Method and device for generating industrial design model based on size-constrained text, and medium

Abstract

The application discloses an industrial design model generation method and device based on size-constrained text and a medium, and belongs to the field of multi-modal generation and computer graphics. The method obtains input text containing morphological description, size interval and proportional relationship, generates a constraint set through analysis, normalization and consistency verification; fuses size, proportion and morphological semantic features to obtain unified text feature representation; generates single-view and multi-view reference graphs verified for consistency according to the unified text feature representation, and reconstructs a three-dimensional grid model; performs size measurement and deviation evaluation on the model, and iteratively optimizes when exceeding the limit; finally, material texture is generated and an editable.obj / .glb file is exported. The application realizes end-to-end automatic generation of natural language size constraints to editable three-dimensional models, provides an explainable deviation metric while ensuring that key sizes are preferentially satisfied, significantly improves modeling efficiency and model geometric credibility, and is suitable for furniture modeling, industrial design and three-dimensional asset production scenes.

Description

Technical Field

[0001] This invention relates to the fields of computer-aided design (CAD), computer graphics and 3D modeling, and in particular to a method, apparatus, electronic device and storage medium for generating and correcting 3D models based on natural language dimension constraint analysis and 3D mesh geometric consistency verification for product structure modeling. Background Technology

[0002] With the development of industrial design, intelligent manufacturing, and digital content production, product appearance and structural solutions have gradually evolved from "two-dimensional drawing-driven" to "parametric three-dimensional model-driven." In scenarios such as furniture, consumer electronics, equipment components, and customized products, design input is often not strictly engineering drawings, but rather comes from natural language text such as requirement documents, procurement specifications, tender documents, patent descriptions, communication emails, or voice transcription. This type of text typically describes dimensional requirements in ways such as "approximately X width, approximately Y height," "range from A to B," "not exceeding / not less than," "at a 1:2 scale," and "equal to the width of a certain component," accompanied by features such as mixed units (mm, cm, m, inches), approximate terms (approximately, around), conditional terms (when, at least, at most), and implicit reference objects (flush with the door frame, equal to the height of the countertop). How to reliably convert the above unstructured descriptions into a computable, verifiable, and traceable set of dimensional constraints, and further drive the generation of three-dimensional models, is one of the key issues in the current digitalization of engineering and design.

[0003] In existing technologies, one approach relies on manual processing of requirement text into parameter tables, followed by the construction of 3D models in CAD software using sketch constraints, dimensioning, and feature modeling. While this approach offers high controllability, it suffers from low efficiency, heavy reliance on engineering experience, poor reusability, and is prone to human error or omissions during multiple rounds of communication and iteration. Another approach attempts to extract and structure dimensional information through natural language processing, but most can only extract superficial information such as "numerical value + unit," struggling to handle complex expressions like intervals, comparisons, combined constraints, and cross-sentence references. Furthermore, when dimensions involve proportions, equality, or correlated derivations, single-point dimension extraction alone is insufficient to form a consistent constraint system. Still other approaches directly use text parsing results to generate or retrieve 3D models, but lack a rigorous geometric consistency verification mechanism, easily leading to discrepancies between the generated model and specifications, resulting in high rework costs.

[0004] Furthermore, in terms of 3D model representation, common parametric modeling uses feature trees as its core, which facilitates the editability of dimensional constraints and geometric topology. However, it has high requirements for the completeness and consistency of input constraints: once constraints conflict (e.g., the lower bound is greater than the upper bound, the same dimension is assigned values ​​in multiple places and inconsistently, proportional relationships and absolute values ​​conflict simultaneously), the solver may fail to converge or produce unstable solutions. In contrast, 3D mesh models are more versatile in terms of vision and rendering, but meshes inherently lack explicit parametric semantics, and size adjustments often rely on scaling, deformation, or reconstruction operations. Without a reliable "measurement-evaluation-iteration" closed-loop mechanism, it is difficult to ensure that the mesh model meets preset requirements in terms of overall width, height, depth, key hole spacing, thickness, chamfers, and other indicators. Especially in multi-part assembly scenarios, the dimensional deviation of a single part will be amplified by the cumulative assembly, leading to interference, gap deviations, or assembly failure.

[0005] The uncertainty and ambiguity in dimensional requirements texts are also a challenge for engineering implementation. Terms such as "approximately," "around," "not less than," "not exceeding," "as thin as possible," and "close to" need to be mapped to executable tolerance models or optimization objectives. Texts from different sources may give different values ​​for the same indicator, or there may be conflicts in unit conversion, rounding, and differences in caliber (external dimensions versus installation dimensions). Existing technologies often lack unified conflict resolution strategies and interpretable deviation measurements, making it impossible to ensure that key constraints are met while making reasonable concessions to non-key constraints; it is also difficult to provide traceable evidence to explain "why the model was judged as qualified / unqualified" and "from which constraints the deviation came."

