A simulation system and method based on digital intelligent IP creative product design

Abstract

This invention discloses a simulation system and method for designing digital intelligent IP cultural and creative products, relating to the field of virtual simulation. The method includes: constructing a multi-source IP material database; collecting cultural materials for digitization and structuring; analyzing features using a multimodal model to construct an IP knowledge graph; retrieving the graph based on user keywords; generating product concept sketches and descriptions conforming to the IP style using a generative network; converting the scheme into a parametric 3D model for virtual simulation to evaluate aesthetic, physical, and interactive performance; combining a lightweight multi-dimensional objective classification prediction model for classification and rating; adjusting generation conditions through optimization algorithms, iteratively simulating, evaluating, and optimizing to obtain the final product scheme. The advantages of this invention are: relying on multimodal technology and IP knowledge graphs to restore the core of IP culture; and through full-dimensional virtual simulation, classification quantitative prediction, and intelligent closed-loop iteration, efficiently producing cultural and creative schemes, thereby enhancing the cultural quality and market competitiveness of products.

Description

Technical Field

[0001] This invention relates to the field of virtual simulation, and in particular to a simulation system and method based on the design of digital intelligent IP cultural and creative products. Background Technology

[0002] With the maturation of technologies such as artificial intelligence, big data, and virtual simulation, traditional IP development models face limitations in creative expression, user interaction, and market adaptability. Especially in emerging industries such as virtual idols, digital collectibles, and immersive experiences, how to efficiently achieve the dynamic coupling of IP image, core story, and user experience has become a key industry challenge.

[0003] The current traditional IP-based cultural and creative design methods on the market have a low level of intelligence and systematization, with many significant shortcomings. Traditional methods lack a standardized multi-source IP material processing system and structured knowledge graph support. Their analysis of core characteristics such as visual style, narrative connotation, and character traits is fragmented and one-sided, relying heavily on manual experience to select materials and conceptualize solutions. This is highly subjective, inefficient, and prone to problems such as product cultural distortion, element abuse, and design homogenization. Furthermore, existing design methods only perform basic visual verification, lacking comprehensive verification steps such as 3D physical simulation, interactive experience simulation, and cultural semantic verification. This makes it impossible to proactively identify product structural defects, interactive lag, and poor user experience. Moreover, the industry generally lacks a standardized, comprehensive, and objective design quality evaluation system, relying solely on manual experience to judge design quality, resulting in highly subjective and inaccurate evaluation results. In addition, traditional models lack a data-driven closed-loop iteration mechanism, leading to blind rectification and optimization. Extensive physical trial and error is necessary, resulting in high costs, low iteration efficiency, and ultimately, difficulty in guaranteeing the cultural compatibility, user experience, and market competitiveness of the cultural and creative products produced. Summary of the Invention

[0004] To improve existing methods and systems, this paper provides a simulation system and method for designing cultural and creative products based on digital intelligent IP. This method relies on multimodal technology and IP knowledge graph to accurately restore the core of IP culture. Through full-dimensional virtual simulation, multi-dimensional objective classification and evaluation, and intelligent closed-loop iteration, it gets rid of the subjective limitations and trial-and-error drawbacks of traditional design, efficiently produces cultural and creative solutions, and enhances the cultural quality and market competitiveness of products.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A simulation method for designing cultural and creative products based on digital intelligent IP includes:

[0007] Construct a multi-source heterogeneous IP material database, collect the original cultural materials of the target IP, and perform digital cleaning, standardization processing and structured annotation to generate standardized IP material units with multi-dimensional tags, which are then stored in the IP material database;

[0008] Based on a pre-trained multimodal neural network model, standardized IP material units are analyzed to extract visual style features, narrative theme features, character personality features, and emotional tone features of the IP. Based on the features and semantic relationships between materials, a structured IP knowledge graph is constructed.

[0009] Based on user-input keywords, semantic retrieval and association reasoning are performed in the IP knowledge graph to locate IP element clusters. A generative adversarial network model is invoked, and the features of the retrieved IP element clusters are used as constraints to generate preliminary product concept visual sketches and text descriptions that conform to the IP visual style and theme. These are then integrated into a product concept solution.

[0010] The product concept solution is transformed into a parametric 3D digital model. In a virtual simulation environment, the digital model is given materials, physical properties and interactive logic. Static aesthetics and cultural semantics simulation are performed. Physical behavior simulation is performed for physical product concepts. User interaction behavior simulation agents are built for digital interactive product concepts. User operation paths are simulated, and the smoothness of the interaction process and the degree of achievement of experience goals are evaluated.

[0011] A lightweight, multi-dimensional objective classification prediction model is constructed. It is trained based on the objective design attributes and simulation performance data of similar historical IP cultural and creative products. The product schemes and their feature vectors after simulation verification are input into the classification prediction model to classify and rate the products in multiple dimensions, including IP fit, design compliance, cultural semantic consistency, physical performance compliance and interactive experience adaptability.

[0012] Based on simulation performance data and classification prediction data, an optimization algorithm is designed. The input conditions of the generative adversarial network model are adjusted in reverse to drive the generation of new alternative optimization schemes. The simulation, evaluation, and optimization cycle is repeated until the product scheme meets the preset comprehensive evaluation threshold or reaches the upper limit of the number of iterations, and the final product concept scheme is obtained.

