AIGC paper-cut generation system and method based on culture gene multi-modal constraint

Abstract

The invention belongs to the technical field of computers, particularly relates to an AIGC paper-cut generation system and method based on culture gene multi-modal constraints, and solves the dual problems of'culture distortion 'and'physically unachievable' of a traditional AIGC generation result through mathematical fusion (such as a mixed integer programming model) of semantic, topology and process three-level constraints. The improved ControlNet architecture realizes directional fusion of culture gene coding vectors through a culture attention mechanism, and dynamically adjusts feature weights in a generation process, so that the generation precision of pattern details is improved to a submillimeter level; based on the combination of an interactive correction tool and an inverse diffusion inversion algorithm, the modification of a local pattern by a user can be quickly fed back to a system, and a'creativity-generation-optimization 'closed loop is realized.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to an AIGC paper-cutting generation system and method based on multimodal constraints of cultural genes. Background Art

[0002] The intangible cultural heritage of paper-cutting, a living gene carrier of Chinese civilization, possesses a creative system that embodies the three core elements of pattern composition, craftsmanship, and cultural meaning. Currently, this intangible cultural heritage faces two major challenges: First, the traditional master-apprentice model of inheritance struggles to meet the demands of an industrialized society. Research data shows that the average age of existing inheritors is mostly over 55, with youth participation less than 15%. Second, digital preservation efforts often remain limited to two-dimensional scanning and archiving. The 37 provincial-level paper-cutting databases established nationwide suffer from data fragmentation and a lack of genetic understanding.

[0003] Traditional digital preservation techniques use high-precision scanners (such as the CRUSE scanning system) for planar acquisition. While these technologies can preserve physical form, they suffer from limitations such as a single dimension of genetic analysis and a lack of process parameters, making them incapable of supporting innovative transformation. In the context of AIGC (Artificial Intelligence Generated Content) technological innovation, existing generation systems generally lack cultural adaptability. Research in authoritative international journals shows that the semantic accuracy of pattern patterns in paper-cutting works generated by mainstream AI art models is less than 42%, highlighting a significant gap in cultural logic. Summary of the Invention

[0004] The embodiments of the present application provide an AIGC paper-cutting generation system and method based on multimodal constraints of cultural genes to solve the above problems.

[0005] According to one aspect of the present application, an AIGC paper-cutting generation system based on multimodal constraints of cultural memes is provided, comprising:

[0006] A cultural gene library construction module is used to collect and encode the pattern characteristics, process parameters and cultural semantic rules of paper-cutting crafts;

[0007] A multimodal constraint engine, which includes semantic constraint submodules, topological constraint submodules, and process constraint submodules, and is used to jointly optimize the constraints of the generation process;

[0008] The AIGC generation module is based on the ControlNet neural network architecture and embeds a cultural attention mechanism to achieve dynamic injection of cultural genes;

[0009] Human-computer collaborative editing module, providing pattern correction tools and feature inversion interface, supporting users to interactively optimize the generated results; and

[0010] The dynamic evolution module realizes the collaborative iterative update of the cultural gene library and the generation model through a dual-channel reinforcement learning framework.

[0011] Optionally, the cultural gene library construction module includes:

[0012] 3D laser scanning device, used to obtain the three-dimensional deformation data and pattern topology structure of paper-cut works;

[0013] Motion capture equipment, used to record the spatiotemporal motion sequence and process parameters of the cutting process;

[0014] Multispectral imager, used to extract the explicit features and implicit cultural semantic features of patterns.

[0015] Optionally, the multimodal constraint engine implements constraint optimization in the following manner:

[0016] The semantic constraint submodule maps cultural semantic rules to the prompt word space of the CLIP text encoder;

[0017] The topological constraint submodule generates protection masks for non-cut areas through geometric analysis and verifies the physical continuity of the ridges.

[0018] The process constraint submodule converts the paper tear strength threshold into a process feasibility verification condition in the potential diffusion process.

[0019] Optionally, the process feasibility verification conditions include:

[0020] The paper tear strength threshold constraint is defined based on physical experimental data;

[0021] The latent vector is reversely corrected through residual connections to ensure the mathematical expressibility of the Yin-Yang engraving rules;

[0022] A process parameter verification layer is implanted in the potential diffusion process to block the generation path that does not conform to the physical laws of shearing.

[0023] Optionally, the AIGC generation module includes:

[0024] ControlNet architecture, embedding a multi-scale cultural attention mechanism in the UNet decoder;

[0025] Dynamic weight allocation unit, which adjusts the weight ratio of cultural characteristics and generation goals according to the generation stage;

[0026] The genealogy tracing unit records the cultural gene code and the source of the process parameters for each generated pattern.