[0006] In the model generation and verification stages, many existing solutions only verify a single scale, such as verifying only the overall bounding box size, failing to cover multi-level indicators such as scale relationships, local key dimensions, and assembly interface dimensions; or they only perform a one-time verification after generation, lacking an iterative optimization process that feeds the verification results back to layout planning, sketching, or mesh adjustment, making it difficult to obtain stable and feasible model outputs under complex constraints. For scenarios that require automatic generation of sketching instruction sequences and migration between different CAD / modeling kernels, inconsistencies in constraint representation, unit systems, coordinate system definitions, and measurement calibers will further amplify errors and affect the degree of automation.

[0007] Therefore, there is an urgent need for a technical solution that can automatically complete the normalization of size ranges, unit normalization, proportional relationship derivation, constraint consistency checks, and conflict quantification assessment based on natural language specification descriptions, and combine this with measurement and alignment strategies for 3D meshes or geometric models to achieve a closed loop of "constraint-generation-verification-iteration". This solution should be able to transform dimensional information in text into a computable set of constraints, provide interpretable processing results when ambiguity and conflict exist, and output a 3D model that meets the allowable deviation threshold while ensuring that key dimensions are prioritized. This would improve modeling efficiency, reduce rework costs, and enhance the consistency and traceability of the model with specifications. Summary of the Invention

[0008] The present invention proposes an industrial design model generation method, device and storage medium based on size constraint text, which can at least solve one of the technical problems in the background art.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A method for generating industrial design models based on size constraint text, comprising the following steps performed via computer equipment: Obtain dimensional text containing morphological description, absolute size range and proportional relationship; preprocess the dimensional text, normalize the units, parse the values ​​and verify the consistency to generate a constraint set containing allowable deviation thresholds; Based on the constraint set, size constraint semantic features, proportional relationship semantic features and morphological description semantic features are extracted and fused to obtain a unified text feature representation; A single-view product reference image is generated based on the unified text features, and size consistency checks and semantic consistency checks are performed on the single-view product reference image. A multi-view product reference image is generated based on the single-view product reference image that has passed the consistency check and the unified text features, and a cross-view consistency check is performed on the multi-view product reference image. A 3D mesh model is generated based on the multi-view product reference image that has passed cross-view consistency checks; The three-dimensional mesh model is aligned by principal axes and its dimensions are measured. The measured values ​​are matched with the constraint set to calculate the comprehensive deviation. When the comprehensive deviation exceeds the allowable deviation threshold, iterative optimization is performed. Generate editable material parameters, UV unwrapping results, and texture mapping relationships for the 3D mesh model that has passed the deviation evaluation, and export it as an editable 3D file.

[0010] Furthermore, the step of obtaining the size text and performing preprocessing includes: Receive user-inputted text of dimensions and perform sentence segmentation, word segmentation, and numerical fragment detection; Extract size description information, which includes at least one of size range, nominal size, and proportional relationship; The size description information is standardized, including adjusting the order of the endpoints of the size range and converting different units of measurement into the same unit system. Output normalized dimension text or structured dimension description results for subsequent encoding processing.

[0011] Furthermore, the extraction of unified text feature representation based on constraint sets includes: Construct the input text sequence and perform vectorization to obtain the initial embedded feature sequence; Based on the initial embedded feature sequence, size constraint semantic features, proportional relationship semantic features, and morphological description semantic features are extracted respectively. The three types of semantic features are weighted and fused to generate a unified text feature representation, wherein the weighted fusion satisfies that the sum of the corresponding weight coefficients is 1; The unified text feature representation is used as the conditional input for the image generation model, the multi-view generation model, and the 3D reconstruction model.

[0012] Furthermore, the generation of the single-view product reference image includes: The unified text feature representation is mapped to a single-view generated conditional feature; An initial single-view product reference image is generated based on the single-view generation condition features. The initial single-view product reference image includes at least one of the following: the outer contour information of the target industrial product, the main structural partition information, and the size-related morphological expression information. Perform size consistency checks and semantic consistency checks on the initial single-view product reference image to determine whether the single-view product reference image meets the generation requirements corresponding to the size constraint text. When the inspection fails, the single-view product reference image is iteratively corrected to update the outer contour, structural proportions, or local morphological expression until the generation requirements are met or the preset iteration termination condition is reached.

[0013] Furthermore, the generation of multi-view product reference images includes: A multi-view product reference image set is generated based on the single-view product reference image and the unified text feature representation. The multi-view product reference image set includes at least two of the following: front view, side view and top view. Perform a cross-view consistency check on the multi-view product reference image set to determine whether the product reference images from different views are consistent in terms of structural semantics, contour correspondence, and dimensional constraint relationships. When the consistency check fails, at least one viewpoint product reference image is iteratively corrected until the preset consistency condition is met or the preset termination condition is reached.

