VR three-dimensional figure intelligent generation method and system based on AI large model

Abstract

This invention discloses a VR 3D stereoscopic graphics intelligent generation method and system based on an AI large-scale model, belonging to the field of stereoscopic graphics generation technology. It includes: multi-dimensional semantic parsing and feature generation, dynamic color tolerance adjustment, texture mapping optimization judgment and execution, and image output. After obtaining user commands, this invention performs multi-dimensional semantic parsing and generates standardized feature vectors. The AI ​​large-scale model determines whether dynamic color tolerance adaptive adjustment is required, and if so, executes the adjustment and outputs multi-view images; otherwise, it directly outputs and preprocesses them into a standard multi-view image set. It then determines whether to perform 3D structure reconstruction and texture mapping optimization. If so, after reconstruction, it jumps to binocular correction and adaptation rendering; otherwise, after panoramic stitching, it enters a simple parallax matching stage. A uniform horizontal offset is applied to the left and right eye images to correct parallax, and the image is output to the display pipeline.

Description

Technical Field

[0001] This invention relates to the field of stereoscopic graphics generation technology, and in particular to a method and system for intelligent generation of VR 3D stereoscopic graphics based on AI large models. Background Technology

[0002] With the integration of virtual reality technology and generative AI (Artificial Intelligence) large-scale modeling technology, 3D models and panoramic stereoscopic images required for VR (Virtual Reality) can be generated from inputs such as text and images, lowering the barrier to VR content creation. Currently, mainstream solutions rely on pre-trained 2D diffusion large-scale models, multi-view image generation networks combined with neural radiation fields, and 3D Gaussian splashing to complete 3D reconstruction and rendering output. However, many key defects still exist in the image processing stage, affecting the immersive visual experience of VR. Firstly, there are issues with multi-view... Figure 1 The consistency defect, also known as the typical Jenners problem, is that when large models generate 2D images from different viewpoints, they lack global 3D spatial constraints and unified camera parameter calibration. The generated images from different viewpoints have contradictions in texture style, color tone, and structural details, resulting in hallucination distortions such as repetitive features, misaligned structures, and semantic conflicts between front and back views. The accuracy of left and right eye stereo disparity matching is insufficient, and there are local misalignments and abnormal changes in depth information. When observing with a VR headset, it is easy to produce visual incongruity and depth perception disorder, which can lead to visual dizziness and fail to meet the disparity matching specifications of binocular stereo display. Secondly, there are defects in texture mapping and material rendering. Large models generate textures with insufficient resolution, blurred details, and missing high-frequency details. UV (UV Mapping) mapping suffers from distortion, misaligned seams, texture stretching, and blank surfaces, making it difficult to generate material parameters that meet PBR (Physically Based Rendering) standards. This results in an inability to accurately reproduce the realistic optical performance of lighting changes, shadow casting, and transparent / reflective materials. Global illumination and local lighting are inconsistent, with flickering and color drifting under dynamic perspectives. Static textures are ill-suited to the real-time lighting rendering requirements of VR dynamic roaming. There are also challenges in real-time image processing and resolution adaptation. High-precision 3D stereoscopic image rendering, view correction, and binocular image synchronization processing are computationally expensive, making it difficult to achieve high-resolution, high-frame-rate, and low-latency real-time stereoscopic image output on VR headsets, resulting in jagged edges and blurry trails. The stereoscopic image processing effects for complex dynamic VR scenes, panoramic surround views, and large-scale virtual environments are poor, making it difficult to achieve immersive VR 3D stereoscopic image output with consistent global scene textures, stable parallax, and unified lighting. Summary of the Invention

[0003] In view of this, embodiments of the present invention provide a VR 3D stereoscopic graphics intelligent generation method and system based on an AI large model. This system can obtain user commands, perform multi-dimensional semantic parsing, and generate standardized feature vectors. The AI ​​large model determines whether dynamic color tolerance adaptive adjustment is required, and then outputs multi-view images; otherwise, it directly outputs and preprocesses them into a standard multi-view image set. Next, it determines whether to perform 3D structure reconstruction and texture mapping optimization. If so, it jumps to binocular correction and adaptation rendering after reconstruction; otherwise, it enters a simple parallax matching stage after panoramic stitching. A uniform horizontal offset is applied to the left and right eye images to correct parallax, and the image is then output to the display pipeline.

[0004] On the one hand, a VR 3D stereoscopic graphics intelligent generation method based on an AI large model is provided. This method includes: acquiring user-inputted target VR scene generation instructions and performing multi-dimensional semantic parsing to extract multi-dimensional information of the target VR 3D scene and generate standardized structured generation feature vectors; inputting the standardized structured generation feature vectors into a pre-trained AI large model to determine whether dynamic color tolerance adaptive adjustment is needed to optimize rendering quality and frame rate. If so, the initial 2D image sequence of multiple views from a VR panoramic surround view is output after adjustment; otherwise, the initial 2D image sequence of multiple views from a VR panoramic surround view is directly output and preprocessed to obtain the standardized structured generation feature vectors. Standardize the multi-view image set; based on the standardized multi-view image set, determine whether to perform 3D spatial structure reconstruction and texture mapping optimization. If yes, jump to the binocular stereo correction and VR adaptation rendering steps after structure reconstruction and texture mapping optimization. If no, directly generate VR panoramic binocular segmented images based on the standardized multi-view image set and enter the simple parallax matching and image output stage. In the simple parallax matching and image output stage, based on the preset global fixed parallax offset threshold, perform coarse-grained parallax matching processing on the horizontal pixel offset of the left and right eye panoramic images to correct the binocular image parallax deviation, and directly output the matched left and right eye images to the display pipeline.

[0005] On the other hand, a VR 3D stereoscopic graphics intelligent generation system based on an AI large model is provided. This system includes: a multi-dimensional semantic parsing and feature generation module, a dynamic color tolerance adjustment module, a texture mapping optimization judgment and execution module, and an image output module. The multi-dimensional semantic parsing and feature generation module is used to acquire the user-inputted target VR scene generation instructions and perform multi-dimensional semantic parsing to extract multi-dimensional information of the target VR 3D scene and generate standardized structured generation feature vectors. The dynamic color tolerance adjustment module is used to input the standardized structured generation feature vectors into a pre-trained AI large model to determine whether to perform dynamic color tolerance adaptive adjustment to optimize rendering quality and frame rate. If so, it outputs a multi-view initial 2D image sequence with a VR panoramic surround view after adjustment; otherwise, it directly outputs the VR... The system generates a standardized multi-view image set by preprocessing a sequence of initial two-dimensional images from a panoramic surround view perspective. A texture mapping optimization judgment and execution module determines whether to perform 3D spatial structure reconstruction and texture mapping optimization based on the standardized multi-view image set. If so, it jumps to the binocular stereo correction and VR adaptation rendering steps after structure reconstruction and texture mapping optimization. If not, it directly generates VR panoramic binocular segmented images based on the standardized multi-view image set and enters the simple parallax matching and image output stage. The image output module performs coarse-grained parallax matching processing on the left and right eye panoramic images in the simple parallax matching and image output stage, based on a preset global fixed parallax offset threshold, to correct the binocular image parallax deviation and directly output the matched left and right eye images to the display pipeline.

[0006] Beneficial effects The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: By constructing a full-link intelligent generation system encompassing multimodal semantic feature encoding, dynamic color tolerance adaptive adjustment, dual-branch 3D reconstruction optimization, binocular stereo correction and adaptive rendering, and lightweight parallax output, this invention effectively addresses the shortcomings of traditional technologies, such as the inability to simultaneously achieve high rendering quality and frame rate, poor color consistency across multiple viewpoints, low robustness of 3D reconstruction, susceptibility to distortion in weak texture and sparse viewpoint scenes, dizziness caused by binocular imaging, and limited hardware compatibility. This invention generates standardized feature vectors through multi-dimensional semantic parsing, enabling precise structured encoding of cross-modal input information and ensuring the accuracy and controllability of AI large-scale model scene generation. Relying on a negative correlation adaptive mechanism between rendering sampling frequency and color deviation tolerance threshold, it dynamically and differentially adjusts the sampling precision of different regions of the image. This reduces computational consumption and increases rendering frame rate in high-load scenarios, while tightening color constraints and suppressing color drift in high-precision scenarios, achieving a dynamic balance between image quality and frame rate. Meanwhile, this invention achieves intelligent switching of reconstruction branches based on geometric consistency error and texture confidence. It performs high-precision 3D reconstruction and Poisson multi-band texture fusion on high-quality image data, obtaining a topologically complete and texture-seamless 3D scene model. For low-quality image data, it sequentially performs viewpoint densification, color calibration, and generative content completion, significantly improving the fault tolerance of complex scene reconstruction. Furthermore, this invention effectively eliminates VR imaging distortion, chromatic aberration, and dizziness problems through virtual binocular camera registration, lens pre-distortion compensation, parallax comfort zone optimization, and motion compensation rendering, ensuring comfortable stereoscopic viewing. It also adds a lightweight scheme for global fixed threshold coarse-grained parallax matching, enabling rapid generation of compliant binocular images without 3D reconstruction, compatible with both high-end and low-end VR terminal devices. This invention improves the intelligence, image quality, stereoscopic perception, and hardware compatibility of VR 3D scene generation, while reducing the computational cost of content generation, effectively balancing high-precision rendering quality with lightweight real-time output efficiency. Attached Figure Description

[0007] Figure 1 A flowchart of a VR 3D stereoscopic graphics intelligent generation method based on an AI large model provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of the VR 3D stereoscopic graphics intelligent generation system based on AI large model provided in the embodiments of this application. Detailed Implementation

[0008] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0009] like Figure 1 The diagram shown is a flowchart of a VR 3D stereoscopic graphics intelligent generation method based on an AI large model provided in this application embodiment. The method includes the following steps: As the first step of this invention, the target VR scene generation instructions input by the user are obtained and multi-dimensional semantic parsing is performed to extract multi-dimensional information of the target VR 3D scene and generate a standardized structured generation feature vector. The generation instructions include text description instructions and reference 2D image instructions. The multi-dimensional information of the target VR 3D scene includes spatial structure information, texture material information, lighting environment information and binocular stereo disparity parameter information.