[0013] Preferably, the construction of the multi-source heterogeneous IP material database specifically includes:

[0014] Collect text, images, audio, video, and 3D data of IP from official databases, publicly available works, and physical objects, and scan and model non-digital physical objects;

[0015] Remove watermarks and noise from images and videos, correct colors and brightness, reduce noise and unify volume in audio, and unify text encoding, filter garbled characters, and proofread key information.

[0016] All materials are converted to a common format, the resolution and frame rate of images and videos are standardized to a preset standard, the audio sampling rate and bit depth are standardized, the coordinate system of the 3D model is aligned, and all visual materials are converted to the same color space.

[0017] Natural language processing technology is used to extract themes, emotions, and key entities from text, identify scenes, objects, and characters for images and videos, and label visual styles. A multi-dimensional tag index is built for each material unit, and the relationship between different material units is recorded.

[0018] The processed material units, along with their complete tags and associations, are stored in the IP material database.

[0019] Preferably, the construction of the structured IP knowledge graph specifically includes:

[0020] The standardized material units are input into the corresponding pre-trained deep neural network encoder to analyze the material content and output high-dimensional feature vectors to form a digital representation of IP elements.

[0021] The multimodal neural network model includes a visual encoder, a text encoder, and a cross-modal attention fusion module; the visual style features are extracted using a convolutional neural network to extract color histograms, texture features, and shape descriptors; the narrative theme features and character personality features are extracted using a natural language processing model by analyzing text descriptions and dialogue content to extract keywords, sentiment tendencies, and character relationships.

[0022] By using cross-modal attention mechanisms, we can analyze the correlation between feature vectors of different material units and uncover strong semantic, narrative, or visual association rules.

[0023] The identified core elements such as style, theme, character, scene, and item are used as nodes. The discovered association rules are transformed into directed edges connecting the nodes, and the relationship types are defined to construct a structured IP knowledge graph.

[0024] Preferably, the integrated product concept solution specifically includes:

[0025] The system receives textual requests from users and performs semantic retrieval in a structured IP knowledge graph using graph traversal algorithms and node vector similarity calculations to locate IP elements and their associated clusters that are relevant to the request.

[0026] The retrieved IP element clusters are converted into structured conditional feature vectors. A generative adversarial network model is called, and the conditional vectors are input along with random noise to generate a preliminary product concept visual sketch that conforms to the IP's visual style and theme. A text generation model is then used to generate product description text based on the same conditions.

[0027] The generated sketch is input into a preset rule engine, which calls the design rule library to perform aesthetic rule checks on the shape, proportion, and color matching of the concept. If the check fails, the rule engine outputs specific modification suggestions.

[0028] The modification suggestions are transformed into adjusted condition vectors, which drive the generative adversarial network model to generate a new round of concepts or make local fine-tuning of the original image until it passes the basic compliance verification and forms an optimized preliminary product concept scheme.

[0029] Preferably, the step of converting the product concept scheme into a parametric three-dimensional digital model, assigning materials, physical properties, and interactive logic to the digital model in a virtual simulation environment, and performing static aesthetic and cultural semantic simulation specifically includes:

[0030] The generated conceptual visual sketches are transformed into three-dimensional digital models through parametric modeling. In a virtual simulation environment, the three-dimensional digital models are given surface materials, physical properties, and interactive logic triggers.

[0031] In a virtual environment, application scenarios that match the IP worldview are built. By setting different virtual light sources, camera angles and rendering pipelines, high-fidelity static rendering images and dynamic display videos of the product are generated to simulate static aesthetics and cultural semantics.

[0032] For physical products, a physics engine is used to simulate their movement, stress, and assembly / disassembly processes under normal use conditions, and to observe their structural stability, smoothness of movement, and whether there is any interference between components.

[0033] For digital product concepts, user behavior simulation agents are deployed in interactive prototypes to perform click, drag, and swipe operations according to a set typical task flow. The entire process records the task completion time, number of operation steps, errors or stutters, and quantitatively evaluates the efficiency of the interaction process.

[0034] Preferably, the construction of the lightweight multi-dimensional objective classification prediction model specifically includes:

[0035] Collect multi-dimensional objective design attributes and simulation performance data of similar historical IP cultural and creative products, extract features such as product style matching degree, IP element restoration degree, design parameter compliance, physical simulation pass rate, and interaction process compliance rate, as well as corresponding expert objective rating labels, and train a lightweight multi-classification machine learning model based on the collected data.

[0036] The preliminary product concept scheme after simulation verification is analyzed into structured feature vectors and input into the trained classification prediction model. The model outputs the quantitative classification rating results of the product in various dimensions, including IP fit level, design compliance level, cultural semantic consistency level, physical performance compliance level, and interactive experience adaptability level.

[0037] In a dedicated online testing platform, a virtual product prototype test page is created, a standardized interactive task flow is deployed, and automated test scripts for different design versions are run simultaneously using blind testing. Objective interactive performance data is collected in real time to verify the accuracy of the classification prediction results.

[0038] Preferably, the method for obtaining the final product concept specifically includes:

[0039] By integrating simulation performance data, multi-dimensional classification prediction data, and automated test data, a multi-dimensional evaluation report is generated to obtain the performance and shortcomings of the preliminary product concept solution in various dimensions.

[0040] Based on the multidimensional evaluation report, key shortcomings below the preset threshold are identified, and corresponding optimization strategies are implemented. If it is an IP fit problem, the weight of the core IP element in the generation conditions is adjusted. If it is a cultural semantic consistency problem, the semantic association reasoning rules of the knowledge graph are adjusted. If it is a physical performance or interactive experience problem, the parameters of the three-dimensional model and the constraints of the interactive logic are adjusted.

