An artificial intelligence-based 3D model real-time rendering method and system

Abstract

This invention discloses a real-time rendering method and system for 3D models based on artificial intelligence, comprising the following steps: Step 1: acquiring real-time rendering source data of the target 3D model; Step 2: generating a rendering state token sequence through model region partitioning and state token mapping; Step 3: encoding rendering state features of the rendering state token sequence using an improved MambaVision network; Step 4: predicting the rendering quality and rendering time of different candidate rendering configurations; Step 5: executing real-time rendering strategy decisions using an improved NSGA-II algorithm; Step 6: performing layered real-time rendering of the target 3D model; Step 7: performing pixel-level inter-frame compensation using a depth-constrained optical flow compensation mechanism to obtain the target real-time rendered image frames. This invention improves the stability of real-time rendering of 3D models through an improved MambaVision network and a depth-constrained optical flow compensation mechanism.

Description

Technical Field

[0001] This invention relates to the field of computer graphics processing and model rendering technology, and in particular to a method and system for real-time rendering of 3D models based on artificial intelligence. Background Technology

[0002] With the increasing demand for 3D visualization, digital twins, virtual reality, online 3D display, and industrial simulation applications, technologies for improving real-time rendering quality, controlling rendering time, and stabilizing inter-frame images for 3D models have received widespread attention. Existing real-time rendering methods for 3D models mainly rely on fixed rendering pipelines, preset detail level switching, texture compression, view frustum clipping, and empirical rule-based resource scheduling for image generation. However, these methods commonly suffer from the following problems in practical applications: The geometry, texture, lighting conditions, and camera viewpoint of the target 3D model continuously change during real-time interaction. Traditional methods typically adjust the mesh detail level according to a fixed distance threshold or static rendering configuration, making it difficult to simultaneously represent the complexity of the model region, viewpoint visibility, and device load. This results in excessive reduction of details in key visible areas or the continued consumption of rendering resources in invisible areas. There are multiple conflicting objectives among different candidate rendering configurations, including rendering quality, rendering time, and rendering frame rate stability. Existing configuration switching methods based on a single frame rate threshold or human experience struggle to achieve a target rendering configuration that balances image quality and real-time performance in complex scenes, easily leading to frame rate fluctuations, abnormal memory usage, and increased rendering latency. After reducing the mesh detail level, texture sampling resolution, shadow calculation accuracy, and reflection calculation accuracy, the initial rendered image frame is prone to texture blurring, edge flickering, shadow jumps, and cross-frame jitter. Traditional post-processing methods do not adequately utilize depth changes, motion displacement, and historical frame alignment relationships, making it difficult to guarantee pixel-level continuity and visual stability of real-time rendered image frames.

[0003] Therefore, how to provide a real-time rendering method and system for 3D models based on artificial intelligence is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a real-time rendering method and system for 3D models based on artificial intelligence. This invention fully utilizes an improved MambaVision network, an improved NSGA-II algorithm, and a depth-constrained optical flow compensation mechanism. It details the process of rendering state feature encoding based on rendering state token sequences, prediction of candidate rendering configuration quality and time consumption, real-time rendering strategy decision-making, layered real-time rendering, and pixel-level inter-frame compensation. This results in advantages such as high accuracy in rendering configuration decisions, strong stability of real-time rendering frame rates, good preservation of image details, and high continuity across frames.

[0005] A real-time rendering method for 3D models based on artificial intelligence according to an embodiment of the present invention includes the following steps: Step 1: During real-time rendering, take the current rendering frame as the processing object and obtain the real-time rendering source data of the target 3D model; Step 2: Based on real-time rendering source data, generate the rendering state token sequence for the current rendering frame through model region division and state token mapping; Step 3: Encode the rendering state token sequence using an improved MambaVision network to obtain the rendering state feature vector of the current rendering frame; the improved MambaVision network introduces a dual-axis state gating scan guided by visibility coupling priority and a feature-level inter-frame residual compensation mechanism. Step 4: Based on the rendering state feature vector, predict the rendering quality and rendering time of different candidate rendering configurations, and obtain the rendering quality score sequence and the rendering time score sequence. Step 5: Based on the rendering quality score sequence and the rendering time score sequence, a real-time rendering strategy decision is made using the improved NSGA-II algorithm to obtain the target rendering configuration; the improved NSGA-II algorithm introduces group constraint crossover and load-aware neighborhood variation; Step 6: Perform layered real-time rendering on the target 3D model according to the target rendering configuration to obtain the initial rendered image frames; Step 7: Perform pixel-level inter-frame compensation on the initial rendered image frame using a depth-constrained optical flow compensation mechanism to obtain the target real-time rendered image frame.

[0006] Optionally, the real-time rendering source data includes the geometric structure parameters, material texture parameters, camera viewpoint parameters, lighting parameters, interactive motion parameters, device load parameters, previous rendering frame state parameters, rendering configuration parameters, and rendering cache data of the target 3D model.

[0007] Optionally, step two specifically includes: Based on geometric structural parameters, the target 3D model is divided into multiple model regions, and the region number of each model region is generated according to the hierarchical order in the model hierarchy and the connection relationship of the facets in the same level. Based on geometric structure parameters, generate a geometric complexity feature vector for each model region in the current rendering frame; based on camera view parameters, generate a view visibility feature vector for each model region in the current rendering frame. Based on material texture parameters and lighting parameters, generate material lighting response vectors; based on camera view parameters, interaction motion parameters and previous rendering frame state parameters, generate inter-frame motion state vectors; based on device load parameters, generate device rendering load vectors. Perform linear mapping and dimension alignment on the geometric complexity feature vector and view visibility feature vector corresponding to each model region to obtain the geometric state token and view state token of each model region, and concatenate them into a region state token group. Linear mapping and dimension alignment are performed on the material lighting response vector, inter-frame motion state vector, and device rendering load vector respectively to obtain the material lighting state token, inter-frame motion state token, and device load state token. Arrange the area status token groups in ascending order of area number to form area status token segments; By concatenating the material lighting state token, inter-frame motion state token, and device load state token sequentially into the region state token segment, the rendering state token sequence of the current rendering frame is obtained.

[0008] Optionally, step three specifically includes: Extract the geometric state tokens and view state tokens of each model region from the rendering state token sequence; The weight mapping values ​​of the geometric state token and the view state token are calculated by linear mapping and Sigmoid activation function, respectively, to obtain the geometric complexity scalar and the view visibility scalar. Multiply the geometric complexity scalar and the view visibility scalar, and normalize them using the Softmax function to obtain the visibility coupling priority value for each model region. Couple and map the geometric state tokens and view state tokens of each model region to generate a geometric-view coupling token; The geometric view coupling tokens of each model region are sorted according to the visibility coupling priority value to obtain a dynamic scan token sequence; Perform a dual-axis state-gated scan guided by visibility coupling priority value on the dynamic scan token sequence to obtain the forward scan hidden state and reverse scan hidden state of each model region; The forward scan hidden state, reverse scan hidden state, and geometric viewpoint coupling token of the same model region are fused to obtain the visibility coupling scan feature; The material lighting state token and visibility coupled scanning features are concatenated, and the texture sensitive component, shadow sensitive component, specular sensitive component and reflection sensitive component of each model region are calculated by four sensitive component mapping units, and the material lighting sensitive vector is formed. Linear mapping is performed on the material's illumination sensitivity vector to obtain the material's illumination recalibration weights; By recalibrating the visibility coupling scan features using material lighting recalibration weights, the material lighting sensitive region features of each model region are obtained. The device load status token and material illumination sensitive area features are concatenated, and the device load gating coefficient is generated by linear mapping and Sigmoid activation function. Based on the device load gating coefficient, the load constraint adjustment is performed on the material lighting sensitive area features to obtain the load constraint rendering state features of each model area. The load-constrained rendering state features are compensated for inter-frame continuity through a feature-level inter-frame residual compensation mechanism to obtain inter-frame compensated rendering state features. The inter-frame compensation rendering state features of each model region are used to calculate the region fusion weight using the Softmax function; based on the region fusion weight, the inter-frame compensation rendering state features of each model region are weighted and summed to obtain the region-level rendering fusion features. The region-level rendering fusion features, material lighting state tokens, inter-frame motion state tokens, and device load state tokens are concatenated to obtain cross-dimensional rendering fusion features. The cross-dimensional rendering fusion features are then linearly mapped to generate the rendering state feature vector of the current rendering frame.