[0010] It should be understood that, firstly, the system acquires the target VR scene generation instructions input from the user terminal in real time. These instructions are compatible with both text descriptions in natural language and externally imported reference 2D image instructions. The text descriptions include detailed custom descriptions of the VR scene type, scene scale, object structure, texture style, lighting atmosphere, and stereoscopic visual effects. The reference 2D image instructions include user-uploaded scene reference images, material reference images, composition reference images, and other 2D reference materials of different dimensions. The system then calls a pre-trained multimodal semantic parsing and encoding network to perform comprehensive, multi-dimensional deep semantic parsing and scene element decomposition on the acquired generation instructions. For text instructions, it performs word segmentation, semantic understanding, keyword feature extraction, scene attribute classification, and semantic vector embedding encoding. For reference 2D image instructions, it performs image preprocessing, pixel feature extraction, scene structure feature recognition, texture and material feature clustering, and lighting feature extraction. This accurately decomposes and separates the four core multi-dimensional information corresponding to the target VR 3D scene. The information includes spatial structure information for constraining scene spatial layout, object size, structural topology, and spatial positional relationships; texture material information for limiting scene surface texture style, material properties, texture resolution, and texture distribution characteristics; light and shadow environment information for characterizing scene illumination intensity, light source direction, color temperature parameters, shadow features, and ambient light atmosphere; and binocular stereo disparity parameter information for adapting to VR binocular imaging rules, constraining binocular disparity range, interpupillary distance matching parameters, and stereo depth perception intervals. After extracting multi-dimensional scene information, the system performs dimensional unification, invalid noise feature filtering, cross-modal feature alignment, and standardization normalization on various heterogeneous feature information to eliminate dimensional differences and semantic deviations between text features and image features. The system also performs structured recombination, feature weight allocation, and feature sequence regularization on discretized scene features, ultimately generating standardized structured feature vectors that are dimensionally unified, semantically accurate, structurally standardized, and can be directly input into multi-modal AI large models for 3D scene generation.

[0011] As the second step of this invention, the standardized structured generated feature vector is input into the pre-trained AI large model to determine whether dynamic color tolerance adaptive adjustment is performed to optimize rendering quality and frame rate. If so, the initial two-dimensional image sequence of the VR panoramic surround view is output after adjustment. If not, the initial two-dimensional image sequence of the VR panoramic surround view is directly output and preprocessed to obtain a standardized multi-view image set.

[0012] It needs to be explained that the specific process for determining whether to perform dynamic color tolerance adaptive adjustment is as follows: Obtain the real-time rendering parameters of the current VR scene, including the current image rendering sampling frequency, the current frame rendering load, and the target frame rate; Image rendering sampling frequency is used to characterize the number of sampling points for discretizing continuous 3D scene signals per unit area; image color deviation tolerance threshold is used to characterize the maximum allowable color difference between the rendered pixel color and the theoretical real color when judging the color quality of the rendered image. There is a negative correlation between image color deviation tolerance threshold and image rendering sampling frequency; specifically, image rendering sampling frequency is the number of samples per pixel or the sampling density per unit length of screen space; when the sampling frequency decreases, the noise level and color deviation in the rendered image increase, and the color deviation tolerance threshold needs to be increased accordingly to avoid resource waste caused by oversampling; when the sampling frequency increases, the quality of the rendered image improves and the color deviation decreases, and the color deviation tolerance threshold needs to be decreased accordingly to maintain the image quality standard.

[0013] Calculate the predicted color deviation value at the current sampling frequency and compare the predicted color deviation value with the current color deviation tolerance threshold; If the predicted color deviation value is greater than the current color deviation tolerance threshold, it is determined that dynamic color tolerance adaptive adjustment is required. Dynamic color tolerance adaptive adjustment specifically includes: calculating an updated color deviation tolerance threshold based on the ratio of the current sampling frequency to the target sampling frequency, according to a preset adaptive adjustment function, where the adaptive adjustment function is a monotonically decreasing function; its mathematical expression is: ; in, The updated color deviation tolerance threshold. Based on the basic color deviation tolerance threshold, The current image is rendered at the sampling frequency. Render sampling frequency for the target image. This is the scene complexity adjustment coefficient, which is dynamically adjusted based on the scene complexity index. The higher the scene complexity, the higher the adjustment coefficient. The larger the value, the more sensitive the color deviation tolerance threshold becomes to changes in sampling frequency in highly complex scenarios.

[0014] The updated color deviation tolerance threshold is input into the rendering pipeline scheduling module of the AI ​​large model. The rendering pipeline scheduling module dynamically adjusts the rendering sampling frequency of each image region based on the updated color deviation tolerance threshold, specifically: For image regions where the color deviation is below the updated tolerance threshold, reduce the rendering sampling frequency to reduce computational overhead; for image regions where the color deviation is above the updated tolerance threshold, increase the rendering sampling frequency to ensure image quality.

[0015] In this embodiment, a complete set of real-time rendering parameters is first collected during the current VR scene rendering process. These real-time rendering parameters specifically include three core control parameters: the current image rendering sampling frequency, the current frame GPU rendering load, and the preset scene target frame rate. This provides real-time data support for subsequent threshold adaptive adjustment. The image rendering sampling frequency defined in this embodiment is used to accurately characterize the number of effective sampling points for discretized sampling of continuous 3D scene optical signals within a unit screen pixel area. It can be specifically quantified as the number of samples per pixel or the texture and light sampling density within a unit length of screen space. This is a core rendering parameter that directly determines the VR rendering image's detail, noise level, and computational cost. A higher sampling frequency value results in more sampling points per pixel, higher image rendering detail, and more accurate color reproduction, but also higher model inference and rendering computational cost. Conversely, a lower sampling frequency results in lower image sampling accuracy, faster rendering speed, but is prone to pixel noise, color shift, and local color distortion. Correspondingly, the image color deviation tolerance threshold in this embodiment is used to quantitatively characterize the allowable actual rendering pixel RG during multi-view image color consistency verification and image quality judgment by the AI ​​large model rendering pipeline. The maximum permissible color difference between the color value, brightness value, and the theoretical true pixel color value is the core control indicator for the tightness of color constraints. This threshold establishes a strict monotonically negative correlation with the image rendering sampling frequency, meaning the sampling frequency and the color tolerance threshold exhibit an inverse adaptation rule. Specifically, when the image rendering sampling frequency decreases, the number of sampling points per pixel is insufficient, the discrete sampling error in the rendering process is amplified, and the inherent noise level of the image and the color deviation from multiple viewpoints naturally increase. At this time, the system adaptively increases the color deviation tolerance threshold, relaxes the color consistency constraints, and avoids excessive error correction, repeated sampling, and wasted computing power caused by the strict color threshold constraints under low sampling conditions, thereby improving rendering efficiency while ensuring no significant visual distortion for the human eye. When the image rendering sampling frequency increases, the single-pixel sampling density is sufficient, the sampling discrete error is significantly reduced, the basic rendering quality of the image is significantly improved, and the native color deviation is greatly reduced. At this time, the system adaptively decreases the color deviation tolerance threshold, tightens the multi-viewpoint color consistency constraint standard, and accurately suppresses subtle color drift, color gamut shift, and color inconsistency between viewpoints under high sampling precision, maintaining a unified image quality standard for high-precision VR stereoscopic images.Based on this, the system calculates the predicted color deviation value of the image under the current sampling accuracy using a preset color deviation prediction model based on the real-time acquisition of the current image rendering sampling frequency. This predicted color deviation value can accurately quantify the pixel color error level of the current frame as a whole and in local areas. Then, the real-time predicted color deviation value is compared and verified with the color deviation tolerance threshold currently enabled by the system to complete the adaptive adjustment trigger judgment. The specific judgment logic is as follows: if the predicted color deviation value is greater than the current color deviation tolerance threshold, it indicates that the image color error under the current sampling matching relationship exceeds the preset qualified standard, and there are obvious color distortion, uneven color of viewing angle, or color drift defects, which cannot meet the VR stereoscopic imaging quality requirements. In this case, the system determines that the dynamic color tolerance adaptive adjustment process needs to be started to iteratively optimize and update the color deviation tolerance threshold in real time. In the specific adaptive adjustment process, the system relies on the ratio of the current actual image rendering sampling frequency to the preset scene target image rendering sampling frequency, and combines the preset monotonically decreasing adaptive adjustment function to complete the precise update of the threshold. For stable texture areas, solid color areas, and distant weakened areas in the image where the actual color deviation is lower than the updated tolerance threshold, it is determined that the color quality of the area meets the standard and there is too much redundant sampling. The local rendering sampling frequency of such areas is actively and appropriately reduced to reduce the computational cost of ineffective sampling and improve the overall rendering frame rate. For texture detail areas, light and shadow transition areas, object edge areas, and binocular parallax core areas in the image where the actual color deviation is higher than the updated tolerance threshold, it is determined that there is a risk of color distortion in the area. The local rendering sampling frequency of such areas is actively and adaptively increased to enhance the detail sampling accuracy and strictly control the local color error. Finally, the system achieves dynamic color tolerance and adaptive optimization of the whole frame VR multi-view image under different rendering loads, different scene complexities, and different sampling accuracies. This not only eliminates the waste of computational resources in high-load scenarios, but also ensures the color consistency of binocular stereo images and the immersive visual experience in high-precision rendering scenarios.

[0016] As the third step of this invention, based on a standardized multi-view image set, it is determined whether to perform three-dimensional spatial structure reconstruction and texture mapping optimization. If so, after structure reconstruction and texture mapping optimization, the process jumps to the binocular stereo correction and VR adaptation rendering steps. If not, the process directly generates VR panoramic binocular segmented images by stitching together the standardized multi-view image set, and then enters the simple parallax matching and image output stage.