[0041] Based on the optimization strategy, the input conditions of the generative adversarial network model and the constraint weights of the rule engine are adjusted in reverse. The adjusted model and rules are invoked to drive a new round of concept generation and preliminary optimization, resulting in optimized new alternative solutions.

[0042] The newly generated alternative solutions are then subjected to simulation verification and automated classification testing again until the comprehensive evaluation score of the new solution reaches the preset comprehensive evaluation threshold or the maximum number of iterations, at which point the final product concept solution is output.

[0043] Furthermore, a simulation system based on digital intelligent IP cultural and creative product design is proposed, including:

[0044] IP Material Processing Module: Collects original IP materials of various types, completes digital cleaning, format standardization and multi-dimensional structured annotation, and builds a compliant and complete multi-source heterogeneous IP material database;

[0045] IP Knowledge Graph Construction Module: Extracts core IP features, mines semantic relationships between materials, builds a structured IP knowledge graph, and realizes digital association representation of IP elements;

[0046] Product concept generation module: Based on the semantic retrieval of user keywords, related IP elements are retrieved, and visual sketches and text schemes for cultural and creative products are generated through a generative model. After verification and fine-tuning according to aesthetic rules, a preliminary concept scheme is formed.

[0047] Virtual simulation verification module: Transforms conceptual solutions into three-dimensional parametric models, endows them with material, physical properties and interactive logic, and completes full-dimensional simulation testing of aesthetics, culture, physical performance and user interaction experience;

[0048] Multi-dimensional objective classification prediction module: Based on a lightweight classification prediction model that has been trained, combined with automated test data, the module performs multi-dimensional objective classification and rating of products based on IP fit, design compliance, cultural semantic consistency, etc., and outputs quantitative classification results.

[0049] Intelligent Iterative Optimization Module: Integrates simulation and classification prediction data to pinpoint the shortcomings of the solution, reversely adjusts the model input conditions and design constraints, and iteratively generates an optimized solution until the preset evaluation threshold is met.

[0050] Compared with the prior art, the advantages of the present invention are:

[0051] By standardizing the processing of multi-source heterogeneous IP materials and constructing a structured multimodal IP knowledge graph, core cultural characteristics such as IP visuals, narratives, and personalities are accurately extracted, eliminating cultural distortion and element misuse in creative design from the source. Relying on generative models combined with aesthetic rule verification, compliant and high-quality product concept solutions can be efficiently produced. Simultaneously, a full-dimensional virtual simulation system is built, covering comprehensive verification of the physical performance, aesthetic semantics, and interactive experience of physical products, coupled with a lightweight, multi-dimensional objective classification prediction model and automated testing to quantify design quality. Finally, a data-driven closed-loop iterative optimization mechanism continuously addresses shortcomings in the solution, significantly reducing the cost of trial and error in physical products and significantly improving the cultural adaptability, experiential quality, and market success rate of IP-based creative products. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of a simulation method for designing cultural and creative products based on digital intelligent IP proposed in this invention;

[0053] Figure 2 This is a schematic diagram illustrating the construction of a structured IP knowledge graph proposed in this invention;

[0054] Figure 3 This is a schematic diagram of the integrated product concept scheme proposed in this invention. Detailed Implementation

[0055] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0056] A simulation system for designing digital intelligent IP cultural and creative products includes:

[0057] IP Material Processing Module: Collects original IP materials of various types, completes digital cleaning, format standardization and multi-dimensional structured annotation, and builds a compliant and complete multi-source heterogeneous IP material database;

[0058] IP Knowledge Graph Construction Module: Extracts core IP features, mines semantic relationships between materials, builds a structured IP knowledge graph, and realizes digital association representation of IP elements;

[0059] Product concept generation module: Based on the semantic retrieval of user keywords, related IP elements are retrieved, and visual sketches and text schemes for cultural and creative products are generated through a generative model. After verification and fine-tuning according to aesthetic rules, a preliminary concept scheme is formed.

[0060] Virtual simulation verification module: Transforms conceptual solutions into three-dimensional parametric models, endows them with material, physical properties and interactive logic, and completes full-dimensional simulation testing of aesthetics, culture, physical performance and user interaction experience;

[0061] Multi-dimensional objective classification prediction module: Based on a lightweight classification prediction model that has been trained, combined with automated test data, the module performs multi-dimensional objective classification and rating of products based on IP fit, design compliance, cultural semantic consistency, etc., and outputs quantitative classification results.

[0062] Intelligent Iterative Optimization Module: Integrates simulation and classification prediction data to pinpoint the shortcomings of the solution, reversely adjusts the model input conditions and design constraints, and iteratively generates an optimized solution until the preset evaluation threshold is met.

[0063] See Figure 1 As shown, a simulation method for designing cultural and creative products based on digital intelligent IP includes:

[0064] Construct a multi-source heterogeneous IP material database, collect the original cultural materials of the target IP, and perform digital cleaning, standardization processing and structured annotation to generate standardized IP material units with multi-dimensional tags, which are then stored in the IP material database;

[0065] Based on a pre-trained multimodal neural network model, standardized IP material units are analyzed to extract visual style features, narrative theme features, character personality features, and emotional tone features of the IP. Based on the features and semantic relationships between materials, a structured IP knowledge graph is constructed.