[0009] Optionally, the visibility coupling priority-guided dual-axis state-gated scan includes a forward state-gated scan and a reverse state-gated scan. The forward state-gated scan is performed in the direction of decreasing visibility coupling priority value, specifically as follows: The forward scan hidden state of the previous model region and the geometric view coupling token of the current model region are coupled by state transition mapping and tanh activation to generate a forward candidate hidden state; Using the visibility coupling priority value of the current model region as the gating coefficient, the forward scan hidden state of the previous model region and the forward candidate hidden state of the current model region are gatingly fused to obtain the forward scan hidden state of the current model region. The reverse state-gated scan is performed in the direction of increasing visibility coupling priority value, specifically as follows: The reverse scan hidden state of the previous model region and the geometric view coupling token of the current model region are coupled with state transition mapping and tanh activation to generate reverse candidate hidden states; Using the visibility coupling priority value of the current model region as the gating coefficient, the inverse scan hidden state of the previous model region and the inverse candidate hidden state of the current model region are gated and fused to obtain the inverse scan hidden state of the current model region. The feature-level inter-frame residual compensation mechanism is specifically as follows: Inter-frame motion state tokens are linearly mapped to the HardSigmoid activation function to generate inter-frame compensation weights. Linear mapping is performed on the inter-frame motion state tokens to obtain inter-frame motion compensation features; The inter-frame motion compensation features are multiplied element-wise with the inter-frame compensation weights to obtain the inter-frame residual compensation features. The inter-frame residual compensation features are then residually concatenated with the load constraint rendering state features to obtain the inter-frame compensation rendering state features for each model region.

[0010] Optionally, step four specifically includes: Based on the rendering configuration parameters, a set of candidate rendering configurations is determined by discrete combination and load constraint filtering. The candidate rendering configuration set includes several candidate rendering configurations, each of which includes mesh detail level, texture sampling resolution, shadow calculation accuracy, reflection calculation accuracy, anti-aliasing level, and compensation intensity parameters; Perform a linear mapping on the current candidate rendering configuration to obtain the configuration feature vector; The rendering state feature vector and the configuration feature vector are concatenated, and then predicted and mapped using the quality score prediction head and the time consumption score prediction head respectively, to obtain the rendering quality score and rendering time consumption score of the current candidate rendering configuration. Both the quality score prediction head and the time consumption score prediction head are composed of a linear mapping, a GELU activation function, and a linear mapping. Traverse all candidate rendering configurations in the candidate rendering configuration set to obtain the rendering quality score sequence and the rendering time score sequence.

[0011] Optionally, the improved NSGA-II algorithm specifically includes: Use the candidate rendering configuration set as the initial population; Based on device load parameters, read the current rendering frame rate and target rendering frame rate of the target 3D model; Based on the rendering quality score sequence, rendering time score sequence, current rendering frame rate and target rendering frame rate, calculate the multi-objective fitness value of each candidate rendering configuration, and then weight and fuse them to obtain a comprehensive decision value. Based on multi-objective fitness values, a fast non-dominated sort is performed on candidate rendering configurations to obtain several non-dominated layers, and a non-dominated level is assigned to each candidate rendering configuration. Calculate the crowding distance for candidate rendering configurations within the same non-dominated layer; Based on the non-dominance level and crowding distance, multiple parent candidate rendering configurations are selected from the current population to form a parent population. Perform grouping constraint crossover and load-aware neighborhood mutation on each parent candidate rendering configuration in the parent population to generate multiple mutated offspring candidate rendering configurations. The specific meaning of the grouping constraint cross is: The parent generation candidate rendering configurations are divided into geometric texture parameter groups, lighting and reflection parameter groups, and edge compensation parameter groups according to the parameter function relationship. Two parent generation candidate rendering configurations are randomly selected from the parent generation population; The geometric texture parameter set and lighting reflection parameter set in the parent candidate rendering configuration with higher rendering quality score are used as the geometric texture parameter set and lighting reflection parameter set in the child candidate rendering configuration, and the edge compensation parameter set in the parent candidate rendering configuration with lower rendering time score is used as the edge compensation parameter set in the child candidate rendering configuration to generate the child candidate rendering configuration. The load-aware neighborhood variation specifically refers to: Calculate the difference between the current rendering frame rate and the target rendering frame rate to obtain the frame rate deviation; When the frame rate deviation is less than 0, the adjacent values ​​of the offspring candidate rendering configuration are adjusted according to the load reduction direction. When the frame rate deviation is greater than or equal to 0, the adjacent values ​​of the offspring candidate rendering configuration are adjusted according to the quality improvement direction to obtain the mutated offspring candidate rendering configuration. The legality of multiple candidate rendering configurations for mutated offspring is repaired to obtain a repaired offspring population; The parent population and the repaired offspring population are merged into a joint population. The joint population is then re-executed with fast non-dominated sorting, crowding distance calculation, parent population preservation, group constraint crossover, load-aware neighborhood mutation, and legality repair until the maximum number of iterations is reached to obtain the final population. Perform fast non-dominated sorting on the final population and determine the candidate rendering configuration with the largest comprehensive decision value in the first non-dominated layer as the target rendering configuration. The target rendering configuration includes target mesh detail level, target texture sampling resolution, target shadow calculation accuracy, target reflection calculation accuracy, target anti-aliasing level, and target compensation intensity parameters.

[0012] Optionally, the depth-constrained optical flow compensation mechanism specifically includes: Based on the rendering cache data, obtain the initial rendered image frame and depth cache of the current rendering frame, and the target real-time rendered image frame and depth cache of the previous rendering frame; Perform grayscale mapping between the initial rendered image frame of the current rendered frame and the target real-time rendered image frame of the previous rendered frame to obtain the current grayscale image and the historical grayscale image. A local pixel window is constructed at each pixel position in the current rendering frame. By minimizing the brightness difference between the local pixel window in the current grayscale image and the corresponding local pixel window in the historical grayscale image, the optical flow vector at each pixel position is calculated to obtain the optical flow field between the current rendering frame and the previous rendering frame. Based on the optical flow field, the target real-time rendered image frame and depth cache of the previous rendering frame are mapped to the image coordinate system of the current rendering frame to obtain the historical aligned image frame and historical aligned depth cache. Based on the depth cache of the current rendering frame and the historical aligned depth cache, the depth consistency weight of each pixel position is calculated using a negative exponential function. The minimum value between the constant 1 and the product of the target compensation intensity parameter and the depth consistency weight is obtained by using the minimum value function to obtain the pixel-level compensation weight for each pixel position. Based on pixel-level compensation weights, the initial rendered image frame and the historical aligned image frame of the current rendered frame are fused at the pixel level to obtain the target real-time rendered image frame of the current rendered frame.

