Method and system for AI aided design workflow based on diffusion model

Abstract

The invention relates to the technical field of diffusion model AI application, and discloses an AI aided design workflow method and system based on a diffusion model, and the method comprises the steps: receiving multi-modal data; vectorizing the sketch, and extracting key set features; analyzing the text into structured semantics; parameter constraints are converted into mathematical expressions or boundary conditions; mapping the multi-modal data to a unified space; selecting a generation path with the minimum loss function from the candidate paths; gaussian noise is gradually added into the submerged space, and noise distribution is predicted; injecting the parameter constraint into a denoising process to generate a path; and the optimized latent vector is converted into a vector graph through a decoder, and a geometric topological relation is reserved. The system comprises a unified multi-modal data module, a path generation module and a user interaction module. According to the method, the actual availability of a design result is improved; the design period is shortened, the process is simplified, and human resource investment is reduced.

Description

Technical Field

[0001] The present invention relates to the field of diffusion model AI application technology, and in particular to a method and system for AI-assisted design workflow based on diffusion model. Background Art

[0002] Diffusion models have demonstrated strong potential and advantages in a variety of fields, including image generation, text-to-image conversion, and architectural design. In architectural design, diffusion models are used to optimize design parameters and generate creative layouts. For example, by combining conditional generation models with diffusion models, high-quality design generation based on specific conditions (such as physical constraints or user input) can be achieved. Furthermore, diffusion models have been applied to specific scenarios such as bicycle design and interior scene generation, optimizing design parameters and improving design efficiency through diffusion and denoising techniques.

[0003] Prior art 1, a Chinese patent with patent number 202411597702.2, belongs to the field of software engineering and discloses a software development method and system based on a low-code platform. The method includes automatically generating an initialization project based on project requirements; designing a workflow based on the initialization project using predefined workflow nodes and a graphical interface; synchronizing and recording all workflow changes in real time; conducting experimental development of sub-branches based on change records, reviewing the experimental development of sub-branches, and merging them into a new workflow based on the review results; launching the new workflow for testing, debugging it based on the test results, and obtaining the final workflow; deploying the final workflow to a production environment, monitoring the final workflow's running status in real time, and continuously iterating and improving the final workflow based on monitoring data and user feedback. Although conducting experimental development of sub-branches based on change records, reviewing, and synchronizing and recording all workflow changes in real time improve team collaboration efficiency and version control capabilities, the low-code platform's automatic workflow generation and version control suffers from low multimodal collaboration efficiency and is unable to effectively integrate text descriptions, hand-drawn sketches, and parameter constraints.

[0004] Prior art two, a Chinese patent with patent number 202410229243.6, relates to a modeling method based on the collaboration of computing components in an optimized design workflow, belonging to the field of system modeling. The method comprises: extending a computationally independent model modeling language, establishing a collaboration mechanism for computing components in the computationally independent model modeling language, constructing an initial workflow model based on the extended modeling language and the collaboration mechanism, and presetting variable values ​​for computing components in the model; extending a platform-independent model modeling language, establishing semantic mapping rules, and automatically optimizing the initial workflow model based on the extended modeling language and the semantic mapping rules to obtain an optimized workflow model. While establishing a collaboration mechanism for computing components solves the problem of computing components often being unable to connect when integrated into a workflow, and by adding parallel semantic support to the modeling language to construct a complex parallel workflow model, and supporting automatic optimization of the workflow model through semantic mapping rules, modeling efficiency and the model's computational performance are improved, optimizing the workflow model through the extended modeling language and the collaboration mechanism does not address the issue of effectively integrating multimodal inputs such as text, sketches, and parameters.

[0005] Prior art three, Chinese patent, patent number: 202410229244.0 relates to a parallel semantics enhanced optimization design workflow collaborative modeling and simulation method, belonging to the field of system modeling and simulation. The method includes the following steps: extending the computation-independent model modeling language, and constructing an initial workflow model based on the extended computation-independent model modeling language; extending the platform-independent model modeling language, establishing semantic mapping rules, and automatically optimizing the initial workflow model based on the extended platform-independent model modeling language and the semantic mapping rules to obtain an optimized workflow model; and automatically deploying the optimized workflow model to achieve simulation. Although the modeling efficiency and the computing performance of the model are improved by adding parallel semantics to the modeling language to support the construction of a complex parallel workflow model, supporting the automatic optimization of the workflow model through semantic mapping rules, and automatically deploying the workflow model through BPMN mapping rules to achieve simulation, the parallel semantics enhancement and automatic deployment of the simulation efficiency are still unable to achieve the collaborative expression and generation of cross-modal data (text, sketches, parameters).

[0006] Currently, due to semantic heterogeneity and differences in expression, multimodal data such as text descriptions, hand-drawn sketches, and parameter constraints in existing technologies 1, 2, and 3 are difficult to integrate within a unified framework. This leads to distorted transmission of design intent and low automation generation efficiency. To address these issues, the present invention provides a method and system for AI-assisted design workflows based on a diffusion model. Summary of the Invention

[0007] The main purpose of the present invention is to provide a method and system for AI-assisted design workflow based on a diffusion model to solve the problem in the prior art that multimodal data such as text descriptions, hand-drawn sketches, parameter constraints, etc. are difficult to integrate in a unified framework due to semantic heterogeneity and differences in expression forms, resulting in distorted design intent transmission and low efficiency of automated generation.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A method for an AI-assisted design workflow based on a diffusion model includes receiving multimodal data, extracting key set features, and mapping the multimodal data into a unified space; generating a joint embedding vector for the multimodal data, generating multiple groups of candidate paths, and selecting the smallest generated path; injecting parameter constraints into a denoising process to generate a path; and converting the optimized latent vector into a vector graphic through a decoder to preserve geometric topological relationships.

[0010] As a further improvement of the present invention, multimodal data of text description, hand-drawn sketch and parameter constraints are received; the sketch is vectorized and key set features are extracted; the text is parsed into structured semantics; the parameter constraints are converted into mathematical expressions or boundary conditions; and the text semantics, sketch features and parameter constraints are mapped to a unified space.