[0014] Furthermore, the generation of the 3D mesh model based on the multi-view product reference image includes: Convert the set of multi-view product reference images that have passed cross-view consistency checks into 3D modeling input representations; The 3D modeling input representation is input into the 3D generation model to generate an initial 3D mesh model; The initial 3D mesh model is subjected to mesh quality optimization processing, which includes at least one of topology repair, noise reduction, hole filling, surface smoothing and mesh simplification, in order to obtain the target 3D mesh model.

[0015] Furthermore, the principal axis alignment and dimensional measurement of the 3D mesh model include: The three-dimensional mesh model is aligned along its principal axes to establish an object coordinate system, and the three-dimensional mesh model is transformed to a uniform orientation. Measure the overall dimensional indicators and key dimensional indicators in the object coordinate system. The overall dimensional indicators include the overall width, overall height and overall depth. The overall dimensional indicators and key dimensional indicators are matched with the target values ​​in the constraint set, and the relative deviation and weighted comprehensive deviation of each dimensional indicator are calculated. When the overall deviation exceeds the allowable deviation threshold, an iterative optimization action is determined. The iterative optimization action includes resampling the 3D modeling input and reconstructing the 3D mesh model, or performing global scaling and / or local deformation correction on the current 3D mesh model. The iteration terminates when the overall deviation meets the allowable deviation threshold or reaches the preset maximum number of iterations.

[0016] Furthermore, the generation of editable material parameters, UV unwrapping results, and texture mapping relationships includes: An overall material slot is created for the three-dimensional mesh model and editable material parameters are generated. The material parameters include at least one of the following: base color, roughness, metallicity, and transparency. Generate or optimize the overall UV unwrapping results, establish texture map reference relationships and texture coordinates, and establish the mapping relationship between texture maps and the surface of the 3D mesh model; Based on the material parameters, UV unwrapping results, and texture mapping relationships, an editable appearance description of the 3D mesh model is constructed. Output editable 3D model data containing 3D mesh model, material slots, material parameters and texture mapping information.

[0017] On the other hand, the present invention also discloses an industrial design 3D model generation device based on size constraint text, including a processor and a memory; the memory is used to store computer-executable instructions; the processor is used to execute the computer-executable instructions to implement the above-described method.

[0018] In another aspect, the present invention also discloses a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, it implements the above-described method to generate a target 3D model file that meets dimensional constraints and can be edited in 3D editing software. The target 3D model file includes .obj files and / or .glb files.

[0019] In summary, this invention discloses a method, apparatus, and storage medium for generating industrial design models based on size-constrained text, belonging to the fields of multimodal generation and computer graphics. The method acquires input text containing morphological descriptions, absolute size ranges, and proportional relationships; performs unit normalization, numerical analysis, and consistency verification on the text to generate a constraint set containing allowable deviation thresholds; based on the constraint set, a layout-based positioning generation strategy is used to generate a sketch construction instruction sequence, and a single-view image conforming to the size description is generated accordingly; further, text prompts and the single-view image are used as input constraints to generate a multi-view intermediate representation; based on this, multi-view... Figure 1 A 3D mesh model is constructed using a consistent generative model, or a multi-frame view sequence is generated using a video generative model and then reconstructed in 3D to obtain the 3D mesh model. The 3D mesh model is then subjected to dimensional measurement and deviation evaluation. When the deviation exceeds a threshold, iterative optimization is performed to update the sketch construction instruction sequence and / or reconstruct the 3D mesh model. Subsequently, editable material information is generated and material slots and texture mapping information are established. Finally, .obj and / or .glb files are exported for shape and material editing in 3D editing software. This invention can improve structural stability and editability while meeting dimensional constraints, and is applicable to scenarios such as furniture modeling, industrial design, and 3D asset production.

[0020] Specifically, the beneficial effects of this invention are as follows: This invention achieves automated generation of editable 3D models from natural language dimensional descriptions by constructing an end-to-end technical closed loop of "text parsing—constraint embedding—layered generation—geometric verification—iterative optimization—material encapsulation". This method deeply integrates and embeds dimensional constraints, proportional relationships, and morphological semantics into the entire generation process. Combined with single-view semantic verification, multi-view consistency constraints, and 3D mesh size measurement and deviation feedback, it effectively solves key problems in existing technologies such as the disconnect between generated models and specification requirements, multi-view structural drift, and the inability to quantify and resolve constraint conflicts. While ensuring that key dimensions are prioritized, it provides interpretable deviation measurements and traceable optimization paths. The final output is an editable .obj / .glb file containing complete material and texture mappings, which can be directly processed in 3D editing software, significantly improving the efficiency of industrial design modeling, the geometric reliability of the model, and its usability in downstream production. Attached Figure Description