[0017] It should be further explained that the specific process for determining whether to perform 3D spatial structure reconstruction and texture mapping optimization is as follows: Obtain disparity maps and depth maps of images from each viewpoint in a standardized multi-view image set, and calculate the geometric consistency error and texture confidence between each viewpoint. The geometric consistency error is used to characterize the reprojection error of the projection position of the same 3D point in different viewpoint images, and the texture confidence is used to characterize the clarity and recognizability of texture details in each viewpoint image. If the geometric consistency error is less than the preset geometric consistency threshold and the texture confidence is greater than the preset texture confidence threshold, then it is determined that 3D spatial structure reconstruction and texture mapping optimization will be performed directly, which includes the following sub-steps: S100, based on a standardized multi-view image set, performs multi-view stereo matching through the 3D reconstruction module of the AI ​​large model, calculates the disparity value of each pixel and generates a global depth map; the 3D reconstruction module adopts a depth estimation network based on cost volume construction and cost aggregation, projects the feature maps of each view image onto the 3D space through homography transformation to construct a 3D cost volume, and generates a high-precision depth map through cost aggregation and depth regression. S200 maps the global depth map to three-dimensional space through back projection transformation, generating an initial three-dimensional point cloud; S300 performs filtering and downsampling on the initial 3D point cloud to remove outlier noise points and simplify the point cloud, resulting in an optimized point cloud. S400, based on optimized point cloud, generates an initial 3D mesh model through Poisson surface reconstruction algorithm; S500 performs mesh optimization processing on the initial 3D mesh model. Mesh optimization processing includes non-manifold edge repair, hole filling and mesh smoothing to obtain a 3D mesh model with a complete topological structure.

[0018] The determination to directly perform 3D spatial structure reconstruction and texture mapping optimization also includes: S600 performs texture mapping optimization on the 3D mesh model. Texture mapping optimization includes: calculating the optimal texture source view for each triangular facet based on the visibility of each view image in the standardized multi-view image set and the viewing angle between each view image and the view. The optimal texture source view is the view with the smallest viewing angle between the view normals and the highest texture clarity. Based on the optimal texture source view, texture pixels are extracted from the view image to generate the initial texture fragments for each triangular facet. S700 performs color consistency fusion and seam elimination processing on the initial texture fragments of each triangular facet. The seam elimination processing adopts a multi-band fusion algorithm based on Poisson editing to smoothly transition the color differences of texture fragments from different perspectives at the boundary, generating a global seamless texture map. The global seamless texture map is mapped to the 3D mesh model to obtain a textured VR 3D scene model. If the geometric consistency error is greater than or equal to the preset geometric consistency threshold, or the texture confidence is less than or equal to the preset texture confidence threshold, then it is determined that the three-dimensional spatial structure reconstruction and texture mapping optimization will not be performed, and the multi-view image enhancement and view interpolation optimization will be performed first. The S800 uses the view synthesis module of the AI ​​large model to perform sparse view densification processing based on a standardized multi-view image set. The view synthesis module adopts a new view synthesis network based on neural radiation field or diffusion model. It takes any two adjacent view images and their corresponding camera pose parameters as input, samples along the interpolation path between the two views to generate at least one intermediate view synthesized image and its corresponding synthesized depth map, and iterates this process until the angular spacing between any adjacent views is less than the preset angular spacing threshold, thus obtaining an enhanced dense multi-view image set.

[0019] The specific steps for performing multi-view image enhancement and viewpoint interpolation optimization are as follows: S900 performs inter-view color consistency calibration on the enhanced dense multi-view image set. The color consistency calibration includes: using the color distribution of the reference view image as the standard, performing color histogram matching on all view images and the composite image to obtain a color-aligned dense multi-view image set. S1000: Based on the dense multi-view image set with color alignment, recalculate the geometric consistency error and texture confidence. If the recalculated geometric consistency error is less than the preset geometric consistency threshold and the texture confidence is greater than the preset texture confidence threshold, then jump to step S100 to continue to perform three-dimensional spatial structure reconstruction and texture mapping optimization. If, after the enhancement processing in steps S800 to S1000, the recalculated geometric consistency error is still greater than or equal to the preset geometric consistency threshold, or the texture confidence is still less than or equal to the preset texture confidence threshold, then it is determined that generative filling-based 3D content completion will be performed, specifically including: The generative completion module of the AI ​​large model identifies the missing or blurred regions in the standardized multi-view image set. Guided by the corresponding regions in the adjacent view images, the completed view images are generated to replace the image content of the original missing or blurred regions, and the completed standardized multi-view image set is obtained. Then, the process jumps to step S100 to perform three-dimensional spatial structure reconstruction and texture mapping optimization. The final textured VR 3D scene model is input into the VR scene rendering output module to generate stereoscopic rendering data adapted to the VR display device and output to the user terminal.

[0020] In this embodiment, after completing the preprocessing of the standardized multi-view image set, the intelligent judgment and hierarchical execution process for 3D spatial structure reconstruction and texture mapping optimization will be automatically initiated. This process adaptively selects the optimal VR 3D scene construction strategy based on the data quality, geometric matching accuracy, and texture effectiveness of the current multi-view images. This effectively solves the technical defects of traditional fixed reconstruction modes, such as 3D model topology damage, texture blurring, obvious seams, and geometric distortion caused by sparse viewpoints, missing textures, and excessive reprojection deviations. This achieves a two-way improvement in reconstruction accuracy and generation stability. Specifically, the system first retrieves the preprocessed standardized multi-view image set, batches the dense disparity map and accurate depth map corresponding to each viewpoint image, and then... The system calculates the geometric consistency error and texture confidence score point-by-point between images from different viewpoints based on multi-view geometric constraints. Geometric consistency error specifically refers to the pixel reprojection deviation of the same real 3D spatial feature point projected onto the 2D image plane under different camera viewpoints. This directly characterizes the spatial matching accuracy and geometric reliability of the current multi-view data. Texture confidence score quantifies the clarity, richness, and feature discernibility of texture details in local regions of each viewpoint image. It effectively distinguishes between high-definition effective texture regions, weak texture regions, solid-color invalid regions, and blurred / distorted regions, providing a quantitative basis for intelligent switching of subsequent reconstruction branches. The system pre-configures fixed geometric consistency and texture confidence thresholds as judgment benchmarks. When the overall geometric consistency error among the measured viewpoints is less than the preset geometric consistency threshold, and the overall global texture confidence is greater than the preset texture confidence threshold, the current multi-view image data is judged to be of excellent quality, with accurate geometric matching and sufficient texture features. A high-precision 3D spatial structure reconstruction and texture mapping optimization process can be directly executed. Specifically, the multi-view stereo matching operation is first performed using the 3D reconstruction module built into the AI ​​large model. This module employs a high-precision depth estimation network based on cost volume construction and multi-scale cost aggregation. It transforms the image feature maps from different viewpoints according to camera pose parameters, projects them uniformly onto the global 3D spatial coordinate system to construct a dense 3D cost volume, and optimizes feature matching through multi-layer cost aggregation operations. To ensure accuracy, the system combines depth regression algorithms to output a global dense depth map with sub-pixel precision, guaranteeing the continuity and accuracy of depth information. The global depth map is then mapped to 3D physical space via pixel-space backprojection, restoring the 3D spatial coordinates of each pixel and generating an initial 3D point cloud covering the entire VR scene. Addressing the common issues of discrete outliers, redundant dense points, and local anomalies in the initial 3D point cloud, the system automatically performs pass-through filtering, statistical filtering, and uniform downsampling to precisely remove noise outliers generated during acquisition and projection. Simultaneously, while preserving key geometric contours and detailed features of the scene, the system simplifies and optimizes the point cloud, resulting in a high-quality optimized point cloud with uniform distribution, extremely low noise, and precise structure.Based on optimized point cloud data, the Poisson surface reconstruction algorithm is used to achieve implicit surface fitting and mesh reconstruction, generating an initial 3D mesh model with basic spatial topological relationships, effectively ensuring the smoothness, continuity, and structural integrity of the model surface. Addressing common topological defects in the initial 3D mesh model, such as non-manifold edges, mesh holes, local patch overlap, and surface bumps, the system performs comprehensive mesh optimization and repair processing. This precisely repairs non-manifold structures, fills holes of varying sizes, and smooths abrupt changes in mesh surface curvature, ultimately obtaining a high-precision 3D mesh model with complete topological structure, accurate geometry, and regular boundary contours. Building upon the 3D geometric reconstruction, the system further performs refined texture mapping optimization processing. First, based on the spatial pose, viewing direction, and spatial normal vectors of the mesh triangular faces from each viewpoint in a standardized multi-view image set, the system calculates the visibility weight and viewing angle of each viewpoint image for each triangular face, selecting the optimal model with the smallest viewing angle, highest texture clarity, and least occlusion interference. The texture source perspective ensures that each mesh facet can be matched with the optimal texture material, avoiding problems such as blurred textures from oblique views and disordered textures due to occlusion. After determining the optimal texture perspective, the texture pixel information in the corresponding perspective image is accurately extracted to generate independent initial texture fragments adapted to each triangular facet. To address the defects that are prone to occur when splicing multi-view textures, such as uneven texture color, boundary discontinuities, abrupt seams, and brightness jumps, this embodiment uses a multi-band fusion algorithm based on Poisson editing to perform global color consistency fusion and seam elimination processing on all initial texture fragments. Through multi-scale texture decomposition, boundary gradient constraints, and smooth transition fusion, the color difference and brightness difference between textures from different perspectives are eliminated, achieving a natural transition of texture boundaries. Finally, a global seamless texture map with no obvious seams, unified color, and continuous details is generated. The global seamless texture map is then accurately mapped onto the surface of the optimized 3D mesh model to complete texture coordinate calibration and texture adaptation, ultimately resulting in a textured VR 3D scene model with both accurate geometric structure and high-quality texture effects. Furthermore, when the system detects that the geometric consistency error of the current multi-view image is greater than or equal to the preset geometric consistency threshold, or the global texture confidence is less than or equal to the preset texture confidence threshold, it determines that the current multi-view data has problems such as sparse viewpoints, large geometric matching deviations, blurred and missing textures, and feature matching failures. Direct reconstruction will produce serious model distortion and texture distortion. Therefore, the system automatically skips the direct reconstruction process and prioritizes the multi-view image enhancement and viewpoint interpolation optimization operations.Specifically, the sparse viewpoint densitying process is achieved through the viewpoint synthesis module of the AI ​​large model. This module is compatible with the neural radiation field implicit rendering synthesis mechanism and the diffusion model new viewpoint generation mechanism. Taking any two sets of adjacent viewpoint images and their corresponding precise camera pose parameters as input, it uniformly samples along the camera interpolation trajectory between the two viewpoints, automatically generating at least one intermediate viewpoint synthesis image and its matching synthesis depth map. The missing viewpoints are continuously supplemented through segment-by-segment iterative interpolation and densification until the camera angular spacing between any adjacent viewpoints is less than a preset angular spacing threshold. This completely solves the problem of insufficient geometric solution accuracy caused by the sparseness of the original viewpoints, resulting in a dense multi-view map with uniform viewpoint distribution and complete data. After the viewpoint densification is completed, the system performs global color consistency calibration on all original viewpoint images and newly added synthetic viewpoint images. Using the color distribution, average brightness, and contrast characteristics of a preset standard reference viewpoint image as a benchmark, color histogram matching and color gain correction are performed frame-by-frame on all images to unify the global color gamut, brightness, and hue distribution, eliminating color shifts and brightness differences between different viewpoints, synthetic images, and original images, resulting in a densely color-aligned multi-viewpoint image set. Subsequently, based on the enhanced and optimized dense multi-viewpoint image set, the global geometric consistency error and texture confidence index are recalculated. If the updated index meets the threshold requirements, the system automatically jumps to the aforementioned high-precision calibration. The 3D reconstruction process fully executes depth estimation, point cloud generation, mesh reconstruction, and texture optimization. If, after multiple enhancement and optimization processes such as viewpoint densification and color calibration, the geometric consistency error and texture confidence of the multi-view images still fail to meet the reconstruction threshold, it indicates that the scene suffers from extreme problems such as large-area texture loss, severe occlusion, and feature failure. In this case, the system will automatically activate a generative incomplete 3D content mechanism. Through the generative incomplete module of the AI ​​large model, it intelligently identifies texture-deficient areas, blurred and distorted areas, and structurally defective areas in the multi-view image set. Using the effective feature content of adjacent clear view images as constraints and guiding information, it intelligently generates realistic 3D content completion for the missing areas. The scene textures and structural content are automatically replaced with original blurry, missing, and invalid image areas to obtain a complete, standardized multi-view image set with clear textures and satisfactory geometric matching. This set then jumps to a high-precision 3D reconstruction and texture optimization process to ensure the effectiveness and integrity of the generated 3D model in extreme scenarios. Finally, the high-quality textured VR 3D scene model, obtained through multi-layered quality control, optimization, and reconstruction, is input into the VR scene rendering output module. This module performs binocular parallax correction, image adaptation, and rendering parameter matching, generating binocular stereo rendering data fully adapted to the optical parameters and display specifications of VR display devices. This data is then output in real-time to the user terminal for immersive VR scene display.

