adaptive rendering prior and mapping the adaptive rendering prior into an intermediate rendering state, constructing a dynamic calculation graph according to the state, determining a sparse execution strategy through path importance propagation to generate a hierarchical rendering control sequence, finally performing non-uniform sampling rendering according to the hierarchical rendering control sequence to obtain a target image and jointly updating model parameters. The application significantly improves the processing efficiency of image rendering under the premise of ensuring the consistency of the rendering style and the content fidelity.

Description

Technical Field

[0001] This invention relates to image processing technology, and more particularly to a method and system for intelligent image rapid rendering processing for the cultural and creative industries. Background Technology

[0002] In the field of image rendering processing within the cultural and creative industries, existing technologies typically rely on a series of relatively fixed processes. The conventional approach is to directly process the raw image data holistically, for example, using a unified style transfer algorithm or a global rendering pipeline. These methods often treat the image as an indivisible entity, directly generating the rendering result through an end-to-end neural network model. During processing, the system usually adjusts the rendering effect based on preset general parameters or a single optimization objective (such as style similarity or content loss). To improve processing speed, common acceleration strategies include uniform downsampling of the image, reducing overall computational precision, or employing static model pruning techniques. These methods constitute the mainstream technical path for rapid rendering of cultural and creative images.

[0003] However, the aforementioned conventional approaches have significant limitations. A major drawback lies in their lack of deep adaptability and refined control to the specific intentions of cultural and creative content. Due to the adoption of a holistic and homogeneous processing strategy, the system struggles to effectively distinguish the differences in semantic importance and stylistic expression requirements among different regions of the image during rendering. This results in insufficient or distorted rendering of key cultural and creative elements, while unnecessary computational resources are consumed in non-critical background areas. Another drawback is that the efficiency optimization strategies of existing methods are often disconnected from the semantics of the content. Acceleration methods based on uniform sampling or static model compression cannot dynamically allocate computational resources according to the specific image's content complexity and rendering intentions. This easily leads to wasted computational power in simple areas and insufficient resources affecting rendering quality in complex areas. This rigid processing mode makes it difficult to achieve a dynamically adjustable balance between rendering effect and processing efficiency for specific cultural and creative scenarios. Summary of the Invention

[0004] This invention provides a method and system for intelligent image rapid rendering processing for the cultural and creative industries, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a method for rapid intelligent image rendering processing for the cultural and creative industries, comprising: Obtain the original image data to be rendered and the textual and graphic representation information associated with the original image data; A dual-domain decoupled representation model containing spatial topological consistency constraints and semantic coherence constraints is constructed. The original image data is then subjected to frequency domain decomposition and semantic field reconstruction to extract a decoupled feature tensor with a hierarchical structure. Based on the cross-modal semantic alignment of the creative graphic representation information and the decoupled feature tensor, an adaptive rendering prior distribution with style consistency guarantee is generated through joint optimization of adversarial style transfer constraints and content fidelity constraints, and the adaptive rendering prior distribution is mapped to the rendering operation space to form an intermediate rendering state representation. A dynamic computation graph is constructed based on the intermediate rendering state representation. The sparse execution strategy of rendering operations is determined through path importance propagation mechanism and gradient sensitivity analysis, and a layered rendering control sequence containing selective skip instructions and precision adaptive instructions is generated. Non-uniform sampling rendering is performed on the original image data according to the hierarchical rendering control sequence to obtain the target rendered image and rendering efficiency feedback index. The constraint weights of the dual-domain decoupled representation model and the activation threshold of the path importance propagation mechanism are jointly updated using the rendering efficiency feedback index.

[0006] A dual-domain decoupled representation model containing spatial topological consistency constraints and semantic coherence constraints is constructed. The original image data is then subjected to frequency domain decomposition and semantic field reconstruction to extract a decoupled feature tensor with a hierarchical structure, including: By constructing spatial topological consistency constraints based on geometric manifold embedding, the spatial structure in the original image data is encoded with topological invariance, generating a spatial topological representation that preserves local neighborhood relationships and global connectivity, and using the spatial topological representation to constrain spatial distortion in the frequency domain decomposition process. Based on the spatial topological representation, frequency domain orthogonal decomposition is performed on the original image data to map the original image data to the frequency domain representation space, thereby obtaining the frequency domain decomposition result; By constructing semantic coherence constraints based on semantic propagation graphs, cross-regional dependency models are performed on semantic elements in the frequency domain decomposition results to generate a semantic field distribution with semantic consistency guarantees, and the semantic field distribution is used to guide the semantic field reconstruction process. Based on the semantic field distribution, semantically driven feature recombination is performed on the frequency domain decomposition result. Semantic field reconstruction is completed through semantic affinity and semantic boundary alignment to obtain semantically reconstructed features. The spatial topological representation and the semantic reconstruction features are coupled in two domains. Intermediate decoupling features are generated through cross-domain consistency optimization. The intermediate decoupling features are then hierarchically encoded according to topological complexity and semantic abstraction level to form a decoupling feature tensor with a hierarchical structure.

[0007] By constructing spatial topological consistency constraints based on geometric manifold embedding, the spatial structure in the original image data is topologically invariantly encoded, generating spatial topological representations that preserve local neighborhood relationships and global connectivity, including: Spatial sampling is performed on the original image data to extract a set of spatial samples containing pixel coordinates and pixel attributes. A manifold metric space reflecting the intrinsic geometric structure is constructed for the set of spatial samples. The set of spatial samples is then embedded in the manifold metric space through iterative optimization of tangent space projection and coordinate transformation to generate a manifold embedding result that maintains geometric invariance. Based on the manifold embedding results, a topology containing local neighborhood connections is constructed, and global connectivity repair is performed on the topology to obtain the repaired topology. The repaired topology is used as a constraint in the manifold metric space. The manifold embedding result is corrected for topological consistency through the backpropagation mechanism of topological constraints, generating a spatial topological representation that maintains both local neighborhood relations and global connectivity.

[0008] Based on the cross-modal semantic alignment of the creative graphic representation information and the decoupled feature tensor, an adaptive rendering prior distribution with style consistency guarantee is generated through joint optimization of adversarial style transfer constraints and content fidelity constraints, including: The textual and creative graphic representation information is projected onto the same semantic representation space as the decoupled feature tensor. A cross-modal semantic alignment function is constructed to calculate the semantic mapping relationship and generate aligned semantic features. Based on the aligned semantic features, a semantic guided transformation is performed on the decoupled feature tensor to obtain the initial rendering features. A joint optimization objective is constructed for the initial rendering features. The style transfer process is driven by an adversarial game mechanism, and content fidelity constraints are applied simultaneously to limit the degree of content deviation. The initial rendering features are iteratively optimized and adjusted to obtain the optimized rendering features. Based on the optimized rendering features and the aligned semantic features, an adaptive rendering prior distribution is generated through probability distribution modeling. The adaptive rendering prior distribution is used to perform style consistency verification on the optimized rendering features, and the verification result is fed back to the cross-modal semantic alignment function for dynamic adjustment.