[0021] Figure 1 This is a flowchart of a method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the process of generating a product reference drawing and creating a 3D model under dimensional constraints according to an embodiment of the present invention. Figure 3 This is a sample of the final generated effect of an embodiment of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0023] like Figure 1 As shown, an embodiment of the present invention provides a method for generating an industrial design model based on size constraint text, comprising the following steps: S1: Obtain the dimension text input by the user, wherein the dimension text includes at least a dimensional constraint description of the target industrial product; S2: Input the size text into the text encoder to extract the corresponding text features; S3: Input the text features into the pre-trained image generation model to generate a single-view product reference image of the target industrial product; S4: The text features and single-view reference image are input into a pre-trained multi-view generation model to generate a multi-view product reference image of the target industrial product. S5: Constrain or guide the generation process of the multi-view product reference image based on the single-view product reference image to improve the consistency between the multi-view product reference image and the single-view product reference image. S6: Input the text features and / or the multi-view product reference image into the pre-trained 3D reconstruction generation model to generate a three-dimensional model of the target industrial product; S7: Export the 3D model as an editable 3D file, the editable 3D file including .obj files and / or .glb files; S8: Output the single-view product reference image, the multi-view product reference image, and the editable 3D file to the user as product design output.

[0024] Specifically, step S1 includes steps S1.1 to S1.4: S1.1: Receive the size text input by the user and preprocess the size text, the preprocessing including sentence segmentation, word segmentation and numerical fragment detection; S1.2: Extract dimensional description information related to the target industrial product from the preprocessed dimensional text, wherein the dimensional description information includes at least one of dimensional range, nominal size, and proportional relationship; S1.3: The extracted size description information is normalized. The normalization process includes adjusting the order of the endpoints of the size range and uniformly converting the size values ​​corresponding to different units of measurement to obtain size description information under a unified unit system. S1.4: Output the normalized dimension text or structured dimension description results for encoding processing by the text encoder in step S2.

[0025] Specifically, step S2 includes steps S2.1 to S2.5: S2.1: Obtain the normalized dimension text representation or structured dimension description result output from step S1, and construct the input text sequence:

[0026] in, This represents the input text sequence. Indicates the first in the input text sequence Each text unit This indicates the length of the input text sequence.

[0027] S2.2: The input text sequence is input into a text encoder for vectorization representation to obtain an initial embedded feature sequence:

[0028] in, This represents the initial embedded feature sequence. Indicates the first Embedding vectors corresponding to each text unit This indicates an embedded mapping function.

[0029] S2.3: Based on the initial embedded feature sequence, extract size constraint semantic features, proportional relationship semantic features, and morphological description semantic features, as follows:

[0030] in, Represents semantic features of size constraints. Semantic features representing proportional relationships It represents the semantic features of morphological description.

[0031] S2.4: The size constraint semantic features, proportional relationship semantic features, and morphological description semantic features are fused and encoded to generate a unified text feature representation:

[0032] in, This represents a unified representation of text features. The feature fusion function is defined as follows: In one embodiment, the feature fusion function is a weighted fusion function that satisfies the following:

[0033]

[0034] in, , and These represent the weight coefficients of the corresponding semantic features.

[0035] S2.5: The unified text feature representation is output to the subsequent generation module as a conditional input for the image generation model, multi-view generation model, and 3D reconstruction model, expressed as:

[0036] in, Indicates the first One generation module, This represents the output result of the corresponding generation module; where, This is a single-view product reference image. This is a multi-view product reference image. It is a three-dimensional model or its corresponding three-dimensional representation.

[0037] Specifically, step S3 includes steps S3.1 to S3.5: S3.1: Input the unified text feature representation output from step S2 into the single-view product reference image generation module to extract visual conditional features characterizing the appearance and structural features of the target industrial product, and construct the single-view generation input, represented as:

[0038] in, This represents the viewpoint condition features corresponding to the single-viewpoint generation module. This represents the unified text feature representation output by step S2. This indicates that the unified text feature representation is mapped to a feature projection function with single-view generation conditions.

[0039] S3.2: Generate an initial image representation of the single-view product reference image based on the aforementioned viewpoint condition features, represented as:

[0040] in, This represents the initially generated single-view product reference image. This represents a single-view image generation function; the initial image represents at least one of the following: the outer contour information of the target industrial product, the main structural partition information, and the size-related morphological expression information.

[0041] S3.3: Perform size consistency checks and semantic consistency checks on the initially generated single-view product reference image to determine whether the generated single-view product reference image meets the generation requirements corresponding to the size constraint text, as shown below:

[0042] in, This indicates the inspection results of a single-view product reference image. This indicates the first [item] extracted from the single-view product reference image. Item size characterization parameter, and They represent the first The lower and upper bounds corresponding to the item size parameters, The image semantic features representing the single-view product reference image. A function representing the similarity calculation between image semantic features and text feature representations. This represents the single-view semantic consistency threshold; it is set when both size consistency and semantic consistency checks are satisfied. If true, it indicates that the single-view product reference image has passed the consistency check.

[0043] S3.4: When the single-view product reference image fails the consistency check, the single-view product reference image generation process is iteratively corrected to update the outer contour, structural proportions, or local morphological expressions in the image, and step S3.3 is repeated until the generation requirements are met or the preset iteration termination condition is reached, as shown below:

[0044] in, This represents the single-view product reference image obtained in the t-th iteration. Indicates the first The single-view product reference image after the second iteration correction. This represents the iterative correction function for a single-view product reference image.