[0027] Optionally, the dynamic evolution module includes:

[0028] Offline training channel, which updates the encoding rules of the cultural gene library based on the generation quality assessment model;

[0029] Online optimization channel, which adjusts multimodal constraint weights in real time based on user interaction data;

[0030] Experience replay pool, which stores the mutation trajectory of cultural genes to optimize reinforcement learning strategies.

[0031] According to another aspect of the present application, a method for generating AIGC paper-cuts based on multimodal constraints of cultural memes is provided, comprising the following steps:

[0032] Gene feature extraction: extracting pattern dominant features and implicit process parameters through multispectral imaging and frequency domain decomposition algorithm;

[0033] Multimodal constraint injection: converting semantic, topological, and process constraints into mathematical specifications for generating models;

[0034] Collaborative generation and optimization: Dynamically update the gene library by combining the cultural adaptability assessment of the initial generation results and user interaction correction.

[0035] Optionally, the gene feature extraction step includes:

[0036] Construct a metadata template containing 17-dimensional cultural gene features, covering pattern composition, craft techniques and semantic rules;

[0037] The spatial and temporal dual-domain features of the cutting action are extracted through a temporal convolutional network;

[0038] A cross-modal contrastive learning algorithm is used to fuse multi-source data and generate cultural gene encoding vectors.

[0039] Optionally, the multimodal constraint injection step includes:

[0040] Map the Yin-Yang engraving rules into vector offsets in the Stable Diffusion latent space;

[0041] The three-level constraints are jointly optimized through a mixed integer programming model to ensure the cultural and logical consistency of the generated results;

[0042] A process feasibility verification layer is inserted into the potential diffusion process to block physically infeasible intermediate generation states.

[0043] Optionally, the collaborative generation and optimization step includes:

[0044] Initial generation results were screened based on a cultural adaptability evaluation model. Model evaluation metrics included semantic accuracy, topological integrity, and process feasibility.

[0045] Receive the user's modification instructions for local patterns through the brush stroke correction tool, and invert the correction data to the cultural gene library;

[0046] A dynamic weight allocation algorithm is used to adjust the weights of constraints to achieve closed-loop optimization of the generation system.

[0047] The AIGC paper-cutting generation system and method based on multimodal constraints of cultural genes provided by this application have the following beneficial effects:

[0048] 1. Multimodal Constrained Joint Optimization: Through the mathematical fusion of semantic, topological, and process constraints (e.g., mixed integer programming models), this approach addresses the dual issues of "cultural distortion" and "physical impracticality" in traditional AIGC generation results.

[0049] 2. Dynamic meme injection mechanism: The improved ControlNet architecture uses a cultural attention mechanism (multi-head cross attention) to achieve targeted fusion of meme encoding vectors. It dynamically adjusts feature weights during the generation process, improving the generation accuracy of pattern details (such as the continuity of embroidery patterns) to submillimeter levels.

[0050] 3. Real-time human-computer collaborative correction capability: Based on the combination of interactive correction tools (such as pressure-sensitive pens) and inverse diffusion inversion algorithms, users' modifications to local patterns (such as adjusting the sawtooth density) can be quickly fed back to the system, realizing a "creativity-generation-optimization" closed loop. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0052] in:

[0053] Figure 1 1 is a schematic diagram of an AIGC paper-cutting generation system based on multimodal constraints of cultural genes according to an embodiment of the present application;

[0054] Figure 2 This is a flowchart of an AIGC paper-cutting generation method based on multimodal constraints of cultural genes according to an embodiment of the present application;

[0055] Figure 3 This is a flow chart of gene feature extraction in the AIGC paper-cut generation method shown in one embodiment of the present application;

[0056] Figure 4This is a flowchart of multimodal constraint injection in the AIGC paper-cut generation method according to an embodiment of the present application;

[0057] Figure 5 This is a flowchart of collaborative generation and optimization in the AIGC paper-cut generation method shown in one embodiment of the present application;

[0058] Figure 6 It is a structural diagram of the terminal device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0059] In the following description, specific details such as specific system structures and techniques are provided for purposes of illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid obscuring the description of the present application with unnecessary detail.

[0060] The embodiment of the present application provides an AIGC paper-cutting generation system based on multimodal constraints of cultural genes, such as Figure 1 As shown, the AIGC paper-cutting generation system includes a cultural gene library construction module 10, a multimodal constraint engine 20, an AIGC generation module 30, a human-computer collaborative editing module 40 and a dynamic evolution module 50.