[0009] Based on the intermediate rendering state representation, a dynamic computation graph is constructed. A sparse execution strategy for rendering operations is determined through path importance propagation and gradient sensitivity analysis. This generates a layered rendering control sequence containing selective skip instructions and precision adaptive instructions, including: Based on the intermediate rendering state representation, the rendering operation nodes and the dependencies between nodes are extracted, and a dynamic computation graph representing the rendering process is constructed. For each rendering operation node in the dynamic computation graph, the contribution weight of each node to the final rendering result is calculated along the dependency path through the path importance propagation mechanism, and the node importance distribution is generated. Gradient sensitivity analysis is performed on the dynamic computation graph to calculate the gradient response intensity of each rendering operation node to the rendering quality, and the gradient sensitivity distribution is obtained. Based on the node importance distribution and the gradient sensitivity distribution, the nodes that need to be executed and the redundant nodes that can be skipped are determined by comparing the node importance and the gradient sensitivity, and a sparse execution strategy is generated. Selective skip instructions are generated based on the redundant nodes identified in the sparse execution strategy, and precision adaptive instructions are generated based on the gradient sensitivity distribution of the nodes identified in the sparse execution strategy. The selective skip instruction and the precision adaptive instruction are organized hierarchically according to the dependency relationship of the dynamic computation graph to form a hierarchical rendering control sequence containing selective skip instructions and precision adaptive instructions.

[0010] Based on the node importance distribution and the gradient sensitivity distribution, the nodes that need to be executed and the redundant nodes that can be skipped are determined by comparing node importance and gradient sensitivity, and a sparse execution strategy is generated, including: A two-dimensional evaluation space is constructed for each node importance value in the node importance distribution and each node gradient sensitivity value in the gradient sensitivity distribution. The comprehensive execution necessity metric of each rendering operation node is calculated through the two-dimensional evaluation space to obtain the node execution necessity distribution. An adaptive decision boundary is constructed based on the node execution necessity distribution. The position of the adaptive decision boundary is determined by analyzing the joint statistical characteristics of the node importance distribution and the gradient sensitivity distribution. The adaptive decision boundary is then used to divide each rendering operation node into nodes that need to be executed and redundant nodes that can be skipped. For the nodes that need to be executed and the redundant nodes that are allowed to be skipped, a sparse execution strategy containing node identifiers and execution instructions is constructed.

[0011] According to the layered rendering control sequence, non-uniform sampling rendering is performed on the original image data to obtain the target rendered image and rendering efficiency feedback indicators, including: The layered rendering control sequence is analyzed, the selective skip instructions and precision adaptive instructions are extracted, and the rendering execution path is established according to the layered structure. Based on the selective skip instruction, the spatial region of the original image data is allocated with differential sampling density. By skipping sampling points in redundant regions and increasing sampling points in the sampling regions, a non-uniform sampling point distribution is generated. Based on the precision adaptive instruction, a rendering precision level is assigned to each sampling point in the non-uniform sampling point distribution to form a sampling rendering configuration that includes sampling position and precision level; According to the sampling rendering configuration, the original image data is rendered in regions, and the image is reconstructed by using a rendering calculation granularity that matches the precision level of different regions, so as to obtain the target rendered image. During the regional rendering process, the computational resource consumption and rendering time of each region are recorded, and the rendering efficiency feedback index is calculated based on the number of skipped regions of the selective skip instruction and the precision adjustment range of the precision adaptive instruction.

[0012] A second aspect of this invention provides an intelligent image rapid rendering and processing system for the cultural and creative industries, comprising: The data acquisition unit is used to acquire the original image data to be rendered and the textual and creative graphic representation information associated with the original image data; The feature decoupling unit is used to construct a dual-domain decoupling representation model that includes spatial topological consistency constraints and semantic coherence constraints. It performs frequency domain decomposition and semantic field reconstruction on the original image data to extract a decoupling feature tensor with a hierarchical structure. The prior generation unit is used to generate an adaptive rendering prior distribution with style consistency guarantee by cross-modal semantic alignment based on the cultural and creative graphic representation information and the decoupled feature tensor, through joint optimization of adversarial style transfer constraints and content fidelity constraints, and to map the adaptive rendering prior distribution to the rendering operation space to form an intermediate rendering state representation. The control sequence unit is used to construct a dynamic computation graph based on the intermediate rendering state representation, determine the sparse execution strategy of rendering operations through path importance propagation mechanism and gradient sensitivity analysis, and generate a layered rendering control sequence containing selective skip instructions and precision adaptive instructions. The rendering execution unit is used to perform non-uniform sampling rendering on the original image data according to the hierarchical rendering control sequence to obtain the target rendered image and the rendering efficiency feedback index, and to jointly update the constraint weights of the dual-domain decoupled representation model and the activation threshold of the path importance propagation mechanism using the rendering efficiency feedback index.

[0013] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] The beneficial effects of this application are as follows: This method significantly improves the efficiency and quality of image rendering for the cultural and creative industries. By constructing a dual-domain decoupled representation model that integrates spatial topological consistency constraints and semantic coherence constraints, it achieves accurate separation and reconstruction of the original image data in the frequency and semantic domains. This model can effectively extract decoupled feature tensors with hierarchical structures, providing a clear and semantically explicit basic feature representation for subsequent cross-modal alignment and style transfer, avoiding rendering distortion caused by feature entanglement.

[0016] By employing cross-modal semantic alignment of creative image representation information and decoupled feature tensors, and combining adversarial style transfer with content fidelity optimization, an adaptive rendering prior distribution with guaranteed style consistency can be generated. This process ensures that the rendered result accurately matches the artistic style of the creative image while maintaining a high degree of fidelity to the original image in terms of content structure, effectively solving the common problems of style spillover or content distortion in style transfer. Mapping this prior distribution to the intermediate rendering state representation formed in the rendering operation space provides quantifiable and actionable guidance for the rendering process.