[0045] S3.5: Output a single-view product reference image that meets the size constraint text requirements, as one of the inputs to the multi-view product reference image generation module, represented as:

[0046] in, This represents the final output single-view product reference image. This represents the target single-view product reference image obtained when the consistency check conditions are met or the preset termination conditions are reached. This function represents the output function of a single-view product reference image.

[0047] Specifically, step S4 includes steps S4.1 to S4.3: S4.1: Obtain the single-view product reference image output in step S3, and combine it with the unified text feature representation output in step S2 to generate a multi-view product reference image set, represented as:

[0048] in, This represents a collection of product reference images from multiple perspectives. This represents the single-view product reference image output in step S3. This represents the unified text feature representation output by step S2. This represents a function for generating multi-view product reference images; the set of multi-view product reference images includes at least two or more view product reference images from the front view, side view, and top view.

[0049] S4.2: Perform a cross-view consistency check on the multi-view product reference image set to determine whether the product reference images at each view are consistent in terms of structural semantics, contour correspondence, and dimensional constraint relationships; when the consistency check fails, iteratively correct at least one view product reference image and repeat the consistency check until a preset consistency condition is met or a preset termination condition is reached, as shown below:

[0050] in, This indicates the consistency check results for the multi-view product reference image set. This represents the cross-view consistency evaluation value of the multi-view product reference image set. Indicates the cross-perspective consistency threshold; when Not lower than the threshold hour, This indicates that the multi-view product reference image set has passed the consistency check; otherwise... .

[0051] S4.3: Output a set of multi-view product reference images that pass the consistency check as input for the subsequent 3D geometry reconstruction module, represented as:

[0052] in, This represents the final output set of multi-view product reference images. This function represents the output function for multi-view product reference images.

[0053] Specifically, step S5 includes steps S5.1 to S5.4: S5.1: Receive the set of multi-view product reference images output in step S4, and convert the set of multi-view product reference images into a 3D modeling input representation, expressed as:

[0054] in, This represents the input representation for 3D modeling. This represents the set of multi-view product reference images output in step S4. This represents the encoding function that translates multi-view product reference images into a 3D modeling space.

[0055] S5.2: Input the 3D modeling input representation into the 3D generation model to generate an initial 3D mesh model, represented as:

[0056] in, This represents the initial 3D mesh model. This represents a 3D mesh generation function.

[0057] S5.3: Perform mesh quality optimization processing on the initial 3D mesh model to obtain the target 3D mesh model. The mesh quality optimization processing includes at least one or more of the following: topology repair, noise reduction, hole filling, surface smoothing, and mesh simplification, expressed as:

[0058] in, Represents the target 3D mesh model. This represents the mesh optimization processing function.

[0059] S5.4: Output the target 3D mesh model as input for subsequent dimensional measurement, structural analysis, or model export processing, expressed as:

[0060] in, This represents the output 3D industrial design model data. This represents the output function of the 3D model.

[0061] Specifically, step S6 includes steps S6.1 to S6.6: S6.1: Align the principal axes of the 3D mesh model output in step S5 or establish an object coordinate system to transform the 3D mesh model into an object coordinate system with a unified orientation, as shown below:

[0062]

[0063] in, Represents the first point in the three-dimensional mesh model point set. Coordinates of a point, Indicates the number of points. Denotes the centroid of the point set. This represents the rotation matrix used for principal axis alignment. Represents the coordinates of a point in the object's coordinate system.

[0064] S6.2: Measure the overall dimensional parameters and key dimensional parameters of the 3D mesh model in the object coordinate system. The overall dimensional parameters include at least the overall width, overall height, and overall depth, expressed as:

[0065]

[0066]

[0067] in, , and These represent the components of the point coordinates in the object's coordinate system along the three axes. Indicates the overall width. Indicates the overall height. Indicates the overall depth.

[0068] S6.3: Match the overall dimensional indicators and key dimensional indicators with the dimensional constraint set, and calculate the relative deviation and comprehensive deviation of each dimensional indicator, expressed as:

[0069]

[0070] in, Indicates the first Dimensional measurements Indicates the first The target value corresponding to the item size index, Indicates the first The relative deviation of the item size index, Indicates a preset positive number. Indicates the first The weight of the item size index, This indicates the overall deviation.

[0071] S6.4: When the overall deviation does not exceed the allowable deviation threshold, the three-dimensional mesh model is determined to meet the size constraints and proceeds to the subsequent processing steps.

[0072] S6.5: When the overall deviation exceeds the allowable deviation threshold, determine the iterative optimization action, which includes at least one of the following: resample the 3D modeling input and return to step S5 to rebuild the 3D mesh model; perform global scaling and / or local deformation correction on the current 3D mesh model.