[0061] The cultural gene library construction module 10 is used to collect and encode the pattern features, process parameters and cultural semantic rules of paper-cutting crafts. The multimodal constraint engine 20 includes a semantic constraint submodule, a topological constraint submodule and a process constraint submodule, which are used to jointly optimize the constraints of the generation process. The AIGC generation module 30 is based on the ControlNet neural network architecture and embeds a cultural attention mechanism to achieve dynamic injection of cultural genes. The human-computer collaborative editing module 40 provides pattern correction tools and feature inversion interfaces to support users' interactive optimization of generation results. The dynamic evolution module 50 realizes the collaborative iterative update of the cultural gene library and the generation model through a dual-channel reinforcement learning framework.

[0062] Specifically, the cultural gene library construction module 10 uses a knowledge graph (Neo4j) and an adversarial feature extraction network to structure the storage of traditional cultural elements such as patterns, symbols, and techniques, establishing a "gene-phenotype" mapping relationship. It outputs gene encoding vectors (e.g., "moir: curvature ∈ [0.2, 0.5], ≥ 3 axes of symmetry").

[0063] The multimodal constraint engine 20 fuses user input (text / sketch) with the gene library rules to generate differentiable constraints: style constraints (CLIP cross-modal alignment), structural constraints (graph neural network topology verification), and process constraints (physical simulation pre-verification). It outputs a constraint tensor (with dimensions aligned with the AIGC latent space).

[0064] The AIGC generation module 30 generates candidate designs based on latent space projections (Latent Diffusion) of the diffusion model, guided by constraint tensors. It employs a phased noise reduction strategy (steps 0-30 focus on cultural gene expression, and steps 30-50 optimize physical feasibility).

[0065] The human-computer collaborative editing module 40 provides a mixed reality (MR) interface that allows users to modify the generated results at a semantic level (e.g., “add symmetry” instead of pixel editing) through gestures / voice. The core algorithm is an intent understanding model based on contrastive learning (user command → latent space gradient direction).

[0066] The dynamic evolution module 50 dynamically updates the gene library weights and constraint generation strategy through reinforcement learning (PPO algorithm), achieving system self-optimization. Update trigger conditions: user adoption rate, physical simulation pass rate, and cultural similarity index joint evaluation.

[0067] The cultural attention mechanism consists of a multimodal feature fusion network, an interpretability enhancement unit, and a dynamic update interface. The multimodal feature fusion network integrates pattern, craft, and semantic features to generate an attention weight matrix. The interpretability enhancement unit outputs a heat map of the influence of cultural features on the generated results. The dynamic update interface iteratively adjusts the attention focus area based on the gene library version.

[0068] The technical details of the cultural attention mechanism are as follows:

[0069] Multimodal fusion: Using a cross-attention architecture, the query vector Q comes from image features, and the key-value pair K / V is generated by cultural gene encoding. The attention weight calculation formula is:

[0070]

[0071] The mask matrix M blocks the attention propagation of non-pruned areas.

[0072] Enhanced interpretability: A heat map is generated using the Grad-CAM algorithm to calculate the contribution of cultural features (for example, if the mean gradient of region A is greater than 0.6, it is marked as a high-influence area).

[0073] Dynamic update: For example, when the Qin and Jin style codes are added to the gene library, the attention heads are automatically expanded to 10, and the proportion of cultural characteristics increases from 45% to 60%.

[0074] This system realizes the intelligent paper-cutting generation constrained by cultural genes through a five-layer architecture. The cultural gene library construction module 10 can collect the geometric topological structure of traditional paper-cutting works, cutting process parameters (such as tool angle ±2°, paper tension threshold 15N / m) by integrating a 3D scanner (accuracy 0.01mm) and a hyperspectral camera (400-2500nm band), and obtain the data of the paper-cutting works. 2 ) and cultural semantic symbols (such as the symbolic rules of the twelve zodiac patterns) to build a gene database containing 17-dimensional feature vectors. The multimodal constraint engine 20 uses a mixed integer programming algorithm to encode semantic rules (CLIP similarity ≥ 0.8), topological constraints (Delaunay triangulation connectivity) and process constraints (such as line width ≥ 0.3mm) as a joint loss function, and dynamically optimizes it during the potential diffusion process. The AIGC generation module 30 is based on the improved ControlNet architecture, and embeds a cultural attention gating unit in the jump connection layer of UNet, through a multi-head cross attention mechanism (8 heads, scaling factor ) achieves dynamic fusion of genetic and image features. The human-computer collaboration module supports users in real-time adjustment of the curvature (error tolerance ±5%) and topology of the generated results by providing a pattern correction interface based on a pressure-sensitive pen, and feeds the modified data back to the gene library through the inverse diffusion model. The dynamic evolution module 50 deploys a dual-channel reinforcement learning framework. The offline channel uses the PPO algorithm (learning rate 3e-4) to optimize the genetic encoding strategy, and the online channel uses Bayesian optimization (Gaussian process kernel function) to adjust the constraint weights in real time to achieve system iterative upgrades.

