3D art model intelligent generation system and method based on multi-source data fusion

Abstract

The invention discloses a 3D art model intelligent generation system and method based on multi-source data fusion, and relates to the technical field of computer aided design and graph generation, and the system comprises a multi-source data input module which is used for receiving a conceptual design drawing, text description and a reference 3D model; the feature extraction module comprises an image feature extraction unit used for extracting a style feature vector and a structure contour vector of the conceptual design drawing; and the text semantic analysis unit is used for outputting a key attribute tag vector described by the text. According to the 3D art model intelligent generation system and method provided by the invention, through multi-source data input including the concept design drawing, the text description and the reference 3D model, the design intention is comprehensively fused, the limitation of a single data source generation model is solved, and the dynamic weight distribution strategy and the closed-loop optimization mechanism of the 3D art model intelligent generation system and method are high in practicability. The geometric shape and style characteristics of the generated model are highly consistent with the design intention, and the workload of manual later correction is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer-aided design and graphics generation, in particular to a 3D art model intelligent generation system and method based on multi-source data fusion. BACKGROUND

[0002] In the field of digital content creation, the generation of 3D art models is a key link in game development, film production and virtual scene construction. Traditional 3D modeling relies on manual operation, which needs to go through multiple links such as concept design, mesh construction, texture mapping, etc., which not only consumes time and effort but also requires high professional skills of the creators. With the development of artificial intelligence technology, data-driven 3D model generation technology has gradually emerged. Existing solutions are mostly based on a single data source (such as text description or two-dimensional image) for model generation, which has improved modeling efficiency to some extent, but it is difficult to fully integrate multi-dimensional design intent, limiting the diversity and accuracy of the generated model.

[0003] However, in actual creation, designers often convey design requirements through concept design drawings, text descriptions and reference models, etc. The existing technology has problems such as lack of dynamic adjustment mechanism for fusion strategy, imperfect arbitration logic for feature conflicts, and insufficient optimization capability for geometry and style consistency when processing multi-source data, resulting in deviation between the generated model and the design intent, which often requires a lot of manual post-correction. At the same time, the existing system has limited support for interactive optimization, making it difficult to efficiently integrate real-time feedback in the creation process. In view of this, we propose a 3D art model intelligent generation system and method based on multi-source data fusion. SUMMARY

[0004] To solve the above technical problems, a 3D art model intelligent generation system and method based on multi-source data fusion are provided, which solves the above problems of generating 3D models based on a single data source, lack of dynamic adjustment of fusion strategy when processing multi-source data, imperfect arbitration of feature conflicts, insufficient optimization of geometry and style, large deviation between generated model and design intent, and limited support for interactive optimization.

[0005] To achieve the above purposes, the technical scheme adopted by the present application is as follows: The 3D art model intelligent generation system based on multi-source data fusion comprises: A multi-source data input module for receiving concept design drawings, text descriptions and reference 3D models; A feature extraction module comprising: an image feature extraction unit for extracting style feature vectors and structure contour vectors of the concept design drawings; a text semantic analysis unit for outputting key attribute label vectors of the text description; and a model structure analysis unit for deconstructing the mesh topology relationship matrix and material parameter set of the reference 3D model; The conditional generation module, the encoder of which inputs the normalized spliced multi-source feature vector into the generative adversarial network, and the generator outputs the vertex coordinates and UV texture coordinates of the initial 3D model, and the discriminator verifies the logical consistency between the generated model and the input features; The closed-loop optimization module includes: a geometric error analysis unit that calculates the structural contour deviation value of the generated model grid from the input conceptual design drawing; and a style consistency unit that quantifies the color similarity of the material map and the conceptual design drawing; The interactive optimization module receives user correction instructions and converts them into parameter adjustment signals to feed back to the conditional generator.