[0066] Based on user-input keywords, semantic retrieval and association reasoning are performed in the IP knowledge graph to locate IP element clusters. A generative adversarial network model is invoked, and the features of the retrieved IP element clusters are used as constraints to generate preliminary product concept visual sketches and text descriptions that conform to the IP visual style and theme. These are then integrated into a product concept solution.

[0067] The product concept solution is transformed into a parametric 3D digital model. In a virtual simulation environment, the digital model is given materials, physical properties and interactive logic. Static aesthetics and cultural semantics simulation are performed. Physical behavior simulation is performed for physical product concepts. User interaction behavior simulation agents are built for digital interactive product concepts. User operation paths are simulated, and the smoothness of the interaction process and the degree of achievement of experience goals are evaluated.

[0068] A lightweight, multi-dimensional objective classification prediction model is constructed. It is trained based on the objective design attributes and simulation performance data of similar historical IP cultural and creative products. The product schemes and their feature vectors after simulation verification are input into the classification prediction model to classify and rate the products in multiple dimensions, including IP fit, design compliance, cultural semantic consistency, physical performance compliance and interactive experience adaptability.

[0069] Based on simulation performance data and classification prediction data, an optimization algorithm is designed. The input conditions of the generative adversarial network model are adjusted in reverse to drive the generation of new alternative optimization schemes. The simulation, evaluation, and optimization cycle is repeated until the product scheme meets the preset comprehensive evaluation threshold or reaches the upper limit of the number of iterations, and the final product concept scheme is obtained.

[0070] Specifically, we comprehensively collect original cultural materials, including text, 2D images, 3D models, audio, and video, from the target IP's official setting library, published works, licensed media archives, and physical merchandise channels. For non-digital physical items, we use 3D scanners or structured light scanning equipment to digitally reconstruct them, generating 3D mesh models and texture maps.

[0071] The collected raw data undergoes automated preprocessing to eliminate noise, repair defects, and unify the basic format. Image / video frame denoising and enhancement employ adaptive filtering techniques, using Gaussian filter kernels to perform convolution operations on the image to smooth noise and preserve edges. After applying short-time Fourier transform to the audio signal, a spectral subtraction algorithm is used in the frequency domain to estimate and suppress noise spectral components, followed by an inverse transform to obtain the purified audio signal. Irrelevant characters and garbled text are automatically deleted through regular expression matching and specific dictionary filtering, and the text encoding format is unified.

[0072] To achieve cross-modal correlation analysis, materials from different sources need to be unified to a standard technical parameter system, the resolution of all visual materials needs to be scaled to a preset standard pixel, the video frame rate needs to be unified, the sampling rate of all audio materials needs to be unified, the coordinate system of the 3D model needs to be unified, and necessary topology optimization needs to be performed.

[0073] The color space is unified so that all visual materials are converted from the original color space to the standard sRGB color space through a color conversion matrix.

[0074] Structured semantic annotation and association mining employ a semi-automated pipeline of model pre-annotation and manual verification. Automated feature extraction and label generation utilize pre-trained natural language processing models to perform named entity recognition and sentiment analysis on text, automatically extracting tags such as topic, sentiment polarity, and key characters / items. Simultaneously, pre-trained computer vision models are used to perform object detection and scene classification on images / videos, automatically labeling them with tags including object, visual style, and scene category.

[0075] The manual verification and in-depth annotation design tool presents automatically generated tags to domain experts for review, correction, and refinement. Experts can supplement deep cultural semantic tags that are difficult for the automated model to recognize, such as the origin of cultural allusions, characteristics of specific art movements, and symbolic meanings.

[0076] Based on the co-occurrence of materials, semantic similarity, and explicit relationships indicated by manual annotation, various types of association edges are established between different material units, including description-described, narrative association, and visual reference. The strength of this association network... Computation and updates are performed using a graph attention network mechanism:

[0077]

[0078] in, , Let i be the feature vectors of data units i and j, W be the learnable weight matrix, and Attention be the attention function used to calculate the correlation strength between them.

[0079] The cleaned, standardized, and deeply labeled materials will be stored in the IP material database, using standardized IP material units as the basic recording unit.

[0080] See Figure 2 As shown, the specific steps for constructing a structured IP knowledge graph include:

[0081] The standardized material units are input into the corresponding pre-trained deep neural network encoder to analyze the material content and output high-dimensional feature vectors to form a digital representation of IP elements.

[0082] The multimodal neural network model includes a visual encoder, a text encoder, and a cross-modal attention fusion module; the visual style features are extracted using a convolutional neural network to extract color histograms, texture features, and shape descriptors; the narrative theme features and character personality features are extracted using a natural language processing model by analyzing text descriptions and dialogue content to extract keywords, sentiment tendencies, and character relationships.

[0083] By using cross-modal attention mechanisms, we can analyze the correlation between feature vectors of different material units and uncover strong semantic, narrative, or visual association rules.

[0084] The identified core elements such as style, theme, character, scene, and item are used as nodes. The discovered association rules are transformed into directed edges connecting the nodes, and the relationship types are defined to construct a structured IP knowledge graph.

[0085] Specifically, different types of standardized material units are input into corresponding pre-trained deep neural network encoders, mapping them into high-dimensional semantic feature vectors of a unified dimension, forming the digital representation basis of IP elements. This process can be formally represented as:

[0086]

[0087] in, This represents the original data of the i-th material unit in mode m; For the pre-trained encoder corresponding to this mode; This refers to the extracted high-dimensional feature vector.