[0017] A dynamic computation graph is constructed based on intermediate rendering state representations, and a sparse execution strategy for rendering operations is determined through path importance propagation and gradient sensitivity analysis, enabling the intelligent generation of layered rendering control sequences. These sequences include selective skip instructions and precision-adaptive instructions, allowing rendering computational resources to be non-uniformly and selectively allocated to critical and complex regions of the image. This significantly reduces redundant computation in non-critical or simple regions, achieving a substantial leap in rendering efficiency while ensuring no loss in final visual quality. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the intelligent image fast rendering processing method for the cultural and creative industries according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the sparse execution strategy based on node importance and gradient sensitivity in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0021] Figure 1 This is a flowchart illustrating the intelligent image fast rendering processing method for the cultural and creative industries according to an embodiment of the present invention. Figure 1 As shown, the method includes: Obtain the original image data to be rendered and the textual and graphic representation information associated with the original image data; A dual-domain decoupled representation model containing spatial topological consistency constraints and semantic coherence constraints is constructed. The original image data is then subjected to frequency domain decomposition and semantic field reconstruction to extract a decoupled feature tensor with a hierarchical structure. Based on the cross-modal semantic alignment of the creative graphic representation information and the decoupled feature tensor, an adaptive rendering prior distribution with style consistency guarantee is generated through joint optimization of adversarial style transfer constraints and content fidelity constraints, and the adaptive rendering prior distribution is mapped to the rendering operation space to form an intermediate rendering state representation. A dynamic computation graph is constructed based on the intermediate rendering state representation. The sparse execution strategy of rendering operations is determined through path importance propagation mechanism and gradient sensitivity analysis, and a layered rendering control sequence containing selective skip instructions and precision adaptive instructions is generated. Non-uniform sampling rendering is performed on the original image data according to the hierarchical rendering control sequence to obtain the target rendered image and rendering efficiency feedback index. The constraint weights of the dual-domain decoupled representation model and the activation threshold of the path importance propagation mechanism are jointly updated using the rendering efficiency feedback index.

[0022] In one optional implementation, a dual-domain decoupled representation model containing spatial topological consistency constraints and semantic coherence constraints is constructed. The original image data is then subjected to frequency domain decomposition and semantic field reconstruction to extract a decoupled feature tensor with a hierarchical structure, including: By constructing spatial topological consistency constraints based on geometric manifold embedding, the spatial structure in the original image data is encoded with topological invariance, generating a spatial topological representation that preserves local neighborhood relationships and global connectivity, and using the spatial topological representation to constrain spatial distortion in the frequency domain decomposition process. Based on the spatial topological representation, frequency domain orthogonal decomposition is performed on the original image data to map the original image data to the frequency domain representation space, thereby obtaining the frequency domain decomposition result; By constructing semantic coherence constraints based on semantic propagation graphs, cross-regional dependency models are performed on semantic elements in the frequency domain decomposition results to generate a semantic field distribution with semantic consistency guarantees, and the semantic field distribution is used to guide the semantic field reconstruction process. Based on the semantic field distribution, semantically driven feature recombination is performed on the frequency domain decomposition result. Semantic field reconstruction is completed through semantic affinity and semantic boundary alignment to obtain semantically reconstructed features. The spatial topological representation and the semantic reconstruction features are coupled in two domains. Intermediate decoupling features are generated through cross-domain consistency optimization. The intermediate decoupling features are then hierarchically encoded according to topological complexity and semantic abstraction level to form a decoupling feature tensor with a hierarchical structure.

[0023] In practical deployment, local neighborhood-preserving embeddings are obtained by performing Laplacian feature mapping on the original image data. Image pixels are treated as sampling points on a manifold, and the geodesic distance between each pixel in the high-dimensional feature space is calculated. Specifically, for a 256×256 resolution input image, a k-nearest neighbor graph is first constructed, where k is set to 8. Affinity weights between adjacent pixels are calculated using a Gaussian kernel function, with a weight decay coefficient typically set to 0.5. The first 32 non-zero eigenvectors of the graph Laplacian matrix are used to construct a spatial topological representation, which preserves the local geometry and global connectivity of the original image.

[0024] In the frequency domain decomposition stage, a two-dimensional discrete cosine transform is performed on the original image data to transform the RGB three channels to the frequency domain space. To prevent edge blurring caused by the loss of high-frequency components, an adaptive frequency mask is introduced. This mask dynamically adjusts the cutoff frequency based on the curvature information in the spatial topological representation. For regions with large curvature, up to the 48th frequency component is retained, while for flat regions, it is truncated to the 24th component. After sorting the frequency domain coefficients according to their energy percentage, the top N coefficients with a cumulative energy of 92% are selected to form the frequency domain decomposition result. In a typical scenario, N is approximately 512.

[0025] Semantic coherence constraints are implemented by constructing a semantic propagation graph, where superpixels are nodes, and the connection strength between nodes is determined by both color similarity and spatial adjacency. Approximately 200 superpixel regions are generated using the SLIC algorithm. The Euclidean distance between adjacent superpixels in the Lab color space is calculated, and nodes with a distance less than a threshold of 15 are connected. A random walk diffusion is performed on the propagation graph, with a diffusion step count of 5 steps, to obtain the semantic attribution probability distribution of each superpixel. This distribution is then upsampled to the original resolution to form a dense semantic field distribution, where each pixel carries a semantic vector of length K, where K is determined based on the cultural and creative type: K=16 for illustration and K=24 for product design.

[0026] In the semantic field reconstruction process, the frequency components of the frequency domain decomposition result are fused with the semantic field distribution at the channel level. Specifically, for the i-th frequency component, its spatial activation map is extracted and weighted and summed with the corresponding semantic vector. The weights are given by the semantic affinity matrix. This matrix is ​​obtained by calculating the cosine similarity between semantic vectors. Regions with a similarity greater than 0.7 undergo feature aggregation, while regions with a similarity less than 0.3 remain independent. Simultaneously, semantic boundary positions are detected, and smoothing constraints are applied within ±3 pixels on both sides of the boundary to avoid semantic breaks during reconstruction. After layer-by-layer processing by a 12-layer convolutional network, a 512-dimensional semantic reconstruction feature is output.

[0027] In the dual-domain coupling stage, the 32-dimensional spatial topological representation is extended to 512 dimensions through a fully connected layer and fused with the semantic reconstruction features through element-wise multiplication. To ensure cross-domain consistency, a contrastive loss function is introduced, requiring topologically similar pixel pairs to have a distance of less than the boundary value of 1.2 in the fused feature space, and topologically unrelated pixel pairs to have a distance of greater than 2.8. Stable intermediate decoupling features are obtained through 100 iterations of optimization.

[0028] The final hierarchical encoding adopts a three-level pyramid structure: the first level corresponds to low-frequency global layout, containing 64-dimensional features; the second level corresponds to mid-frequency texture details, containing 256-dimensional features; and the third level corresponds to high-frequency edge information, containing 192-dimensional features. Features at each level are arranged from simple to complex in terms of topological complexity, and the semantic abstraction level progresses from concrete objects to abstract styles, forming a hierarchical decoupled feature tensor with a total dimension of 512. This tensor enables fine-grained rendering control in subsequent style transfer.

[0029] In one optional implementation, by constructing spatial topological consistency constraints based on geometric manifold embedding, the spatial structure in the original image data is topologically invariantly encoded to generate a spatial topological representation that preserves local neighborhood relationships and global connectivity, including: Spatial sampling is performed on the original image data to extract a set of spatial samples containing pixel coordinates and pixel attributes. A manifold metric space reflecting the intrinsic geometric structure is constructed for the set of spatial samples. The set of spatial samples is then embedded in the manifold metric space through iterative optimization of tangent space projection and coordinate transformation to generate a manifold embedding result that maintains geometric invariance. Based on the manifold embedding results, a topology containing local neighborhood connections is constructed, and global connectivity repair is performed on the topology to obtain the repaired topology. The repaired topology is used as a constraint in the manifold metric space. The manifold embedding result is corrected for topological consistency through the backpropagation mechanism of topological constraints, generating a spatial topological representation that maintains both local neighborhood relations and global connectivity.