[0006] The 3D art model intelligent generation method based on multi-source data fusion includes the following steps: Receiving multi-source input data, including a conceptual design drawing, a text description, and a reference 3D model; Performing feature extraction operations, extracting visual style feature vectors and structural contour feature vectors of the conceptual design drawing through image analysis steps, extracting semantic attribute feature vectors of the text description through text analysis steps, and extracting topological structure feature matrices and material parameter feature sets of the reference 3D model through model analysis steps; Performing feature fusion operations, inputting the visual style feature vectors, structural contour feature vectors, semantic attribute feature vectors, topological structure feature matrices, and material parameter feature sets into a conditional generation model, which fuses the features according to a dynamic weight distribution strategy and generates initial 3D model grid data and texture maps; Performing closed-loop optimization operations, calculating the structural contour deviation value of the generated model grid from the input conceptual design drawing through a geometric error analysis step, and comparing the color distribution similarity of the generated model texture map and the conceptual design drawing through a style consistency analysis step; When the deviation value or similarity does not reach the threshold value, an optimization step is triggered to adjust the feature fusion weight and regenerate the model; If a user modification instruction is received, semantic features are extracted through an instruction analysis step to update the generated parameters; the generation-evaluation-optimization operation is iteratively performed until the final 3D model is output.

[0007] Preferably, the dynamic weight distribution strategy specifically includes: The confidence scores of five types of input features are calculated respectively: the visual style feature vector confidence is based on the image clarity value of the conceptual design drawing and the proportion of the style saliency area in the total image area; the structural contour feature vector confidence is based on the percentage of continuous edge pixels in the total edge pixels in the edge detection result; the semantic attribute feature vector confidence is calculated by comprehensively calculating the appearance frequency of entity keywords and the density of modifying words in the text description; the topological structure feature matrix confidence is determined by the grid closure index of the reference 3D model and the inverse relationship between the number of non-manifold edges; and the material parameter feature set confidence is related to the resolution value of the material map and the physical parameter integrity index. The five types of confidence scores are converted into weight coefficients by using a normalized exponential function, so that the values of each weight coefficient are in the interval of zero to one and the sum is always one. Finally, the features are fused by weighted summation operation, that is, each type of feature vector is multiplied by the corresponding weight coefficient and then added to generate a fused feature vector.

[0008] Preferably, a dimension alignment operation is performed before feature fusion: The visual style feature vector and the structural contour feature vector are spliced into a joint image feature vector; the semantic attribute feature vector is mapped to the target dimension space through a fully connected neural network layer; the topological structure feature matrix is compressed by a graph convolution network to generate a topological feature vector of fixed dimension; and the material parameter feature set is encoded into a material feature vector by a multilayer perceptron; A three-level feature pyramid fusion structure is adopted: the first level fusion layer inputs the joint image feature vector and the topological feature vector into a cross-attention mechanism module to output a geometric enhancement feature vector; the second level fusion layer performs element-wise multiplication operation on the geometric enhancement feature vector and the material feature vector to generate a material binding feature vector; and the third level fusion layer fuses the material binding feature vector and the mapped semantic attribute feature vector through a gating mechanism to output a unified dimension condition tensor required by the conditional generation model.

[0009] Preferably, the structural contour deviation value calculation process comprises: The initial 3D model grid data is projected onto the corresponding view plane of the conceptual design drawing to generate a binary contour mask; Edge detection is performed on the conceptual design drawing to generate a reference contour mask; The deviation value calculation adopts a standardized Hausdorff distance algorithm: the average value of the Euclidean distance from each pixel point in the generated contour mask to the nearest pixel point in the reference contour mask is calculated, and then the average value of the Euclidean distance from each pixel point in the reference contour mask to the nearest pixel point in the generated contour mask is calculated, and finally the arithmetic sum of the two average values is taken as the deviation value.

