Method for generating stylized textures for three-dimensional objects

Abstract

The present application relates to the technical field of three-dimensional computer graphics image generation, and discloses a three-dimensional object stylized texture generation method, comprising: generating a high-resolution prior view by denoising an image using a latent diffusion model; constructing a feature correlation tensor according to a grid model to reversely determine the spatial coordinates and normal vectors of invalid pixel points in a texture map; constructing a distance operator using the coordinates and normal vectors and driving feature propagation under topological constraints to adjust a weight factor and update color parameters; and calculating loss to optimize map parameters through a differential rendering loop, wherein the present application anchors the repair area to a three-dimensional manifold geometry field, constructs a feature damping barrier at island boundaries and wrinkles, eradicates cross-border leakage and pattern adhesion, and guarantees the spatial continuity of the surface material texture of a complex grid model.

Description

Technical Field

[0001] This invention belongs to the field of three-dimensional computer graphics and image generation technology, and particularly relates to a method for generating stylized textures for three-dimensional objects. Background Technology

[0002] Currently, pre-trained diffusion models are used to generate multi-view style images, which are then written into the mesh model coordinate system using back projection matrices to construct multi-dimensional textures. This is a conventional technique. This approach relies on the global semantic association of the generated image fields under different viewing perspectives to maintain the texture continuity of the spatial geometric surface. Its physical mechanism depends on the geometric mapping between two-dimensional pixel coordinates and three-dimensional manifold elements. The spatial depth and color of vertices are projected onto the control channel of the discrete plane through the view frustum transformation matrix, so that the prior data of the multi-dimensional images converge into a unified texture flow of the mesh surface. However, complex mesh models contain a large number of mutually occluding concave cavities and folded thin walls. The structure and multi-view camera sampling stream cannot achieve full coverage of the mesh model surface, resulting in observation blind spots containing noise in such spatial dead zones. At the same time, when unfolding the continuous three-dimensional manifold surface into a two-dimensional texture space, the transformation cuts off the original spatial proximity of the physical surface, generating multiple isolated texture islands and introducing geometric stretching and topological breaks at the edges. When the conventional neighborhood pixel extrapolation method is used to fill the observation blind spots, the technical path assumes that the proximity of the planar texture is equivalent to the proximity of the physical manifold surface, causing the calculation logic to ignore the intrinsic geometric features of the manifold surface, causing color channel data to escape across the island boundaries into the anisotropic normal region.

[0003] Industry experts have attempted to cover blind spots by increasing the number of viewing angles of the observation camera or improving the resolution of texture maps. However, such adjustments increase system sampling overhead and computational load, and cannot overcome the topological barrier of the opposing normals on both sides of the folded wall. They also struggle to resolve texture seam misalignment caused by unfolding and stretching. Changing the viewpoint arrangement and other hardware sampling configurations further increase system computational load and acquisition costs, and are difficult to resolve seam misalignment when facing strongly twisted or folded walls. Software methods such as post-processing data reconstruction also have shortcomings in addressing blind spot repair and texture alignment. For example, Chinese invention patent application CN116863101A discloses a method based on adaptive mesh refinement... The proposed reconstruction model geometry and texture optimization method utilizes generative adversarial networks (GANs) for texture regeneration. It identifies artifacts through a discriminator and drives the generator to refill pixels. However, this end-to-end image reconstruction based on GANs relies on implicit distribution matching or two-dimensional image spatial interpolation. The data calculation is detached from the physical manifold topological constraints and local curvature abrupt changes of the three-dimensional mesh. When faced with complex and non-ideal geometric boundary conditions such as dense folds and deep cavities, due to the lack of explicit geometric field continuity barriers, the neural network model is prone to texture blurring and non-physical color adhesion at hard edges with anisotropic normals, resulting in cross-boundary leakage of features at the seams.

[0004] Therefore, how to convert planar texture coordinates into three-dimensional manifold topological geometric constraints, and simultaneously repair the observation blind zone and constrain the feature dissipation across the seams in the multi-view denoising convergence flow, has become the technical problem to be solved by this invention. Summary of the Invention

[0005] This invention aims to solve the problem of cross-boundary leakage in the rendered image field caused by the smooth expansion of planar pixels cutting off the topological continuity of three-dimensional manifolds.

[0006] In this technical solution, a method for generating stylized textures for three-dimensional objects includes: Step S1: Obtain the 3D mesh model and initial guiding semantics, and perform two-stage denoising processing through the latent diffusion model. In the first stage, high-frequency noise suppression processing is performed using spatial topological features. In the second stage, high-resolution image reconstruction processing is performed by incorporating detailed features through the upsampling module to generate a high-resolution prior view. Step S2: Based on the topological adjacency matrix of the three-dimensional mesh model, construct a geometric feature association tensor that has a pixel mapping relationship with the high-resolution two-dimensional texture map. Based on the barycentric coordinate interpolation, reversely retrieve and determine the spatial coordinate vector and normalized normal vector of the invalid pixel in the high-resolution two-dimensional texture map on the geometric surface of the three-dimensional mesh model. Step S3: Construct a comprehensive structural distance operator using spatial coordinate vectors and normalized normal vectors, transforming planar neighborhood expansion into feature value propagation under three-dimensional mesh space topological constraints. When the local curvature change rate exceeds the set curvature threshold, adjust the geometric weight factor of the comprehensive structural distance operator and update the color feature parameters of invalid pixels. Step S4: Using the high-resolution prior view as the prior reference image, calculate the multi-view loss through the differential rendering loop, update the color parameters of the high-resolution two-dimensional texture map, and generate a stylized texture map of the target object.