[0030] When processing the raw image data, spatial domain sampling is first performed on the image. The sampling process employs a multi-scale grid partitioning strategy, decomposing the image into spatial blocks of different resolution levels. For a W×H resolution raw image, key pixels are extracted at each scale level, recording their two-dimensional coordinates (x_i, y_i) and corresponding multi-dimensional attribute vectors. These attribute vectors contain information such as RGB color components, gradient magnitude, and texture feature responses. This method constructs a spatial sample set, where each sample point carries a composite description of both location information and visual features.

[0031] For the constructed spatial sample set, a manifold metric space reflecting its inherent geometric structure is established. This metric space is based on the geodesic distance definition between sample points, which comprehensively considers the differences in Euclidean space positions and the similarity in the feature space. By calculating the local distance metrics between sample point pairs, a global manifold structure representation is gradually formed. In the manifold metric space, a tangent space projection method is used for dimensionality reduction. Specifically, a local neighborhood is selected in the high-dimensional sample space, and principal component decomposition is performed on this neighborhood to obtain the tangent space basis vectors. The sample points are then projected onto this tangent space to obtain a low-dimensional coordinate representation. By iteratively optimizing the weighted sum of projection error and reconstruction error and adjusting the coordinate transformation parameters, an embedding representation in the low-dimensional manifold space is finally obtained, which maintains the geometric invariance of the original data.

[0032] A topological representation is constructed based on the manifold embedding results. A k-nearest neighbor (kNN) strategy is employed, searching for the k nearest neighbors of each sample point in the embedding space and establishing edge connections. A distance threshold mechanism is also introduced to filter out weak connections exceeding a set threshold, retaining strong connections that reflect real-world proximity. The initial topological structure contains isolated nodes or broken connected components, therefore a global connectivity repair operation is performed. The repair process involves detecting broken connected components and adding bridging edges between different components, selecting these edges based on minimizing the geodesic distance between cross-component node pairs. After repair, a topological structure with complete global connectivity is obtained.

[0033] The repaired topology is used as an explicit constraint, applied inversely to the embedding representation in the manifold metric space. A topological constraint backpropagation mechanism is established, defining a topological constraint loss function that measures the deviation between the current embedding representation and the desired topology. The loss function consists of two components: a local neighborhood preservation term penalizing changes in the distance between adjacent nodes in the embedding space, and a global connectivity preservation term penalizing the path length deviation between component nodes. The gradient of the constraint loss with respect to the embedding coordinates is calculated through backpropagation, iteratively updating the node coordinates in the manifold embedding result to ensure topological consistency while maintaining geometric invariance. After multiple rounds of iterative correction, a final spatial topological representation is generated. This representation preserves both the local neighborhood relationships of the original image and maintains the stability of the global connectivity structure in the manifold embedding space, providing a topologically invariant spatial structural foundation for subsequent semantic field reconstruction and style transfer operations.

[0034] In one optional implementation, based on the cross-modal semantic alignment of the creative graphic representation information and the decoupled feature tensor, an adaptive rendering prior distribution with style consistency guarantee is generated through joint optimization of adversarial style transfer constraints and content fidelity constraints, including: The textual and creative graphic representation information is projected onto the same semantic representation space as the decoupled feature tensor. A cross-modal semantic alignment function is constructed to calculate the semantic mapping relationship and generate aligned semantic features. Based on the aligned semantic features, a semantic guided transformation is performed on the decoupled feature tensor to obtain the initial rendering features. A joint optimization objective is constructed for the initial rendering features. The style transfer process is driven by an adversarial game mechanism, and content fidelity constraints are applied simultaneously to limit the degree of content deviation. The initial rendering features are iteratively optimized and adjusted to obtain the optimized rendering features. Based on the optimized rendering features and the aligned semantic features, an adaptive rendering prior distribution is generated through probability distribution modeling. The adaptive rendering prior distribution is used to perform style consistency verification on the optimized rendering features, and the verification result is fed back to the cross-modal semantic alignment function for dynamic adjustment.

[0035] The process transforms the textual and creative image representation information from its original form to a semantic representation space compatible with the decoupled feature tensor. The textual and creative image representation information is stored as a multi-dimensional vector, with dimensions ranging from 512 to 2048, represented using floating-point numbers, and its numerical range normalized to -1 to +1. The decoupled feature tensor originates from the image encoding process and contains decoupled representations of content and style features. The tensor shape is batch size multiplied by the number of channels multiplied by the height multiplied by the width; the number of channels is typically 256 or 512, and the spatial resolution is 16 x 16 to 64 x 64.

[0036] To achieve cross-modal semantic alignment, a mapping network was constructed, incorporating multi-layer fully connected transformations and non-linear activations. This network receives creative graphic representation information as input, and its output dimension matches the channel dimension of the decoupled feature tensor. The mapping network consists of three fully connected layers, with 1024 neurons in the middle layer, using rectified linear units as the activation function. The cross-modal semantic alignment function quantifies the semantic mapping relationship by calculating the cosine similarity between the mapped creative graphic features and the decoupled feature tensor in the semantic space, with a similarity threshold set between 0.7 and 0.9.

[0037] The aligned semantic features are aligned with the spatial dimensions of the decoupled feature tensor via a broadcast mechanism. The aligned semantic features are then copied and expanded spatially to have the same height and width as the decoupled feature tensor. Subsequently, a semantic guided transformation is performed through element-wise multiplication or addition. The semantic guided transformation employs adaptive fusion weights, which are dynamically adjusted based on the content complexity of each spatial location within the decoupled feature tensor. Content complexity is estimated by calculating the magnitude of local gradients. Regions with high complexity are assigned smaller fusion weights to preserve original content details, while regions with low complexity are assigned larger fusion weights to enhance stylistic expression, resulting in the initial rendered features.

[0038] The joint optimization objective constructed for the initial rendered features comprises two components: adversarial style transfer constraints and content fidelity constraints. The adversarial style transfer constraints are implemented through an adversarial game mechanism between the generator and discriminator. The generator receives the initial rendered features and outputs style-transferred feature representations, while the discriminator determines whether the input features conform to the target style distribution. The generator employs a residual connection structure containing four residual blocks, each containing two convolutional layers with a kernel size of 3x3. The discriminator uses a multi-scale architecture, performing discrimination at three different spatial resolutions. Each scale discriminator contains five convolutional layers, with the number of channels doubling from 64 to 512 layer by layer.