[0011] As a further improvement of the present invention, the process of mapping text semantics, sketch features, and parameter constraints to a unified space includes the following steps:

[0012] Receive multimodal data including text description, hand-drawn sketch, and parametric constraints; convert the hand-drawn sketch into vector format using a vector algorithm, extract collective features such as intersections and lines; detect off-center intersections and generate straight line vectors aligned with the sketch edges;

[0013] Analyze text descriptions to extract key design parameters and structured semantics of material characteristics; identify rules and grammatical structures in unstructured text and convert them into triples and store them in the database; convert parameter constraints into mathematical expressions or boundary conditions;

[0014] Self-supervised learning tasks are used to align text semantics with sketch geometric features; the mathematical expressions of parameter constraints are integrated as boundary conditions and jointly embedded; and multimodal features are fused using a 768-dimensional joint embedding vector.

[0015] As a further improvement of the present invention, a 768-dimensional joint embedding vector is generated for text semantics, sketch features, and parameter constraints; multiple groups of candidate paths are generated, and the generation path with the smallest loss function is selected among the candidate paths; Gaussian noise is gradually added to the latent space to predict the noise distribution; and parameter constraints are injected into the denoising process to generate paths.

[0016] As a further improvement of the present invention, the process of generating multiple groups of candidate paths includes the following steps:

[0017] Encode the semantics of structured text to generate 768-dimensional text; map the vectorized hand-drawn sketch into a 768-dimensional vector; parse mathematical expressions and boundary conditions into symbolic representations and map them to a 768-dimensional vector space aligned with the text / hand-drawn sketch through learning;

[0018] The separate embeddings of 768-dimensional text, 768-dimensional sketch, and 768-dimensional parameters are concatenated into a 2304-dimensional vector, which is then reduced to 768 dimensions through a fully connected layer. Mathematical boundary conditions for parameter fusion are added during the fusion process.

[0019] Generate multiple sets of candidate paths and select the path with the smallest loss function among the candidate paths; gradually add Gaussian noise in the latent space and predict the noise distribution; inject parameter constraints into the denoising process to generate paths.

[0020] As a further improvement of the present invention, the process of generating 768-dimensional text includes the following steps:

[0021] The text is converted into a token sequence through a word segmenter and input into a program with a 12-layer hidden structure. Each token generates a 768-dimensional context-related vector. The global semantic vector of the entire text is extracted through an average pooling operation to generate a 768-dimensional text embedding.

[0022] The hand-drawn sketch after vectorization is subjected to feature extraction, and the vectorized path point coordinate sequence is input into the linear projection layer to map the geometric topological relationship into a 768-dimensional vector;

[0023] Mathematical expressions and boundary conditions are parsed into symbolic representations and mapped into a 768-dimensional vector space aligned with text and hand-drawn sketches through a learned embedding layer; numerical constraints and semantic features are operated in a unified space.

[0024] As a further improvement of the present invention, the process of adding mathematical boundary conditions for parameter fusion during the fusion process includes the following steps:

[0025] The 2304-dimensional concatenated vector is input to the fully connected layer; the weight matrix of the fully connected layer is set to 768x2304, and the bias vector dimension is 768; matrix multiplication is used to implement linear transformation, mapping each 2304-dimensional input sample to a 768-dimensional space;

[0026] During the matrix multiplication process, the parameters of the weight matrix are integrated with the mathematical boundary conditions. Through the regularization term and the constrained optimization objective function, the parameter distribution of the fully connected layer is made to meet the pre-defined mathematical boundary conditions.

[0027] Each input dimension is calculated by linearly combining all 2304 input dimensions to obtain each element in the output vector; through the encoder, the high-dimensional embedding is mapped into a low-dimensional space.

[0028] As a further improvement of the present invention, the process of gradually adding Gaussian noise in the latent space includes the following steps:

[0029] Initialize multiple latent vector samples in the latent space. Each sample is generated by gradually adding Gaussian noise of different intensities to form a different initial noise state. Perform multiple steps of denoising operations on each candidate path.

[0030] Use the prediction network to predict the noise distribution in the current latent variable and gradually remove the noise based on the prediction results; encode the mathematical expression and parameter constraints such as boundary conditions into a 768-dimensional vector and inject it into the value denoising process; when predicting the noise, align the parameter constraints with the latent variable characteristics;

[0031] At each denoising step, multiple possible noise adjustment directions are generated. By sampling different noise prediction results, the intermediate states of multiple candidate paths are generated. A function is calculated for the candidate paths generated at each step. If the cumulative loss of a candidate path exceeds a preset threshold, dynamic planning is performed and the branch is reselected.

[0032] As a further improvement of the present invention, the optimized latent vector is converted into a vector graphic through a decoder, preserving the geometric topological relationship; the output is a format compatible with CAD software, and metadata is generated; and the user's modification opinions on the design results are collected through an interactive interface.

[0033] To achieve the above object, the present invention also provides the following technical solutions:

[0034] A system for AI-assisted design workflow based on a diffusion model, which is applied to the method for AI-assisted design workflow based on a diffusion model, and the system for AI-assisted design workflow based on a diffusion model includes:

[0035] The unified multimodal data module is used to receive text descriptions, hand-drawn sketches, and parameter constraint multimodal data; vectorize the sketches and extract key set features; parse the text into structured semantics; convert parameter constraints into mathematical expressions or boundary conditions; and map the text semantics, sketch features, and parameter constraints into a unified space.

[0036] The path generation module is used to generate a 768-dimensional joint embedding vector for text semantics, sketch features, and parameter constraints; generate multiple sets of candidate paths and select the one with the lowest loss function among the candidate paths; gradually add Gaussian noise to the latent space and predict the noise distribution; and inject parameter constraints into the denoising process to generate paths;

[0037] The user interaction module is used to convert the optimized latent vector into vector graphics through a decoder, preserving the geometric topological relationship; output it to a format compatible with CAD software and generate metadata; and collect user modification opinions on the design results through an interactive interface.