[0010] Preferably, the color distribution similarity is calculated by: extracting the concept design image and generating 256-level histogram data of the texture map in L channel, a channel and b channel respectively; The minimum value of the two histograms in each gray level is calculated independently for each color channel, and the arithmetic mean of the three channel accumulations is taken as the final similarity. At the same time, the style feature similarity is calculated: the pre-trained VGG-19 network is used to extract the feature map of the specified layer, and the Gram matrix is calculated respectively, and then the square sum of the difference of all elements of the two matrices is calculated, and then divided by the product of the square of the number of feature maps and the square of the size of the feature map.

[0011] Preferably, a threshold is set for the geometric error and the style consistency respectively, and when the geometric error or the style consistency does not reach the threshold, the optimization step performs the following operations: According to the structural contour deviation value, the fusion weight coefficient of the structural contour feature vector is increased by a preset proportion; according to the style feature similarity, the fusion weight coefficient of the visual style feature vector is increased by a preset proportion, and the weight coefficients of the other three types of features of the non-visual style and the structural contour are compressed by the original proportion, and the compression factor is one minus the sum of the weight increase of the two types divided by the sum of the weight coefficients of the other three types of features; The adjusted weight coefficients are re-input into the conditional generation model to perform model generation.

[0012] Preferably, the instruction parsing step adopts a multi-level processing flow: The semantic embedding vector of the user modification instruction is extracted by a pre-trained language model; The vector is input into an intent classifier to identify the operation type, which includes three types of geometric adjustment, material change and style transfer; The geometric adjustment branch outputs the vertex displacement vector field parameter, the material change branch outputs the material property modification amount containing 12-dimensional parameters, and the style transfer branch outputs the 512-dimensional style feature correction term; The final generated parameter set contains the aforementioned displacement vector field parameter, material property modification amount and style feature correction term, which is used to control model regeneration.

[0013] Preferably, the conditional generation model includes a parallel mesh generation branch and a texture generation branch: The mesh generation branch adopts a graph attention mechanism, and outputs the vertex coordinate matrix and the face index matrix after processing the fused features through an eight-layer residual graph convolutional network; The texture generation branch adopts a U-Net structure, and generates an RGBA texture map of a specified resolution through an encoder-decoder architecture; The two branches share a conditional encoder and exchange information through a cross-modal attention mechanism at the intermediate feature layer.

[0014] Preferably, when different input source features exist conflicts, a conflict resolution process is started: The Euclidean distance between the structure profile feature vector of the conceptual design drawing and the reference model topology feature vector is calculated, and then divided by the larger norm value to obtain the structure difference degree; If the structure difference degree exceeds the warning value, a conflict arbitration mechanism is started, the conceptual design drawing features are preferentially adopted, and the reference model weight coefficient is reduced to one fifth of the original value; For the conflict of text and visual features, the cosine similarity between the semantic feature vector and the visual feature vector is calculated, and if the similarity is less than 0.6, the text semantic feature weight coefficient is modified to (0.5+0.5*original value).

[0015] Compared with the prior art, the beneficial effects of the present application are: The 3D art model intelligent generation system and method provided by the present application comprehensively integrates design intentions through multi-source data input, including conceptual design drawings, text descriptions and reference 3D models, solves the limitations of single data source model generation, and ensures that the generated model is highly consistent with the design intention in terms of geometric shape and style characteristics, reduces the workload of manual post-correction, and at the same time, the interactive optimization module supports real-time feedback of users, further improves the accuracy and individuality of the model, in addition, the conflict resolution process effectively handles the contradictions between different input source features, enhances the robustness of the system, and provides strong support for the rapid and accurate generation of 3D art models. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The system module framework of the present application is shown in the figure; Figure 2 The method flowchart of the present application is shown in the figure. DETAILED DESCRIPTION

[0017] The following description is used to disclose the present application so that those skilled in the art can implement the present application. The preferred embodiments in the following description are only as examples, and other obvious modifications can be thought of by those skilled in the art.