[0039] Content fidelity constraints quantify content deviation by calculating the distance between the initial rendered features and intermediate features during optimization within the content space. The content space is defined using the intermediate layer output of a pre-trained feature extraction network. The joint optimization objective combines adversarial style transfer constraints and content fidelity constraints through a weighted summation. The weight coefficients are dynamically adjusted based on the application scenario. In style-dominant scenarios, the adversarial constraint weights are set to 0.6 to 0.8, and the content fidelity constraint weights are set to 0.2 to 0.4. Iterative optimization uses gradient descent, with an initial learning rate between 0.0001 and 0.001, decaying with each iteration using a cosine annealing strategy. The number of iterations ranges from 100 to 500. During optimization, the numerical changes of the joint optimization objective are monitored. When the change in the objective value is less than 0.001 for ten consecutive iterations, convergence is determined, and the iteration terminates. The optimized rendered features are then output.

[0040] Probabilistic distribution modeling is performed based on optimized rendering features and aligned semantic features to generate an adaptive rendering prior distribution. The probabilistic distribution modeling employs a variational autoencoder architecture. The encoder maps the optimized rendering features to mean and variance parameters in a latent space, with the latent space dimension set between 128 and 512. Prior distribution samples are generated from the latent space using reparameterization techniques. The decoder reconstructs the distribution parameters of the rendering feature space from the sampled latent space representation, including the mean tensor and covariance tensor. The covariance tensor adopts the diagonal covariance assumption. The adaptive rendering prior distribution is represented as a probability density function, providing the probability distribution of the feature values ​​at each location in the rendering feature space.

[0041] Style consistency is verified using an adaptive rendering prior distribution on the optimized rendering features. The log-likelihood of the optimized rendering features under the prior distribution is calculated. A log-likelihood value higher than a set threshold (between -5 and -2) indicates that style consistency is met. Spatial locations with log-likelihood values ​​below the threshold are marked as style inconsistencies. The verification result includes a style consistency score and a spatial mask of the inconsistent regions. The style consistency score is calculated by averaging the log-likelihood values ​​across all spatial locations, and the spatial mask is a binary tensor. The verification result is fed back to the cross-modal semantic alignment function for dynamic adjustment. The mapping network parameters are adjusted using gradient backpropagation, with the style consistency score as the optimization objective. The learning rate is set between 0.00001 and 0.0001, and the number of adjustment iterations is between 10 and 50. The fusion weights are adjusted based on the spatial masks of the inconsistent regions. The fusion weights for inconsistent regions increase by 0.1 to 0.3, while the fusion weights for consistent regions remain unchanged or decrease by 0.05 to 0.1.

[0042] In one optional implementation, a dynamic computation graph is constructed based on the intermediate rendering state representation. A sparse execution strategy for rendering operations is determined through a path importance propagation mechanism and gradient sensitivity analysis. This generates a layered rendering control sequence containing selective skip instructions and precision adaptive instructions, including: Based on the intermediate rendering state representation, the rendering operation nodes and the dependencies between nodes are extracted, and a dynamic computation graph representing the rendering process is constructed. For each rendering operation node in the dynamic computation graph, the contribution weight of each node to the final rendering result is calculated along the dependency path through the path importance propagation mechanism, and the node importance distribution is generated. Gradient sensitivity analysis is performed on the dynamic computation graph to calculate the gradient response intensity of each rendering operation node to the rendering quality, and the gradient sensitivity distribution is obtained. Based on the node importance distribution and the gradient sensitivity distribution, the nodes that need to be executed and the redundant nodes that can be skipped are determined by comparing the node importance and the gradient sensitivity, and a sparse execution strategy is generated. Selective skip instructions are generated based on the redundant nodes identified in the sparse execution strategy, and precision adaptive instructions are generated based on the gradient sensitivity distribution of the nodes identified in the sparse execution strategy. The selective skip instruction and the precision adaptive instruction are organized hierarchically according to the dependency relationship of the dynamic computation graph to form a hierarchical rendering control sequence containing selective skip instructions and precision adaptive instructions.

[0043] After obtaining the intermediate rendering state representation, it needs to be transformed into an executable rendering control strategy. First, the information of each rendering operation node contained in the intermediate rendering state representation is analyzed. These nodes include basic rendering units such as texture mapping, lighting calculation, shadow generation, and color blending. By analyzing the input-output data dependencies of each operation node in the rendering state representation, a dynamic computation graph with a directed acyclic graph structure is established, where nodes represent specific rendering operations and edges represent the direction of data flow.

[0044] The path importance propagation mechanism employs a backpropagation approach, starting from the output node of the dynamic computation graph and propagating importance values ​​forward along dependent paths. Specifically, the importance weight of the output node is initialized to 1. For each intermediate node, its importance weight is equal to the weighted sum of the importance weights of all subsequent nodes, with the weight coefficient determined by the proportion of the node's output to the input of its successor nodes. After complete propagation, an importance distribution vector covering all nodes is obtained.

[0045] Gradient sensitivity analysis quantifies the sensitivity by calculating the partial derivatives of the rendering quality loss function with respect to the output of each node. Peak signal-to-noise ratio (PSNR) or structural similarity is selected as the rendering quality metric. Automatic differentiation is performed on the dynamic computation graph to obtain the gradient tensor of the quality metric with respect to the output features of each node. The L2 norm of the gradient tensor is calculated as the gradient sensitivity value for that node; nodes with high sensitivity have a significant impact on the final rendering quality.

[0046] The determination of the sparse execution strategy relies on a comprehensive assessment of importance and sensitivity. An importance threshold θ_I and a sensitivity threshold θ_G are set. When a node's importance weight is lower than θ_I and its gradient sensitivity is lower than θ_G, the node is marked as redundant. For redundant nodes, a selective skip instruction is generated, containing a node identifier and skip condition logic. For nodes that are retained for execution, a computational precision level is assigned based on their gradient sensitivity value. Nodes with high sensitivity use high-precision floating-point operations, while nodes with medium sensitivity use half-precision operations. The generated precision-adaptive instruction includes a node identifier, precision level parameters, and a numerical truncation strategy.

[0047] The construction of the layered rendering control sequence must ensure that the order of instruction execution conforms to the topological dependencies of the dynamic computation graph. A topological sort is performed on the dynamic computation graph, dividing nodes into multiple execution levels. Nodes within the same level have no mutual dependencies and can be executed in parallel. Selective skip instructions and precision adaptive instructions are organized according to the hierarchical order, with each level forming an instruction package. Within the instruction package, skip instructions are executed first to filter redundant operations, and then precision adaptive instructions are executed to configure and reserve the computational resources of the nodes. The final generated layered rendering control sequence is stored sequentially by level, with each level containing a set of control instructions for all nodes in that level, ensuring that rendering execution satisfies dependency constraints while accelerating computation.