[0088] After obtaining a unified feature representation, the core features of the IP are extracted from the high-dimensional feature vector:

[0089] For visual feature vectors By analyzing its statistical properties and projection onto the style feature subspace, the dominant color distribution, texture patterns, and shape regularities are quantitatively extracted. For example, style is characterized by calculating the higher-order moments of the eigenvectors.

[0090]

[0091] in, For style induction function, The mean of the eigenvectors. [⋅] represents the expectation, n is the order of the moment, and the final output is a style descriptor such as low polygon or ink wash rendering;

[0092] For text feature vectors Using topic modeling and sentiment analysis models, we extract high-frequency topic distributions and character sentiment and personality vectors from plot texts and character dialogues.

[0093] By fusing multimodal features, a comprehensive sentiment tendency is calculated through a feature fusion layer:

[0094]

[0095] Where [⋅;⋅] represents vector concatenation, and The weights and biases are learnable, σ is the activation function, and the output is the sentiment classification and its strength.

[0096] To establish semantic relationships between elements, cross-modal association analysis based on an attention mechanism is employed to calculate the association weight between feature vectors of any two different modal material units. :

[0097]

[0098] Here, score(⋅) is a scoring function, such as scaled dot product attention. Related weights Quantify the strength of the semantic correlation between element i and element j. Automatically mine... Strong association rules that exceed a preset threshold.

[0099] Based on the output of the aforementioned steps, a structured knowledge graph G=(V,E) centered on IP is constructed:

[0100] The identified semantic elements, such as core styles, themes, characters, key items, and classic scenes, are used as nodes in the knowledge graph. The attributes of each node consist of its corresponding feature vector and label description.

[0101] The discovered strong association rules are transformed into directed edges connecting the corresponding nodes. Each edge has a type attribute, such as symbol, occurrence, and usage, and its weight can be initialized as the association weight.

[0102] Graph neural networks are used to learn and optimize the representation of the initial graph, updating the node representation so that it contains both its own attributes and local graph structure information.

[0103] See Figure 3 As shown, the integrated product concept solution specifically includes:

[0104] The system receives textual requests from users and performs semantic retrieval in a structured IP knowledge graph using graph traversal algorithms and node vector similarity calculations to locate IP elements and their associated clusters that are relevant to the request.

[0105] The retrieved IP element clusters are converted into structured conditional feature vectors. A generative adversarial network model is called, and the conditional vectors are input along with random noise to generate a preliminary product concept visual sketch that conforms to the IP's visual style and theme. A text generation model is then used to generate product description text based on the same conditions.

[0106] The generated sketch is input into a preset rule engine, which calls the design rule library to perform aesthetic rule checks on the shape, proportion, and color matching of the concept. If the check fails, the rule engine outputs specific modification suggestions.

[0107] The modification suggestions are transformed into adjusted condition vectors, which drive the generative adversarial network model to generate a new round of concepts or make local fine-tuning of the original image until it passes the basic compliance verification and forms an optimized preliminary product concept scheme.

[0108] Specifically, the system receives the user's natural language design requirement Q, uses a semantic encoder to map requirement Q into a query vector, and performs semantic retrieval in the IP knowledge graph. The retrieval combines graph traversal and vector similarity calculation. From the graph node set V, several core nodes with the highest cosine similarity to the query vector are identified. Starting from these core nodes, a multi-hop traversal is performed through graph edges to expand and retrieve strongly related node clusters C. Finally, a subgraph of IP elements related to the semantics of the requirement is formed.

[0109] The IP element cluster C retrieved from the knowledge graph is transformed into a conditional feature vector c that the generative model can understand through graph attention network encoding subgraph aggregation, as shown in the formula:

[0110]

[0111] in, The condition vector c represents the features of node i, and Pooling is the pooling function. This conditional vector c condenses the visual style, theme, and semantic information of the desired IP elements. Using the conditional vector c and the random noise vector z as input, the generator G of the Conditional Generative Adversarial Network (cGAN) generates a preliminary visual sketch of the product concept. :

[0112]

[0113] Simultaneously, using a Transformer-based conditional language model, guided by the same conditional vector c, product descriptions are generated. :

[0114]

[0115] Obtain a preliminary graphic and textual concept scheme. .

[0116] The initial conceptual design will be fed into a pre-defined rules engine for multi-dimensional validation. This engine incorporates a design rules library (RR) that includes principles of formal beauty, cultural taboos, production feasibility, and IP licensing specifications. The validation process can be formalized as a series of constraint functions. Evaluation. For example, color harmony constraints can be used to check the generated image. Does the color scheme match the IP main color scheme? :

[0117]

[0118] in, Calculate the color histogram. Calculate the histogram distance. If If a violation occurs, it triggers a rule violation. The rule engine synthesizes the evaluation results of all constraints, and if a violation is found, it generates a specific set of modification suggestions.

[0119] The modification suggestions output by the rules engine are transformed into adjustments to the input conditions of the generative model. For example, if color adjustment is suggested, the system will modify the weights of the corresponding color style dimension in the condition vector, or add a bias term to the noise vector to guide color adjustment, forming new input conditions. Subsequently, the generator G is called again to generate conditions, or the original image is modified. Fine-tuning based on latent space editing is performed, and the newly generated solution will re-enter the rule engine verification process. This closed-loop iteration of generation-verification-adjustment will continue until the solution meets all preset rule constraints or reaches the maximum number of iterations, thereby outputting an optimized preliminary product concept solution.