[0021] It's important to understand that the specific steps involved in binocular stereo calibration and VR adaptation rendering include: The textured VR 3D scene model obtained after 3D spatial structure reconstruction and texture mapping optimization, as well as the current viewpoint position and orientation information in the virtual space, are obtained. The current viewpoint position and orientation information include the left eye viewpoint position and the right eye viewpoint position. The distance between the left eye viewpoint position and the right eye viewpoint position is equal to the preset human eye interpupillary distance value. Based on the current viewpoint position and orientation information, a left-eye virtual camera and a right-eye virtual camera are set in the VR 3D scene model. The left-eye virtual camera and the right-eye virtual camera have the same focal length, field of view and near-far clipping plane parameters, and their optical axes are parallel on the horizontal plane and aligned with the same gaze point, so that the left-eye virtual camera and the right-eye virtual camera form a standard orthogonal parallax stereo camera pair. The VR 3D scene model is rendered using the left-eye virtual camera to generate the original left-eye rendered image; the same scene is rendered synchronously using the right-eye virtual camera to generate the original right-eye rendered image; the original left-eye rendered image and the original right-eye rendered image constitute the original binocular image pair; during the rendering process, the disparity of each pixel is calculated based on the depth value of each object in the VR 3D scene model. The disparity is the lateral offset of the projection position of the same spatial point in the original left-eye rendered image and the original right-eye rendered image; the magnitude of the disparity is inversely proportional to the object's depth value, that is, the closer the object is to the viewpoint, the greater the disparity, and the farther the object is from the viewpoint, the smaller the disparity. Binocular stereo correction is performed on the original binocular images, specifically: lens distortion inverse compensation processing is performed on the original rendered images of the left and right eyes respectively. The lens distortion inverse compensation processing applies pre-distortion to the original rendered images according to the optical lens characteristic parameters of the VR display device. The distortion direction of the pre-distortion is opposite to the distortion direction generated by the optical lens of the VR display device. In addition, chromatic aberration correction is performed on the original rendered images of the left and right eyes to eliminate the separation of red, green and blue channels caused by the chromatic dispersion of the VR display device lens. After stereo correction, disparity range optimization is performed on the left and right eye images. Specifically, for pixels with disparity greater than or equal to a preset disparity threshold, their disparity is compressed towards zero disparity to within a preset maximum comfortable disparity threshold; for pixels with disparity less than a preset disparity threshold, their disparity is stretched towards zero disparity to above a preset minimum comfortable disparity threshold. Furthermore, color and brightness consistency calibration is performed on corresponding pixels in the left and right eye images. Using the left eye image as a reference, the chromaticity and brightness of each pixel in the right eye image are matched with the corresponding pixels in the left eye image to eliminate binocular color inconsistencies caused by differences in the rendering pipeline. The optimized result is a comfortable binocular image pair.

[0022] The binocular stereo calibration and VR adaptation rendering steps also include: The parallax-comfortable binocular images are input into the VR adaptation rendering pipeline. The VR adaptation rendering pipeline includes: scaling the left-eye and right-eye images according to the panel resolution of the VR display device to generate left-eye and right-eye display images adapted to the display panel resolution; performing sub-pixel rendering mapping on the left-eye and right-eye display images according to the sub-pixel arrangement of the VR display device to generate display data for each sub-pixel channel driving the VR display panel; performing asynchronous time warp or asynchronous spatial warp processing on the display data; and when a change in the user's head posture is detected, performing a reprojection transformation on the current frame image based on the depth map and color map rendered in the previous frame to generate an interpolated compensation frame. The left-eye and right-eye display images, after being processed by the VR adaptation rendering pipeline, are frame-synthesized and encapsulated into a stereoscopic frame data stream according to the transmission protocol of the VR display device. The stereoscopic frame data stream is then output to the VR display device for display through the video transmission interface. The stereoscopic frame data stream is synchronized with the user's interactive input, and the virtual camera parameters for the next frame are updated according to the user's real-time viewpoint changes. A standardized multi-view image set is stitched together panoramically according to the acquisition pose parameters of each viewpoint to generate a panoramic background image with cube maps or equidistant cylindrical projections. This panoramic background image is then used as static VR scene content. Based on the current user's left and right eye viewpoint positions, the viewport areas corresponding to the left and right eyes are cropped respectively to generate left-eye and right-eye panoramic viewport images. Simple parallax matching is performed on objects within the overlapping areas of the left-eye and right-eye panoramic viewport images. Specifically, local parallax offset is calculated based on optical flow or block matching algorithms, and the left and right eye viewport images are adjusted according to the local parallax offset to generate a simple binocular image with basic stereoscopic effect. This simple binocular image is then directly output to a VR display device for display.