[0038] The present invention generates a 768-dimensional joint embedding vector for text semantics, sketch features, and parameter constraints, simultaneously generates multiple groups of candidate paths, and selects the path with the smallest loss function; through the generation of high-dimensional embedding vectors, the model's ability to understand input information is improved, while the path selection mechanism is optimized to ensure the quality and efficiency of generated paths; Gaussian noise is gradually added to the latent space, and the noise distribution is predicted, while parameter constraints are injected into the denoising process to generate paths; it can also be directly applied to CAD software, improving the practical usability of design results; enhancing the flexibility and user-friendliness of the design process, making the final result more in line with the actual needs of users; shortening the design cycle, simplifying the process, and reducing human resource investment. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a schematic flow chart of steps of an embodiment of a method for AI-assisted workflow design based on a diffusion model of the present invention;

[0040] Figure 2 This is a flowchart of the steps for mapping text semantics, sketch features, and parameter constraints into a unified space in one embodiment of the AI-assisted design workflow method based on the diffusion model of the present invention;

[0041] Figure 3 A schematic flow chart of steps for generating multiple sets of candidate paths according to an embodiment of the method for AI-assisted design of a workflow based on a diffusion model of the present invention;

[0042] Figure 4 A schematic flow chart of steps for generating 768-dimensional text according to an embodiment of the method for AI-assisted design workflow based on a diffusion model of the present invention;

[0043] Figure 5 This is a flowchart of the steps for adding mathematical boundary conditions for parameter fusion during the fusion process of an embodiment of the AI-assisted design workflow method based on a diffusion model of the present invention;

[0044] Figure 6 A schematic flow chart of steps for generating multiple sets of candidate paths according to an embodiment of the method for AI-assisted design of a workflow based on a diffusion model of the present invention;

[0045] Figure 7 A schematic flow chart of a multi-step denoising operation for each candidate path according to an embodiment of the AI-assisted design workflow method based on a diffusion model of the present invention;

[0046] Figure 8 A schematic flow chart of steps for aligning parameter constraints with latent variable features in an embodiment of a method for AI-assisted design of a workflow based on a diffusion model of the present invention;

[0047] Figure 9 A flowchart showing the steps of dynamic programming in an embodiment of the AI-assisted workflow design method based on a diffusion model of the present invention;

[0048] Figure 10 This is a functional module diagram of an embodiment of a system for AI-assisted design workflow based on a diffusion model of the present invention;

[0049] Figure 11 This is a schematic structural diagram of an embodiment of an electronic device of the present invention;

[0050] Figure 12 This is a schematic structural diagram of an embodiment of a storage medium of the present invention. DETAILED DESCRIPTION

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0052] The terms "first", "second" and "third" in the present invention are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first", "second" and "third" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "multiple" is at least two, for example, two, three, etc., unless otherwise clearly and specifically defined. All directional indications in the embodiments of the present invention (such as up, down, left, right, front, back...) are only used to explain the relative position relationship, movement, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or devices.

[0053] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present invention. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute a separate or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0054] like Figure 1 As shown, this embodiment provides an embodiment of a method for AI-assisted design workflow based on a diffusion model. In this embodiment, the method for AI-assisted design workflow based on a diffusion model specifically includes the following steps:

[0055] Step S1: Receive multimodal data such as text description, hand-drawn sketch, and parameter constraints; vectorize the sketch and extract key set features; parse the text into structured semantics; convert the parameter constraints into mathematical expressions or boundary conditions; and map the text semantics, sketch features, and parameter constraints into a unified space.

[0056] Step S2: Generate a 768-dimensional joint embedding vector for text semantics, sketch features, and parameter constraints; generate multiple sets of candidate paths and select the path with the smallest loss function among the candidate paths; gradually add Gaussian noise to the latent space and predict the noise distribution; inject parameter constraints into the denoising process to generate paths;

[0057] Step S3: The optimized latent vector is converted into vector graphics through a decoder, preserving the geometric topological relationship; the output is a format compatible with CAD software and metadata is generated; and user modification suggestions on the design results are collected through an interactive interface.

[0058] Preferably, this embodiment receives multimodal data such as text descriptions, hand-drawn sketches, and parameter constraints, and finally maps this information into a unified space by vectorizing the sketches, parsing the text into structured semantics, and converting the parameter constraints into mathematical expressions or boundary conditions; it realizes the fusion of different modal data, provides comprehensive information support for subsequent generation tasks, and improves the accuracy and consistency of the generated content; generates a 768-dimensional joint embedding vector for text semantics, sketch features, and parameter constraints, and simultaneously generates multiple sets of candidate paths and selects the path with the smallest loss function; through the generation of high-dimensional embedding vectors, the model's ability to understand input information is improved , while optimizing the path selection mechanism to ensure the quality and efficiency of the generated path; gradually adding Gaussian noise to the latent space and predicting the noise distribution, while injecting parameter constraints into the denoising process to generate a path; converting the optimized latent vector into a vector graphic through a decoder while retaining the geometric topological relationship, and outputting it in a format compatible with CAD software; ensuring that the generated vector graphics not only meet the design requirements, but can also be directly applied to CAD software, thereby improving the practical usability of the design results; enhancing the flexibility and user-friendliness of the design process, making the final result closer to the actual needs of users, shortening the design cycle, simplifying the process and reducing human resource investment.

[0059] Furthermore, if Figure 2 As shown, the process of mapping text semantics, sketch features, and parameter constraints to a unified space in step S1 specifically includes the following steps:

[0060] Step S11: receiving multimodal data such as text description, hand-drawn sketch, and parameter constraints; converting the hand-drawn sketch into a vector format using a vector algorithm, extracting collective features such as intersections and lines; detecting off-center intersections, and generating straight line vectors aligned with the sketch edges;

[0061] Step S12: Analyze the text description and extract structured semantics such as key design parameters and material characteristics; identify the rules and grammatical structures in the unstructured text and convert them into triples and store them in the database; convert parameter constraints into mathematical expressions or boundary conditions;

[0062] Step S13: Use self-supervised learning tasks to align text semantics with sketch geometric features; integrate the mathematical expressions of parameter constraints as boundary conditions and jointly embed them; use a 768-dimensional joint embedding vector to fuse multimodal features.