[0073] S6.6: Terminate the iteration when the overall deviation meets the allowable deviation threshold, or output the 3D mesh model with the smallest deviation when the preset iteration limit is reached, as follows:

[0074]

[0075] in, Indicates the first The overall deviation corresponding to the next iteration. Indicates the allowable deviation threshold. This indicates the maximum number of preset iterations.

[0076] Specifically, step S7 includes steps S7.1 to S7.4: S7.1: Establish an overall material slot for the 3D mesh model output in step S6, and generate editable material parameters for the 3D mesh model. The material parameters include at least one of the following: base color, roughness, metallicity, and transparency, expressed as follows:

[0077] in, This represents a 3D mesh model that satisfies dimensional constraints. This represents the set of material parameters corresponding to the three-dimensional mesh model. This represents the function that generates material properties.

[0078] S7.2: Generate or optimize the overall UV unwrapping result for the 3D mesh model, and establish texture map reference relationships and texture coordinates, represented as follows:

[0079] in, Represents the three-dimensional coordinates of points on the surface of a three-dimensional mesh model. This represents the two-dimensional texture coordinates corresponding to the three-dimensional coordinates. This represents the mapping function from three-dimensional surface coordinates to two-dimensional texture space; the mapping relationship between the texture map and the surface of the three-dimensional mesh model is established through the UV unwrapping and texture coordinates.

[0080] S7.3: Based on the material parameter set, UV unwrapping results, and texture mapping relationships, construct the editable appearance description information of the 3D mesh model, represented as:

[0081] in, This represents the texture and appearance description information of a 3D mesh model. This represents the UV unwrapping result of the three-dimensional mesh model. This represents the function that constructs the texture and appearance description.

[0082] S7.4: Outputs editable 3D model data containing 3D mesh model, material slots, material parameters, and texture mapping information, represented as:

[0083] in, This represents editable 3D model data. This represents the output function for enhancing the appearance of the model.

[0084] Specifically, step S8 includes: The editable 3D model data output in step S7 is encapsulated and output to generate a target 3D model file. The target 3D model file contains at least a 3D mesh model, material parameter information, and texture mapping information, represented as follows:

[0085] in, This represents the target 3D model file. This represents a 3D mesh model that satisfies dimensional constraints. Represents a set of material parameters. Represents a set of texture mapping information.

[0086] In a specific application example, using industrial products such as office chairs, beds, and sofas as target objects, the user inputs a text description including product category, overall appearance, main structure, and size requirements. The system first extracts and uniformly encodes the product morphological semantics, size values, size ranges, and proportional relationships in the text description, and generates a single-view product reference image of the target industrial product based on the encoding results. The single-view product reference image is used to determine the overall appearance, outer contour, main structural divisions, and basic proportional relationships between the components of the target industrial product.

[0087] Furthermore, using the single-view product reference image and unified text feature representation as constraints, the product reference image is expanded to multiple views, resulting in a set of multi-view product reference images representing the same target industrial product from different viewing directions. During the multi-view expansion process, consistency checks are performed on the product category, main components, contour correspondence, structural connection relationship, and size ratio relationship in each view product reference image to reduce situations where components are missing, structures drift, proportions are mismatched, or appearance semantics are inconsistent between different views.

[0088] like Figure 2 As shown, the single-view product reference image of the target industrial product is expanded into a multi-view product reference image set after being expanded to multiple views, and then used for the reconstruction of the 3D mesh model. The generated 3D mesh model can be exported as .obj files and / or .glb files, and can be imported into 3D design or editing software such as Blender and Autodesk Fusion 360 for display, rotation, and subsequent editing of shape, mesh structure, or materials. For different categories of industrial products, users can also add overall width, overall height, overall depth, or other key dimension annotations to the input text, so that the generation process of the product reference image and 3D model is constrained by the corresponding dimensional information.

[0089] like Figure 3 As shown, the above method can generate corresponding product reference images and 3D model effects for different product design needs. The generated results correspond to the input text description in terms of product category, overall outline, main structural components, and appearance, and can form a 3D product representation that can be viewed from different viewing directions. Compared with the method of directly generating 3D models based solely on text, this embodiment determines the overall appearance and main structural features of the product through single-view product reference images, supplements the outline and geometric information under different viewing directions through multi-view product reference images, and reduces the impact of multi-view structural inconsistencies on the 3D model reconstruction results through cross-view consistency checks.

[0090] To verify the generation efficiency and consistency of the multi-view generation results of the method described in this invention, tests were conducted in a computing environment with an NVIDIA RTX A6000 GPU (48GB VRAM), dual Intel Xeon Gold 6148 processors, and approximately 251 GiB of memory. The computing environment used was Ubuntu 22.04. During the tests, each generation task utilized only one NVIDIA RTX A6000 GPU, without employing multi-GPU parallel computing. Under the conditions of complete model preloading and sequential single-task execution, two industrial products, beds and sofas, were selected as test objects. For each test object, six multi-view product reference images with a resolution of 512×512 were generated: a front view, a left view, a rear view, a right view, a top view, and a bottom view. A 3D mesh model was then generated based on these multi-view product reference images. The multi-view product reference image generation process employed 50 sampling steps, while the 3D model reconstruction process employed 30 sampling steps.