[0048] In one optional implementation, based on the node importance distribution and the gradient sensitivity distribution, the nodes to be executed and the redundant nodes that can be skipped are determined by comparing node importance and gradient sensitivity, and a sparse execution strategy is generated, including: A two-dimensional evaluation space is constructed for each node importance value in the node importance distribution and each node gradient sensitivity value in the gradient sensitivity distribution. The comprehensive execution necessity metric of each rendering operation node is calculated through the two-dimensional evaluation space to obtain the node execution necessity distribution. An adaptive decision boundary is constructed based on the node execution necessity distribution. The position of the adaptive decision boundary is determined by analyzing the joint statistical characteristics of the node importance distribution and the gradient sensitivity distribution. The adaptive decision boundary is then used to divide each rendering operation node into nodes that need to be executed and redundant nodes that can be skipped. For the nodes that need to be executed and the redundant nodes that are allowed to be skipped, a sparse execution strategy containing node identifiers and execution instructions is constructed.

[0049] A two-dimensional evaluation space is constructed for the node importance values ​​in the node importance distribution and the gradient sensitivity values ​​in the gradient sensitivity distribution. The node importance distribution is stored as a one-dimensional array with a length equal to the total number of rendering nodes. Each element corresponds to the importance value of one node, and the values ​​are normalized to the range of 0 to 1. The gradient sensitivity distribution is also stored as a one-dimensional array with the same length as the node importance distribution. Each element corresponds to the gradient sensitivity value of one node, and the values ​​are normalized to the range of 0 to 1.

[0050] The two-dimensional evaluation space is constructed using node importance and gradient sensitivity values ​​as two-dimensional coordinates. The horizontal axis represents node importance, and the vertical axis represents gradient sensitivity. Each rendering operation node corresponds to one coordinate point in this space. When calculating the comprehensive execution necessity metric for each rendering operation node, the node's importance and gradient sensitivity values ​​are weighted and combined. The weighting coefficients are determined based on the characteristics of the rendering task. For tasks that prioritize visual quality, the weight of the importance value is set to 0.6 to 0.7, and the weight of the gradient sensitivity value is set to 0.3 to 0.4. The comprehensive execution necessity metric is obtained by multiplying the node's importance value by its corresponding weight and then adding the gradient sensitivity value multiplied by its corresponding weight. The result ranges from 0 to 1. The node execution necessity distribution is stored in a one-dimensional array, with the array length equal to the total number of rendering operation nodes.

[0051] When constructing an adaptive decision boundary based on the node execution necessity distribution, the location of the adaptive decision boundary is determined by analyzing the joint statistical properties of the node importance distribution and the gradient sensitivity distribution. The joint statistical properties include the joint probability distribution of node importance values ​​and gradient sensitivity values, the correlation coefficient, and the mean and variance of the marginal distributions. The joint probability distribution is obtained by counting the number of nodes falling into each unit in the two-dimensional evaluation space and dividing by the total number of nodes. The correlation coefficient is obtained by calculating the Pearson correlation coefficient between node importance values ​​and gradient sensitivity values, with a value ranging from -1 to +1. The adaptive decision boundary is represented by a curve in the two-dimensional evaluation space, dividing the space into two regions. Nodes located above or to the right of the boundary are classified as nodes that need to be executed, while nodes located below or to the left of the boundary are classified as redundant nodes that can be skipped.

[0052] The location of the adaptive decision boundary is determined by finding the boundary between the high-density and low-density regions of the joint probability distribution. Specifically, this is achieved by calculating the cumulative distribution function of the joint probability distribution and selecting contour lines whose cumulative probability values ​​equal a preset threshold as the decision boundary. The preset threshold ranges from 0.3 to 0.7. The mathematical representation of the decision boundary is parametric. Points on the boundary are determined by a series of parameter values, which are uniformly sampled between 0 and 1, with the number of sampling points set between 50 and 200. When dividing each rendering operation node using the adaptive decision boundary, the coordinates of all nodes in the two-dimensional evaluation space are traversed to determine the relative positional relationship between the point and the decision boundary. The determination method is to calculate the directed distance from the point to the decision boundary. A positive directed distance indicates that the point is above or to the right of the boundary, and a negative directed distance indicates that the point is below or to the left of the boundary. The division result is stored in the form of a 2-value array, with the array length equal to the total number of rendering operation nodes. Each element has a value of 1 to indicate a node that needs to be executed, and a value of 0 to indicate a redundant node that can be skipped.

[0053] A sparse execution strategy, comprising node identifiers and execution instructions, is constructed for nodes that need to be executed and redundant nodes that can be skipped. Node identifiers are integers ranging from 0 to the total number of nodes minus 1. Execution instructions are represented by Boolean values: true indicates the node needs to be executed, and false indicates the node can be skipped. The sparse execution strategy is stored as a dictionary, with node identifiers as keys and execution instructions as values. The sparse execution strategy is serialized using a compact binary format. For nodes that need to be executed, only their node identifiers are stored; for redundant nodes that can be skipped, no information is stored. The application of the sparse execution strategy involves querying the execution instruction of the corresponding node before a rendering operation is executed. If the instruction is true, the operation is executed normally; if the instruction is false, the operation is skipped and a cached result or an approximation is used instead.

[0054] In one optional implementation, non-uniform sampling rendering is performed on the original image data according to the layered rendering control sequence to obtain the target rendered image and rendering efficiency feedback indicators, including: The layered rendering control sequence is analyzed, the selective skip instructions and precision adaptive instructions are extracted, and the rendering execution path is established according to the layered structure. Based on the selective skip instruction, the spatial region of the original image data is allocated with differential sampling density. By skipping sampling points in redundant regions and increasing sampling points in the sampling regions, a non-uniform sampling point distribution is generated. Based on the precision adaptive instruction, a rendering precision level is assigned to each sampling point in the non-uniform sampling point distribution to form a sampling rendering configuration that includes sampling position and precision level; According to the sampling rendering configuration, the original image data is rendered in regions, and the image is reconstructed by using a rendering calculation granularity that matches the precision level of different regions, so as to obtain the target rendered image. During the regional rendering process, the computational resource consumption and rendering time of each region are recorded, and the rendering efficiency feedback index is calculated based on the number of skipped regions of the selective skip instruction and the precision adjustment range of the precision adaptive instruction.

[0055] like Figure 2 As shown, the method includes: During the non-uniform sampling rendering stage, the layered rendering control sequence is first structurally parsed. This sequence is organized in a nested data structure, where selective skip instructions are encoded as binary marker vectors. A marker value of 1 indicates that rendering calculations are required for that spatial region, while a marker value of 0 indicates that detailed calculations for that region can be skipped. Precision adaptive instructions represent the precision level as floating-point numbers, typically ranging from 0.1 to 1.0, with larger values ​​indicating higher required computational precision. Based on the instruction priorities of different levels in the layered structure, a rendering execution path is constructed from coarse-grained to fine-grained, ensuring that skip decisions from higher layers are passed to lower layers.