[0007] Preferably, the two-stage denoising process using the latent diffusion model in step S1 includes: obtaining the initial latent space image corresponding to the initial guiding semantics, calculating the global gradient vector of the initial latent space image in the diffusion model, and decomposing the global gradient vector into low-frequency topological components and high-frequency detail components; in the first stage, denoising is performed using the low-frequency topological components; in the second stage, the feature map processed in the first stage is input into the upsampling module, which consists of multiple transposed convolutional layers, and the high-frequency detail components are fused for denoising to generate a high-resolution prior view.

[0008] Preferably, step S2, which involves reversely retrieving and determining the spatial coordinate vectors and normalized normal vectors of invalid pixels in the high-resolution 2D texture map on the geometric surface of the 3D mesh model based on centroid coordinate interpolation, includes: retrieving invalid pixel regions in the high-resolution 2D texture map based on geometric feature association tensors, and mapping the invalid pixel regions onto polygonal patches of the 3D mesh model; and using the spatial coordinate vectors and normalized normal vectors of the polygonal patch vertices, performing linear interpolation through centroid coordinates to determine the spatial coordinate vectors and normalized normal vectors corresponding to the invalid pixels.

[0009] Preferably, in step S3, when the local curvature change rate exceeds the set curvature threshold, adjusting the geometric weight factor of the integrated structural distance operator and updating the color feature parameters of invalid pixels includes: detecting the local curvature change rate of the region where the invalid pixel is located; when the local curvature change rate exceeds the set curvature threshold of 0.05, adjusting the geometric weight factor in the integrated structural distance operator from 0.5 to 2.0, increasing the weight value of pixels with high surface normal direction deviation in the integrated structural distance operator, and restricting the mixing of color features across the mesh edge.

[0010] Preferably, in step S4, calculating the multi-view loss through a differential rendering loop and updating the parameters of the high-resolution 2D texture map includes: projecting the high-resolution 2D texture map onto multiple discrete 3D viewing angles using a differential renderer to obtain multiple multi-view rendered views; calculating the absolute color deviation and perceptual structure similarity loss between the multi-view rendered views and the high-resolution prior views to construct a composite view loss function; and using a backpropagation mechanism to inversely project the gradient of the composite view loss function onto the parameter plane of the high-resolution 2D texture map to update the feature channel color values ​​of each pixel in the high-resolution 2D texture map.

[0011] Preferably, in step S1, the denoising process performed in the first stage using spatial topological features and the denoising process performed in the second stage by incorporating detail features through an upsampling module includes: in the early stage of denoising, the time step is set to be in the low-frequency semantic range, and the initial latent spatial image is denoised using the denoising network of the latent diffusion model; in the later stage of denoising, the time step is set to be in the high-frequency detail range, the feature map is spatially upgraded using an upsampling module, and detail texture denoising is performed through a spatial feature fusion attention mechanism.

[0012] Preferably, step S3, which converts planar neighborhood expansion into feature value propagation under the topological constraints of a three-dimensional mesh space, includes: establishing the surface topological neighborhood distance and local surface normal direction deviation between invalid pixels and valid neighboring pixels determined by the geometric feature association tensor on the geometric surface of the three-dimensional mesh model based on the topological adjacency matrix of the facets; using the surface topological neighborhood distance and local surface normal direction deviation to jointly constrain the weight allocation of color features, and restricting feature value propagation at the boundaries of texture regions and mesh wrinkles.

[0013] Preferably, step S2, based on the topological adjacency matrix of the 3D mesh model, constructs a geometric feature association tensor that has a pixel-to-pixel mapping relationship with the high-resolution 2D texture map, including the following steps: Step S21, analyze the polygonal topology of the 3D mesh model, extract the adjacency topology relationships between patches and between vertices, and construct a mesh topological adjacency matrix; Step S22, establish a barycentric coordinate mapping table between the ultraviolet coordinate space of the high-resolution 2D texture map and the geometric surface of the 3D mesh model; Step S23, combine the mesh topological adjacency matrix and the barycentric coordinate mapping table to assign each pixel in the high-resolution 2D texture map its spatial coordinate vector and normalized normal vector on the geometric surface of the 3D mesh model, forming a third-order geometric feature association tensor.

[0014] Preferably, step S4, using the high-resolution prior view as the prior reference image, includes: optimizing the high-resolution prior view as the prior reference image; assigning fixed weight coefficients to each pixel of the high-resolution prior view when calculating the multi-view loss in the differential rendering loop; and updating the color parameters of the high-resolution two-dimensional texture map through multi-view loss gradient backpropagation.

[0015] Compared with existing technologies, the three-dimensional object stylized texture generation method of the present invention has the following advantages: 1. In the stylized texture generation of 3D objects, depth features generated by rendering the vertex coordinates of the 3D model are introduced into the multi-view diffusion process. Combined with the control channel that constrains the geometric configuration of the 3D object surface, the 30-step denoising process is divided into 25 steps of original denoising and 5 steps of alternating upsampling denoising. A fast decoder with a fixed linear transformation matrix is ​​used to transform the latent features to the pixel space. After the resolution is improved, the latent space is returned by transformation for continuous denoising. This ensures that the high-frequency material texture features are consistent with the global semantic structure of the multi-view, avoids the high-resolution view from becoming disconnected from the latent space during generation, and avoids pixel misalignment. It also stably controls the distribution of spatial frequencies in the high-resolution prior view.