[0091] Test results show that the generation times for the multi-view product reference images of the bed and sofa test cases were approximately 15.06 seconds and 12.81 seconds, respectively, with an average of approximately 13.94 seconds; the corresponding 3D model reconstruction times were approximately 49.95 seconds and 49.24 seconds, respectively, with an average of approximately 49.60 seconds. The total time for the two stages of multi-view product reference image generation and 3D model reconstruction was approximately 65.01 seconds and 62.05 seconds, respectively, with an average total time of approximately 63.53 seconds, both of which can be completed within 2 minutes. The above times do not include the one-time initialization time required for the initial loading of the model.

[0092] To evaluate the cross-view consistency of the generated results, a pre-trained image feature extraction model was used to extract the visual features of the front, left, rear, and right views of each test object. Pairwise cosine similarity scores were calculated between product reference images from different horizontal viewpoints, and the average of these similarities was used as the cross-view feature consistency evaluation value. This evaluation value characterizes the degree of consistency between product reference images from different viewpoints in terms of product category, overall appearance, and main structural features. Since there are significant differences in viewing direction between the top and bottom views and the horizontal views, this embodiment uses the front, left, rear, and right views for feature consistency statistics. The top and bottom views are mainly used to supplement the geometric information of the product in the vertical viewing direction.

[0093] Under the test conditions of this embodiment, the average feature cosine similarity between the four horizontal view product reference images of the bed and sofa examples is approximately 0.7993 and 0.8217, respectively, with an average of approximately 0.8105 for the two test examples. This test result illustrates that, after single-view product reference image constraints and cross-view consistency processing, the generated results from different horizontal views can maintain the consistency of the target industrial product's product category, overall appearance, and main structural features. The feature cosine similarity is only used as a reference indicator for evaluating the multi-view generation results in this embodiment and does not constitute a limitation on the technical solution or generation results of this invention.

[0094] In terms of 3D model generation, multi-view product reference images, after cross-view consistency checks, are input into the 3D reconstruction model. This generates a 3D mesh model that corresponds to the product reference images in terms of overall outline, main structural components, and appearance. The generated 3D model can be viewed from different viewing directions and can serve as an initial 3D asset in the conceptual design phase of industrial products. It should be noted that the generated 3D model is primarily used for product appearance design demonstration and subsequent design editing, and is not intended as an engineering model that can be directly used for manufacturing without verification. The generated 3D model is output in .glb format and supports importing into 3D design or editing software such as Blender for secondary editing, including rotation, mesh structure adjustment, overall scale adjustment, material parameter modification, and texture replacement. In other embodiments, the 3D model can also be exported in .obj format to adapt to the data input requirements of different 3D design software.

[0095] The above-mentioned operational efficiency and cross-view feature consistency test results show that the present invention can complete the generation of multi-view product reference images and 3D models in a short time. The generation results of different horizontal viewpoints can maintain the overall appearance and main structural features of the product. The generated 3D model can be further imported into 3D design or editing software for secondary processing, thereby reducing the operation required for manually drawing multi-view reference images and manually building initial 3D models, improving the efficiency, generation consistency and subsequent editing usability of industrial product concept design and initial 3D asset construction.

[0096] In another aspect, the present invention also discloses an industrial design 3D model generation device based on size-constrained text, including a processor and a memory; the memory is used to store computer-executable instructions; the processor is used to execute the computer-executable instructions to implement the above-described method.

[0097] In another aspect, the present invention also discloses a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, it implements the above-described method to generate a target 3D model file that meets dimensional constraints and can be edited in 3D editing software. The target 3D model file includes .obj files and / or .glb files.

[0098] It is understood that the systems, devices, and storage media provided in the embodiments of the present invention correspond to the methods provided in the embodiments of the present invention, and the explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

[0099] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0100] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0101] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0102] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating industrial design models based on size constraint text, characterized in that, Includes the following steps: Obtain dimensional text containing morphological description, absolute size range and proportional relationship; preprocess the dimensional text, normalize the units, parse the values ​​and verify the consistency to generate a constraint set containing allowable deviation thresholds; Based on the constraint set, size constraint semantic features, proportional relationship semantic features and morphological description semantic features are extracted and fused to obtain a unified text feature representation; A single-view product reference image is generated based on the unified text features, and size consistency checks and semantic consistency checks are performed on the single-view product reference image. A multi-view product reference image is generated based on the single-view product reference image that has passed the consistency check and the unified text features, and a cross-view consistency check is performed on the multi-view product reference image. A 3D mesh model is generated based on the multi-view product reference image that has passed cross-view consistency checks; The three-dimensional mesh model is aligned by principal axes and its dimensions are measured. The measured values ​​are matched with the constraint set to calculate the comprehensive deviation. When the comprehensive deviation exceeds the allowable deviation threshold, iterative optimization is performed. Generate editable material parameters, UV unwrapping results, and texture mapping relationships for the 3D mesh model that has passed the deviation evaluation, and export it as an editable 3D file.