[0056] In the differential sampling density allocation process, the original image data is divided into several spatial grid units. For redundant regions marked as 0 (selectively skipped), the sampling density is set to 0.2 times the baseline density, retaining only sparse sampling points for boundary continuity maintenance. For critical sampling regions marked as 1, the sampling density is increased to 1.5 to 2.0 times the baseline density, with the specific multiple determined by the importance score of the region in the intermediate rendering state representation. Sampling point positions are generated using a quasi-random sequence, ensuring spatial coverage uniformity while avoiding aliasing artifacts caused by regular grids. The generated non-uniform sampling point distribution is stored in the form of a coordinate list, with each sampling point containing two-dimensional spatial coordinates and an identifier of its region.

[0057] When assigning precision levels, the floating-point values ​​of the precision adaptive command are discretized into three specific calculation modes: low precision level corresponds to fast approximation calculation, medium precision level uses the standard lighting calculation process, and high precision level enables multiple reflections and global illumination simulation. For each sampling point in a non-uniform sampling point distribution, the calculation mode is determined based on the precision command value of its region. For sampling points with a precision value below 0.4, a low precision level is assigned and single ray tracing is used; for sampling points with a precision value between 0.4 and 0.7, a medium precision level is assigned and standard rendering including shadow calculation is performed; for sampling points with a precision value above 0.7, a high precision level is assigned and path tracing algorithms are enabled. The sampling rendering configuration is organized in the form of a data structure array, with each element associated with the sampling point coordinates, precision level identifier, and corresponding set of rendering parameters.

[0058] During region-based rendering, the original image data is divided into blocks according to the sampling rendering configuration. Low-precision regions use a 4-pixel × 4-pixel calculation granularity, and pixel filling is completed through bilinear interpolation; medium-precision regions use a 2-pixel × 2-pixel calculation granularity, and detail enhancement is achieved by combining normal maps; high-precision regions use a single-pixel calculation granularity and execute a complete physically-based rendering process. The rendering results of each region are combined using a weighted fusion method, and a 3-pixel wide gradient transition band is applied at the region boundaries to eliminate block artifacts. The image reconstruction process adopts a layer-by-layer accumulation strategy, starting from the coarsest level of rendering results and gradually stacking detail layers until the final target rendered image is generated.

[0059] The calculation of rendering efficiency feedback metrics involves multi-dimensional statistics. Computational resource consumption is obtained by monitoring GPU utilization and memory allocation, and rendering time is recorded in milliseconds for the processing time of each region. The ratio of the number of skipped regions to the total number of regions constitutes the skip rate metric, and the precision adjustment range is calculated as the average of the difference between the actual precision level and the highest precision level at each sampling point. Combining these values, a weighted summation formula is used to calculate a comprehensive efficiency metric reflecting the rendering acceleration effect. The higher the metric value, the more significant the computational savings achieved while maintaining image quality.

[0060] A second aspect of this invention provides an intelligent image rapid rendering and processing system for the cultural and creative industries, comprising: The data acquisition unit is used to acquire the original image data to be rendered and the textual and creative graphic representation information associated with the original image data; The feature decoupling unit is used to construct a dual-domain decoupling representation model that includes spatial topological consistency constraints and semantic coherence constraints. It performs frequency domain decomposition and semantic field reconstruction on the original image data to extract a decoupling feature tensor with a hierarchical structure. The prior generation unit is used to generate an adaptive rendering prior distribution with style consistency guarantee by cross-modal semantic alignment based on the cultural and creative graphic representation information and the decoupled feature tensor, through joint optimization of adversarial style transfer constraints and content fidelity constraints, and to map the adaptive rendering prior distribution to the rendering operation space to form an intermediate rendering state representation. The control sequence unit is used to construct a dynamic computation graph based on the intermediate rendering state representation, determine the sparse execution strategy of rendering operations through path importance propagation mechanism and gradient sensitivity analysis, and generate a layered rendering control sequence containing selective skip instructions and precision adaptive instructions. The rendering execution unit is used to perform non-uniform sampling rendering on the original image data according to the hierarchical rendering control sequence to obtain the target rendered image and the rendering efficiency feedback index, and to jointly update the constraint weights of the dual-domain decoupled representation model and the activation threshold of the path importance propagation mechanism using the rendering efficiency feedback index.

[0061] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0062] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0063] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for rapid intelligent image rendering processing for the cultural and creative industries, characterized in that, include: Obtain the original image data to be rendered and the textual and graphic representation information associated with the original image data; A dual-domain decoupled representation model containing spatial topological consistency constraints and semantic coherence constraints is constructed. The original image data is then subjected to frequency domain decomposition and semantic field reconstruction to extract a decoupled feature tensor with a hierarchical structure. Based on the cross-modal semantic alignment of the creative graphic representation information and the decoupled feature tensor, an adaptive rendering prior distribution with style consistency guarantee is generated through joint optimization of adversarial style transfer constraints and content fidelity constraints, and the adaptive rendering prior distribution is mapped to the rendering operation space to form an intermediate rendering state representation. A dynamic computation graph is constructed based on the intermediate rendering state representation. The sparse execution strategy of rendering operations is determined through path importance propagation mechanism and gradient sensitivity analysis, and a layered rendering control sequence containing selective skip instructions and precision adaptive instructions is generated. Non-uniform sampling rendering is performed on the original image data according to the hierarchical rendering control sequence to obtain the target rendered image and rendering efficiency feedback index. The constraint weights of the dual-domain decoupled representation model and the activation threshold of the path importance propagation mechanism are jointly updated using the rendering efficiency feedback index.

2. The method according to claim 1, characterized in that, A dual-domain decoupled representation model containing spatial topological consistency constraints and semantic coherence constraints is constructed. The original image data is then subjected to frequency domain decomposition and semantic field reconstruction to extract a decoupled feature tensor with a hierarchical structure, including: By constructing spatial topological consistency constraints based on geometric manifold embedding, the spatial structure in the original image data is encoded with topological invariance, generating a spatial topological representation that preserves local neighborhood relationships and global connectivity, and using the spatial topological representation to constrain spatial distortion in the frequency domain decomposition process. Based on the spatial topological representation, frequency domain orthogonal decomposition is performed on the original image data to map the original image data to the frequency domain representation space, thereby obtaining the frequency domain decomposition result; By constructing semantic coherence constraints based on semantic propagation graphs, cross-regional dependency models are performed on semantic elements in the frequency domain decomposition results to generate a semantic field distribution with semantic consistency guarantees, and the semantic field distribution is used to guide the semantic field reconstruction process. Based on the semantic field distribution, semantically driven feature recombination is performed on the frequency domain decomposition result. Semantic field reconstruction is completed through semantic affinity and semantic boundary alignment to obtain semantically reconstructed features. The spatial topological representation and the semantic reconstruction features are coupled in two domains. Intermediate decoupling features are generated through cross-domain consistency optimization. The intermediate decoupling features are then hierarchically encoded according to topological complexity and semantic abstraction level to form a decoupling feature tensor with a hierarchical structure.