[0016] 2. Before the texture diffusion stage, based on the topological adjacency matrix of the 3D mesh model, a geometric feature association tensor with a pixel-to-pixel mapping relationship with the high-resolution 2D texture map is pre-constructed. Based on the centroid coordinate interpolation, the spatial coordinate vector and normalized normal vector of invalid pixels on the 3D physical surface are retrieved and determined in reverse. The planar topological proximity expansion is transformed into adaptive texture propagation under the constraints of manifold geometric space. The physical space absolute distance and local normal convergence are combined into a comprehensive structural distance operator. Physically impermeable feature damping barriers are automatically constructed at the boundaries of isolated 2D texture islands and thin-walled geometric folds to eliminate the seam breaks induced by texture cross-boundary leakage.

[0017] 3. By real-time detection of the local curvature change rate of invalid pixel nodes, when the local curvature change rate exceeds the preset sensitivity threshold of 0.05, the region is determined to be a high-frequency topological cutoff point. This spontaneously triggers a step-by-step adjustment based on curvature gradient, which linearly amplifies the geometric weight control factor in the integrated structural distance operator from a fixed 0.5 to 2.0 in real time. By strongly suppressing the participation of pixels with anisotropic surface normals in color mixing, the transmission of irrelevant texture features at geometric boundaries is blocked, achieving hard color cutoff protection at complex sharp edges and narrow cavities. This solves the problem of non-physical pattern adhesion that is prone to occur when traditional texture filling technology faces strongly distorted surfaces and folded walls. Attached Figure Description

[0018] Fig. 1 This is a flowchart illustrating the overall process of the three-dimensional object stylized texture generation method of the present invention. Fig. 2 This is a structural diagram of the execution workstation module of the three-dimensional object stylized texture generation method of the present invention; Fig. 3 This is a diagram illustrating the internal operating principle of the three-dimensional object stylized texture generation method of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0020] A method for generating stylized textures for 3D objects includes the following steps: Step S1: Obtain the 3D mesh model and initial guiding semantics, and perform two-stage denoising processing through the latent diffusion model. In the first stage, high-frequency noise suppression processing is performed using spatial topological features. In the second stage, high-resolution image reconstruction processing is performed by incorporating detailed features through the upsampling module to generate a high-resolution prior view. Step S2: Based on the topological adjacency matrix of the three-dimensional mesh model, construct a geometric feature association tensor that has a pixel mapping relationship with the high-resolution two-dimensional texture map. Based on the barycentric coordinate interpolation, reversely retrieve and determine the spatial coordinate vector and normalized normal vector of the invalid pixel in the high-resolution two-dimensional texture map on the geometric surface of the three-dimensional mesh model. Step S3: Construct a comprehensive structural distance operator using spatial coordinate vectors and normalized normal vectors, transforming planar neighborhood expansion into feature value propagation under three-dimensional mesh space topological constraints. When the local curvature change rate exceeds the set curvature threshold, adjust the geometric weight factor of the comprehensive structural distance operator and update the color feature parameters of invalid pixels. Step S4: Using the high-resolution prior view as the prior reference image, calculate the multi-view loss through the differential rendering loop, update the color parameters of the high-resolution two-dimensional texture map, and generate a stylized texture map of the target object.

[0021] Preferably, the two-stage denoising process using the latent diffusion model in step S1 includes: obtaining the initial latent space image corresponding to the initial guiding semantics, calculating the global gradient vector of the initial latent space image in the diffusion model, and decomposing the global gradient vector into low-frequency topological components and high-frequency detail components; in the first stage, denoising is performed using the low-frequency topological components; in the second stage, the feature map processed in the first stage is input into the upsampling module, which consists of multiple transposed convolutional layers, and the high-frequency detail components are fused for denoising to generate a high-resolution prior view.

[0022] Preferably, step S2, which involves reversely retrieving and determining the spatial coordinate vectors and normalized normal vectors of invalid pixels in the high-resolution 2D texture map on the geometric surface of the 3D mesh model based on centroid coordinate interpolation, includes: retrieving invalid pixel regions in the high-resolution 2D texture map based on geometric feature association tensors, and mapping the invalid pixel regions onto polygonal patches of the 3D mesh model; and using the spatial coordinate vectors and normalized normal vectors of the polygonal patch vertices, performing linear interpolation through centroid coordinates to determine the spatial coordinate vectors and normalized normal vectors corresponding to the invalid pixels.

[0023] Preferably, in step S3, when the local curvature change rate exceeds the set curvature threshold, adjusting the geometric weight factor of the integrated structural distance operator and updating the color feature parameters of invalid pixels includes: detecting the local curvature change rate of the region where the invalid pixel is located; when the local curvature change rate exceeds the set curvature threshold of 0.05, adjusting the geometric weight factor in the integrated structural distance operator from 0.5 to 2.0, increasing the weight value of pixels with high surface normal direction deviation in the integrated structural distance operator, and restricting the mixing of color features across the mesh edge.

[0024] Preferably, in step S4, calculating the multi-view loss through a differential rendering loop and updating the parameters of the high-resolution 2D texture map includes: projecting the high-resolution 2D texture map onto multiple discrete 3D viewing angles using a differential renderer to obtain multiple multi-view rendered views; calculating the absolute color deviation and perceptual structure similarity loss between the multi-view rendered views and the high-resolution prior views to construct a composite view loss function; and using a backpropagation mechanism to inversely project the gradient of the composite view loss function onto the parameter plane of the high-resolution 2D texture map to update the feature channel color values ​​of each pixel in the high-resolution 2D texture map.