[0018] Reference Figure 1As shown, the 3D art model intelligent generation system based on multi-source data fusion constructs a highly integrated creation framework, the core of which is to efficiently coordinate multiple heterogeneous data sources and accurately convert them into three-dimensional digital assets. The system starts with a multi-source data input module, which not only receives concept design drawings, text descriptions, and reference 3D models, but also includes a preprocessing mechanism to ensure data quality. The concept design drawings are adjusted to a standardized size and converted from RGB to Lab color space, with noise suppression processing; the text description is processed through a natural language processing pipeline for part-of-speech tagging and dependency analysis, removing irrelevant particles and marking core entities such as material and shape; the reference 3D model performs triangle mesh retopology optimization and global coordinate system unification to establish a benchmark for subsequent feature extraction.

[0019] The feature extraction module uses a multi-modal deep analysis architecture. The image feature extraction unit deploys an improved VGG-19 network, applies global average pooling on the activation map of a specific convolutional layer to extract a 128-dimensional style feature vector, and combines a phase consistency edge detection algorithm with a Hough line transform to generate a 512-dimensional contour vector describing the main structure. The text semantic analysis unit uses a pre-trained BERT model for bidirectional encoding, focusing on key modifiers through an attention mechanism, and outputs a 256-dimensional label vector containing semantic attributes such as material type and size ratio. The model structure analysis unit performs multi-scale analysis on the reference model, uses a topological traversal algorithm based on half-edge data structure to construct a grid adjacency relationship matrix, and separates subsets of material parameters such as diffuse reflection, highlight, and normal through spectral clustering algorithm.

[0020] The condition generation module as the core engine of the system, its encoder first performs feature standardization and dimension alignment. After Z-score normalization of different dimensional feature data, the visual style and structural contour vector are spliced into a 640-dimensional joint image feature; the semantic attribute vector is mapped to a 512-dimensional target space through a two-layer fully connected network; the topological structure matrix is compressed into a 512-dimensional feature vector through a three-layer graph convolution network; and the material parameter set is encoded into a 256-dimensional vector through a multi-layer perceptron with residual connection. A three-level pyramid fusion strategy is adopted: the first level generates a geometric enhancement feature by using a cross-attention mechanism to make the joint image feature and the topological feature interact; the second level binds the physical attributes by element-wise multiplication of the feature and the material vector; and the third level outputs a unified condition tensor by fusing the semantic attributes through a gated recurrent unit. The generator adopts a double-branch architecture: the grid generation branch includes an eight-layer residual graph convolution network, each layer of which integrates a graph attention mechanism to process the vertex neighborhood relationship, and finally outputs a vertex coordinate matrix and a triangular patch index; the parallel texture generation branch adopts a U-Net with skip connection, the encoder of which is down-sampled by five times of convolution with a step size of 2, the bottleneck layer fuses the condition tensor, and the decoder gradually reconstructs a 2048x2048 resolution RGBA texture map through transpose convolution. The two branches exchange information through cross-modal attention at the fourth level feature map, ensuring the spatial correspondence of geometry and texture. The discriminator constructs a three-scale evaluation network to verify the local geometric continuity, global morphological proportion and texture style consistency at resolutions of 64x64, 128x128 and 256x256 respectively.

[0021] The closed-loop optimization module establishes a quantitative feedback mechanism to continuously optimize the generation quality. The geometric error analysis unit implements accurate spatial comparison: first, project the generation model to a two-dimensional plane according to the viewpoint parameters of the conceptual design drawing, and generate a binary contour mask through an improved Sobel-Feldman operator; at the same time, perform Canny edge detection on the original design drawing with an adaptive threshold to obtain the reference contour. The deviation calculation uses the standardized bidirectional Hausdorff distance algorithm to calculate the average Euclidean distance of each pixel of the generated contour to the nearest reference contour pixel D_gen-ref and the average Euclidean distance of each pixel of the reference contour to the nearest generated contour pixel D_ref-gen, and the final geometric deviation value is the arithmetic sum of the two. The style consistency unit implements two-dimensional evaluation: the color dimension extracts the Lab space histogram of the design drawing and the generated map respectively, and calculates the average of the cumulative sum of the minimum values of the 256-level gray histograms of L, a and b channels as the color similarity; the artistic style dimension inputs both into the VGG-19 network to extract the relu4_1 layer feature map, calculates the square difference of the Frobenius norm of its Gram matrix, and then normalizes it by dividing by the square of the number of feature map channels and the square of the spatial size.