[0063] Preferably, this embodiment receives multimodal data such as text descriptions, hand-drawn sketches and parameter constraints, converts the hand-drawn sketches into vector format through a vector algorithm, and extracts collective features such as intersections and lines; detects eccentric intersections and generates straight line vectors aligned with the edges of the sketch; improves the digitization accuracy of the hand-drawn sketch, facilitating subsequent geometric feature extraction and analysis; through vectorization processing, realizes the conversion of the sketch from unstructured to structured, providing a basis for subsequent cross-modal learning; analyzes the text description, extracts structured semantics such as key design parameters and material characteristics; identifies rules and grammatical structures in unstructured text, and converts them into triple data and stores them in a database; converts parameter constraints into mathematical expressions or boundary conditions; uses the mathematical expressions of parameter constraints as boundary conditions, integrates and embeds them into a 768-dimensional joint embedding vector, and fuses multimodal features; achieves effective alignment of text semantics with sketch geometric features, and improves the cross-modal understanding ability of the model; uses the 768-dimensional joint embedding vector to fuse multimodal features, enhances the model's understanding of complex scenes, and at the same time improves prediction accuracy and generalization ability.

[0064] Furthermore, if Figure 3 As shown, the process of generating multiple groups of candidate paths in step S2 specifically includes the following steps:

[0065] Step S21: Encode the structured text semantics to generate 768-dimensional text; map the vectorized hand-drawn sketch into a 768-dimensional vector; parse the mathematical expressions and boundary conditions into symbolic representations, and map them into a 768-dimensional vector space aligned with the text / hand-drawn sketch through learning;

[0066] Step S22: Concatenate the 768-dimensional text, 768-dimensional sketch, and 768-dimensional parameters into a 2304-dimensional vector, and then reduce the dimensionality to 768 dimensions through a fully connected layer; add mathematical boundary conditions for parameter fusion during the fusion process;

[0067] Step S23: Generate multiple groups of candidate paths and select the generated path with the smallest loss function among the candidate paths; gradually add Gaussian noise in the latent space to predict the noise distribution; inject parameter constraints into the denoising process to generate paths.

[0068] Preferably, this embodiment encodes the semantics of structured text into a 768-dimensional vector, maps the hand-drawn sketch into a 768-dimensional vector through vectorization, and parses the mathematical expressions and boundary conditions into symbolic representations and maps them to a 768-dimensional vector space; by unifying different types of input (text, sketches, mathematical expressions) into the same 768-dimensional vector space, cross-modal semantic alignment is achieved. The three 768-dimensional vectors are spliced ​​into a 2304-dimensional vector, and then reduced to 768 dimensions through a fully connected layer, and mathematical boundary conditions are added in the fusion process; through splicing and dimensionality reduction operations, effective integration of multimodal features is achieved, and mathematical boundary conditions are introduced to ensure the rationality of the generated path. In addition, the use of the fully connected layer improves the efficiency of feature extraction, enabling the model to better capture the association between different modalities. Generate multiple groups of candidate paths and select the path with the smallest loss function; gradually add Gaussian noise in the latent space to predict the noise distribution; inject parameter constraints into the denoising process to generate the final path; optimize the loss function to select the optimal path, thereby improving the quality and accuracy of the generated path.

[0069] Furthermore, if Figure 4 As shown, the process of generating 768-dimensional text in step S21 specifically includes the following steps:

[0070] Step S211: The text is converted into a token sequence through a word segmenter and input into a program with a 12-layer hidden structure. Each token generates a 768-dimensional contextual correlation vector. The global semantic vector of the entire text is extracted through an average pooling operation to generate a 768-dimensional text embedding.

[0071] Step S212: extract features from the vectorized hand-drawn sketch, input the vectorized path point coordinate sequence into the linear projection layer, and map the geometric topological relationship into a 768-dimensional vector;

[0072] Step S213: Parse the mathematical expressions and boundary conditions into symbolic representations, map them to a 768-dimensional vector space aligned with text / hand-drawn sketches through a learning embedding layer, and operate the numerical constraints and semantic features in a unified space.

[0073] Preferably, this embodiment converts the text into a token sequence through a word segmenter, and each token is mapped to a 768-dimensional vector; a global semantic vector is extracted through average pooling, and local information is integrated into a global representation, thereby capturing the overall semantic features of the text; a 768-dimensional text embedding vector is generated, which can capture the semantics and contextual relationships of the text, and at the same time, the model's ability to understand the overall semantics of the text is improved through the global average pooling operation; the sketch is converted into a vectorized path point coordinate sequence; the geometric topological relationship is mapped to a 768-dimensional vector through a linear projection layer; by converting the geometric information of the hand-drawn sketch into a high-dimensional vector, the alignment of geometric features and text features is achieved, so that data from different modalities can be compared and fused in the same vector space. This technology can significantly improve the expressiveness and generalization capabilities of the model in multimodal tasks. Parsing mathematical expressions and boundary conditions into symbolic representations is the core technology of mathematical expression parsing; mapping mathematical expressions and boundary conditions into a 768-dimensional vector space by learning the embedding layer, which is aligned with the representation of text and hand-drawn sketches; operating numerical constraints and semantic features in a unified space, which shows that the model can process symbolic and semantic features at the same time, achieving efficient cross-modal calculation and fusion; by converting mathematical expressions and boundary conditions into a 768-dimensional vector space that is aligned with the representation of text and hand-drawn sketches, the model can realize efficient cross-modal calculation and fusion. Figure 1 The consistent vector representation realizes unified representation and operation across modalities.

[0074] Furthermore, if Figure 5 As shown, the process of adding mathematical boundary conditions for parameter fusion in the fusion process in step S22 specifically includes the following steps:

[0075] Step S221: Input the 2304-dimensional concatenated vector into the fully connected layer; set the weight matrix dimension of the fully connected layer to 768x2304 and the bias vector dimension to 768; use matrix multiplication to implement linear transformation to map each 2304-dimensional input sample to a 768-dimensional space;

[0076] Step S222: During the matrix multiplication process, the parameters of the weight matrix are integrated with the mathematical boundary conditions. Through the regularization term and the constrained optimization objective function, the parameter distribution of the fully connected layer is made to meet the pre-defined mathematical boundary conditions.