[0120] The optimized conceptual sketches are transformed into 3D digital models with precise geometry through parametric modeling. This process can be formalized as an optimization problem seeking a 3D mesh M=(V,F) consistent with the visual representation of the input image, where V is the vertex set and F is the face set. During modeling, key dimensions and scales are parameterized. The model is imported into a game engine, where material properties, physical properties, and interaction logic are precisely assigned to each part of the 3D digital model within this environment.

[0121] Construct typical application scenarios consistent with the IP's worldview within a virtual environment. Render by configuring different high dynamic range lighting environment maps, directional light sources, and virtual camera paths, and then utilizing the physical rendering pipeline. Generate high-fidelity static rendering sequences and dynamic display videos from multiple perspectives.

[0122] During the evaluation, key aesthetic indicators are automatically calculated, such as color harmony scores, by comparing the model's primary color tone with the IP standard color palette. Evaluate using Euclidean distance in the Lab color space:

[0123]

[0124] Where {ci} is the model's main color scheme set. Combined with the IP knowledge graph, the model's shape, color, and decorative elements are judged to be consistent with the core cultural semantics of the IP, and a semantic consistency report is output.

[0125] For physical products, a physics engine in a virtual environment is activated to simulate the product's physical behavior by applying virtual forces, torques, or constraints. Its core is solving the rigid body dynamics equations.

[0126]

[0127] Where x is the position of the center of mass and ω is the angular velocity. and The simulation includes the resultant external force and resultant external moment.

[0128] Stability testing simulates the product's placement on different planes, calculating the relationship between its center of gravity projection and the supporting surface; motion mechanism testing, for products with moving parts, drives their joint movements to check for interference or jamming; simplified drop and force testing releases the product from a preset height, simulating a collision with the ground, assessing structural strength based on the impulse theorem, and checking for component detachment or damage. The simulation process records data such as maximum stress, displacement, and number of collisions, automatically generating a physical plausibility report.

[0129] For digital interactive products, a script-based user behavior simulation agent is deployed in an interactive prototype. The agent performs automated testing based on a pre-defined typical user task flow. Each task is defined as a series of atomic operations. During the agent's simulation, key performance indicators are recorded: task completion time, operation path efficiency, and error rate. The operation path efficiency is calculated as the ratio of the actual number of steps to the optimal number of steps. Errors include clicking on unresponsive areas and process freezes. Simultaneously, the interface rendering frame rate and interaction latency are recorded to assess smoothness. All data is aggregated to generate an interactive experience evaluation report, clearly indicating process bottlenecks and usability issues. The simulation results are finally archived together with static aesthetic and physical simulation data to form a comprehensive simulation verification report.

[0130] Construct a training dataset by collecting multi-dimensional data on similar IP-based cultural and creative products that have passed expert review and entered production, or have undergone complete expert review, to form a sample set. The input feature vector for each sample. include:

[0131] Product style matching degree is obtained by calculating the cosine similarity between the product's visual features and the standard style feature vectors in the IP knowledge graph.

[0132] IP element fidelity is calculated by weighting the number of core IP elements identified in the product, their morphological similarity, and their compatibility with the combined logic.

[0133] Design parameter compliance indicates whether the product's size, proportion, color values, etc., conform to the preset design specifications. It is a Boolean value or a percentage of compliance.

[0134] Physical simulation pass rate refers to the percentage of tests that pass in physical behavior simulation.

[0135] Interaction flow compliance rate refers to the percentage of preset task flows that are successfully completed without significant delays in user interaction behavior simulation.

[0136] Therefore, eigenvectors The corresponding output labels are a five-dimensional objective rating vector given by experts based on the same product, with each dimension typically represented by a discrete level.

[0137] Based on dataset D, a lightweight multi-output classification model is trained with the goal of minimizing the loss between predicted ratings and real expert ratings. An ensemble approach or a neural network with shared-specific association structures is employed to achieve multi-task learning, controlling model complexity while maintaining accuracy. The loss function L can be defined as a weighted sum of the classification losses for each dimension:

[0138]

[0139] in, For model parameters, This represents the model's output for the j-th rating dimension. Let CE be the true rank of the i-th sample in the j-th dimension, and let CE be the cross-entropy loss function. These are the weighting coefficients for the corresponding dimensions, used to balance the importance of different dimension ratings.

[0140] For the new preliminary product concept scheme after simulation verification Extract its objective feature vector Inputting this into a pre-trained model, the model outputs its five-dimensional probability distribution of predicted ratings:

[0141]

[0142] Here, p is a probability vector representing the likelihood of belonging to each level. The final quantitative classification rating result is obtained by taking the expected value of the probability vectors of each dimension or the level corresponding to the highest probability.

[0143] To verify the accuracy of the predictive model and continuously optimize it, a virtual product prototype test page was deployed on a dedicated online testing platform. The solutions that received predictive ratings were placed on the platform along with other design solutions, and automated test scripts for different design versions were run simultaneously using a blind testing approach.

[0144] The automated test scripts collect objective interaction performance data in real time, such as task completion time, number of operation steps, and number of erroneous clicks. Using predefined performance thresholds, these objective metrics are converted into a measured interaction experience fit level. This measured level is then compared with the model's predictions to calculate a consistency index, verifying the accuracy of the prediction model in this dimension.

[0145] Integrate all the data into a visual, multi-dimensional evaluation report to intuitively show the performance and shortcomings of the solution in each dimension;

[0146] By analyzing multidimensional evaluation reports, a set of key shortcomings that perform below a preset threshold are identified. Optimization actions are taken for different types of shortcomings. If the problem is IP fit, the weight of the core IP elements in the generation conditions is adjusted. If the problem is cultural semantic consistency, the semantic association reasoning rules of the knowledge graph are adjusted. If the problem is physical performance or interactive experience, the parameters of the 3D model and the constraints of the interactive logic are adjusted.