[0075] When the AIGC paper-cutting generation system based on multimodal constraints of cultural genes in the embodiment of the present application is used, the forward generation process is as follows:

[0076] 1. Gene bank initialization

[0077] The cultural gene library construction module 10 extracts typical pattern features from the intangible cultural heritage database (such as the Forbidden City pattern library) and generates a gene coding dictionary (Key: cultural category, Value: gene vector + constraint boundary).

[0078] 2. Constraint Fusion Generation

[0079] When the user inputs "dragon and phoenix pattern, suitable for laser engraving", the multimodal constraint engine 20 executes:

[0080] Semantic parsing: calling the "dragon and phoenix" related gene codes in the gene library (curvature constraint, scale density threshold);

[0081] Process mapping: converting “laser engraving” into physical constraints such as minimum line width (≥0.1mm) and maximum hollowing area (≤40%);

[0082] Output: Fused constraint tensor (dimensions: 256×256×32).

[0083] 3.AIGC Generation and Interactive Correction

[0084] The AIGC module generates the initial design based on the constraint tensor. The human-machine collaboration module detects the user's intention to adjust the "dragon tail length" and provides feedback through the following paths:

[0085] graph LR

[0086] User gesture → Intent recognition model → Latent space gradient Δz → AIGC module reprojection → Update design

[0087] A modified trajectory log is generated for each interaction and used for strategy optimization of the dynamic evolution module 50 .

[0088] 4. Dynamic evolution trigger

[0089] When the user correction rate is greater than 30% in 10 consecutive generations, the dynamic evolution module 50 is activated:

[0090] Adjust the weight of "dragon and phoenix" genes in the gene pool (for example, lower the priority of complex scale features);

[0091] The fusion strategy of the multimodal constraint engine 20 is updated (the loss function coefficient of the process constraint is increased).

[0092] In one embodiment, the cultural gene library construction module 10 includes a 3D laser scanning device, a motion capture device, and a multispectral imager. The 3D laser scanning device is used to obtain the three-dimensional deformation data and pattern topology structure of the paper-cut works. The motion capture device is used to record the spatiotemporal motion sequence and process parameters of the cutting process. The multispectral imager is used to extract the explicit features and implicit cultural semantic features of the pattern. Specifically:

[0093] 3D laser scanning: A blue light structured light scanner (accuracy ±5μm) is used to obtain the three-dimensional deformation data of paper cutting. The multi-view point cloud is aligned through the ICP algorithm to extract the pattern key points (SIFT feature matching error is less than 0.1px) and the curvature parameters of the hinge connection (R≥0.5mm).

[0094] Motion capture system: Use an IMU sensor array (200Hz sampling rate) to record the six-axis data of the shearing action (acceleration ±16g, angular velocity ±2000° / s), extract spatiotemporal features through the LSTM network (time window 50ms), and establish a process parameter library (such as the tool tip pressure-pattern line width mapping table).

[0095] Multispectral imaging: Pattern reflectance spectra were collected in the visible light (400-700nm) and near-infrared (900-1700nm) bands. After dimensionality reduction using principal component analysis (PCA), a CNN network (ResNet-18 architecture) was used to separate explicit pattern features (classification accuracy of 92%) from implicit cultural semantic codes.

[0096] In one embodiment, the multimodal constraint engine 20 implements constraint optimization in the following manner: The semantic constraint submodule maps cultural semantic rules to the prompt word space of the CLIP text encoder. The topological constraint submodule generates protection masks for non-clipped areas through geometric analysis and verifies the physical continuity of the ribbed pattern. The process constraint submodule converts the paper tear strength threshold into a process feasibility verification condition for the potential diffusion process. Specifically:

[0097] Semantic Constraint: Cultural rules (such as "lotus symbolizes integrity") are converted into text prompt pairs (positive / negative prompts) for the CLIP model. The cosine similarity between the generated image and text is constrained through contrastive learning (threshold ≥ 0.75).