[0077] Step S223: Each input dimension is calculated by linearly combining all 2304 input dimensions to obtain each element in the output vector; through the encoder, the high-dimensional embedding is mapped into the low-dimensional space.

[0078] Preferably, in this embodiment, the input vector (dimension is 2304) is matrix multiplied with the weight matrix (dimension is 768x2304) to map the input sample from the high-dimensional space to the low-dimensional space (768 dimensions); during the matrix multiplication process, the parameters of the weight matrix are constrained by mathematical boundary conditions, and the objective function is optimized through regularization terms and constraints to ensure that the parameter distribution conforms to predefined rules; an encoder is used to map the high-dimensional input data to the low-dimensional space, thereby reducing the computational complexity and retaining key features; through matrix multiplication, the fully connected layer can efficiently extract the key features of the input data and map it to the target dimension (768 dimensions); through the fusion of regularization constraints and mathematical boundary conditions, the parameter distribution of the fully connected layer is more reasonable; after mapping the high-dimensional input data to the low-dimensional space, the computational complexity is significantly reduced while retaining the key features of the data; the application of regularization constraints and boundary conditions enhances the model's resistance to noise.

[0079] Furthermore, if Figure 6 As shown, the process of gradually adding Gaussian noise in the latent space in step S23 specifically includes the following steps:

[0080] Step S231: Initialize multiple latent vector samples in the latent space, each sample is generated by gradually adding Gaussian noise of different intensities to form a different initial noise state; perform a multi-step denoising operation on each candidate path;

[0081] Step S232: Use the prediction network to predict the noise distribution in the current latent variable, and gradually remove the noise based on the prediction results; encode the mathematical expression and parameter constraints such as boundary conditions into a 768-dimensional vector and inject it into the value denoising process; when predicting the noise, align the parameter constraints with the latent variable characteristics;

[0082] Step S233: At each denoising step, multiple possible noise adjustment directions are generated. By sampling different noise prediction results, the intermediate states of multiple candidate paths are generated. A function is calculated for the candidate paths generated at each step. If the cumulative loss of a candidate path is higher than a preset threshold, dynamic programming is performed to reselect the branch.

[0083] Preferably, this embodiment initializes multiple latent vector samples in the latent space and generates an initial noise state by gradually adding Gaussian noise of varying intensities. A multi-step denoising operation is performed on each candidate path. In each step, the prediction network predicts the noise distribution in the current latent variable and gradually removes the noise based on the predicted results. Parameter constraints, such as mathematical expressions and boundary conditions, are encoded into a 768-dimensional vector and injected into the denoising process. During each denoising step, multiple possible noise adjustment directions are generated, and by sampling different noise prediction results, intermediate states of multiple candidate paths are generated. If the cumulative loss of a candidate path exceeds a preset threshold, dynamic programming is performed to reselect the branch. The generated path is more closely aligned with the actual data distribution while satisfying the specified mathematical expressions and boundary conditions. The dynamic programming mechanism allows for timely adjustments during the path selection process, avoiding failures caused by excessive cumulative loss. It can also reduce unnecessary computation while ensuring generation quality. By encoding high-dimensional features, such as mathematical expressions and boundary conditions, and injecting them into the denoising process, the generated path better integrates multiple information sources, thereby improving the overall performance of the generation task.

[0084] Furthermore, if Figure 7 As shown, the process of performing the multi-step denoising operation on each candidate path in step S231 specifically includes the following steps:

[0085] Step S2311: Expand the potential connections of the candidate special nodes and / or combine different nodes to form new paths to generate diversified path branches; based on the processed candidate special nodes, calculate the feasibility or performance index of each path to form a candidate special path set;

[0086] Step S2312: Combine the candidate special paths with the candidate main paths to generate a set of candidate test paths. Among all the candidate paths, select the path with the smallest loss function value as the generated path;

[0087] Step S2313: gradually add Gaussian noise in the latent space to predict the noise distribution; inject parameter constraints into the denoising process to generate a path.

[0088] Preferably, this embodiment generates diversified path branches by expanding special nodes (such as concurrent process blocks) and combining different nodes. At the same time, the depth-first algorithm is combined to calculate the path, and a set of candidate special paths is screened out according to performance indicators; a diversified path set covering the test case requirements can be generated, thereby improving the coverage and comprehensiveness of the test cases. In addition, through performance indicator screening, it is ensured that the generated path has high practicality and effectiveness; the candidate special paths are combined with the main path to form a set of candidate test paths, and the pros and cons of each path are evaluated by the loss function, and the optimal path is finally selected; the path selection is optimized by the loss function to ensure that the generated path has higher efficiency and performance while meeting the test requirements. Gaussian noise is gradually added to the latent space, and the denoising process is optimized using parameter constraints to generate high-quality paths; this step can effectively improve the quality of the generated path through latent space noise injection and parameter constraint optimization. The latent space representation reduces the amount of calculation, while the noise prediction and denoising process improve the stability and reliability of the generated path.

[0089] Furthermore, if Figure 8 As shown, the process of aligning the parameter constraints with the latent variable features in step S232 specifically includes the following steps:

[0090] Step S2321: The prediction network receives the latent variable state at the current time step, generates a 768-dimensional parameter constraint vector encoded as a conditional input, and the latent variable gradually extracts multi-scale spatial features through the predicted encoder part;

[0091] Step S2322: The parameter constraint vector interacts with the predicted intermediate features; a cross-attention mechanism is inserted into the predicted encoder layer to dynamically align the parameter constraint semantic information with the spatial features of the latent variable through attention weight calculation;

[0092] Step S2323: The last layer of the prediction network outputs the predicted noise distribution, and the parameter constraint vector is integrated into the prediction process through the conditional mechanism; multi-scale information is integrated through residual connections and skip connections, and the noise intensity is dynamically adjusted according to the predicted noise.