[0025] Preferably, in step S1, the denoising process performed in the first stage using spatial topological features and the denoising process performed in the second stage by incorporating detail features through an upsampling module includes: in the early stage of denoising, the time step is set to be in the low-frequency semantic range, and the initial latent spatial image is denoised using the denoising network of the latent diffusion model; in the later stage of denoising, the time step is set to be in the high-frequency detail range, the feature map is spatially upgraded using an upsampling module, and detail texture denoising is performed through a spatial feature fusion attention mechanism.

[0026] Preferably, step S3, which converts planar neighborhood expansion into feature value propagation under the topological constraints of a three-dimensional mesh space, includes: establishing the surface topological neighborhood distance and local surface normal direction deviation between invalid pixels and valid neighboring pixels determined by the geometric feature association tensor on the geometric surface of the three-dimensional mesh model based on the topological adjacency matrix of the facets; using the surface topological neighborhood distance and local surface normal direction deviation to jointly constrain the weight allocation of color features, and restricting feature value propagation at the boundaries of texture regions and mesh wrinkles.

[0027] Preferably, step S2, based on the topological adjacency matrix of the 3D mesh model, constructs a geometric feature association tensor that has a pixel-to-pixel mapping relationship with the high-resolution 2D texture map, including the following steps: Step S21, analyze the polygonal topology of the 3D mesh model, extract the adjacency topology relationships between patches and between vertices, and construct a mesh topological adjacency matrix; Step S22, establish a barycentric coordinate mapping table between the ultraviolet coordinate space of the high-resolution 2D texture map and the geometric surface of the 3D mesh model; Step S23, combine the mesh topological adjacency matrix and the barycentric coordinate mapping table to assign each pixel in the high-resolution 2D texture map its spatial coordinate vector and normalized normal vector on the geometric surface of the 3D mesh model, forming a third-order geometric feature association tensor.

[0028] Preferably, step S4, using the high-resolution prior view as the prior reference image, includes: optimizing the high-resolution prior view as the prior reference image; assigning fixed weight coefficients to each pixel of the high-resolution prior view when calculating the multi-view loss in the differential rendering loop; and updating the color parameters of the high-resolution two-dimensional texture map through multi-view loss gradient backpropagation.

[0029] Example 1: In applications involving the generation of textures for 3D mesh models, the system faces challenges such as texture breaks and white space caused by viewpoint occlusion in local areas of the model. By utilizing diffusion-guided progressive upsampling texture optimization and the collaborative operation of a view-aware texture repair and refinement module, the system achieves quality repair of the 3D model texture mapping. The system acquires the 3D mesh model to be processed and the initial guiding semantics, and performs two-stage denoising processing through a latent diffusion model. In the first stage, the system utilizes spatial topological features to perform high-frequency noise suppression, controlling the denoising process within the low-frequency semantic range to maintain the stability of the global semantic structure of the 3D mesh model. In the second stage, the system utilizes... An upsampling module containing multiple transposed convolutional layers is used to incorporate high-frequency detail components for high-resolution image reconstruction, generating a high-resolution prior view. Specifically, during execution, the aforementioned spatial feature fusion attention mechanism linearly transforms the intermediate feature map output from the transposed convolutional layer to construct a query matrix Q. Simultaneously, the 3D spatial topological feature matrix derived from the vertex coordinates and patch indices of the 3D mesh model is mapped to the same feature dimension through a corresponding projection matrix, thereby generating a key matrix K and a value matrix V. The spatial correlation weights between the 2D pixel space and the 3D manifold geometry are captured by calculating the dot product of the query matrix Q and the transpose of the key matrix K. The attention weight matrix is ​​obtained by function normalization. Finally, the attention weight matrix is ​​multiplied by the value matrix V in a weighted manner, so as to accurately integrate the low-frequency manifold space topological skeleton features into the high-frequency detail channel of the two-dimensional image in the upsampling module, thereby realizing the detail texture denoising and reconstruction under the global semantic structure constraint.

[0030] The system constructs a geometric feature association tensor based on the topological adjacency matrix of the 3D mesh model. This tensor establishes a mapping relationship between high-resolution 2D texture map pixels and the coordinate vectors and normalized normal vectors of the 3D physical surface. Based on the barycentric coordinate interpolation rule, the system inversely searches for invalid pixels with missing pixels in the high-resolution 2D texture map and determines the spatial coordinate vectors and normalized normal vectors of these pixels on the 3D physical surface. The system uses the spatial coordinate vectors and normalized normal vectors to construct a comprehensive structural distance operator, transforming the original planar neighborhood expansion algorithm into a 3D mesh spatial topology algorithm. Under constraints, when the local curvature change rate of the region containing an invalid pixel exceeds a set threshold of 0.05, the system determines that the region is a high-frequency topological truncation and automatically increases the geometric weight factor in the integrated structural distance operator from 0.5 to 2.0. This enhances the weight of pixels with high deviation in the surface normal direction, thereby limiting the mixing of color features across mesh boundaries. The feature propagation is based on the manifold heterogeneous diffusion equation, treating the surface of the 3D mesh model as a heat conduction field. The damping of color feature diffusion is controlled by the angle between spatial distance and normal. The calculation formula for the integrated structural distance operator is as follows: ,in, Invalid pixels with a value greater than 0 with effective neighbor points The combined structural distance between them and This represents the coordinate vector of the corresponding pixel in the three-dimensional grid space. Let Euclidean distance be the spatial distance between the two. The spatial distance normalization control coefficient ranges from 0.1 to 1.0. and This is the normalized normal vector at the corresponding pixel. For the dot product of the normal directions taking values ​​from -1.0 to 1.0, This is the geometric weighting factor; when the local curvature change rate is 0.01, Maintaining the default value of 0.5, the characteristic values ​​diffuse uniformly across the manifold surface; when the local rate of curvature change increases to 0.08, Increasing the value to 2.0 reduces the weight of pixels with high surface normal deviation, limiting color propagation to regions with opposite normals. To ensure that the above operators accurately represent the geometric barrier intention in the numerical calculation of the discrete manifold heterogeneous diffusion equation, the comprehensive structural distance output by the above operators is used in the actual diffusion solution process. The transmission coefficient weight, explicitly defined as the propagation weight of color feature values ​​between two points, is calculated when the two points are on opposite sides of a thin-walled fold and their normal directions are completely opposite. Approaching -1, so that the terms in the operator As the value approaches 0, the numerical output of the entire operator approaches the minimum value of 0. In the physical conduction network of heterogeneous diffusion, the conduction coefficient approaching 0 is mathematically directly mapped to the manifold conduction impedance approaching infinity. Thus, without changing the mesh continuity, a physically impermeable zero-transmission characteristic damping barrier is constructed at the abrupt boundary of the anisotropic normal, achieving a closed-loop mechanism to block cross-boundary leakage.