[0013] According to an embodiment of the present invention, a real-time rendering system for 3D models based on artificial intelligence includes: The data acquisition module is used to acquire real-time rendering source data of the target 3D model during the real-time rendering process, taking the current rendering frame as the processing object. The rendering state token generation module is used to generate a sequence of rendering state tokens for the current rendering frame based on real-time rendering source data by mapping model region division with state tokens. The rendering state feature encoding module is used to encode the rendering state token sequence using an improved MambaVision network to obtain the rendering state feature vector of the current rendering frame. The rendering score prediction module is used to predict the rendering quality and rendering time of different candidate rendering configurations based on the rendering state feature vector, and to obtain the rendering quality score sequence and the rendering time score sequence. The rendering strategy decision module is used to execute real-time rendering strategy decisions based on the rendering quality score sequence and the rendering time score sequence, and obtain the target rendering configuration by using the improved NSGA-II algorithm. The layered real-time rendering module is used to perform layered real-time rendering on the target 3D model according to the target rendering configuration to obtain the initial rendered image frame; The pixel-level inter-frame compensation module is used to perform pixel-level inter-frame compensation on the initial rendered image frame through the depth-constrained optical flow compensation mechanism to obtain the target real-time rendered image frame.

[0014] The beneficial effects of this invention are: This invention introduces a visibility-coupling priority-guided dual-axis state-gated scanning and a feature-level inter-frame residual compensation mechanism. This allows the improved MambaVision network to encode rendering state features around the geometric complexity, viewpoint visibility, material lighting sensitivity, and inter-frame motion changes of the model region, thereby enhancing the feature representation capabilities of key visible areas, complex geometric areas, and specular and shadow-sensitive areas. In the rendering strategy decision-making stage, the rendering quality and rendering time of different candidate rendering configurations are predicted. An improved NSGA-II algorithm incorporating group constraint cross-validation and load-aware neighborhood variation is used to perform multi-objective optimization, enabling the target rendering configuration to simultaneously consider image quality, rendering time, and frame rate stability, reducing frame rate fluctuations caused by abnormal memory usage, increased rendering latency, and unreasonable configuration switching. In the image output stage, a depth-constrained optical flow compensation mechanism performs pixel-level inter-frame compensation on the initial rendered image frame. Combined with the depth cache and image alignment relationship between the current and previous rendered frames, texture blurring, edge flickering, shadow jumps, and cross-frame jitter are suppressed, thereby improving the detail integrity, pixel-level continuity, and visual stability of the target real-time rendered image frame. Therefore, this invention is of great significance for improving the real-time rendering quality of 3D models, reducing rendering resource waste, enhancing frame rate stability in complex interactive scenarios, and improving the real-time visualization experience in digital twin, virtual reality, and industrial simulation scenarios. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of a real-time rendering method and system for 3D models based on artificial intelligence proposed in this invention. Figure 2 This invention presents a flowchart of an improved MambaVision network structure in a real-time rendering method for 3D models based on artificial intelligence and a system thereof. Figure 3 This is a flowchart of a real-time rendering method for 3D models based on artificial intelligence and a depth-constrained optical flow compensation mechanism in the system proposed in this invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figures 1-3 A real-time rendering method for 3D models based on artificial intelligence includes the following steps: Step 1: During real-time rendering, take the current rendering frame as the processing object and obtain the real-time rendering source data of the target 3D model; Step 2: Based on real-time rendering source data, generate the rendering state token sequence for the current rendering frame through model region division and state token mapping; Step 3: Encode the rendering state token sequence using an improved MambaVision network to obtain the rendering state feature vector of the current rendering frame; the improved MambaVision network introduces a dual-axis state gating scan guided by visibility coupling priority and a feature-level inter-frame residual compensation mechanism. Step 4: Based on the rendering state feature vector, predict the rendering quality and rendering time of different candidate rendering configurations, and obtain the rendering quality score sequence and the rendering time score sequence. Step 5: Based on the rendering quality score sequence and the rendering time score sequence, a real-time rendering strategy decision is made using the improved NSGA-II algorithm to obtain the target rendering configuration; the improved NSGA-II algorithm introduces group constraint crossover and load-aware neighborhood variation; Step 6: Perform layered real-time rendering on the target 3D model according to the target rendering configuration to obtain the initial rendered image frame; wherein, layered real-time rendering specifically includes: performing geometric detail clipping based on the target mesh detail level, performing texture sampling based on the target texture sampling resolution, performing shadow rendering based on the target shadow calculation accuracy, performing reflection rendering based on the target reflection calculation accuracy, and performing edge smoothing based on the target anti-aliasing level. Step 7: Perform pixel-level inter-frame compensation on the initial rendered image frame using a depth-constrained optical flow compensation mechanism to obtain the target real-time rendered image frame.

[0018] In this embodiment, the real-time rendering source data includes the geometric structure parameters, material texture parameters, camera viewpoint parameters, lighting parameters, interactive motion parameters, device load parameters, previous rendering frame state parameters, rendering configuration parameters, and rendering cache data of the target 3D model.

[0019] The geometric parameters include the vertex coordinates, facet connectivity, normal direction, texture coordinates, model hierarchy, and model region division parameters of the target 3D model; the material texture parameters include the material identifier, diffuse reflection coefficient, specular reflection coefficient, roughness parameter, metallicity parameter, transparency parameter, and texture sampling resolution corresponding to each model region in the target 3D model; and the camera viewpoint parameters include the camera position, camera orientation, frustum range, viewport parameters, display resolution, and screen center coordinates. Lighting parameters include light source position, light intensity, light direction, and shadow casting parameters; interactive motion parameters include the change in camera viewpoint, model pose, model scaling, screen space displacement, user interaction commands, and the screen space position of the current rendering frame relative to the previous rendering frame; device load parameters include device graphics processing load, device video memory usage, rendering queue length, single-frame rendering time, current rendering frame rate, and target rendering frame rate. The previous rendering frame's state parameters include the camera view parameters, model pose parameters, model scaling parameters, screen space position parameters, and motion vector cache. The rendering configuration parameters include the range of mesh detail level values, the range of texture sampling resolution values, the range of shadow calculation precision values, the range of reflection calculation precision values, the range of anti-aliasing level values, and the range of compensation intensity parameter values. The rendering cache data includes the depth cache of the current rendering frame, the depth cache of the previous rendering frame, and the target real-time rendered image frame of the previous rendering frame.