[0093] Preferably, the prediction network in this embodiment receives the state of the latent variable at the current time step and generates a 768-dimensional parameter constraint vector as a conditional input. The latent variable is gradually extracted with multi-scale spatial features through the encoder portion of the prediction. Multi-scale feature extraction can capture spatial information at different levels, thereby enhancing the model's ability to understand complex data. The parameter constraint vector interacts with the intermediate features of the prediction, and a cross-attention mechanism is inserted in the encoder layer of the prediction. Attention weight calculation dynamically aligns the semantic information of the parameter constraints with the spatial features of the latent variable. The parameter constraint vector interacts with the intermediate features of the prediction, and a cross-attention mechanism is inserted in the encoder layer of the prediction. Attention weight calculation dynamically aligns the semantic information of the parameter constraints with the spatial features of the latent variable. Residual connections help alleviate the vanishing gradient problem and improve the training efficiency of deep networks. Skip connections allow low-level features to be directly transferred to higher levels, enhancing the model's expressive power. The final layer of the prediction network outputs the predicted noise distribution and incorporates the parameter constraint vector into the prediction process through a conditional mechanism. This conditional mechanism enables the model to better utilize external constraint information, thereby improving prediction accuracy and robustness.

[0094] Furthermore, if Figure 9 As shown, the process of dynamic programming in step S233 specifically includes the following steps:

[0095] Step S2331: In each denoising step, the noise distribution in the current latent variable is predicted by the prediction network, and multiple possible noise adjustment directions are generated; the predicted noise distribution is compared with the historical noise distribution to obtain the noise prediction difference;

[0096] Step S2332: Encode the mathematical expression and boundary conditions into a 768-dimensional vector and align it with the latent variable features to measure the degree of satisfaction of the parameter constraints; the loss function of all candidate paths is accumulated at each time step; if the cumulative loss of a time path exceeds a preset threshold, the dynamic programming mechanism is triggered;

[0097] Among them, the conditional guidance vector is introduced to manipulate the noise prediction network in real time:

[0098] Noise prediction and difference analysis (step S2331), noise multi-directional prediction, noise prediction network ε θ Generate N candidate noise directions simultaneously:

[0099]

[0100] in: represents the i-th candidate noise direction; x t represents the latent variable at time step t; y represents the constraint encoding vector;

[0101] Historical comparison mechanism:

[0102] Maintain a sliding window to store the noise distribution of the last K steps:

[0103] H t =[ε t-K ,ε t-K+1 ,…,ε t-1 ]

[0104] Calculate the Mahalanobis distance Δ between the candidate noise and the historical distribution (i) :

[0105]

[0106] Where: μ H , Represents the sliding window H t The mean and covariance matrix of ;

[0107] Constraint encoding (step S2332), mathematical expression encoding, using the BERT-base architecture to convert the constraint y into a 768-dimensional vector:

[0108]

[0109] Constraint satisfaction calculation, aligning latent variables with constraints through the attention mechanism:

[0110]

[0111] Satisfaction score:

[0112] s t =σ(W s ·(αV(v constraint )))∈[0,1]

[0113] Where: Q, K, V represent query, key, and value matrices; W s represents the learnable weight matrix; σ represents the Sigmoid activation function;

[0114] Dynamic loss accumulation, constructing a composite loss function:

[0115]

[0116] in: represents the total variation regularization term, which suppresses high-frequency noise; λ1:λ2:λ3=3:2:1 represents the loss weight ratio;

[0117] Conditional guidance mechanism, dynamic parameter adjustment. Constraint-driven affine transformation:

[0118]

[0119] Latent variable update formula:

[0120] x t-1 =W T (y)·x t +b T (y)

[0121] Path screening algorithm, maintains a priority queue \(\mathcal{Q}\) to record the top-K paths:

[0122]

[0123] (Update with constrained gradients)

[0124] Calculate L total (path∪Δx)

[0125] (New Path)

[0126] Output:

[0127] Dynamic programming trigger logic, when any of the following conditions are met, trigger path resampling:

[0128] Sudden loss:

[0129] Constraint Stagnation:

[0130] From n(μ H ,∑ h ) Sample M new directions and supplement the candidate pool;

[0131] Multi-path exploration searches for the optimal denoising path in parallel across N candidate noise directions. Dynamic stability control uses a sliding window to monitor noise distribution shifts, and a total variation regularization term to suppress unphysical oscillations. Constraint fusion uses BERT to encode semantic constraints, and an attention mechanism dynamically aligns latent variables with constraints. A closed-loop optimization priority queue enables dynamic path selection and resampling. This approach achieves the synergy between noise prediction, constraint fusion, and path optimization, while ensuring mathematical rigor while clarifying the physical meaning of each symbol and the engineering control logic.

[0132] Step S2333: Filter the branch path with the lowest cumulative loss and eliminate the path with high loss; when reselecting a branch, adjust the noise prediction direction in combination with the gradient information of the parameter constraint.

[0133] Preferably, in each denoising step, this embodiment predicts the noise distribution in the current latent variable through the prediction network and generates multiple possible noise adjustment directions; by predicting the noise distribution, the noise characteristics in the current state are understood more accurately, thereby generating more effective noise adjustment directions and improving the accuracy and efficiency of denoising; the predicted noise distribution is compared with the historical noise distribution to obtain the noise prediction difference; by comparing the historical noise distribution, the model can identify the changing trend and abnormal conditions of the noise, thereby dynamically adjusting the denoising strategy to improve the robustness and adaptability of the denoising; the mathematical expression and boundary conditions are encoded into a 768-dimensional vector, aligned with the latent variable characteristics, and the degree of satisfaction of the parameter constraints is measured; by encoding the mathematical expression and boundary conditions, the model can better understand and satisfy the parameter constraints, ensure that the generated path meets the actual needs, and improve the accuracy and reliability of the generated results; the loss function of all candidate paths is accumulated by time step; if the cumulative loss of a certain time path exceeds the preset threshold value, the dynamic programming mechanism is triggered; through the dynamic programming mechanism, the model can effectively screen out the optimal path, avoid unnecessary calculation and resource waste, and improve overall efficiency and performance; introduce conditional guidance vectors to manipulate the noise prediction network in real time; through the conditional guidance vectors, the model can dynamically adjust the noise prediction direction according to the current state and conditions, improving the flexibility and adaptability of denoising; use dynamic adjustment matrices W_T and b_T to adjust the noise prediction direction; through the dynamic adjustment matrix, the model can adjust the noise prediction direction in real time, improving the accuracy and efficiency of denoising; screen the branch path with the lowest cumulative loss and eliminate the path with high loss; by screening the branch path with the lowest cumulative loss, the model can effectively select the optimal path and improve the quality and efficiency of the generated results; when reselecting a branch, the noise prediction direction is adjusted in combination with the gradient information of the parameter constraints; by combining the gradient information of the parameter constraints, the model can more accurately adjust the noise prediction direction, improving the accuracy and reliability of the generated results.