3. The method according to claim 2, characterized in that, By constructing spatial topological consistency constraints based on geometric manifold embedding, the spatial structure in the original image data is topologically invariantly encoded, generating spatial topological representations that preserve local neighborhood relationships and global connectivity, including: Spatial sampling is performed on the original image data to extract a set of spatial samples containing pixel coordinates and pixel attributes. A manifold metric space reflecting the intrinsic geometric structure is constructed for the set of spatial samples. The set of spatial samples is then embedded in the manifold metric space through iterative optimization of tangent space projection and coordinate transformation to generate a manifold embedding result that maintains geometric invariance. Based on the manifold embedding results, a topology containing local neighborhood connections is constructed, and global connectivity repair is performed on the topology to obtain the repaired topology. The repaired topology is used as a constraint in the manifold metric space. The manifold embedding result is corrected for topological consistency through the backpropagation mechanism of topological constraints, generating a spatial topological representation that maintains both local neighborhood relations and global connectivity.

4. The method according to claim 1, characterized in that, Based on the cross-modal semantic alignment of the creative graphic representation information and the decoupled feature tensor, an adaptive rendering prior distribution with style consistency guarantee is generated through joint optimization of adversarial style transfer constraints and content fidelity constraints, including: The textual and creative graphic representation information is projected onto the same semantic representation space as the decoupled feature tensor. A cross-modal semantic alignment function is constructed to calculate the semantic mapping relationship and generate aligned semantic features. Based on the aligned semantic features, a semantic guided transformation is performed on the decoupled feature tensor to obtain the initial rendering features. A joint optimization objective is constructed for the initial rendering features. The style transfer process is driven by an adversarial game mechanism, and content fidelity constraints are applied simultaneously to limit the degree of content deviation. The initial rendering features are iteratively optimized and adjusted to obtain the optimized rendering features. Based on the optimized rendering features and the aligned semantic features, an adaptive rendering prior distribution is generated through probability distribution modeling. The adaptive rendering prior distribution is used to perform style consistency verification on the optimized rendering features, and the verification result is fed back to the cross-modal semantic alignment function for dynamic adjustment.

5. The method according to claim 1, characterized in that, Based on the intermediate rendering state representation, a dynamic computation graph is constructed. A sparse execution strategy for rendering operations is determined through path importance propagation and gradient sensitivity analysis. This generates a layered rendering control sequence containing selective skip instructions and precision adaptive instructions, including: Based on the intermediate rendering state representation, the rendering operation nodes and the dependencies between nodes are extracted, and a dynamic computation graph representing the rendering process is constructed. For each rendering operation node in the dynamic computation graph, the contribution weight of each node to the final rendering result is calculated along the dependency path through the path importance propagation mechanism, and the node importance distribution is generated. Gradient sensitivity analysis is performed on the dynamic computation graph to calculate the gradient response intensity of each rendering operation node to the rendering quality, and the gradient sensitivity distribution is obtained. Based on the node importance distribution and the gradient sensitivity distribution, the nodes that need to be executed and the redundant nodes that can be skipped are determined by comparing the node importance and the gradient sensitivity, and a sparse execution strategy is generated. Selective skip instructions are generated based on the redundant nodes identified in the sparse execution strategy, and precision adaptive instructions are generated based on the gradient sensitivity distribution of the nodes identified in the sparse execution strategy. The selective skip instruction and the precision adaptive instruction are organized hierarchically according to the dependency relationship of the dynamic computation graph to form a hierarchical rendering control sequence containing selective skip instructions and precision adaptive instructions.

6. The method according to claim 5, characterized in that, Based on the node importance distribution and the gradient sensitivity distribution, the nodes that need to be executed and the redundant nodes that can be skipped are determined by comparing node importance and gradient sensitivity, and a sparse execution strategy is generated, including: A two-dimensional evaluation space is constructed for each node importance value in the node importance distribution and each node gradient sensitivity value in the gradient sensitivity distribution. The comprehensive execution necessity metric of each rendering operation node is calculated through the two-dimensional evaluation space to obtain the node execution necessity distribution. An adaptive decision boundary is constructed based on the node execution necessity distribution. The position of the adaptive decision boundary is determined by analyzing the joint statistical characteristics of the node importance distribution and the gradient sensitivity distribution. The adaptive decision boundary is then used to divide each rendering operation node into nodes that need to be executed and redundant nodes that can be skipped. For the nodes that need to be executed and the redundant nodes that are allowed to be skipped, a sparse execution strategy containing node identifiers and execution instructions is constructed.

7. The method according to claim 1, characterized in that, According to the layered rendering control sequence, non-uniform sampling rendering is performed on the original image data to obtain the target rendered image and rendering efficiency feedback indicators, including: The layered rendering control sequence is analyzed, the selective skip instructions and precision adaptive instructions are extracted, and the rendering execution path is established according to the layered structure. Based on the selective skip instruction, the spatial region of the original image data is allocated with differential sampling density. By skipping sampling points in redundant regions and increasing sampling points in the sampling regions, a non-uniform sampling point distribution is generated. Based on the precision adaptive instruction, a rendering precision level is assigned to each sampling point in the non-uniform sampling point distribution to form a sampling rendering configuration that includes sampling position and precision level; According to the sampling rendering configuration, the original image data is rendered in regions, and the image is reconstructed by using a rendering calculation granularity that matches the precision level of different regions, so as to obtain the target rendered image. During the regional rendering process, the computational resource consumption and rendering time of each region are recorded, and the rendering efficiency feedback index is calculated based on the number of skipped regions of the selective skip instruction and the precision adjustment range of the precision adaptive instruction.

8. A smart image rapid rendering and processing system for the cultural and creative industries, used to implement the method of any one of claims 1-7, characterized in that, include: The data acquisition unit is used to acquire the original image data to be rendered and the textual and creative graphic representation information associated with the original image data; The feature decoupling unit is used to construct a dual-domain decoupling representation model that includes spatial topological consistency constraints and semantic coherence constraints. It performs frequency domain decomposition and semantic field reconstruction on the original image data to extract a decoupling feature tensor with a hierarchical structure. The prior generation unit is used to generate an adaptive rendering prior distribution with style consistency guarantee by cross-modal semantic alignment based on the cultural and creative graphic representation information and the decoupled feature tensor, through joint optimization of adversarial style transfer constraints and content fidelity constraints, and to map the adaptive rendering prior distribution to the rendering operation space to form an intermediate rendering state representation. The control sequence unit is used to construct a dynamic computation graph based on the intermediate rendering state representation, determine the sparse execution strategy of rendering operations through path importance propagation mechanism and gradient sensitivity analysis, and generate a layered rendering control sequence containing selective skip instructions and precision adaptive instructions. The rendering execution unit is used to perform non-uniform sampling rendering on the original image data according to the hierarchical rendering control sequence to obtain the target rendered image and the rendering efficiency feedback index, and to jointly update the constraint weights of the dual-domain decoupled representation model and the activation threshold of the path importance propagation mechanism using the rendering efficiency feedback index.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images