[0020] In this embodiment, step two specifically includes: Based on geometric structural parameters, the target 3D model is divided into multiple model regions, and the region number of each model region is generated according to the hierarchical order in the model hierarchy and the connection relationship of the facets in the same level. Based on geometric structure parameters, a geometric complexity feature vector is generated for each model region in the current rendering frame. The components of each dimension of the geometric complexity feature vector are: number of vertices, number of faces, normal change intensity, curvature change intensity, boundary contour ratio, and model level number. Based on camera view parameters, a view visibility feature vector is generated for each model region in the current rendering frame. The components of each dimension of the view visibility feature vector are: visibility ratio, occlusion ratio, screen projection area, view frustum inside / outside identifier, viewpoint distance, and screen center distance. Based on material texture parameters and lighting parameters, generate material lighting response vectors; based on camera view parameters, interaction motion parameters and previous rendering frame state parameters, generate inter-frame motion state vectors; based on device load parameters, generate device rendering load vectors. Perform linear mapping and dimension alignment on the geometric complexity feature vector and view visibility feature vector corresponding to each model region to obtain the geometric state token and view state token of each model region, and concatenate them into a region state token group. Linear mapping and dimension alignment are performed on the material lighting response vector, inter-frame motion state vector, and device rendering load vector respectively to obtain the material lighting state token, inter-frame motion state token, and device load state token. Arrange the area status token groups in ascending order of area number to form area status token segments; By concatenating the material lighting state token, inter-frame motion state token, and device load state token sequentially into the region state token segment, the rendering state token sequence of the current rendering frame is obtained.

[0021] In this embodiment, step three specifically includes: Extract the geometric state tokens and view state tokens of each model region from the rendering state token sequence; The weight mapping values ​​of the geometric state token and the view state token are calculated by linear mapping and Sigmoid activation function, respectively, to obtain the geometric complexity scalar and the view visibility scalar. Multiply the geometric complexity scalar and the view visibility scalar, and normalize them using the Softmax function to obtain the visibility coupling priority value for each model region. Couple and map the geometric state tokens and view state tokens of each model region to generate a geometric-view coupling token; The geometric view coupling tokens of each model region are sorted according to the visibility coupling priority value to obtain a dynamic scan token sequence; Perform a dual-axis state-gated scan guided by visibility coupling priority value on the dynamic scan token sequence to obtain the forward scan hidden state and reverse scan hidden state of each model region; The forward scan hidden state, reverse scan hidden state, and geometric viewpoint coupling token of the same model region are fused to obtain the visibility coupling scan feature; The material lighting state token and visibility coupled scan features are concatenated, and the texture sensitive component, shadow sensitive component, specular sensitive component and reflection sensitive component of each model region are calculated by four sensitive component mapping units to form the material lighting sensitive vector; wherein, each of the four sensitive component mapping units is composed of linear mapping, LayerNorm normalization and Sigmoid activation. Linear mapping is performed on the material's illumination sensitivity vector to obtain the material's illumination recalibration weights; By recalibrating the visibility coupling scan features using material lighting recalibration weights, the material lighting sensitive region features of each model region are obtained. The device load status token and material illumination sensitive area features are concatenated, and the device load gating coefficient is generated by linear mapping and Sigmoid activation function. Based on the device load gating coefficient, the load constraint adjustment is performed on the material lighting sensitive area features to obtain the load constraint rendering state features of each model area. The load-constrained rendering state features are compensated for inter-frame continuity through a feature-level inter-frame residual compensation mechanism to obtain inter-frame compensated rendering state features. The inter-frame compensation rendering state features of each model region are used to calculate the region fusion weight using the Softmax function; based on the region fusion weight, the inter-frame compensation rendering state features of each model region are weighted and summed to obtain the region-level rendering fusion features. The region-level rendering fusion features, material lighting state tokens, inter-frame motion state tokens, and device load state tokens are concatenated to obtain cross-dimensional rendering fusion features. The cross-dimensional rendering fusion features are then linearly mapped to generate the rendering state feature vector of the current rendering frame.

[0022] In this embodiment, the visibility coupling priority-guided dual-axis state-gated scan includes forward state-gated scan and reverse state-gated scan. The forward state-gated scan is performed in the direction of decreasing visibility coupling priority value, specifically as follows: The forward scan hidden state of the previous model region and the geometric view coupling token of the current model region are coupled by state transition mapping and tanh activation to generate a forward candidate hidden state; Using the visibility coupling priority value of the current model region as the gating coefficient, the forward scan hidden state of the previous model region and the forward candidate hidden state of the current model region are gatingly fused to obtain the forward scan hidden state of the current model region. The reverse state-gated scan is performed in the direction of increasing visibility coupling priority value, specifically as follows: The reverse scan hidden state of the previous model region and the geometric view coupling token of the current model region are coupled with state transition mapping and tanh activation to generate reverse candidate hidden states; Using the visibility coupling priority value of the current model region as the gating coefficient, the inverse scan hidden state of the previous model region and the inverse candidate hidden state of the current model region are gated and fused to obtain the inverse scan hidden state of the current model region. The feature-level inter-frame residual compensation mechanism is specifically as follows: Inter-frame motion state tokens are linearly mapped to the HardSigmoid activation function to generate inter-frame compensation weights. Linear mapping is performed on the inter-frame motion state tokens to obtain inter-frame motion compensation features; The inter-frame motion compensation features are multiplied element-wise with the inter-frame compensation weights to obtain the inter-frame residual compensation features. The inter-frame residual compensation features are then residually concatenated with the load constraint rendering state features to obtain the inter-frame compensation rendering state features for each model region.

[0023] In this invention, the improved MambaVision network has the same basic network framework as the standard MambaVision network. The standard MambaVision network typically divides the input image or visual features into several visual tokens, obtains a token representation with a unified dimension through embedding mapping, and then uses the Mamba state space modeling structure to perform long-range dependency modeling on the token sequence. At the same time, it combines local feature enhancement, normalization layers, nonlinear activation layers and output mapping layers to generate visual feature representations.

[0024] The improved MambaVision network does not directly use the standard MambaVision network to extract visual features from ordinary image tokens. Instead, it features structural improvements specifically for real-time 3D model rendering tasks. The network introduces a visibility coupling priority value, dynamically sorting geometric-viewpoint coupling tokens after coupling the geometric complexity scalar with the viewpoint visibility scalar, and further performing a dual-axis state-gated scan guided by the visibility coupling priority value. The improved MambaVision network adds a material lighting-sensitive component mapping structure, recalibrating region features through texture-sensitive, shadow-sensitive, specular-sensitive, and reflection-sensitive components. It also introduces a device load gating coefficient to adjust the load constraints on material lighting-sensitive region features. Simultaneously, the improved MambaVision network adds a feature-level inter-frame residual compensation mechanism, generating inter-frame residual compensation features through inter-frame motion state tokens to enhance feature continuity between adjacent rendering frames.

[0025] Through the aforementioned improvements, the visibility coupling priority-guided dual-axis state-gated scanning enables the network to prioritize model regions with high geometric complexity and strong visual visibility in the current rendering frame, reducing the interference of invisible and low-contribution regions on rendering strategy decisions. The material lighting-sensitive component mapping and region recalibration structure enhance the network's feature representation capabilities for regions with significant changes in texture, shadows, highlights, and reflections, improving the accuracy of subsequent rendering quality score predictions. The device load gating structure allows the network encoding results to reflect the current device computing power and frame rate pressure, providing a more reliable state basis for balancing quality and time consumption in the subsequent improved NSGA-II algorithm. The feature-level inter-frame residual compensation mechanism utilizes inter-frame motion information to reduce state abrupt changes between adjacent rendering frames, ensuring that the obtained rendering state feature vector simultaneously possesses region visibility representation, material lighting-sensitive representation, load constraint representation, and inter-frame continuity representation, thereby improving the accuracy of real-time rendering configuration decisions, frame rate stability, and the visual continuity of the target real-time rendered image frames.