[0031] The system uses a high-resolution prior view as a benchmark, calculates multi-view loss through a differential rendering loop, and uses a backpropagation mechanism to project the gradient backward onto the parameter plane of the high-resolution two-dimensional texture map, updates the feature channel color values ​​of each pixel, and performs nonlinear texture propagation on the observation blind area through the view-aware texture repair and refinement module to generate a complete texture map with continuous texture space.

[0032] Example 2: This experiment verifies the repair performance of the 3D object stylized texture generation method in dealing with viewpoint occlusion and curvature abrupt changes in texture maps. The experiment uses an industrial-grade graphics workstation equipped with a graphics processor with 24GB of video memory. A standard test dataset containing 50 3D mesh models is used as input, including test samples with high-frequency topological features and complex occlusion structures to ensure that the verification process covers the technical challenges under real-world conditions. The experiment design includes two sets of comparative samples. The control group uses a planar neighborhood feature filling scheme without curvature-guided weights. When dealing with sharp wrinkles on the model surface, the geometric weight factor is fixed at 0.5. The experimental group, according to the scheme of this application, detects local curvature changes in real time, and when the rate of change exceeds a set threshold of 0.05, the geometric weight factor is linearly adjusted from 0.5 to 2.0.

[0033] During the experiment, the system acquired the topology of the 3D mesh model, generated a high-resolution prior view using a pre-trained diffusion model, and back-projected the 2D texture map to the 3D observation view through a differential rendering loop. It then performed centroid coordinate interpolation calculation on the spatial coordinates and normal vectors of invalid pixels and constructed a comprehensive structural distance operator accordingly. In flat areas, the mean color deviation of texture pixels generated by the control group and the experimental group was controlled within 3%, both showing good smoothness.

[0034] In sharp edge regions with high-frequency topological truncation, the texture restoration results of the control group showed obvious color leakage. According to the image quality assessment model, the root mean square error of color at the edge was 12.4%, which visually manifested as texture features sticking together at the seam. In contrast, the experimental group triggered a dynamic compensation action of the geometric weight factor to 2.0 through the curvature change rate detection mechanism, which improved the control weight of pixels with high normal direction deviation in feature value propagation. Test data showed that the root mean square error of color in the experimental group in similar edge regions decreased to 2.8%, and the spatial continuity at the texture boundary was enhanced. When the local curvature change rate was less than 0.03, the geometric weight factor remained at the default setting of 0.5. At this time, the texture restoration process mainly relied on the spatial Euclidean distance, and the resulting texture smoothness met the normal visual requirements. When the local curvature change rate rose to 0.08, the system automatically performed a step-by-step increase in the geometric weight factor. The feature damping effect generated by this adjustment effectively suppressed color sampling interference across the grid boundary, confirming that the curvature threshold is the performance window for balancing smooth texture transition and maintaining sharp topological edges.

[0035] Example 3: This example combines Figs. 1 to 3 The method for generating stylized textures for 3D objects is explained, such as... Fig. 1 The diagram illustrates the overall process of this 3D object stylized texture generation method, which includes steps S1 to S4. Step S1 involves acquiring a 3D mesh model and initial guiding semantics, followed by two-stage denoising using a latent diffusion model. In the first stage, high-frequency noise is suppressed using spatial topological features. In the second stage, detailed features are incorporated through an upsampling module for image reconstruction, generating a high-resolution prior view. Step S2 involves constructing a geometric feature association tensor that maps pixels to the high-resolution 2D texture map based on the topological adjacency matrix of the 3D mesh model, and then interpolating based on the barycentric coordinates. Step S3 involves reversing the search to determine the spatial coordinate vector and normalized normal vector of invalid pixels on the mesh surface. Step S4 involves constructing a comprehensive structural distance operator using the spatial coordinate vector and normalized normal vector to convert planar neighborhood expansion into feature value propagation under spatial topological constraints. When the local curvature change rate exceeds the set curvature threshold, the geometric weight factor is adjusted, and the color feature parameters of invalid pixels are updated. Step S5 involves using the high-resolution prior view as the prior reference image, calculating the multi-view loss through the differential rendering loop, updating the color parameters of the high-resolution two-dimensional texture map, and generating a stylized texture map of the target object.