[0023] In this embodiment, after completing the 3D spatial structure reconstruction and texture mapping optimization to obtain a complete textured VR 3D scene model, the system sequentially executes the entire process of binocular stereo calibration and VR adaptation rendering. It simultaneously supports two parallel branches: high-precision 3D model stereo rendering and lightweight panoramic binocular rendering, adapting to different computing power terminals and different image quality requirements. The overall execution logic is layered and refined, parameters are quantifiable, and the entire chain is controllable. Specific implementation details are as follows: The system first retrieves the textured VR 3D scene model after geometric reconstruction, mesh repair, and seamless texture mapping. Simultaneously, it collects the current viewpoint position and orientation information in the virtual space, which is transmitted in real-time by the VR headset tracking module. These viewpoint parameters are then refined and decomposed... The system is divided into two independent spatial point sets: the left-eye viewpoint 3D coordinates and the right-eye viewpoint 3D coordinates. It strictly ensures that the straight-line distance between the left and right eye viewpoints equals the device's preset standard human interpupillary distance (IPD). This IPD parameter can be adaptively assigned based on user manual configuration and the headset's factory default parameters, providing a rigid constraint for the physical benchmark of binocular imaging. Based on the acquired left and right eye viewpoint positions and overall line-of-sight orientation, the system completes the spatial mounting and parameter configuration of the left and right virtual cameras within the virtual 3D scene coordinate system. During implementation, the optical imaging parameters of the two virtual cameras are forcibly made completely identical, including lens focal length, horizontal and vertical field of view, and camera proximity. The Z-axis depth values ​​of the clipping plane and the far clipping plane are set, and the optical axes of the two cameras are kept parallel to each other on a horizontal plane. Both optical axes point to the same gaze point in the virtual space, ultimately constructing a standard parallel optical axis orthogonal parallax binocular virtual camera pair. This avoids parallax errors, visual ghosting, and dizziness caused by inconsistent camera parameters and optical axis convergence deviations. After configuring the binocular virtual cameras, the rendering pipeline is split into two paths to simultaneously perform image rendering calculations. The left-eye virtual camera completely traverses all faces, textures, and lighting information of the VR 3D scene model for rasterization rendering, outputting a distortion-free original rendering image for the left eye. The right-eye virtual camera uses the exact same rendering pipeline and sampling parameters to synchronously render the same VR scene. Once the scene rendering is complete, the original rendered image for the right eye is output. These two images together constitute the original binocular image pair for the entire process. During the rasterization pixel-by-pixel rendering process, the system retrieves the 3D depth information of all object vertices and faces within the VR 3D scene model. It calculates the lateral pixel offset of the corresponding 3D point in space on the left and right original images pixel by pixel. This lateral offset is defined as the pixel disparity. It strictly follows the optical law that disparity and depth are inversely proportional. Near foreground objects have smaller depth and larger lateral disparity, while distant objects have larger depth and smaller lateral disparity. This accurately restores the depth and stereoscopic perception law of the human eye when naturally viewing the real world, providing pixel-level quantitative basis for subsequent disparity adjustment.After obtaining the original binocular image pair, the system initiates a layered binocular stereo correction process. The first layer of correction is VR optical lens distortion inverse compensation processing. It retrieves a complete set of optical lens characteristic parameters calibrated before the VR display device leaves the factory, covering radial distortion coefficient, tangential distortion coefficient, and red, green, and blue three-color channel dispersion offset parameters. For the original rendered images of the left and right eyes, inverse pre-distortion mapping is applied pixel by pixel. The distortion trend and distortion amplitude of the pre-distortion are completely opposite to the barrel distortion produced by normal imaging of the VR lens, ensuring that the distortion cancels each other out after the image is refracted by the VR lens, and finally the human eye receives a regular and undistorted planar image. While correcting geometric distortion, RGB three-color channel color difference correction is performed simultaneously. According to the difference in refractive index of different wavelengths of light from the lens, the coordinates of red, green, and blue pixels are independently offset and corrected, completely eliminating the three-color channel misalignment and color fringing problems at the edge of the image caused by VR lens dispersion. After completing distortion and chromatic aberration correction, the system enters the second layer of parallax range comfort optimization. The system pre-programs parallax thresholds, maximum comfortable parallax thresholds, and minimum comfortable parallax thresholds that are adapted to the visual tolerance limits of the human eye. It then performs zoned control by traversing all parallax data of the binocular image pixel by pixel: if the pixel parallax is greater than or equal to the preset parallax threshold, it means that the parallax at that point is too large, which can easily cause difficulty in eye fusion, eye strain, and dizziness. The pixel parallax is compressed towards zero parallax, limiting the final parallax to not exceed the maximum comfortable parallax threshold; if the pixel parallax is less than the preset parallax threshold, it means that the parallax is too small, resulting in poor stereoscopic depth. To address any perceived parallax deficiency, the pixel's parallax is moderately stretched towards zero parallax to ensure it remains above the minimum comfortable parallax threshold. After parallax adjustment, binocular image color and brightness consistency calibration is performed. Throughout the process, the calibrated left-eye image is used as the color reference. Chroma channel mapping, brightness gain matching, and contrast alignment are performed on all pixels of the right-eye image. Pixel-level histogram matching is used to eliminate binocular color casts and uneven brightness caused by differences in lighting calculations between the left and right paths of the rendering pipeline. After all pixels are calibrated, a comfortable binocular image pair with compliant parallax range, unified color, and visual comfort is generated.Once stereoscopic correction is completed, the binocular images are fed into a standardized VR adaptation rendering pipeline for terminal adaptation processing. The pipeline performs multiple rounds of image adaptation in stages: First, it performs bilinear scaling on the left and right eye images according to the fixed resolution of the current VR display panel hardware, adapting to the inherent pixel size of the screen, and generating left and right eye display images adapted to the panel specifications; Second, it completes subpixel rendering mapping by combining the subpixel arrangement rules such as RGB stripes, BGR arrangement, and Delta arrangement of the VR screen, splitting the three-color subpixel positions of each pixel in the image, accurately matching the subpixel light-emitting points of the screen, and improving the edge sharpness of the image. In addition to color precision, complete display driving data is generated that can directly drive the three-color sub-pixels of the display panel. The third step is to equip the entire set of driving data with an asynchronous time warp or asynchronous space warp motion compensation mechanism. The head-mounted display IMU monitors the user's head posture rotation and translation in real time. Once a change in head position is detected, the complete scene depth map and color rendering map of the previous frame are retrieved. Based on the rotation Euler angles and translation vectors of the head posture change, the image plane homography transformation matrix is ​​solved. The matrix is ​​used to perform a global reprojection transformation on the previous frame to generate a motion transition compensation frame, which effectively fills the screen stuttering and ghosting problems caused by hardware rendering frame rate fluctuations. After all adaptation processing is completed, the system performs frame format synthesis on the adapted images for the left and right eyes. It can either merge and encapsulate them into single-frame split-screen or left-right split-screen binocular composite frames, or retain the independent frame format of the two channels. It strictly follows the established protocols for VR device HDMI and Type-C video transmission to package and encapsulate them into standardized binocular stereo frame data streams, which are then sent to the VR display device in real time through the hardware video transmission interface to complete photoelectric imaging. At the same time, the entire rendering chain continuously connects to the user's head tracking interaction input, and collects the dynamic changes of the viewpoint position and orientation in real time. Each point update synchronously refreshes the intrinsic and extrinsic parameters of the binocular virtual camera, and iterates to complete a new round of rendering, correction, adaptation and output, so as to achieve a smooth and immersive interaction where the image follows the head rotation.This embodiment also includes a lightweight panoramic alternative rendering path, adapted to low-end computing power hardware. The system pre-retrieves a complete and standardized multi-view image set, and relies on the pre-calibrated camera pose parameters for each viewpoint, including the camera's 3D spatial coordinates and three-axis rotation orientation angles, to complete multi-view spherical panoramic stitching. The stitching can flexibly choose a six-sided cube texture format or a latitude and longitude equidistant cylindrical projection panoramic format. The cube texture consists of six square textures: front, back, left, right, top, and bottom. The equidistant cylindrical projection uses a rectangular panoramic canvas. The resulting panoramic background image serves as the static VR scene base for the lightweight solution, without further 3D mesh reconstruction. The system still reads the real-time left and right eye viewpoint parameters and crops the panoramic image according to the spherical horizontal angle corresponding to the binocular interpupillary distance. The system generates a left-eye panoramic viewport image and a right-eye panoramic viewport image by defining dedicated viewport areas for the left and right eyes. For the overlapping scene areas of the two panoramic viewport images, a simplified two-dimensional parallax matching is performed. A dense optical flow field algorithm or an N×N window block matching algorithm can be used to calculate the pixel-level parallax offset of each local region. Based on the calculated offset, a global horizontal remapping is performed on the right-eye panoramic viewport image to adjust the lateral position of the binocular images to match human depth perception, forming a simplified binocular image with basic stereoscopic effects. This lightweight solution eliminates all 3D reconstruction, fine meshing, and texture calculations. Finally, the simplified binocular image is directly packaged and output to the VR display device, balancing the smoothness of operation on low-configuration hardware with a basic stereoscopic visual experience. The entire binocular correction and VR adaptation rendering process is divided into two complete technical branches: high-precision 3D reconstruction rendering and lightweight panoramic stitching rendering. The two branches share the same underlying logic for interpupillary distance, lens distortion, and color calibration. This ensures that high-performance devices have the ultimate stereoscopic image quality and color consistency, while also enabling smooth operation on low-end terminals. It completely solves multiple defects of traditional VR rendering, such as parallax discomfort, lens distortion, chromatic aberration, frame rate stuttering, uneven binocular color, and poor hardware compatibility.

[0024] As the fourth step of the present invention, in the simplified parallax matching and image output stage, based on a preset global fixed parallax offset threshold, coarse-grained parallax matching processing of horizontal pixel offset is performed on the panoramic images of the left and right eyes to correct the parallax deviation of the binocular images, and the matched left and right eye images are directly output to the display pipeline.

[0025] It should be understood that the specific process for coarse-grained disparity matching processing with horizontal pixel offset is as follows: Acquire the left-eye panoramic viewport image and the right-eye panoramic viewport image, set the left-eye panoramic viewport image as the reference image and keep it unchanged, and set the right-eye panoramic viewport image as the target image and use it as the object for parallax adjustment. A preset global fixed disparity offset threshold is obtained. Based on the global fixed disparity offset threshold, a horizontal pixel offset processing is performed on each pixel in the right-eye panoramic viewport image. Specifically, in the horizontal direction, the horizontal coordinate of each pixel in the right-eye panoramic viewport image is uniformly increased or decreased by the number of pixels specified by the global fixed disparity offset threshold to generate a right-eye panoramic offset image. This ensures that a constant horizontal disparity offset is formed between the right-eye panoramic offset image and the left-eye panoramic viewport image across the entire image. The global fixed disparity offset threshold is a predefined horizontal pixel offset constant. Its value is preset based on the standard human interpupillary distance parameters of the VR display device, the spherical mapping resolution of the panoramic background image, and the preset comfortable viewing distance range. The global fixed disparity offset threshold does not change with scene content, pixel position, or scene depth, so that the disparity between the left and right eye images remains constant across the entire image. Edge filling processing is performed on the blank areas at the edges of the right-eye panoramic offset image generated after pixel offset. Specifically, the pixels inside the boundary of the right-eye panoramic offset image are copied and extended outward along the boundary to fill the blank areas at the edges, so as to maintain the complete resolution of the right-eye panoramic offset image.