[0026] In this embodiment, step four specifically includes: Based on the rendering configuration parameters, a set of candidate rendering configurations is determined by discrete combination and load constraint filtering. The candidate rendering configuration set includes several candidate rendering configurations, each of which includes mesh detail level, texture sampling resolution, shadow calculation accuracy, reflection calculation accuracy, anti-aliasing level, and compensation intensity parameters; The load constraint screening is as follows: discrete combinations that do not meet the constraints of video memory usage, single frame rendering time, and rendering frame rate are eliminated. Perform a linear mapping on the current candidate rendering configuration to obtain the configuration feature vector; The rendering state feature vector and the configuration feature vector are concatenated, and then predicted and mapped using the quality score prediction head and the time consumption score prediction head respectively, to obtain the rendering quality score and rendering time consumption score of the current candidate rendering configuration. Both the quality score prediction head and the time consumption score prediction head are composed of a linear mapping, a GELU activation function, and a linear mapping. Traverse all candidate rendering configurations in the candidate rendering configuration set to obtain the rendering quality score sequence and the rendering time score sequence.

[0027] In this embodiment, the improved NSGA-II algorithm specifically includes: The candidate rendering configuration set is used as the initial population, where each candidate rendering configuration in the candidate rendering configuration set is a population individual in the initial population; Based on device load parameters, read the current rendering frame rate and target rendering frame rate of the target 3D model; Based on the rendering quality score sequence, rendering time score sequence, current rendering frame rate, and target rendering frame rate, a multi-objective fitness value is calculated for each candidate rendering configuration, and then weighted and fused to obtain a comprehensive decision value. The multi-objective fitness value includes a quality optimization target value, a time optimization target value, and a frame rate stability target value. The quality optimization target value is the rendering quality score; the time optimization target value is the negative of the rendering time score; and the frame rate stability target value is the negative of the absolute difference between the current rendering frame rate and the target rendering frame rate. Based on multi-objective fitness values, a fast non-dominated sort is performed on candidate rendering configurations to obtain several non-dominated layers, and a non-dominated level is assigned to each candidate rendering configuration. Calculate the crowding distance for candidate rendering configurations within the same non-dominated layer; Based on the non-dominance level and crowding distance, multiple parent candidate rendering configurations are selected from the current population to form a parent population. Specifically, two candidate rendering configurations are randomly selected from the current population for a binary tournament comparison. If the non-dominance levels of the two candidate rendering configurations are different, the candidate rendering configuration with the smaller non-dominance level is selected as the parent candidate rendering configuration. If the non-dominance levels of the two candidate rendering configurations are the same, the candidate rendering configuration with the larger crowding distance is selected as the parent candidate rendering configuration. The binary tournament comparison is repeated until a preset number of parent candidate rendering configurations are selected to form the parent population. Perform grouping constraint crossover and load-aware neighborhood mutation on each parent candidate rendering configuration in the parent population to generate multiple mutated offspring candidate rendering configurations. The specific meaning of the grouping constraint cross is: The parent generation candidate rendering configurations are divided into three groups based on the relationship between parameters: geometric texture parameters, lighting and reflection parameters, and edge compensation parameters. The geometric texture parameters include mesh detail level and texture sampling resolution, the lighting and reflection parameters include shadow calculation accuracy and reflection calculation accuracy, and the edge compensation parameters include anti-aliasing level and compensation intensity parameters. Two parent generation candidate rendering configurations are randomly selected from the parent generation population; The geometric texture parameter set and lighting reflection parameter set in the parent candidate rendering configuration with higher rendering quality score are used as the geometric texture parameter set and lighting reflection parameter set in the child candidate rendering configuration, and the edge compensation parameter set in the parent candidate rendering configuration with lower rendering time score is used as the edge compensation parameter set in the child candidate rendering configuration to generate the child candidate rendering configuration. The load-aware neighborhood variation specifically refers to: Calculate the difference between the current rendering frame rate and the target rendering frame rate to obtain the frame rate deviation; When the frame rate deviation is less than 0, the adjacent values ​​of the offspring candidate rendering configuration are adjusted according to the load reduction direction. When the frame rate deviation is greater than or equal to 0, the adjacent values ​​of the offspring candidate rendering configuration are adjusted according to the quality improvement direction to obtain the mutated offspring candidate rendering configuration. The load reduction approach involves lowering the mesh detail level, texture sampling resolution, shadow calculation accuracy, reflection calculation accuracy, and anti-aliasing level, while increasing the compensation intensity parameter; the quality improvement approach involves increasing the mesh detail level, texture sampling resolution, shadow calculation accuracy, reflection calculation accuracy, and anti-aliasing level, while lowering the compensation intensity parameter. The legality of multiple candidate rendering configurations for mutated offspring is repaired to obtain a repaired offspring population. The legality repair includes verification of video memory usage constraints, verification of single-frame rendering time constraints, verification of rendering frame rate constraints, and verification of parameter value range constraints. The parent population and the repaired offspring population are merged into a joint population. The joint population is then re-executed with fast non-dominated sorting, crowding distance calculation, parent population preservation, group constraint crossover, load-aware neighborhood mutation, and legality repair until the maximum number of iterations is reached to obtain the final population. Perform fast non-dominated sorting on the final population and determine the candidate rendering configuration with the largest comprehensive decision value in the first non-dominated layer as the target rendering configuration. The target rendering configuration includes target mesh detail level, target texture sampling resolution, target shadow calculation accuracy, target reflection calculation accuracy, target anti-aliasing level, and target compensation intensity parameters.

[0028] In this invention, the improved NSGA-II algorithm introduces group constraint crossover and load-aware neighborhood mutation on the basis of the standard NSGA-II algorithm. The group constraint crossover is not a random crossover of parameters in the candidate rendering configurations, but rather divides the mesh detail level and texture sampling resolution into a geometric texture parameter group, the shadow calculation accuracy and reflection calculation accuracy into a lighting reflection parameter group, and the anti-aliasing level and compensation intensity into an edge compensation parameter group, based on the functional correlation of the parameters. Furthermore, it selects the dominant parameter group from different parent generations based on the rendering quality score and rendering time score to generate the child generation candidate rendering configurations, thereby maintaining the functional synergy between parameters within the rendering configuration. Load-aware neighborhood mutation dynamically selects the direction of load reduction or quality improvement to adjust adjacent values ​​based on the frame rate deviation between the current rendering frame rate and the target rendering frame rate, enabling the offspring candidate rendering configuration to adapt to changes in device load and real-time frame rate. Through the above improvements, the improved NSGA-II algorithm further enhances the directionality and constraint adaptation capability of candidate rendering configuration search, reduces the number of invalid configurations caused by random crossover and blind mutation, and improves the overall balance between image quality, time consumption, and frame rate stability of the target rendering configuration.