[0147] Based on the selected optimization strategy, the generation parameters or model weights affecting relevant indicators are modified in reverse. If the goal is to improve the overall score, this can be formalized as an optimization problem, with the objective function defined as the negative of the current evaluation score. In each iteration, the system updates the parameters using a gradient-based method or evolutionary algorithm. After the parameters are updated, the adjusted generation process is invoked to generate new alternative solutions.

[0148] The new scheme will begin a new round of simulation and automated classification testing, forming a closed loop of generation, simulation / evaluation, optimization, and regeneration. The iterative process continues until the comprehensive evaluation score of the latest scheme reaches a preset threshold or the number of iterations reaches a preset upper limit. When the iteration terminates, the scheme with the highest comprehensive evaluation score among all iterated schemes is selected as the current optimal output.

[0149] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0150] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0151] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A simulation method for designing cultural and creative products based on digital intelligent IP, characterized in that, include: Construct a multi-source heterogeneous IP material database, collect the original cultural materials of the target IP, and perform digital cleaning, standardization processing and structured annotation to generate standardized IP material units with multi-dimensional tags, which are then stored in the IP material database; Based on a pre-trained multimodal neural network model, standardized IP material units are analyzed to extract visual style features, narrative theme features, character personality features, and emotional tone features of the IP. Based on the features and semantic relationships between materials, a structured IP knowledge graph is constructed. Based on user-input keywords, semantic retrieval and association reasoning are performed in the IP knowledge graph to locate IP element clusters. A generative adversarial network model is invoked, and the features of the retrieved IP element clusters are used as constraints to generate preliminary product concept visual sketches and text descriptions that conform to the IP visual style and theme. These are then integrated into a product concept solution. The product concept solution is transformed into a parametric 3D digital model. In a virtual simulation environment, the digital model is given materials, physical properties and interactive logic. Static aesthetics and cultural semantics simulation are performed. Physical behavior simulation is performed for physical product concepts. User interaction behavior simulation agents are built for digital interactive product concepts. User operation paths are simulated, and the smoothness of the interaction process and the degree of achievement of experience goals are evaluated. A lightweight, multi-dimensional objective classification prediction model is constructed. It is trained based on the objective design attributes and simulation performance data of similar historical IP cultural and creative products. The product schemes and their feature vectors after simulation verification are input into the classification prediction model to classify and rate the products in multiple dimensions, including IP fit, design compliance, cultural semantic consistency, physical performance compliance and interactive experience adaptability. Based on simulation performance data and classification prediction data, an optimization algorithm is designed. The input conditions of the generative adversarial network model are adjusted in reverse to drive the generation of new alternative optimization schemes. The simulation, evaluation, and optimization cycle is repeated until the product scheme meets the preset comprehensive evaluation threshold or reaches the upper limit of the number of iterations, and the final product concept scheme is obtained.

2. The simulation method for designing cultural and creative products based on digital intelligent IP according to claim 1, characterized in that, The construction of the multi-source heterogeneous IP material database specifically includes: Collect text, images, audio, video, and 3D data of IP from official databases, publicly available works, and physical objects, and scan and model non-digital physical objects; Remove watermarks and noise from images and videos, correct colors and brightness, reduce noise and unify volume in audio, and unify text encoding, filter garbled characters, and proofread key information. All materials are converted to a common format, the resolution and frame rate of images and videos are standardized to a preset standard, the audio sampling rate and bit depth are standardized, the coordinate system of the 3D model is aligned, and all visual materials are converted to the same color space. Natural language processing technology is used to extract themes, emotions, and key entities from text, identify scenes, objects, and characters for images and videos, and label visual styles. A multi-dimensional tag index is built for each material unit, and the relationship between different material units is recorded. The processed material units, along with their complete tags and associations, are stored in the IP material database.

3. The simulation method for designing cultural and creative products based on digital intelligent IP according to claim 1, characterized in that, The construction of the structured IP knowledge graph specifically includes: The standardized material units are input into the corresponding pre-trained deep neural network encoder to analyze the material content and output high-dimensional feature vectors to form a digital representation of IP elements. The multimodal neural network model includes a visual encoder, a text encoder, and a cross-modal attention fusion module; the visual style features are extracted using a convolutional neural network to extract color histograms, texture features, and shape descriptors; the narrative theme features and character personality features are extracted using a natural language processing model by analyzing text descriptions and dialogue content to extract keywords, sentiment tendencies, and character relationships. By using cross-modal attention mechanisms, we can analyze the correlation between feature vectors of different material units and uncover strong semantic, narrative, or visual association rules. The identified core elements such as style, theme, character, scene, and item are used as nodes. The discovered association rules are transformed into directed edges connecting the nodes, and the relationship types are defined to construct a structured IP knowledge graph.