2. The method according to claim 1, characterized in that, Obtaining and preprocessing the size text includes: Receive user-inputted text of dimensions and perform sentence segmentation, word segmentation, and numerical fragment detection; Extract size description information, which includes at least one of size range, nominal size, and proportional relationship; The size description information is standardized, including adjusting the order of the endpoints of the size range and converting different units of measurement into the same unit system. Output normalized dimension text or structured dimension description results for subsequent encoding processing.

3. The method according to claim 1, characterized in that, Extracting unified text feature representations based on constraint sets includes: Construct the input text sequence and perform vectorization to obtain the initial embedded feature sequence; Based on the initial embedded feature sequence, size constraint semantic features, proportional relationship semantic features, and morphological description semantic features are extracted respectively. The three types of semantic features are weighted and fused to generate a unified text feature representation, wherein the weighted fusion satisfies that the sum of the corresponding weight coefficients is 1; The unified text feature representation is used as the conditional input for the image generation model, the multi-view generation model, and the 3D reconstruction model.

4. The method according to claim 1, characterized in that, Generating single-view product reference images includes: The unified text feature representation is mapped to a single-view generated conditional feature; An initial single-view product reference image is generated based on the single-view generation condition features. The initial single-view product reference image includes at least one of the following: the outer contour information of the target industrial product, the main structural partition information, and the size-related morphological expression information. Perform size consistency checks and semantic consistency checks on the initial single-view product reference image to determine whether the single-view product reference image meets the generation requirements corresponding to the size constraint text. When the inspection fails, the single-view product reference image is iteratively corrected to update the outer contour, structural proportions, or local morphological expression until the generation requirements are met or the preset iteration termination condition is reached.

5. The method according to claim 1, characterized in that, Generating multi-view product reference images includes: A multi-view product reference image set is generated based on the single-view product reference image and the unified text feature representation. The multi-view product reference image set includes at least two of the following: front view, side view and top view. Perform a cross-view consistency check on the multi-view product reference image set to determine whether the product reference images from different views are consistent in terms of structural semantics, contour correspondence, and dimensional constraint relationships. When the consistency check fails, at least one viewpoint product reference image is iteratively corrected until the preset consistency condition is met or the preset termination condition is reached.

6. The method according to claim 1, characterized in that, Generating a 3D mesh model based on multi-view product reference images includes: Convert the set of multi-view product reference images that have passed cross-view consistency checks into 3D modeling input representations; The 3D modeling input representation is input into the 3D generation model to generate an initial 3D mesh model; The initial 3D mesh model is subjected to mesh quality optimization processing, which includes at least one of topology repair, noise reduction, hole filling, surface smoothing and mesh simplification, in order to obtain the target 3D mesh model.

7. The method according to claim 1, characterized in that, Principal axis alignment and dimensional measurement of a 3D mesh model include: The three-dimensional mesh model is aligned along its principal axes to establish an object coordinate system, and the three-dimensional mesh model is transformed to a uniform orientation. Measure the overall dimensional indicators and key dimensional indicators in the object coordinate system. The overall dimensional indicators include the overall width, overall height and overall depth. The overall dimensional indicators and key dimensional indicators are matched with the target values ​​in the constraint set, and the relative deviation and weighted comprehensive deviation of each dimensional indicator are calculated. When the overall deviation exceeds the allowable deviation threshold, an iterative optimization action is determined. The iterative optimization action includes resampling the 3D modeling input and reconstructing the 3D mesh model, or performing global scaling and / or local deformation correction on the current 3D mesh model. The iteration terminates when the overall deviation meets the allowable deviation threshold or reaches the preset maximum number of iterations.

8. The method according to claim 1, characterized in that, Generating editable material parameters, UV unwrapping results, and texture mapping relationships includes: An overall material slot is created for the three-dimensional mesh model and editable material parameters are generated. The material parameters include at least one of the following: base color, roughness, metallicity, and transparency. Generate or optimize the overall UV unwrapping results, establish texture map reference relationships and texture coordinates, and establish the mapping relationship between texture maps and the surface of the 3D mesh model; Based on the material parameters, UV unwrapping results, and texture mapping relationships, an editable appearance description of the 3D mesh model is constructed. Output editable 3D model data containing 3D mesh model, material slots, material parameters and texture mapping information.

9. A device for generating 3D industrial design models based on size-constrained text, characterized in that, It includes a processor and a memory; the memory is used to store computer-executable instructions; the processor is used to execute the computer-executable instructions to implement the method of any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method of any one of claims 1 to 8 to generate a target 3D model file that meets dimensional constraints and can be edited in 3D editing software, the target 3D model file including .obj files and / or .glb files.

Images