[0029] In this embodiment, the depth-constrained optical flow compensation mechanism specifically includes: Based on the rendering cache data, the initial rendered image frame and depth cache of the current rendering frame, and the target real-time rendered image frame and depth cache of the previous rendering frame are obtained; among them, the depth cache is the set of depth values ​​of the visible model surface points corresponding to each pixel position in the camera coordinate system. Perform grayscale mapping between the initial rendered image frame of the current rendered frame and the target real-time rendered image frame of the previous rendered frame to obtain the current grayscale image and the historical grayscale image. A local pixel window is constructed at each pixel position in the current rendering frame. By minimizing the brightness difference between the local pixel window in the current grayscale image and the corresponding local pixel window in the historical grayscale image, the optical flow vector at each pixel position is calculated to obtain the optical flow field between the current rendering frame and the previous rendering frame. Based on the optical flow field, the target real-time rendered image frame and depth cache of the previous rendering frame are mapped to the image coordinate system of the current rendering frame to obtain the historical aligned image frame and historical aligned depth cache. Based on the depth cache of the current rendering frame and the historical aligned depth cache, the depth consistency weight of each pixel position is calculated using a negative exponential function. The minimum value between the constant 1 and the product of the target compensation intensity parameter and the depth consistency weight is obtained by using the minimum value function to obtain the pixel-level compensation weight for each pixel position. Based on pixel-level compensation weights, the initial rendered image frame and the historical aligned image frame of the current rendered frame are fused at the pixel level to obtain the target real-time rendered image frame of the current rendered frame.

[0030] In this invention, a pixel-level correspondence is established between the optical flow fields of the current rendering frame and the previous rendering frame. A depth consistency weight is calculated using a depth cache, and pixel information in historically aligned image frames that matches the current pixel's spatial position and depth relationship is selectively fused. This avoids erroneous compensation caused by occlusion changes, abrupt depth shifts, or model boundary displacements when relying solely on optical flow compensation. Simultaneously, pixel-level compensation weights are generated by combining compensation intensity parameters and depth consistency weights. This allows the initial rendered image frame of the current rendering frame to absorb stable detail information from the previous rendering frame while preserving the real-time rendering result. This suppresses cross-frame discontinuities such as texture flicker, edge jitter, shadow jumps, and motion blur, improving the pixel-level continuity, detail integrity, and visual stability of the target real-time rendered image frame.

[0031] A real-time rendering system for 3D models based on artificial intelligence, comprising: The data acquisition module is used to acquire real-time rendering source data of the target 3D model during the real-time rendering process, taking the current rendering frame as the processing object. The rendering state token generation module is used to generate a sequence of rendering state tokens for the current rendering frame based on real-time rendering source data by mapping model region division with state tokens. The rendering state feature encoding module is used to encode the rendering state token sequence using an improved MambaVision network to obtain the rendering state feature vector of the current rendering frame. The rendering score prediction module is used to predict the rendering quality and rendering time of different candidate rendering configurations based on the rendering state feature vector, and to obtain the rendering quality score sequence and the rendering time score sequence. The rendering strategy decision module is used to execute real-time rendering strategy decisions based on the rendering quality score sequence and the rendering time score sequence, and obtain the target rendering configuration by using the improved NSGA-II algorithm. The layered real-time rendering module is used to perform layered real-time rendering on the target 3D model according to the target rendering configuration to obtain the initial rendered image frame; The pixel-level inter-frame compensation module is used to perform pixel-level inter-frame compensation on the initial rendered image frame through the depth-constrained optical flow compensation mechanism to obtain the target real-time rendered image frame.

[0032] Example 1: To verify the feasibility of this invention in practice, it was applied to a real-time rendering scene of a 3D assembly in a digital twin operation and maintenance platform for industrial equipment. This platform displays a complex mechanical equipment model containing a motor, reducer, hydraulic pipelines, sensor brackets, and a housing structure. The model has approximately 2.86 million faces, 4.8GB of texture maps, a display resolution of 2560×1440, and a target rendering frame rate of 60 FPS. Maintenance personnel need to rotate, scale, and section the equipment model in real time within the platform to observe the local structural details of bearing seats, seals, pipeline interfaces, and high-temperature alarm areas. Existing real-time rendering methods mainly use fixed LOD switching and conventional anti-aliasing. When users quickly adjust their viewpoint or enter the equipment to observe small parts, problems such as flickering edges of local structures, abrupt shadow changes, texture blurring, and frame rate drops easily occur, especially in high-light areas, occluded boundary areas, and areas with dense pipelines on the model surface. The image stability is insufficient, affecting the maintenance personnel's judgment of the critical structural status.

[0033] In the implementation of this invention, the system takes the current rendering frame as the processing object and obtains the real-time rendering source data of the target 3D model, including the model's vertex coordinates, facet connectivity, normal direction, texture coordinates, material parameters, camera position, camera orientation, lighting intensity, user interaction commands, device graphics processing load, current rendering frame rate, previous rendering frame state parameters, and rendering cache data. Subsequently, the system divides the complex mechanical equipment model into 186 model regions according to the model's hierarchical structure and facet connectivity within the same level, and generates a region number for each model region. For each model region, the system calculates the geometric complexity feature vector and the viewpoint visibility feature vector, and maps them to geometric state tokens and viewpoint state tokens. These are then combined with material lighting state tokens, inter-frame motion state tokens, and device load state tokens to form the rendering state token sequence for the current rendering frame. An improved MambaVision network is used to encode the rendering state token sequence as a rendering state feature. A dual-axis state gating scan guided by visibility coupling priority value is used to prioritize the enhancement of model regions with high visibility and high geometric complexity in the current view. A feature-level inter-frame residual compensation mechanism is used to reduce the state abrupt changes between adjacent rendering frames.

[0034] During the rendering configuration decision-making phase, the system predicts the rendering quality and rendering time for different candidate rendering configurations based on the rendering state feature vector. Candidate rendering configurations include mesh detail level, texture sampling resolution, shadow calculation accuracy, reflection calculation accuracy, anti-aliasing level, and compensation intensity parameters. Subsequently, the improved NSGA-II algorithm executes real-time rendering strategy decisions based on the rendering quality score sequence and the rendering time score sequence. It maintains the functional synergy between the geometric texture parameter group, lighting and reflection parameter group, and edge compensation parameter group through grouped constraint cross-validation, and performs load-aware neighborhood mutation based on the difference between the current rendering frame rate and the target rendering frame rate. When the device load is high and the current frame rate is lower than the target frame rate, the system reduces the mesh detail level and shadow calculation accuracy, and increases the compensation intensity parameters; when the frame rate meets the target requirements, the system increases the texture sampling resolution and reflection calculation accuracy of key visible areas. After completing layered real-time rendering, the system uses a depth-constrained optical flow compensation mechanism, utilizing the depth cache, optical flow field, and historical aligned image frames of the current and previous rendering frames to perform pixel-level inter-frame compensation, thereby outputting the target real-time rendered image frame. To further verify the actual effect of the present invention, a comparative experiment was conducted between the present invention and the LOD real-time rendering scheme and the real-time rendering scheme using the standard MambaVision network and the standard NSGA-II algorithm. The comparison results are shown in Table 1.

[0035] Table 1. Comparison of Effects of Different Real-Time Rendering Schemes

[0036] As shown in Table 1, under the same 3D model, display resolution, and hardware environment, the average rendering frame rate of this invention reaches 61.4 FPS, which is 14.6 FPS higher than the LOD real-time rendering scheme and 7.8 FPS higher than the standard MambaVision + standard NSGA-II scheme. The minimum rendering frame rate reaches 52.8 FPS, indicating that this invention can maintain high real-time performance even when rapidly rotating the viewpoint and observing dense structures inside the device. The average rendering time per frame is reduced from 21.36ms in the LOD real-time rendering scheme to 16.29ms, and the peak memory usage is reduced from 7.9GB to 6.4GB, indicating that the improved NSGA-II algorithm can dynamically select a more reasonable target rendering configuration based on the device load, reducing the consumption of unnecessary rendering resources.