[0026] Coarse-grained disparity matching processing for horizontal pixel offset also includes: Obtain preset parallax offset direction parameters, which are set according to the stereoscopic display mode of the VR display device, including positive parallax mode and negative parallax mode; positive parallax mode shifts the right eye image to the right, making objects appear to be recessed into the screen; negative parallax mode shifts the right eye image to the left, making objects appear to be protruding out of the screen; determine the offset direction in horizontal pixel offset processing based on the parallax offset direction parameters; The right-eye panoramic offset image after pixel offset processing is combined with the left-eye panoramic viewport image that remains unchanged to form a coarse-grained matching binocular image pair. The disparity relationship between the left and right eye images in the coarse-grained matching binocular image pair is uniquely determined only by the globally fixed disparity offset threshold.

[0027] In this embodiment, after cropping the panoramic viewport images for both eyes, a coarse-grained parallax matching process based on a globally fixed horizontal pixel offset is initiated. This process does not rely on 3D depth information or perform pixel-level stereo matching and depth interpolation calculations. It achieves rapid binocular registration solely through a fixed parallax offset pre-calibrated by the device parameters. This enables the construction of a stable, compliant, and human-adaptable basic binocular parallax relationship with extremely low computational overhead. It effectively solves the problems of traditional lightweight VR panoramic displays, such as lack of stereoscopic depth, high overlap of binocular images, lack of stereoscopic sense, or chaotic parallax. (Specific implementation details follow.) The method is as follows: The system first acquires the cropped left-eye panoramic viewport image and the right-eye panoramic viewport image, and assigns roles to the two images. The left-eye panoramic viewport image is constantly set as the reference image, ensuring that the image pixel position and image content remain completely unchanged throughout the entire coarse-grained parallax matching process. At the same time, the right-eye panoramic viewport image is fixed as the target image to be adjusted, serving as the sole object for global parallax offset processing. This ensures a unified binocular matching reference and a single, controllable adjustment logic, avoiding parallax disorder caused by bidirectional offset. Based on this, the system retrieves and fixes the pre-calibrated... The system uses a globally fixed parallax offset threshold for storage. This threshold is a globally constant pixel offset that does not change with scene content, image position, object depth, or image texture. Its specific value is obtained through joint calibration of three types of hardware and human visual parameters: standard human interpupillary distance parameters of VR display devices, spherical mapping resolution of panoramic background images, and preset comfortable viewing distance range. This ensures compatibility with the optical imaging characteristics of current VR devices and the stereoscopic fusion range of the human eye. Based on this globally fixed parallax offset threshold, the system performs a unified horizontal pixel offset operation on the right-eye panoramic viewport image. Specifically, in the horizontal coordinate axis direction of the image, the right... The horizontal coordinates of all effective pixels within the panoramic viewport image are uniformly increased or decreased by the standard pixel offset corresponding to the threshold, achieving synchronous translation of all pixels in the image. This ultimately generates a right-eye panoramic offset image after overall displacement, resulting in a globally unified and constant standard horizontal disparity offset between the left-eye panoramic viewport image and the right-eye panoramic offset image. Unlike the pixel-by-pixel differentiated offset method of fine disparity matching, this coarse-grained matching scheme adopts a uniform offset mechanism across the entire field, making the disparity distribution of the entire image uniform and stable, without the problems of local disparity abrupt changes or depth disorder, greatly reducing the computational power consumption of real-time rendering.Meanwhile, the system is pre-configured with switchable parallax offset direction parameters to adapt to the visual presentation requirements of different VR stereoscopic display modes. Specifically, the parallax offset direction parameters include two configurable working conditions: positive parallax mode and negative parallax mode. In positive parallax mode, the entire right-eye image is horizontally offset to the right, making the overall parallax of the scene present a concave stereoscopic visual effect, which is suitable for the immersive viewing requirements of far-distance panoramic scenes. In negative parallax mode, the entire right-eye image is horizontally offset to the left, making the overall parallax of the scene present a convex stereoscopic visual effect, which is suitable for the stereoscopic presentation requirements of close-range interactive scenes. The system can automatically identify and lock the corresponding offset direction parameters according to the currently enabled VR stereoscopic display mode, accurately determine the positive or negative offset execution logic of this horizontal pixel offset processing, and ensure that the stereoscopic presentation effect meets the preset visual design goals. After completing the horizontal pixel offset of the entire image, the edges of the right-eye panoramic offset image will have regular blank missing areas due to the overall translation. In order to ensure the integrity of the output image resolution, the absence of black borders and defects at the image edges, and the continuity of the panoramic view without breaks, the system automatically performs intelligent filling processing of the edge blank areas. Specifically, it adopts the boundary pixel extension expansion filling method, extracts the effective texture pixels adjacent to the inner side of the right-eye panoramic offset image boundary, and fills the edge blank areas layer by layer from the inside out through pixel copying and smooth extension, ensuring that the texture transition of the filled area is natural, the color connection is continuous, and there are no obvious splicing marks, thus completely preserving the visual integrity and image consistency of the panoramic image. Finally, the right-eye panoramic offset image, after being horizontally translated and edge-completed with a fixed direction and offset, is paired with the left-eye panoramic viewport image, which maintains the same baseline throughout, to form a complete coarse-grained matching binocular image pair. The overall disparity magnitude, disparity direction, and stereoscopic effect of this binocular image pair are uniquely determined by the preset global fixed disparity offset threshold and disparity offset direction parameters. The disparity distribution is uniform throughout the image, the stereoscopic effect is stable, and the calculation process is extremely simple. It can quickly output binocular images with compliant stereoscopic visual effects in lightweight VR panoramic display scenarios with low computing power and low latency requirements, achieving efficient VR stereoscopic imaging output without 3D reconstruction or depth calculation.

[0028] like Figure 2The diagram shows the structure of the VR 3D stereoscopic graphics intelligent generation system based on an AI large model provided in this application embodiment. It includes: a multi-dimensional semantic parsing and feature generation module, a dynamic color tolerance adjustment module, a texture mapping optimization judgment and execution module, and an image output module. The multi-dimensional semantic parsing and feature generation module is used to acquire the user-inputted target VR scene generation command and perform multi-dimensional semantic parsing to extract multi-dimensional information of the target VR 3D scene and generate a standardized structured generation feature vector. The dynamic color tolerance adjustment module is used to input the standardized structured generation feature vector into a pre-trained AI large model, determine whether to perform dynamic color tolerance adaptive adjustment to optimize rendering quality and frame rate, and if so, output a multi-view initial 2D image sequence of VR panoramic surround view after adjustment; otherwise, directly... The system outputs a sequence of initial 2D images from multiple perspectives within a VR panoramic surround view and preprocesses them to obtain a standardized multi-view image set. A texture mapping optimization judgment and execution module determines whether to perform 3D spatial structure reconstruction and texture mapping optimization based on the standardized multi-view image set. If so, it jumps to the binocular stereo correction and VR adaptation rendering steps after structure reconstruction and texture mapping optimization. If not, it directly generates VR panoramic binocular segmented images based on the standardized multi-view image set and enters the simple parallax matching and image output stage. The image output module, in the simple parallax matching and image output stage, performs coarse-grained parallax matching processing on the left and right eye panoramic images based on a preset global fixed parallax offset threshold, correcting the binocular image parallax deviation, and directly outputs the matched left and right eye images to the display pipeline.