[0022] When the geometric deviation or style similarity does not reach the preset threshold, the system activates the dynamic weight redistribution. First, the five types of feature confidence are recalculated: the visual style confidence is determined by the product of the image resolution (0-1 normalization) and the proportion of the style salient area (calculated by saliency detection); the structural contour confidence is based on the edge continuity index (proportion of continuous edge pixels); the semantic attribute confidence is the harmonic mean of the entity keyword frequency (number of occurrences per hundred words) and the modifier density (proportion of adjectives and adverbs); the topological structure confidence is determined by the geometric mean of the grid closure (proportion of closed patches) and the number of non-manifold edges (normalized after taking the reciprocal); the material confidence is related to the texture resolution (0-1 normalization) and the physical parameter completeness (coverage ratio of the preset parameter set). The five types of confidence are converted into initial weight coefficients through the Softmax function. In the optimization stage, if the geometric error exceeds the standard, the structural contour feature weight is increased by the proportion of the error exceeding; if the style consistency is insufficient, the visual style weight is simultaneously increased, and the other three types of feature weights are compressed by the original proportion, with the compression factor being 1 minus the sum of the increments of the first two divided by the remaining weight sum. The updated weights are re-input into the conditional generator for iterative optimization.

[0023] The interactive optimization module provides a fine control channel for human intervention. User correction instructions are first converted into 768-dimensional semantic embedding vectors through the BERT model, and then input into a three-layer fully connected classification network to identify the operation type (geometric adjustment / material change / style transfer). Geometric adjustment instructions are processed through a graph neural network to output vertex displacement vector field parameters (including displacement direction and amplitude); material change instructions are parsed into modification amounts of 12-dimensional parameters including base color, metallic degree, roughness, etc.; style transfer instructions generate 512-dimensional style feature correction terms (obtained through feature decomposition of the style matrix). These parameters directly act on the corresponding control nodes of the generator to achieve controllable regeneration of the model.

[0024] The conflict resolution mechanism adopts a hierarchical arbitration strategy. When the concept graph conflicts with the reference model structure, the Euclidean distance and norm ratio of the feature vectors of the two are calculated. If the structural difference (distance / maximum norm) exceeds the 0.7 threshold, the reference model weight is forced to be reduced to 20% of the original value. For text and visual feature conflicts, the cosine similarity of the semantic feature vector and the visual feature vector is calculated. If the similarity is less than 0.6, a compromise strategy is adopted, and the text semantic weight is corrected to 0.5 plus 0.5 times the original value. The system outputs a 3D art model that meets the artistic and technical standards through multiple generation-evaluation-optimization cycles while ensuring the coordination of multi-source features.

[0025] Reference Figure 2As shown, the use process of the present application is: receiving multi-source data, including conceptual design drawings, text descriptions and reference 3D models; extracting features, including visual and structural features of design drawings, semantic features of text, topological and material features of models; performing feature fusion, inputting various types of features into conditional generation model, generating initial 3D model grid and texture according to dynamic weight distribution strategy; performing closed-loop optimization, calculating structural deviation of model grid and design drawing and texture color similarity; adjusting weight to generate when not reaching threshold; receiving user modification instructions, analyzing and updating parameters; iterative generation-evaluation-optimization process until outputting final 3D model.

[0026] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection required by the present application is defined by the appended claims and their equivalents.