[0037] Furthermore, the percentage of edge flickering pixels decreased from 6.8% to 1.7%, and the percentage of shadow-skipping frames decreased from 9.5% to 2.1%, indicating that the depth-constrained optical flow compensation mechanism can effectively utilize historical frame information and depth consistency weights to improve the problem of cross-frame discontinuity. Texture detail retention increased to 93.8%, and image structure similarity increased to 0.958, demonstrating that the improved MambaVision network can accurately identify key visible areas, complex geometric areas, and specular and shadow-sensitive areas, enabling real-time rendering to maintain good image detail while reducing time consumption. Therefore, this invention can balance rendering quality, rendering time, frame rate stability, and cross-frame image continuity in complex 3D model real-time interactive scenarios, making it suitable for applications such as digital twin operation and maintenance, industrial simulation demonstration, virtual assembly training, and online 3D model display.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An artificial intelligence-based 3D model real-time rendering method, characterized in that, Includes the following steps: Step 1: During real-time rendering, take the current rendering frame as the processing object and obtain the real-time rendering source data of the target 3D model; Step 2: Based on real-time rendering source data, generate the rendering state token sequence for the current rendering frame through model region division and state token mapping; Step 3: Encode the rendering state token sequence using an improved MambaVision network to obtain the rendering state feature vector of the current rendering frame; the improved MambaVision network introduces a dual-axis state gating scan guided by visibility coupling priority and a feature-level inter-frame residual compensation mechanism. Step 4: Based on the rendering state feature vector, predict the rendering quality and rendering time of different candidate rendering configurations, and obtain the rendering quality score sequence and the rendering time score sequence. Step 5: Based on the rendering quality score sequence and the rendering time score sequence, a real-time rendering strategy decision is made using the improved NSGA-II algorithm to obtain the target rendering configuration; the improved NSGA-II algorithm introduces group constraint crossover and load-aware neighborhood variation; Step 6: Perform layered real-time rendering on the target 3D model according to the target rendering configuration to obtain the initial rendered image frames; Step 7: Perform pixel-level inter-frame compensation on the initial rendered image frame using a depth-constrained optical flow compensation mechanism to obtain the target real-time rendered image frame.

2. The real-time rendering method of a 3D model based on artificial intelligence according to claim 1, characterized in that, The real-time rendering source data includes the geometric structure parameters, material texture parameters, camera viewpoint parameters, lighting parameters, interactive motion parameters, device load parameters, previous rendering frame state parameters, rendering configuration parameters, and rendering cache data of the target 3D model.

3. The real-time rendering method of a 3D model based on artificial intelligence according to claim 1, characterized in that, Step two specifically includes: Based on geometric structural parameters, the target 3D model is divided into multiple model regions, and the region number of each model region is generated according to the hierarchical order in the model hierarchy and the connection relationship of the facets in the same level. Based on geometric structure parameters, generate a geometric complexity feature vector for each model region in the current rendering frame; based on camera view parameters, generate a view visibility feature vector for each model region in the current rendering frame. Based on material texture parameters and lighting parameters, generate material lighting response vectors; based on camera view parameters, interaction motion parameters, and previous rendering frame state parameters, generate inter-frame motion state vectors; based on device load parameters, generate device rendering load vectors. Perform linear mapping and dimension alignment on the geometric complexity feature vector and view visibility feature vector corresponding to each model region to obtain the geometric state token and view state token of each model region, and concatenate them into a region state token group. Linear mapping and dimension alignment are performed on the material lighting response vector, inter-frame motion state vector, and device rendering load vector respectively to obtain the material lighting state token, inter-frame motion state token, and device load state token. Arrange the area status token groups in ascending order of area number to form area status token segments; By concatenating the material lighting state token, inter-frame motion state token, and device load state token sequentially into the region state token segment, the rendering state token sequence of the current rendering frame is obtained.

4. The real-time rendering method of a 3D model based on artificial intelligence according to claim 1, characterized in that, Step three specifically includes: Extract the geometric state tokens and view state tokens of each model region from the rendering state token sequence; The weight mapping values ​​of the geometric state token and the view state token are calculated by linear mapping and Sigmoid activation function, respectively, to obtain the geometric complexity scalar and the view visibility scalar. Multiply the geometric complexity scalar and the view visibility scalar, and normalize them using the Softmax function to obtain the visibility coupling priority value for each model region. Couple and map the geometric state tokens and view state tokens of each model region to generate a geometric-view coupling token; The geometric view coupling tokens of each model region are sorted according to the visibility coupling priority value to obtain a dynamic scan token sequence; Perform a dual-axis state-gated scan guided by visibility coupling priority value on the dynamic scan token sequence to obtain the forward scan hidden state and reverse scan hidden state of each model region; The forward scan hidden state, reverse scan hidden state, and geometric viewpoint coupling token of the same model region are fused to obtain the visibility coupling scan feature; The material lighting state token and visibility coupled scanning features are concatenated, and the texture sensitive component, shadow sensitive component, specular sensitive component and reflection sensitive component of each model region are calculated by four sensitive component mapping units, and the material lighting sensitive vector is formed. Linear mapping is performed on the material's illumination sensitivity vector to obtain the material's illumination recalibration weights; By recalibrating the visibility coupling scan features using material lighting recalibration weights, the material lighting sensitive region features of each model region are obtained. The device load status token and material illumination sensitive area features are concatenated, and the device load gating coefficient is generated by linear mapping and Sigmoid activation function. Based on the device load gating coefficient, the load constraint adjustment is performed on the material lighting sensitive area features to obtain the load constraint rendering state features of each model area. The load-constrained rendering state features are compensated for inter-frame continuity through a feature-level inter-frame residual compensation mechanism to obtain inter-frame compensated rendering state features. The inter-frame compensation rendering state features of each model region are used to calculate the region fusion weight using the Softmax function; based on the region fusion weight, the inter-frame compensation rendering state features of each model region are weighted and summed to obtain the region-level rendering fusion features. The region-level rendering fusion features, material lighting state tokens, inter-frame motion state tokens, and device load state tokens are concatenated to obtain cross-dimensional rendering fusion features. The cross-dimensional rendering fusion features are then linearly mapped to generate the rendering state feature vector of the current rendering frame.

5. The real-time rendering method of a 3D model based on artificial intelligence according to claim 4, characterized in that, The visibility coupling priority-guided dual-axis state-gated scan includes forward state-gated scan and reverse state-gated scan; The forward state-gated scan is performed in the direction of decreasing visibility coupling priority value, specifically as follows: The forward scan hidden state of the previous model region and the geometric view coupling token of the current model region are coupled by state transition mapping and tanh activation to generate a forward candidate hidden state; Using the visibility coupling priority value of the current model region as the gating coefficient, the forward scan hidden state of the previous model region and the forward candidate hidden state of the current model region are gatingly fused to obtain the forward scan hidden state of the current model region. The reverse state-gated scan is performed in the direction of increasing visibility coupling priority value, specifically as follows: The reverse scan hidden state of the previous model region and the geometric view coupling token of the current model region are coupled with state transition mapping and tanh activation to generate reverse candidate hidden states; Using the visibility coupling priority value of the current model region as the gating coefficient, the inverse scan hidden state of the previous model region and the inverse candidate hidden state of the current model region are gatingly fused to obtain the inverse scan hidden state of the current model region. The feature-level inter-frame residual compensation mechanism is specifically as follows: Inter-frame motion state tokens are linearly mapped to the HardSigmoid activation function to generate inter-frame compensation weights. Linear mapping is performed on the inter-frame motion state tokens to obtain inter-frame motion compensation features; The inter-frame motion compensation features are multiplied element-wise with the inter-frame compensation weights to obtain the inter-frame residual compensation features. The inter-frame residual compensation features are then residually concatenated with the load constraint rendering state features to obtain the inter-frame compensation rendering state features for each model region.