[0036] like Fig. 2As shown, the workstation module structure for executing this method is as follows: the left side of the diagram contains a data storage and input node, which integrates the 3D mesh model, initial guiding semantics, and standard test dataset. This input node establishes a one-way data transmission link with the core industrial-grade graphics workstation. The industrial-grade graphics workstation integrates a 24GB graphics processor, which sequentially loads and runs a latent diffusion model denoising network, a geometric feature association tensor construction module, a comprehensive structural distance operator calculation core, a differential rendering loop optimization module, a view-aware texture repair and refinement module, a high-resolution prior view, and an upsampling module. After completing the data processing, the industrial-grade graphics workstation outputs the results to the output target storage node on the right, and forms a stylized texture map of the target object within this node.

[0037] like Fig. 3 As shown, the internal operating principle mechanism of the method is illustrated. The top layer of the diagram depicts the two-stage denoising process of the synchronous multi-view diffusion model. Through an iterative denoising mechanism, the latent variable view is input into the U-net network and undergoes a serial processing sequence between the latent variable view and the U-net network. Through upsampling optimization, multi-view results with numbers from i=1 to i=N are generated. The bottom layer of the diagram presents the corresponding texture inpainting stage operation chain, which performs geometric combination processing between the mesh model without texture mapping and the object texture mapping, thereby rendering the corresponding rendered object images with numbers from i=1 to i=N. The system combines the view results generated at the top layer and the rendered object images generated at the bottom layer to calculate the multi-view loss. The obtained loss result triggers the texture parameter update operation and feeds it back to update the object texture mapping node at the bottom layer, which is then input into the view-aware texture inpainting refinement module, and finally drives the output of the repaired object texture mapping.

[0038] Example 4: In the scenario of texture mapping for high-frequency geometric features on the surface of complex 3D models, the system handles the color leakage at texture feature blind spots and topological boundaries caused by incomplete observation from multiple perspectives. When the 3D mesh model contains multiple non-continuous geometric patches and these patches form sharp angles in 3D space, the feature propagation algorithm is prone to causing color features to leak across physical edges, resulting in visual disjointed rendering results.

[0039] During texture generation, the system performs inverse projection calibration on pixels in the high-resolution 2D texture map based on the established mesh model topology. The system extracts the spatial coordinate vectors and normalized normal vectors corresponding to invalid pixels on the geometric surface of the 3D physical model, and constructs a comprehensive structural distance operator. This operator uses the Euclidean distance between invalid pixel nodes and their neighboring valid seed pixels in 3D space as a benchmark, and introduces a geometric weight factor for weighted adjustment. The system detects the local curvature change rate of the region where the invalid pixel node is located in real time and compares this value with a set threshold of 0.05. When the local curvature change rate exceeds 0.05, the system determines that the region is a high-frequency topological truncation point and adjusts the geometric weight factor from the initial value of 0.5 to 2.0. This action increases the control weight of pixels with high normal direction deviation in feature value propagation, and converges the feature propagation path within the mesh patch.

[0040] To address the second-order execution bias caused by data transmission and storage during texture mapping, the system employs a discretized color gradient compensation operator. Specifically, this operator operates by establishing a local difference operator matrix on the two-dimensional discrete mesh space of the texture map. When a boundary anomaly is detected due to accumulated error in the feature channel value of the current center pixel, the operator calculates the discrete color gradient vector between that pixel and its four adjacent effective pixels using first-order forward difference. It then uses the inverse of the color variance of all effective pixels in the current pixel's neighborhood as a dynamic weighting factor, performing a Hadamard product dot product on the inverse variance matrix and the calculated discrete color gradient vector. This adaptively calculates a compensation fine-tuning amount to offset the second-order execution bias within the discrete parameter space, which is then superimposed back into the original feature channel, eliminating the bias through a digital-level local smoothing filtering mechanism. To address discrete artifacts introduced by data mutations, specifically, this discretized color gradient compensation operator operates by establishing a local difference operator matrix on the two-dimensional discrete grid space of the texture map. When a boundary anomaly is detected due to accumulated error in the feature channel value of the current center pixel, the operator calculates the discrete color gradient vector between the pixel and its four adjacent effective pixels using first-order forward difference. It then uses the inverse of the color variance of all effective pixels in the current pixel's neighborhood as a dynamic weighting factor, performing a Hadamard product dot product on the inverse variance matrix and the calculated discrete color gradient vector. This adaptively calculates a compensation fine-tuning amount to offset second-order execution bias within the discrete parameter space and superimposes it back into the original feature channel, using a digital-level local smoothing filtering mechanism to eliminate discrete artifacts introduced by data mutations. The quantitative calculation formula for the color gradient compensation operator is as follows: in, This is a compensation adjustment amount used to offset second-order execution bias. This is the discrete color gradient vector obtained by first-order forward difference calculation. For the Hadama product operator, A dynamic weighted factor matrix system, constructed from the reciprocal of the color variance of effective pixels in the current pixel's neighborhood, monitors the numerical saturation of texture map feature channels. When a feature channel value is in the low-order range below 0.1 or the high-order range above 0.9, a numerical clamping action is performed, and the channel value is weighted and corrected with the color variance of neighboring pixels to eliminate color banding caused by numerical overflow. This compensation logic reduces color dispersion during texture generation by constructing a smooth buffer at the edges of feature channel values. When the local curvature change rate remains below 0.02, the geometric weights... With the factor maintained at an initial setting of 0.5, the texture restoration process mainly relies on spatial Euclidean distance propagation, and the texture pixel deviation is kept within an error range of 2.5%. When the local curvature change rate reaches 0.09, the system triggers the adjustment of the geometric weight factor to 2.0, and the root mean square error of color at the edge decreases from 12.4% in the control group to 2.8%, effectively avoiding the cross-boundary propagation of non-physical features. The above parameter settings and adjustment logic constitute the control mechanism for texture generation. Through curvature adaptive weighting and color numerical clamping, the spatial continuity and fidelity of the surface material texture of complex mesh models are guaranteed.