Claims

1. A 3D art model intelligent generation system based on multi-source data fusion, characterized in that, The application relates to a method for generating a 3D model based on multi-source input data, comprising the following steps: A multi-source data input module is used for receiving a concept design drawing, a text description and a reference 3D model; A feature extraction module comprises an image feature extraction unit for extracting a style feature vector and a structure contour vector of the concept design drawing, a text semantic analysis unit for outputting a key attribute label vector of the text description, and a model structure analysis unit for deconstructing a grid topological relation matrix and a material parameter set of the reference 3D model; An conditional generation module, wherein an encoder inputs normalized and spliced multi-source feature vectors into a generative adversarial network, a generator outputs vertex coordinates and UV texture coordinates of an initial 3D model, and a discriminator verifies the logical consistency of the generated model and the input features; A closed-loop optimization module comprises a geometric error analysis unit for calculating the structure contour deviation value of the generated model grid and the input concept design drawing, and a style consistency unit for quantifying the color similarity of the material map and the concept design drawing; An interactive optimization module receives user correction instructions and converts them into parameter adjustment signals which are fed back to the conditional generator.

2. The 3D art model intelligent generation method based on multi-source data fusion, characterized in that, The method comprises the following steps: Receiving multi-source input data, including a concept design drawing, a text description and a reference 3D model; Performing a feature extraction operation, wherein a visual style feature vector and a structure contour feature vector of the concept design drawing are extracted through an image analysis step, a semantic attribute feature vector of the text description is extracted through a text analysis step, and a topological structure feature matrix and a material parameter feature set of the reference 3D model are extracted through a model analysis step; Performing a feature fusion operation, wherein the visual style feature vector, the structure contour feature vector, the semantic attribute feature vector, the topological structure feature matrix and the material parameter feature set are input into a conditional generation model, the model fuses the features according to a dynamic weight distribution strategy and generates initial 3D model grid data and texture maps; Performing a closed-loop optimization operation, wherein a geometric error analysis step is used to calculate the structure contour deviation value of the generated model grid and the input concept design drawing, and a style consistency analysis step is used to compare the color distribution similarity of the generated model texture map and the concept design drawing; When the deviation value or the similarity does not reach a threshold value, an optimization step is triggered to adjust the feature fusion weight and regenerate the model; If a user modification instruction is received, a semantic feature is extracted through an instruction analysis step to update the generated parameters; Iterative generation-evaluation-optimization operations are performed until a final 3D model is output. 3.The method of claim 2, wherein, The dynamic weight distribution strategy specifically comprises the following steps: Confidence scores of five types of input features are calculated, wherein the visual style feature vector confidence score is based on the image clarity value of the concept design drawing and the proportion of the style saliency area in the total image area, the structure contour feature vector confidence score is based on the percentage of continuous edge pixels in the total edge pixels in the edge detection result, the semantic attribute feature vector confidence score is calculated by comprehensively calculating the entity keyword appearance frequency and the modifier density in the text description, the topological structure feature matrix confidence score is determined by the grid closure index of the reference 3D model and the inverse relationship between the number of non-manifold edges, and the material parameter feature set confidence score is related to the material map resolution value and the physical parameter integrity index. The five types of confidence scores are converted into weight coefficients by using a normalized exponential function, so that each weight coefficient value is within the interval of zero to one and the sum is always one; Finally, the features are fused by weighted summation operation, that is, each type of feature vector is multiplied by the corresponding weight coefficient and then added to generate a fused feature vector.