4. The simulation method for designing cultural and creative products based on digital intelligent IP according to claim 1, characterized in that, The integrated product concept solution specifically includes: The system receives textual requests from users and performs semantic retrieval in a structured IP knowledge graph using graph traversal algorithms and node vector similarity calculations to locate IP elements and their associated clusters that are relevant to the request. The retrieved IP element clusters are converted into structured conditional feature vectors. A generative adversarial network model is called, and the conditional vectors are input along with random noise to generate a preliminary product concept visual sketch that conforms to the IP's visual style and theme. A text generation model is then used to generate product description text based on the same conditions. The generated sketch is input into a preset rule engine, which calls the design rule library to perform aesthetic rule checks on the shape, proportion, and color matching of the concept. If the check fails, the rule engine outputs specific modification suggestions. The modification suggestions are transformed into adjusted condition vectors, which drive the generative adversarial network model to generate a new round of concepts or make local fine-tuning of the original image until it passes the basic compliance verification and forms an optimized preliminary product concept scheme.

5. The simulation method for designing cultural and creative products based on digital intelligent IP according to claim 1, characterized in that, The process of transforming product concept schemes into parametric 3D digital models, assigning materials, physical properties, and interactive logic to the digital models in a virtual simulation environment, and performing static aesthetic and cultural semantic simulation specifically includes: The generated conceptual visual sketches are transformed into three-dimensional digital models through parametric modeling. In a virtual simulation environment, the three-dimensional digital models are given surface materials, physical properties, and interactive logic triggers. In a virtual environment, application scenarios that match the IP worldview are built. By setting different virtual light sources, camera angles and rendering pipelines, high-fidelity static rendering images and dynamic display videos of the product are generated to simulate static aesthetics and cultural semantics. For physical products, a physics engine is used to simulate their movement, stress, and assembly / disassembly processes under normal use conditions, and to observe their structural stability, smoothness of movement, and whether there is any interference between components. For digital product concepts, user behavior simulation agents are deployed in interactive prototypes to perform click, drag, and swipe operations according to a set typical task flow. The entire process records the task completion time, number of operation steps, errors or stutters, and quantitatively evaluates the efficiency of the interaction process.

6. The simulation method for designing cultural and creative products based on digital intelligent IP according to claim 1, characterized in that, The construction of the lightweight, multi-dimensional objective classification prediction model specifically includes: Collect multi-dimensional objective design attributes and simulation performance data of similar historical IP cultural and creative products, extract features such as product style matching degree, IP element restoration degree, design parameter compliance, physical simulation pass rate, and interaction process compliance rate, as well as corresponding expert objective rating labels, and train a lightweight multi-classification machine learning model based on the collected data. The preliminary product concept scheme after simulation verification is analyzed into structured feature vectors and input into the trained classification prediction model. The model outputs the quantitative classification rating results of the product in various dimensions, including IP fit level, design compliance level, cultural semantic consistency level, physical performance compliance level, and interactive experience adaptability level. In a dedicated online testing platform, a virtual product prototype test page is created, a standardized interactive task flow is deployed, and automated test scripts for different design versions are run simultaneously using blind testing. Objective interactive performance data is collected in real time to verify the accuracy of the classification prediction results.

7. The simulation method for designing cultural and creative products based on digital intelligent IP according to claim 1, characterized in that, The specific steps for obtaining the final product concept plan include: By integrating simulation performance data, multi-dimensional classification prediction data, and automated test data, a multi-dimensional evaluation report is generated to obtain the performance and shortcomings of the preliminary product concept solution in various dimensions. Based on the multidimensional evaluation report, key shortcomings below the preset threshold are identified, and corresponding optimization strategies are implemented. If it is an IP fit problem, the weight of the core IP element in the generation conditions is adjusted. If it is a cultural semantic consistency problem, the semantic association reasoning rules of the knowledge graph are adjusted. If it is a physical performance or interactive experience problem, the parameters of the three-dimensional model and the constraints of the interactive logic are adjusted. Based on the optimization strategy, the input conditions of the generative adversarial network model and the constraint weights of the rule engine are adjusted in reverse. The adjusted model and rules are invoked to drive a new round of concept generation and preliminary optimization, resulting in optimized new alternative solutions. The newly generated alternative solutions are then subjected to simulation verification and automated classification testing again until the comprehensive evaluation score of the new solution reaches the preset comprehensive evaluation threshold or the maximum number of iterations, at which point the final product concept solution is output.

8. A simulation system for designing cultural and creative products based on digital intelligent IP, used to implement the simulation method for designing cultural and creative products based on digital intelligent IP as described in any one of claims 1-7, characterized in that, include: IP Material Processing Module: Collects original IP materials of various types, completes digital cleaning, format standardization and multi-dimensional structured annotation, and builds a compliant and complete multi-source heterogeneous IP material database; IP Knowledge Graph Construction Module: Extracts core IP features, mines semantic relationships between materials, builds a structured IP knowledge graph, and realizes digital association representation of IP elements; Product concept generation module: Based on the semantic retrieval of user keywords, related IP elements are retrieved, and visual sketches and text schemes for cultural and creative products are generated through a generative model. After verification and fine-tuning according to aesthetic rules, a preliminary concept scheme is formed. Virtual simulation verification module: Transforms conceptual solutions into three-dimensional parametric models, endows them with material, physical properties and interactive logic, and completes full-dimensional simulation testing of aesthetics, culture, physical performance and user interaction experience; Multi-dimensional objective classification prediction module: Based on a lightweight classification prediction model that has been trained, combined with automated test data, the module performs multi-dimensional objective classification and rating of products based on IP fit, design compliance, cultural semantic consistency, etc., and outputs quantitative classification results. Intelligent Iterative Optimization Module: Integrates simulation and classification prediction data to pinpoint the shortcomings of the solution, reversely adjusts the model input conditions and design constraints, and iteratively generates an optimized solution until the preset evaluation threshold is met.

Images