[0041] Example 5: In a real-world industrial application scenario of texture stylization rendering for high-load 3D complex mesh models, the system addresses the data processing load fluctuation problem caused by uneven density of geometric facets in the model. When the 3D mesh model contains more than a million faces and has densely wrinkled areas, the texture generation method experiences processing delays during pixel projection due to uneven allocation of computational units, causing the rendering cycle to exceed the time limit for real-time interaction. Before starting the texture generation task, the system pre-scans the input 3D mesh model and calculates the compactness index of the geometric distribution of facets in each region. The system divides the mesh model into several processing regions. When the facet density in any region exceeds 1.5 times the baseline preset value, resource pre-allocation logic is executed. Based on the number of available threads in the graphics processor stream processor, the texture resampling task in the high-density region is divided into four independent parallel processing batches, and an independent texture feature calculation buffer is allocated to each batch to balance the load of each parallel processing unit. With adaptive weighting, the computing system performs hardware-level thread scheduling, and the underlying calculation is based on the thread block mapping model of the parallel computing system. The processing unit reads the compactness index of the geometric distribution of facets in each region of the 3D mesh model. When the patch density in a specific region exceeds 1.5 times the baseline preset value, the underlying concurrent stream interface of the graphics processor is invoked to decompose the texture resampling task of the current region into four independent parallel processing threads, and independent video memory calculation buffers are allocated to balance the hardware load of the stream processor; during this process, the aforementioned compactness index... The quantization calculation path is as follows: For each independent processing region of the mesh model, the total number of polygonal faces contained within that region is counted, and the 3D bounding box volume of all faces in that region is calculated. The ratio of the total number of faces to the bounding box volume is defined as the absolute face packing density of that region. This ratio is then divided by the average distribution density of all faces in the entire 3D mesh model, and this normalized dimensionless ratio is used as the compactness index of that region. When a specific area When the value is greater than or equal to 1.5, logically it is directly equivalent to the geometric distribution density of the local area exceeding 1.5 times the baseline preset value. This accurately quantifies and triggers the subsequent four-branch hardware-level thread stream splitting and independent memory buffer allocation. Under normal working conditions, the resampling task is allocated 64 concurrent threads. When the amplitude of the change in surface density reaches 60%, the number of hardware concurrent threads is adjusted to 256, so that the real-time rendering latency is stabilized at 320ms. The controller monitors the texture rendering link processing latency fluctuation. When the real-time rendering latency deviates from the preset baseline latency by more than 15%, the learning rate decay factor of the backpropagation algorithm is adjusted from 0.01 to 0.03, directly reducing the gradient update step size matrix written to the graphics processor memory and suppressing the numerical jump caused by the change in geometric structure.

[0042] To improve processing stability under extreme conditions, the system is configured with a dynamic damping compensation operator for data processing load. The system monitors processing latency fluctuations in the texture rendering chain. When the detected real-time rendering latency deviates from the preset baseline rendering latency by more than 15%, the system triggers damping adjustment logic, adjusting the learning rate decay factor of the backpropagation in the differential rendering loop from 0.01 to 0.03. This action reduces the gradient update step size, suppressing the numerical jumps in texture features caused by abrupt changes in geometric structure. While ensuring texture generation accuracy, it reduces the frequency of abnormal numerical processing within the system from an average of 4 times per hour. With the frequency reduced to within 0.2 times and the surface density fluctuation amplitude remaining within 20% under normal operating conditions, the resource pre-allocation logic is in a silent state, and the texture generation time is stabilized at around 250ms. Under extreme stress test scenarios where the surface density mutation amplitude reaches 60%, through the dual intervention of resource pre-allocation logic and dynamic damping compensation operator, the system reduces the processing latency from a peak of 650ms to 320ms. The execution of the above engineering calibration procedures ensures the load management baseline of the 3D object stylization texture generation system when facing complex geometric topologies, and guarantees the processing efficiency of the texture generation process under multiple operating conditions.

[0043] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. A method for generating stylized textures for three-dimensional objects, characterized in that, include: Step S1: Obtain the 3D mesh model and initial guiding semantics, and perform two-stage denoising processing through the latent diffusion model. In the first stage, high-frequency noise suppression processing is performed using spatial topological features. In the second stage, high-resolution image reconstruction processing is performed by incorporating detailed features through the upsampling module to generate a high-resolution prior view. Step S2: Based on the topological adjacency matrix of the three-dimensional mesh model, construct a geometric feature association tensor that has a pixel mapping relationship with the high-resolution two-dimensional texture map. Based on the barycentric coordinate interpolation, reversely retrieve and determine the spatial coordinate vector and normalized normal vector of the invalid pixel in the high-resolution two-dimensional texture map on the geometric surface of the three-dimensional mesh model. Step S3: Construct a comprehensive structural distance operator using spatial coordinate vectors and normalized normal vectors, transforming planar neighborhood expansion into feature value propagation under three-dimensional mesh space topological constraints. When the local curvature change rate exceeds the set curvature threshold, adjust the geometric weight factor of the comprehensive structural distance operator and update the color feature parameters of invalid pixels. Step S4: Using the high-resolution prior view as the prior reference image, calculate the multi-view loss through the differential rendering loop, update the color parameters of the high-resolution two-dimensional texture map, and generate a stylized texture map of the target object.