4. The 3D art model intelligent generation method based on multi-source data fusion according to claim 3, characterized in that, Dimension alignment operation is performed before feature fusion: The visual style feature vector and the structural contour feature vector are spliced into a joint image feature vector; the semantic attribute feature vector is mapped to the target dimensional space through a fully connected neural network layer; the topological structure feature matrix is compressed by a graph convolution network to generate a topological feature vector with fixed dimensions; the material parameter feature set is encoded into a material feature vector by a multilayer perceptron; A three-level feature pyramid fusion structure is adopted: the first level fusion layer inputs the joint image feature vector and the topological feature vector into the cross-attention mechanism module to output a geometry-enhanced feature vector; the second level fusion layer performs element-wise multiplication operation on the geometry-enhanced feature vector and the material feature vector to generate a material binding feature vector; The third level fusion layer fuses the material binding feature vector and the mapped semantic attribute feature vector through a gating mechanism to output a unified dimensional condition tensor required by the conditional generation model. 5.The method of claim 2, wherein, The structural contour deviation value calculation process includes: Projecting the initial 3D model grid data onto the perspective plane corresponding to the conceptual design drawing to generate a binary contour mask; Performing edge detection on the conceptual design drawing to generate a reference contour mask; The deviation value is calculated using the standardized Hausdorff distance algorithm: calculate the average Euclidean distance of each pixel point in the generated contour mask to the nearest pixel point in the reference contour mask, then calculate the average Euclidean distance of each pixel point in the reference contour mask to the nearest pixel point in the generated contour mask, and finally take the arithmetic sum of the two averages as the deviation value. 6.The method of claim 2, wherein, The color distribution similarity is calculated as follows: Extract the 256-level histogram data of the L channel, a channel and b channel of the conceptual design drawing and the generated texture map respectively; Calculate the cumulative sum of the minimum values of the two histograms at each gray level for each color channel, and take the arithmetic average of the cumulative sums of the three channels as the final similarity; At the same time, the style feature similarity is calculated: input both sides into the pre-trained VGG-19 network to extract the specified layer feature map, calculate the Gram matrix of each, and then calculate the square sum of the difference of all elements of the two matrices, and divide by the product of the square of the number of feature maps and the square of the feature map size. 7.The method of claim 2, wherein, Threshold values are preset for geometric error and style consistency. When the geometric error or style consistency does not reach the threshold value, the optimization step performs the following operations: According to the structural contour deviation value, increase the fusion weight coefficient of the structural contour feature vector by a preset proportion; according to the style feature similarity, increase the fusion weight coefficient of the visual style feature vector by a preset proportion, and at the same time, compress the weight coefficients of the other three types of features of non-visual style and structural contour by the original proportion, with the compression factor being one minus the sum of the weight increases of the two types divided by the sum of the other three types of feature weight coefficients; The adjusted weight coefficients are re-input into the conditional generation model to perform model generation. 8.The method of claim 2, wherein, The instruction analysis step adopts a multi-level processing flow: Extract the semantic embedding vector of the user modification instruction through a pre-trained language model; The vector is input into an intent classifier to identify the operation type, including geometric adjustment, material change, and style transfer; The geometric adjustment branch outputs a vertex displacement vector field parameter, the material change branch outputs a material attribute modification quantity containing 12-dimensional parameters, and the style transfer branch outputs a 512-dimensional style feature correction term; The final parameter set contains the aforementioned displacement vector field parameter, material attribute modification quantity, and style feature correction term, which are used to control model regeneration. 9.The method of claim 2, wherein, The conditional generation model contains a parallel grid generation branch and a texture generation branch: The grid generation branch uses a graph attention mechanism and outputs a vertex coordinate matrix and a face index matrix after processing the fused features through an eight-layer residual graph convolutional network; The texture generation branch uses a U-Net structure and generates an RGBA texture map of a specified resolution through an encoder-decoder architecture; Both branches share a conditional encoder and exchange information through a cross-modal attention mechanism at the intermediate feature layer. 10.The method of claim 2, wherein, When there is a conflict between different input source features, a conflict resolution process is started: Calculate the Euclidean distance between the conceptual design graph structure contour feature vector and the reference model topology feature vector, and then divide by the larger norm value to obtain the structural difference; If the structural difference exceeds the warning value, the conflict arbitration mechanism is started, and the conceptual design graph features are prioritized, with the reference model weight coefficient reduced to one-fifth of the original value; For conflicts between text and visual features, calculate the cosine similarity between the semantic feature vector and the visual feature vector. If the similarity is less than 0.6, the text semantic feature weight coefficient is modified to (0.5 + 0.5 × original value).

Images