6. The real-time rendering method of a 3D model based on artificial intelligence according to claim 1, characterized in that, Step four specifically includes: Based on the rendering configuration parameters, a set of candidate rendering configurations is determined by discrete combination and load constraint filtering. The candidate rendering configuration set includes several candidate rendering configurations, each of which includes mesh detail level, texture sampling resolution, shadow calculation accuracy, reflection calculation accuracy, anti-aliasing level, and compensation intensity parameters; Perform a linear mapping on the current candidate rendering configuration to obtain the configuration feature vector; The rendering state feature vector and the configuration feature vector are concatenated, and then predicted and mapped using the quality score prediction head and the time consumption score prediction head respectively, to obtain the rendering quality score and rendering time consumption score of the current candidate rendering configuration. Both the quality score prediction head and the time consumption score prediction head are composed of a linear mapping, a GELU activation function, and a linear mapping. Traverse all candidate rendering configurations in the candidate rendering configuration set to obtain the rendering quality score sequence and the rendering time score sequence.

7. The real-time rendering method of a 3D model based on artificial intelligence according to claim 1, characterized in that, The improved NSGA-II algorithm specifically includes: Use the candidate rendering configuration set as the initial population; Based on device load parameters, read the current rendering frame rate and target rendering frame rate of the target 3D model; Based on the rendering quality score sequence, rendering time score sequence, current rendering frame rate and target rendering frame rate, calculate the multi-objective fitness value of each candidate rendering configuration, and then weight and fuse them to obtain a comprehensive decision value. Based on multi-objective fitness values, a fast non-dominated sort is performed on candidate rendering configurations to obtain several non-dominated layers, and a non-dominated level is assigned to each candidate rendering configuration. Calculate the crowding distance for candidate rendering configurations within the same non-dominated layer; Based on the non-dominance level and crowding distance, multiple parent candidate rendering configurations are selected from the current population to form a parent population. Perform grouping constraint crossover and load-aware neighborhood mutation on each parent candidate rendering configuration in the parent population to generate multiple mutated offspring candidate rendering configurations. The specific meaning of the grouping constraint cross is: The parent generation candidate rendering configurations are divided into geometric texture parameter groups, lighting and reflection parameter groups, and edge compensation parameter groups according to the parameter function relationship. Two parent generation candidate rendering configurations are randomly selected from the parent generation population; The geometric texture parameter set and lighting reflection parameter set in the parent candidate rendering configuration with higher rendering quality score are used as the geometric texture parameter set and lighting reflection parameter set in the child candidate rendering configuration, and the edge compensation parameter set in the parent candidate rendering configuration with lower rendering time score is used as the edge compensation parameter set in the child candidate rendering configuration to generate the child candidate rendering configuration. The load-aware neighborhood variation specifically refers to: Calculate the difference between the current rendering frame rate and the target rendering frame rate to obtain the frame rate deviation; When the frame rate deviation is less than 0, the adjacent values ​​of the offspring candidate rendering configuration are adjusted according to the load reduction direction. When the frame rate deviation is greater than or equal to 0, the adjacent values ​​of the offspring candidate rendering configuration are adjusted according to the quality improvement direction to obtain the mutated offspring candidate rendering configuration. The legality of multiple candidate rendering configurations for mutated offspring is repaired to obtain a repaired offspring population; The parent population and the repaired offspring population are merged into a joint population. The joint population is then re-executed with fast non-dominated sorting, crowding distance calculation, parent population preservation, group constraint crossover, load-aware neighborhood mutation, and legality repair until the maximum number of iterations is reached to obtain the final population. Perform fast non-dominated sorting on the final population and determine the candidate rendering configuration with the largest comprehensive decision value in the first non-dominated layer as the target rendering configuration. The target rendering configuration includes target mesh detail level, target texture sampling resolution, target shadow calculation accuracy, target reflection calculation accuracy, target anti-aliasing level, and target compensation intensity parameters.

8. The real-time rendering method of a 3D model based on artificial intelligence according to claim 1, characterized in that, The depth-constrained optical flow compensation mechanism specifically includes: Based on the rendering cache data, obtain the initial rendered image frame and depth cache of the current rendering frame, and the target real-time rendered image frame and depth cache of the previous rendering frame; Perform grayscale mapping between the initial rendered image frame of the current rendered frame and the target real-time rendered image frame of the previous rendered frame to obtain the current grayscale image and the historical grayscale image. A local pixel window is constructed at each pixel position in the current rendering frame. By minimizing the brightness difference between the local pixel window in the current grayscale image and the corresponding local pixel window in the historical grayscale image, the optical flow vector at each pixel position is calculated to obtain the optical flow field between the current rendering frame and the previous rendering frame. Based on the optical flow field, the target real-time rendered image frame and depth cache of the previous rendering frame are mapped to the image coordinate system of the current rendering frame to obtain the historical aligned image frame and historical aligned depth cache. Based on the depth cache of the current rendering frame and the historical aligned depth cache, the depth consistency weight of each pixel position is calculated using a negative exponential function. The minimum value between the constant 1 and the product of the target compensation intensity parameter and the depth consistency weight is obtained by using the minimum value function to obtain the pixel-level compensation weight for each pixel position. Based on pixel-level compensation weights, the initial rendered image frame and the historical aligned image frame of the current rendered frame are fused at the pixel level to obtain the target real-time rendered image frame of the current rendered frame.

9. An artificial intelligence-based 3D model real-time rendering system, which executes an artificial intelligence-based 3D model real-time rendering method according to any one of claims 1 to 8. include: The data acquisition module is used to acquire real-time rendering source data of the target 3D model during the real-time rendering process, taking the current rendering frame as the processing object. The rendering state token generation module is used to generate a sequence of rendering state tokens for the current rendering frame based on real-time rendering source data by mapping model region division with state tokens. The rendering state feature encoding module is used to encode the rendering state token sequence using an improved MambaVision network to obtain the rendering state feature vector of the current rendering frame. The rendering score prediction module is used to predict the rendering quality and rendering time of different candidate rendering configurations based on the rendering state feature vector, and to obtain the rendering quality score sequence and the rendering time score sequence. The rendering strategy decision module is used to execute real-time rendering strategy decisions based on the rendering quality score sequence and the rendering time score sequence, and obtain the target rendering configuration by using the improved NSGA-II algorithm. The layered real-time rendering module is used to perform layered real-time rendering on the target 3D model according to the target rendering configuration to obtain the initial rendered image frame; The pixel-level inter-frame compensation module is used to perform pixel-level inter-frame compensation on the initial rendered image frame through the depth-constrained optical flow compensation mechanism to obtain the target real-time rendered image frame.

Images