2. The method for generating stylized textures for three-dimensional objects according to claim 1, characterized in that, Step S1 involves a two-stage denoising process using a latent diffusion model, which includes: acquiring the initial latent space image corresponding to the initial guiding semantics, calculating the global gradient vector of the initial latent space image in the diffusion model, and decomposing the global gradient vector into low-frequency topological components and high-frequency detail components; in the first stage, denoising is performed using the low-frequency topological components; in the second stage, the feature map processed in the first stage is input into the upsampling module, which consists of multiple transposed convolutional layers, and the high-frequency detail components are fused for denoising to generate a high-resolution prior view.

3. The method for generating stylized textures for three-dimensional objects according to claim 1, characterized in that, Step S2, which involves reversibly retrieving and determining the spatial coordinate vectors and normalized normal vectors of invalid pixels in the high-resolution 2D texture map on the geometric surface of the 3D mesh model based on barycentric coordinate interpolation, includes: retrieving invalid pixel regions in the high-resolution 2D texture map based on geometric feature association tensors, and mapping the invalid pixel regions onto polygonal patches of the 3D mesh model; and using the spatial coordinate vectors and normalized normal vectors of the polygonal patch vertices, performing linear interpolation through barycentric coordinates to determine the spatial coordinate vectors and normalized normal vectors corresponding to the invalid pixels.

4. The method for generating stylized textures for three-dimensional objects according to claim 1, characterized in that, In step S3, when the local curvature change rate exceeds the set curvature threshold, the geometric weight factor of the integrated structural distance operator is adjusted, and the color feature parameters of invalid pixels are updated, including: detecting the local curvature change rate of the region where the invalid pixel is located; when the local curvature change rate exceeds the set curvature threshold of 0.05, the geometric weight factor in the integrated structural distance operator is adjusted from 0.5 to 2.0, increasing the weight value of pixels with high surface normal direction deviation in the integrated structural distance operator, and restricting the mixing of color features across the mesh edge.

5. The method for generating stylized textures for three-dimensional objects according to claim 1, characterized in that, In step S4, the multi-view loss is calculated using a differential rendering loop to update the parameters of the high-resolution 2D texture map. This includes: projecting the high-resolution 2D texture map onto multiple discrete 3D viewing angles using a differential renderer to obtain multiple multi-view rendered views; calculating the absolute color deviation and perceptual structure similarity loss between the multi-view rendered views and the high-resolution prior views to construct a composite view loss function; and using a backpropagation mechanism to project the gradient of the composite view loss function backward onto the parameter plane of the high-resolution 2D texture map to update the feature channel color values ​​of each pixel in the high-resolution 2D texture map.

6. The method for generating stylized textures for three-dimensional objects according to claim 1, characterized in that, In step S1, the first stage uses spatial topological features for denoising, and the second stage uses an upsampling module to incorporate detailed features for denoising. This includes: in the early stage of denoising, the time step is set to be in the low-frequency semantic range, and the denoising network of the latent diffusion model is used to denoise the initial latent spatial image for low-frequency feature images; in the later stage of denoising, the time step is set to be in the high-frequency detail range, the upsampling module is used to improve the spatial resolution of the feature map, and the spatial feature fusion attention mechanism is used to denoise the detailed texture.

7. The method for generating stylized textures for three-dimensional objects according to claim 1, characterized in that, Step S3, which transforms planar neighborhood expansion into feature value propagation under the topological constraints of a 3D mesh space, includes: based on the topological adjacency matrix of the 3D mesh model, establishing the surface topological neighborhood distance and local surface normal direction deviation between invalid pixels and valid neighboring pixels determined by the geometric feature association tensor on the geometric surface of the 3D mesh model; using the surface topological neighborhood distance and local surface normal direction deviation to jointly constrain the weight allocation of color features, and restricting feature value propagation at the boundaries of texture regions and mesh wrinkles.

8. The method for generating stylized textures for three-dimensional objects according to claim 1, characterized in that, Step S2 involves constructing a geometric feature association tensor that maps pixels to the high-resolution 2D texture map based on the topological adjacency matrix of the 3D mesh model. This includes the following steps: Step S21: Analyze the polygonal topology of the 3D mesh model, extract the adjacency topological relationships between faces and between vertices, and construct a mesh topological adjacency matrix; Step S22: Establish a barycentric coordinate mapping table between the ultraviolet coordinate space of the high-resolution 2D texture map and the geometric surface of the 3D mesh model; Step S23: Combine the mesh topological adjacency matrix and the barycentric coordinate mapping table to assign each pixel in the high-resolution 2D texture map its spatial coordinate vector and normalized normal vector on the geometric surface of the 3D mesh model, forming a third-order geometric feature association tensor.

9. The method for generating stylized textures for three-dimensional objects according to claim 1, characterized in that, Step S4, which uses the high-resolution prior view as the prior reference image, includes: optimizing the high-resolution prior view as the prior reference image; assigning fixed weight coefficients to each pixel of the high-resolution prior view when calculating the multi-view loss in the differential rendering loop; and updating the color parameters of the high-resolution two-dimensional texture map through multi-view loss gradient backpropagation.

Images