[0134] like Figure 10 As shown, this embodiment further provides an embodiment of a system for an AI-assisted design workflow based on a diffusion model. In this embodiment, the system for an AI-assisted design workflow based on a diffusion model is applied to the method for an AI-assisted design workflow based on a diffusion model in the above embodiment. The system for an AI-assisted design workflow based on a diffusion model includes:

[0135] The unified multimodal data module 1 is used to receive multimodal data such as text descriptions, hand-drawn sketches, and parameter constraints; vectorize the sketches and extract key set features; parse the text into structured semantics; convert parameter constraints into mathematical expressions or boundary conditions; and map the text semantics, sketch features, and parameter constraints into a unified space;

[0136] Generate Path Module 2, which is used to generate a 768-dimensional joint embedding vector for text semantics, sketch features, and parameter constraints; generate multiple sets of candidate paths and select the one with the smallest loss function among the candidate paths; gradually add Gaussian noise to the latent space and predict the noise distribution; inject parameter constraints into the denoising process to generate paths;

[0137] User interaction module 3 is used to convert the optimized latent vector into vector graphics through a decoder, preserving the geometric topological relationship; output it to a format compatible with CAD software and generate metadata; and collect user modification opinions on the design results through an interactive interface.

[0138] Preferably, this embodiment receives multimodal data such as text descriptions, hand-drawn sketches, and parameter constraints, vectorizes the sketches, and extracts key set features; parses the text into structured semantics; converts parameter constraints into mathematical expressions or boundary conditions, and maps this information into a unified space. By mapping data from different modalities (such as text, images, and constraints) into a unified semantic space, cross-modal semantic alignment and fusion are achieved. A 768-dimensional joint embedding vector is generated, and multiple sets of candidate paths are generated based on these vectors. Gaussian noise is gradually added to the latent space; the noise distribution is predicted; and parameter constraints are injected to optimize the denoising process, ultimately generating the target path. By combining the embedding vector and the predicted noise distribution, the model can generate more accurate design paths that meet the constraints. The optimized latent vector is converted into vector graphics through a decoder, preserving geometric topology. Simultaneously, metadata is generated and output in a format compatible with CAD software. Furthermore, user feedback is collected through an interactive interface for further design optimization. This enhances the system's flexibility and user experience. By preserving geometric topology and compatibility with CAD formats, users can intuitively modify the design results, while the generated metadata provides fundamental support for subsequent iterations. This two-way interactive mechanism significantly improves design efficiency and user satisfaction. Through the collaborative work of three modules: unified multimodal data processing, path optimization, and user interaction, data from different modalities is mapped into a unified space. By combining embedding vectors and noise distribution prediction, a high-quality design path that meets the constraints is generated. The user interaction module supports real-time feedback and modifications, making the design process more flexible and efficient. The output results are directly compatible with CAD software, reducing the difficulty of design implementation.

[0139] like Figure 11 As shown, this embodiment provides an embodiment of an electronic device. In this embodiment, the electronic device 4 includes a processor 41 and a memory 42 coupled to the processor 41.

[0140] The memory 42 stores program instructions for implementing the layout method of the AI-assisted design workflow method based on the diffusion model in any of the above embodiments.

[0141] The processor 41 is configured to execute program instructions stored in the memory 42 to perform the layout of the AI-assisted design workflow method based on the diffusion model.

[0142] The processor 41 may also be referred to as a CPU (Central Processing Unit). The processor 41 may be an integrated circuit chip having signal processing capabilities. The processor 41 may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic device, a discrete gate or transistor logic device, or a discrete hardware component. The general-purpose processor may be a microprocessor or any conventional processor.

[0143] Further, Figure 12 This is a schematic diagram of the structure of a storage medium in an embodiment of the present application. The storage medium 5 in the embodiment of the present application stores program instructions 51 that can implement all the above methods, wherein the program instructions 51 can be stored in the above storage medium in the form of a software product, including a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or terminal devices such as a computer, server, mobile phone, and tablet.

[0144] In the several embodiments provided by the present invention, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interface, device or unit, which can be electrical, mechanical or other forms.

[0145] In addition, the functional units in the various embodiments of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated units may be implemented in the form of hardware or in the form of software functional units. The above is only an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

[0146] The above detailed description of the specific embodiments of the invention is intended to be illustrative only, and the present invention is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions to the invention are also within the scope of the present invention. Therefore, equivalent changes, modifications, and improvements made without departing from the spirit and scope of the present invention are also encompassed within the scope of the present invention.

Claims

1. A method for AI-assisted design workflow based on a diffusion model, characterized in that: The method for AI-assisted design of a workflow based on a diffusion model includes receiving multimodal data, extracting key set features, and mapping the multimodal data into a unified space; generating a joint embedding vector for the multimodal data, generating multiple sets of candidate paths, and selecting the smallest generated path; Inject parameter constraints into the denoising process to generate paths; The optimized latent vector is converted into vector graphics through the decoder, preserving the geometric topological relationship.

2. The method for AI-assisted design workflow based on diffusion model according to claim 1, characterized in that: Receive multimodal data including text description, hand-drawn sketch and parameter constraints; vectorize the sketch and extract key set features; parse the text into structured semantics; convert parameter constraints into mathematical expressions or boundary conditions; and map text semantics, sketch features and parameter constraints into a unified space.