[0029] In this embodiment, the system is configured with four core functional modules: a multi-dimensional semantic parsing and feature generation module, a dynamic color tolerance adjustment module, a texture mapping optimization judgment and execution module, and a screen output module. Each module has a clear hierarchical division of labor, tight logical connections, and a closed-loop data flow. Through the collaborative linkage between modules, the entire process of automated intelligent generation—from user command input, intelligent feature encoding, adaptive optimization of rendering quality, intelligent judgment of 3D reconstruction branches, to the final VR binocular screen output—is achieved. This effectively solves the technical shortcomings of traditional VR 3D generation systems, such as high module coupling, inability to dynamically balance image quality and frame rate, single reconstruction strategy, poor adaptability to high and low computing power scenarios, and lack of lightweight output stereoscopic effects. The specific module implementation and collaborative working method are as follows: The multi-dimensional semantic parsing and feature generation module, as the core module for system front-end input and feature encoding, is responsible for undertaking the initial data input and feature structuring processing of the entire system. Specifically, it is used to obtain the target VR scene generation instructions actively input by the user terminal in real time. These generation instructions are compatible with both natural language text description instructions and external reference 2D image instructions. This module is designed to adapt to diverse user scenario customization needs. Internally, it uses a built-in multimodal semantic parsing network to perform comprehensive and multi-dimensional deep semantic decomposition, semantic alignment, and feature deep encoding on unstructured user-generated instructions. It accurately decomposes and extracts a complete set of multi-dimensional scene information corresponding to the target VR 3D scene. This includes spatial structure information used to constrain the overall layout of the scene, object geometry, and spatial topology; texture material information used to limit the surface texture style, material properties, texture resolution, and texture distribution patterns; light and shadow environment information used to characterize scene illumination intensity, light source direction, color temperature and hue, shadow levels, and ambient light atmosphere; and stereoscopic parallax parameter information specifically adapted to VR binocular immersive display needs and used to constrain the basic rules of binocular imaging. The module performs noise filtering, dimensional unification, cross-modal feature alignment, and structural reorganization on the extracted heterogeneous scene information. Finally, it generates standardized structured generated feature vectors that are dimensionally regular, semantically accurate, and adapted to the input specifications of AI large models. This provides standardized pre-feature input support for subsequent image generation and rendering optimization of AI large models.The dynamic color tolerance adjustment module, as the core control module for dynamically balancing the system's rendering quality and frame rate, is electrically and data-connected to the backend of the multi-dimensional semantic parsing and feature generation module. It receives standardized, structured generated feature vectors from the frontend and inputs them completely into a pre-trained and converged multimodal 3D generation AI model. By real-time collection of core parameters such as the current rendering sampling frequency, frame rendering load, and target frame rate, it intelligently determines whether the dynamic color tolerance adaptive adjustment mechanism needs to be activated under the current rendering conditions. This achieves optimal bidirectional adaptation between rendering quality and computing power / frame rate. When it is determined that dynamic color tolerance adaptive adjustment needs to be executed, the module, based on the negative correlation between sampling frequency and color deviation tolerance threshold, adjusts the preset color tolerance threshold. The adaptive adjustment function dynamically updates the color tolerance threshold and differentiates the rendering sampling precision of each region of the image. After completing the image quality optimization and adjustment, it drives the AI ​​large model to output a VR panoramic surround view multi-view initial 2D image sequence with uniform image quality and high color consistency. When it is determined that color tolerance adjustment is not required, the module directly drives the AI ​​large model to output a panoramic surround view multi-view initial 2D image sequence and automatically performs standardized preprocessing operations such as pixel noise reduction, lens distortion correction, cross-view coordinate alignment and global color mean unification on the initial image sequence. Finally, a standardized multi-view image set with unified view semantics, consistent color benchmark and no obvious imaging defects is obtained, providing high-quality image data support for subsequent 3D reconstruction or lightweight image generation. The texture mapping optimization judgment and execution module, as the core branch scheduling module of the system's layered rendering strategy, connects to the output of the dynamic color tolerance adjustment module. Based on the data quality, geometric accuracy, and texture integrity of the pre-processed standardized multi-view image set, it adaptively and intelligently determines whether the system should execute the high-precision 3D spatial structure reconstruction and texture mapping optimization process, achieving intelligent switching between high and low precision rendering strategies. When the system detects that the geometric consistency error of the multi-view images meets the standard, the texture confidence is sufficient, and the scene data quality is excellent, it determines to execute the complete 3D reconstruction and texture optimization process, sequentially completing depth estimation, point cloud optimization, mesh reconstruction, topology repair, optimal viewpoint texture extraction, multi-band texture fusion, and global texture mapping. Seamless mapping generates a complete VR 3D scene model with high-precision textures. After completing the detailed scene construction, it automatically jumps to the binocular stereo calibration and VR adaptation rendering steps to achieve high-precision immersive stereo output. When the system detects data defects such as sparse viewpoints, excessive geometric matching errors, and blurred or missing textures in multi-view images, and determines that it is not suitable to directly perform high-precision 3D reconstruction and texture mapping optimization, it automatically skips the complex 3D modeling and texture optimization process, and directly completes panoramic stitching and binocular segmentation processing based on the existing standardized multi-view image set. It quickly generates VR panoramic binocular segmented images and automatically enters the lightweight processing stage of simple parallax matching and image output, which greatly reduces the operating cost of low-end computing power equipment.The image output module, as the final execution module driving the system's imaging and display, interfaces with the lightweight branch output of the texture mapping optimization judgment and execution module. It is specifically designed for rapid imaging processing in the simple parallax matching and image output stages. The module relies on a preset, globally fixed parallax offset threshold, jointly calibrated by the VR device's interpupillary distance, spherical resolution, and comfortable viewing distance for the human eye. Using the left-eye panoramic viewport image as a fixed reference and the right-eye panoramic viewport image as the adjustment target, it performs uniform horizontal pixel offset coarse-grained parallax matching processing on both the left and right eye panoramic images. By establishing a stable and uniform basic binocular parallax relationship through a fixed pixel offset, it... The system extends and fills the edge blank areas caused by pixel offset to ensure the integrity of the panoramic image without any defects. This effectively corrects the problems of binocular parallax deviation and weak stereoscopic effect in lightweight panoramic images. Finally, the compliant binocular image pairs, after coarse-grained parallax matching and image integrity repair, are directly connected to the VR terminal display pipeline to complete the rapid image output. While ensuring the basic stereoscopic immersive viewing experience, it realizes the display of lightweight VR scenes with low latency and low computing power. This allows the system to be adapted to both high-precision 3D rendering scenes of high-end devices and lightweight panoramic display scenes of low-end devices, greatly improving the system's hardware adaptability and scene universality.

[0030] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the scope and intent of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and variations.

Claims

1. A VR 3D stereoscopic graphics intelligent generation method based on AI large-scale models, characterized in that, Includes the following steps: The system acquires user-inputted instructions for generating the target VR scene and performs multi-dimensional semantic parsing to extract multi-dimensional information from the target VR 3D scene, generating a standardized structured feature vector. The standardized structured feature vector is input into the pre-trained AI large model to determine whether to perform dynamic color tolerance adaptive adjustment to optimize rendering quality and frame rate. If yes, the initial two-dimensional image sequence of VR panoramic surround view is output after adjustment. If no, the initial two-dimensional image sequence of VR panoramic surround view is directly output and preprocessed to obtain a standardized multi-view image set. Based on a standardized multi-view image set, it is determined whether to perform 3D spatial structure reconstruction and texture mapping optimization. If so, after structure reconstruction and texture mapping optimization, it jumps to the binocular stereo correction and VR adaptation rendering steps. If not, it directly generates VR panoramic binocular segmented images by stitching together the standardized multi-view image set and enters the simple parallax matching and image output stage. In the simplified parallax matching and image output stage, based on the preset global fixed parallax offset threshold, coarse-grained parallax matching processing with horizontal pixel offset is performed on the panoramic images of the left and right eyes to correct the parallax deviation of the binocular images, and the matched left and right eye images are directly output to the display pipeline.

2. The VR 3D stereoscopic graphics intelligent generation method based on AI large model as described in claim 1, characterized in that: The specific process for determining whether to perform dynamic color tolerance adaptive adjustment is as follows: Obtain the real-time rendering parameters of the current VR scene, including the current image rendering sampling frequency, the current frame rendering load, and the target frame rate; The image rendering sampling frequency is used to characterize the number of sampling points for discretizing the continuous three-dimensional scene signal per unit area; The image color deviation tolerance threshold is used to characterize the maximum allowable color difference between the rendered pixel color and the theoretical real color when judging the color quality of the rendered image. The image color deviation tolerance threshold has a negative correlation with the image rendering sampling frequency. Calculate the predicted color deviation value at the current sampling frequency and compare the predicted color deviation value with the current color deviation tolerance threshold; If the predicted color deviation value is greater than the current color deviation tolerance threshold, it is determined that dynamic color tolerance adaptive adjustment is required. The dynamic color tolerance adaptive adjustment specifically includes: calculating an updated color deviation tolerance threshold according to the ratio of the current sampling frequency to the target sampling frequency, and the adaptive adjustment function is a monotonically decreasing function. The updated color deviation tolerance threshold is input into the rendering pipeline scheduling module of the AI ​​large model. The rendering pipeline scheduling module dynamically adjusts the rendering sampling frequency of each image region according to the updated color deviation tolerance threshold, specifically: For image regions where the color deviation is below the updated tolerance threshold, reduce the rendering sampling frequency to reduce computational overhead; for image regions where the color deviation is above the updated tolerance threshold, increase the rendering sampling frequency to ensure image quality.

3. The VR 3D stereoscopic graphics intelligent generation method based on AI large model as described in claim 1, characterized in that: The specific process for determining whether to perform 3D spatial structure reconstruction and texture mapping optimization is as follows: Obtain disparity maps and depth maps of images from each viewpoint in a standardized multi-view image set, and calculate the geometric consistency error and texture confidence between each viewpoint; the geometric consistency error is used to characterize the reprojection error of the projection position of the same three-dimensional point in different viewpoint images, and the texture confidence is used to characterize the clarity and recognizability of texture details in each viewpoint image. If the geometric consistency error is less than a preset geometric consistency threshold and the texture confidence is greater than a preset texture confidence threshold, then it is determined that three-dimensional spatial structure reconstruction and texture mapping optimization will be directly performed, specifically including the following sub-steps: S100, based on a standardized multi-view image set, performs multi-view stereo matching through the 3D reconstruction module of the AI ​​large model, calculates the disparity value of each pixel and generates a global depth map; the 3D reconstruction module adopts a depth estimation network based on cost volume construction and cost aggregation, projects the feature maps of each view image onto the 3D space through homography transformation to construct a 3D cost volume, and generates a high-precision depth map through cost aggregation and depth regression. S200 maps the global depth map to three-dimensional space through back projection transformation, generating an initial three-dimensional point cloud; S300, Filtering and downsampling are performed on the initial three-dimensional point cloud to remove outlier noise points and simplify the point cloud to obtain an optimized point cloud; S400, Based on the optimized point cloud, an initial three-dimensional mesh model is generated using the Poisson surface reconstruction algorithm; S500, perform mesh optimization processing on the initial three-dimensional mesh model. The mesh optimization processing includes non-manifold edge repair, hole filling and mesh smoothing to obtain a three-dimensional mesh model with a complete topological structure.

4. The intelligent generation method for VR 3D stereoscopic graphics based on AI large model as described in claim 3, characterized in that: The determination to directly perform three-dimensional spatial structure reconstruction and texture mapping optimization also includes: S600, perform texture mapping optimization on the 3D mesh model. The texture mapping optimization includes: calculating the optimal texture source view for each triangular facet based on the visibility of each view image of the 3D mesh model and the viewing angle between each view image in the standardized multi-view image set, wherein the optimal texture source view is the view with the smallest viewing angle and the highest texture clarity; and extracting texture pixels from the view image based on the optimal texture source view to generate the initial texture fragments of each triangular facet. S700, performs color consistency fusion and seam elimination processing on the initial texture fragments of each triangular facet. The seam elimination processing adopts a multi-band fusion algorithm based on Poisson editing to smoothly transition the color differences of texture fragments from different perspectives at the boundary, generating a global seamless texture map; the global seamless texture map is mapped to the three-dimensional mesh model to obtain a textured VR three-dimensional scene model. If the geometric consistency error is greater than or equal to the preset geometric consistency threshold, or the texture confidence is less than or equal to the preset texture confidence threshold, then it is determined that the three-dimensional spatial structure reconstruction and texture mapping optimization will not be performed, and the multi-view image enhancement and view interpolation optimization will be performed first. S800, through the view synthesis module of the AI ​​large model, performs sparse view densification processing based on a standardized multi-view image set; the view synthesis module adopts a new view synthesis network based on neural radiation field or diffusion model, takes any two adjacent view images and their corresponding camera pose parameters as input, samples along the interpolation path between the two view to generate at least one intermediate view synthesized image and its corresponding synthesized depth map, iterates this process until the angular spacing between any adjacent view is less than the preset angular spacing threshold, and obtains the enhanced dense multi-view image set.