3. The method for AI-assisted design workflow based on diffusion model according to claim 2, characterized in that: The process of mapping text semantics, sketch features, and parameter constraints into a unified space includes the following steps: Receive multimodal data including text description, hand-drawn sketch, and parameter constraints; convert the hand-drawn sketch into vector format using a vector algorithm, extract intersection points and line set features; detect off-center intersection points and generate line vectors aligned with the sketch edge; Analyze text descriptions to extract key design parameters and structured semantics of material characteristics; identify rules and grammatical structures in unstructured text and convert them into triples and store them in the database; convert parameter constraints into mathematical expressions or boundary conditions; Self-supervised learning tasks are used to align text semantics with sketch geometric features; the mathematical expressions of parameter constraints are integrated as boundary conditions and jointly embedded; and multimodal features are fused using a 768-dimensional joint embedding vector.

4. The method for AI-assisted design workflow based on diffusion model according to claim 1, characterized in that: A 768-dimensional joint embedding vector is generated for text semantics, sketch features, and parameter constraints; multiple sets of candidate paths are generated, and the generation path with the smallest loss function is selected among the candidate paths; Gaussian noise is gradually added to the latent space to predict the noise distribution; and parameter constraints are injected into the denoising process to generate paths.

5. The method for AI-assisted design workflow based on diffusion model according to claim 4, characterized in that: The process of generating multiple sets of candidate paths includes the following steps: Encode the semantics of structured text to generate 768-dimensional text; map the vectorized hand-drawn sketch into a 768-dimensional vector; parse mathematical expressions and boundary conditions into symbolic representations and map them to a 768-dimensional vector space aligned with the text / hand-drawn sketch through learning; The separate embeddings of 768-dimensional text, 768-dimensional sketch, and 768-dimensional parameters are concatenated into a 2304-dimensional vector, which is then reduced to 768 dimensions through a fully connected layer. Mathematical boundary conditions for parameter fusion are added during the fusion process. Generate multiple sets of candidate paths and select the path with the smallest loss function among the candidate paths; gradually add Gaussian noise in the latent space and predict the noise distribution; inject parameter constraints into the denoising process to generate paths.

6. The method for AI-assisted design workflow based on diffusion model according to claim 5, characterized in that: The process of generating 768-dimensional text includes the following steps: The text is converted into a token sequence through a word segmenter and input into a program with a 12-layer hidden structure. Each token generates a 768-dimensional context-related vector. The global semantic vector of the entire text is extracted through an average pooling operation to generate a 768-dimensional text embedding. The hand-drawn sketch after vectorization is subjected to feature extraction, and the vectorized path point coordinate sequence is input into the linear projection layer to map the geometric topological relationship into a 768-dimensional vector; Mathematical expressions and boundary conditions are parsed into symbolic representations and mapped into a 768-dimensional vector space aligned with text and hand-drawn sketches through a learned embedding layer; numerical constraints and semantic features are operated in a unified space.

7. The method for AI-assisted design workflow based on diffusion model according to claim 5, characterized in that: The process of adding mathematical boundary conditions for parameter fusion during the fusion process includes the following steps: The 2304-dimensional concatenated vector is input to the fully connected layer; the weight matrix of the fully connected layer is set to 768x2304, and the bias vector dimension is 768; matrix multiplication is used to implement linear transformation, mapping each 2304-dimensional input sample to a 768-dimensional space; During the matrix multiplication process, the parameters of the weight matrix are integrated with the mathematical boundary conditions. Through the regularization term and the constrained optimization objective function, the parameter distribution of the fully connected layer is made to meet the pre-defined mathematical boundary conditions. Each input dimension is calculated by linearly combining all 2304 input dimensions to obtain each element in the output vector; through the encoder, the high-dimensional embedding is mapped into a low-dimensional space.

8. The method for AI-assisted design workflow based on diffusion model according to claim 5, characterized in that: The process of gradually adding Gaussian noise in the latent space includes the following steps: Initialize multiple latent vector samples in the latent space. Each sample is generated by gradually adding Gaussian noise of different intensities to form a different initial noise state. Perform multiple steps of denoising operations on each candidate path. Use the prediction network to predict the noise distribution in the current latent variable and gradually remove the noise based on the prediction results; encode the mathematical expression and boundary condition parameter constraints into a 768-dimensional vector and inject it into the value denoising process; when predicting the noise, align the parameter constraints with the latent variable characteristics; At each denoising step, multiple possible noise adjustment directions are generated. By sampling different noise prediction results, the intermediate states of multiple candidate paths are generated. A function is calculated for the candidate paths generated at each step. If the cumulative loss of a candidate path exceeds a preset threshold, dynamic planning is performed and the branch is reselected.

9. The method for AI-assisted design workflow based on diffusion model according to claim 1, characterized in that: The optimized latent vector is converted into vector graphics through the decoder, preserving the geometric topology relationship; The output is in a format compatible with CAD software and metadata is generated; user comments on the design results are collected through the interactive interface.

10. A system for AI-assisted design workflow based on a diffusion model, which is applied to the method for AI-assisted design workflow based on a diffusion model according to any one of claims 1 to 9, characterized in that: The AI-assisted design workflow system based on the diffusion model includes: The unified multimodal data module is used to receive text descriptions, hand-drawn sketches, and parameter constraint multimodal data; vectorize the sketches and extract key set features; parse the text into structured semantics; convert parameter constraints into mathematical expressions or boundary conditions; and map the text semantics, sketch features, and parameter constraints into a unified space. The path generation module is used to generate a 768-dimensional joint embedding vector for text semantics, sketch features, and parameter constraints; generate multiple sets of candidate paths and select the one with the lowest loss function among the candidate paths; gradually add Gaussian noise to the latent space and predict the noise distribution; and inject parameter constraints into the denoising process to generate paths; The user interaction module is used to convert the optimized latent vector into vector graphics through a decoder, preserving the geometric topological relationship; output it to a format compatible with CAD software and generate metadata; and collect user modification opinions on the design results through an interactive interface.

Images