5. The intelligent generation method for VR 3D stereoscopic graphics based on AI large model as described in claim 4, characterized in that: The specific steps for performing multi-view image enhancement and viewpoint interpolation optimization are as follows: S900, Perform inter-view color consistency calibration on the enhanced dense multi-view image set. The color consistency calibration includes: using the color distribution of the reference view image as a standard, performing color histogram matching on all view images and the composite image to obtain a color-aligned dense multi-view image set. S1000: Based on the dense multi-view image set with color alignment, recalculate the geometric consistency error and texture confidence. If the recalculated geometric consistency error is less than the preset geometric consistency threshold and the texture confidence is greater than the preset texture confidence threshold, then jump to step S100 to continue to perform three-dimensional spatial structure reconstruction and texture mapping optimization. If, after the enhancement processing in steps S800 to S1000, the recalculated geometric consistency error is still greater than or equal to the preset geometric consistency threshold, or the texture confidence is still less than or equal to the preset texture confidence threshold, then it is determined that generative filling-based 3D content completion will be performed, specifically including: The generative completion module of the AI ​​large model identifies the missing or blurred regions in the standardized multi-view image set. Guided by the corresponding regions in the adjacent view images, the completed view images are generated to replace the image content of the original missing or blurred regions, and the completed standardized multi-view image set is obtained. Then, the process jumps to step S100 to perform three-dimensional spatial structure reconstruction and texture mapping optimization. The final textured VR 3D scene model is input into the VR scene rendering output module to generate stereoscopic rendering data adapted to the VR display device and output to the user terminal.

6. The intelligent generation method for VR 3D stereoscopic graphics based on AI large model as described in claim 1, characterized in that: The binocular stereo correction and VR adaptation rendering steps specifically include: The textured VR 3D scene model obtained after 3D spatial structure reconstruction and texture mapping optimization, as well as the current viewpoint position and orientation information in the virtual space, are obtained. The current viewpoint position and orientation information include the left eye viewpoint position and the right eye viewpoint position. The distance between the left eye viewpoint position and the right eye viewpoint position is equal to the preset human interpupillary distance value. Based on the current viewpoint position and orientation information, a left-eye virtual camera and a right-eye virtual camera are set in the VR 3D scene model respectively; The VR 3D scene model is rendered using the left-eye virtual camera to generate the original left-eye rendered image; the same scene is rendered synchronously using the right-eye virtual camera to generate the original right-eye rendered image; the original left-eye rendered image and the original right-eye rendered image constitute the original binocular image pair; during the rendering process, the disparity of each pixel is calculated based on the depth value of each object in the VR 3D scene model, and the disparity is the lateral offset of the projection position of the same spatial point in the original left-eye rendered image and the original right-eye rendered image. Perform stereo correction on the original binocular images, specifically: perform lens distortion inverse compensation processing on the original rendered images of the left eye and the right eye respectively. The lens distortion inverse compensation processing applies pre-distortion to the original rendered images according to the optical lens characteristic parameters of the VR display device. The distortion direction of the pre-distortion is opposite to the distortion direction generated by the optical lens of the VR display device. The disparity range is optimized for the left and right eye images after binocular stereo correction. Specifically, for pixels with disparity greater than or equal to a preset disparity threshold, their disparity is compressed towards zero disparity to within the preset maximum comfortable disparity threshold; for pixels with disparity less than the preset disparity threshold, their disparity is stretched towards zero disparity to above the preset minimum comfortable disparity threshold.

7. The VR 3D stereoscopic graphics intelligent generation method based on AI large model as described in claim 6, characterized in that: The binocular stereo correction and VR adaptation rendering steps also include: The parallax-comfortable binocular images are input into the VR adaptation rendering pipeline. The VR adaptation rendering pipeline includes: scaling the left-eye and right-eye images according to the panel resolution of the VR display device to generate left-eye and right-eye display images adapted to the display panel resolution; performing sub-pixel rendering mapping on the left-eye and right-eye display images according to the sub-pixel arrangement of the VR display device to generate display data for each sub-pixel channel driving the VR display panel; performing asynchronous time warp or asynchronous spatial warp processing on the display data; and when a change in the user's head posture is detected, performing a reprojection transformation on the current frame image based on the depth map and color map rendered in the previous frame to generate an interpolated compensation frame. The left-eye and right-eye display images, after being processed by the VR adaptation rendering pipeline, are frame-synthesized and encapsulated into a stereoscopic frame data stream according to the transmission protocol of the VR display device. The stereoscopic frame data stream is then output to the VR display device for display through a video transmission interface. The stereoscopic frame data stream is synchronized with the user's interactive input, and the virtual camera parameters for the next frame are updated according to the user's real-time viewpoint changes. A standardized multi-view image set is stitched together panoramically according to the acquisition pose parameters of each viewpoint to generate a panoramic background image with cube mapping or equidistant cylindrical projection. Using this panoramic background image as static VR scene content, the viewport areas corresponding to the left and right eyes are cropped according to the current user's left and right eye viewpoint positions, respectively, to generate a left-eye panoramic viewport image and a right-eye panoramic viewport image. Simple parallax matching is performed on objects within the overlapping areas of the left-eye and right-eye panoramic viewport images. Specifically, local parallax offset is calculated based on optical flow field or block matching algorithms, and the left and right eye viewport images are adjusted according to the local parallax offset to generate a simple binocular image with basic stereoscopic effect. The simple binocular image is then directly output to a VR display device for display.

8. The intelligent generation method for VR 3D stereoscopic graphics based on AI large model as described in claim 1, characterized in that: The specific process for coarse-grained disparity matching processing with horizontal pixel offset is as follows: Acquire the left-eye panoramic viewport image and the right-eye panoramic viewport image, set the left-eye panoramic viewport image as the reference image and keep it unchanged, and set the right-eye panoramic viewport image as the target image and use it as the object for parallax adjustment. Obtain a preset global fixed disparity offset threshold. Based on the global fixed disparity offset threshold, perform horizontal pixel offset processing on each pixel in the right-eye panoramic viewport image. Specifically, in the horizontal direction, uniformly increase or decrease the number of pixels specified by the global fixed disparity offset threshold in the horizontal coordinate of each pixel in the right-eye panoramic viewport image to generate a right-eye panoramic offset image, so that a constant horizontal disparity offset is formed between the right-eye panoramic offset image and the left-eye panoramic viewport image in the entire image range. Edge filling processing is performed on the blank areas at the edges of the right-eye panoramic offset image generated after pixel offset. Specifically, the pixels inside the boundary of the right-eye panoramic offset image are copied and extended outward along the boundary to fill the blank areas at the edges, so as to maintain the complete resolution of the right-eye panoramic offset image.

9. The intelligent generation method for VR 3D stereoscopic graphics based on AI large model as described in claim 8, characterized in that: The coarse-grained disparity matching processing for horizontal pixel offset also includes: Obtain preset parallax offset direction parameters, which are set according to the stereoscopic display mode of the VR display device, including positive parallax mode and negative parallax mode; the positive parallax mode shifts the right-eye image to the right, making objects appear recessed into the screen; the negative parallax mode shifts the right-eye image to the left, making objects appear bulging out of the screen; determine the offset direction in the horizontal pixel offset processing based on the parallax offset direction parameters; The right-eye panoramic offset image after pixel offset processing is combined with the left-eye panoramic viewport image that remains unchanged to form a coarse-grained matching binocular image pair. The disparity relationship between the left and right eye images in the coarse-grained matching binocular image pair is uniquely determined only by the globally fixed disparity offset threshold.

10. A system applying the intelligent generation method for VR 3D stereoscopic graphics based on AI large models as described in any one of claims 1-9, characterized in that, include: The module includes a multi-dimensional semantic parsing and feature generation module, a dynamic color tolerance adjustment module, a texture mapping optimization judgment and execution module, and a screen output module. The multi-dimensional semantic parsing and feature generation module is used to obtain the target VR scene generation instructions input by the user and perform multi-dimensional semantic parsing to extract multi-dimensional information of the target VR three-dimensional scene and generate a standardized structured generation feature vector. The dynamic color tolerance adjustment module is used to input the standardized structured generated feature vector into the pre-trained AI large model, determine whether to perform dynamic color tolerance adaptive adjustment to optimize rendering quality and frame rate. If yes, it outputs the initial two-dimensional image sequence of the VR panoramic surround view after adjustment. If no, it directly outputs the initial two-dimensional image sequence of the VR panoramic surround view and performs preprocessing to obtain a standardized multi-view image set. The texture mapping optimization judgment and execution module is used to determine whether to perform three-dimensional spatial structure reconstruction and texture mapping optimization based on a standardized multi-view image set. If yes, it will jump to the binocular stereo correction and VR adaptation rendering steps after structure reconstruction and texture mapping optimization. If no, it will directly generate VR panoramic binocular segmented images based on the standardized multi-view image set and enter the simple parallax matching and image output stage. The image output module is used to perform coarse-grained parallax matching processing on the horizontal pixel offset of the left and right eye panoramic images in the simple parallax matching and image output process, based on a preset global fixed parallax offset threshold, to correct the parallax deviation of the binocular images, and directly output the matched left and right eye images to the display pipeline.

Images