[0098] Topological constraints: The Alpha Shape algorithm is used to generate pattern protection areas (area ≥ 80%), and the Euler characteristics of the cutting path are verified (number of holes ≤ 3). An L2 regularization penalty (λ = 0.01) is applied to topological breakpoints.

[0099] Process constraints: Based on the experimental data of paper tearing strength (Xuan paper critical stress σ c =18MPa), a process verification layer is inserted in each denoising step of the diffusion model, and when a thin line width less than 0.25mm is detected, the latent vector is forced to be corrected by gradient clipping (clipnorm=1.0).

[0100] In a specific embodiment, the process feasibility verification conditions include:

[0101] The paper tear strength threshold constraint is defined based on physical experimental data;

[0102] The latent vector is reversely corrected through residual connections to ensure the mathematical expressibility of the Yin-Yang engraving rules;

[0103] A process parameter verification layer is implanted in the potential diffusion process to block the generation path that does not conform to the physical laws of shearing.

[0104] From the above, we can see that process feasibility verification includes physical and mathematical double verification:

[0105] Physical constraints: A stress distribution model of the shearing path is established through finite element simulation. When the line width corresponding to the potential vector exceeds the threshold, the gradient redirection mechanism is triggered to shift the generation direction to the safe zone (stress <12MPa).

[0106] Residual correction: insert residual check units in the 4th, 8th, and 12th layers of UNet to impose orthogonal projection constraints on feature maps that violate the positive and negative engraving rules (positive engraving area <30%), forcing the latent space to satisfy ∑(x i 2 )≤1.2.

[0107] Real-time blocking: Based on the cutting tool motion parameters (speed v max =5mm / s, acceleration a max =2m / s 2 ), calculate the curvature-velocity matching degree of the generated path (formula: κ≤v2 / (a max ·r min )), perform mask filtering on the substandard areas (maskratio≥0.6).

[0108] In one embodiment, the AIGC generation module 30 includes:

[0109] ControlNet architecture, embedding a multi-scale cultural attention mechanism in the UNet decoder;

[0110] Dynamic weight allocation unit, which adjusts the weight ratio of cultural characteristics and generation goals according to the generation stage;

[0111] The genealogy tracing unit records the cultural gene code and the source of the process parameters for each generated pattern.

[0112] As can be seen from the above, the core innovation of the AIGC generation module 30 lies in:

[0113] Improved ControlNet: A cultural feature injection channel is added to the standard architecture, multi-scale feature fusion is achieved through deformable convolution (offset ±3px), and the cultural attention weight is dynamically adjusted (range 0.2-0.7) by a gating mechanism (sigmoid activation).

[0114] Dynamic weight allocation: A GRU network is used to analyze features in the generation stage (sketch generation → detail refinement), automatically adjust the weights of cultural features (initial value 0.5, ±0.3 dynamic range), and increase the cultural constraint strength to 0.8 in key areas (such as hinged thorn patterns).

[0115] Lineage tracing: A metadata chain is attached to each generated pattern, including the cultural gene hash value (SHA-256), process parameter version number and user modification record, supporting traceability verification (such as copyright ownership verification).

[0116] In one embodiment, the dynamic evolution module 50 includes:

[0117] Offline training channel, which updates the encoding rules of the cultural gene library based on the generation quality assessment model;

[0118] Online optimization channel, which adjusts multimodal constraint weights in real time based on user interaction data;

[0119] Experience replay pool, which stores the mutation trajectory of cultural genes to optimize reinforcement learning strategies.

[0120] As can be seen from the above, the dynamic evolution module 50 realizes system self-optimization through a closed data loop:

[0121] Offline training channel: Use the PPO reinforcement learning algorithm (discount factor γ = 0.99), reward function R = 0.5 × semantic score + 0.3 × craft score + 0.2 × user rating, and update the policy network every 24 hours (batch size = 256).

[0122] Online optimization channel: collects user interaction data in real time (e.g., the number of stroke corrections is greater than 3 times / region), adjusts constraint weights through Bayesian optimization (learning rate 1e-3), and achieves response latency of less than 150ms.

[0123] Experience replay pool: stores cultural gene mutation events (such as pattern style mutation Δ>0.4), uses DBSCAN clustering (eps=0.15) to identify key mutation patterns, and optimizes the exploration rate parameter (ε decays from 0.5 to 0.1).

[0124] In one embodiment, the dual-channel reinforcement learning framework includes:

[0125] Offline policy network, which optimizes the meme encoding strategy based on the generated quality scoring function;

[0126] Online policy network, which adjusts the priority of constraints in real time based on user feedback on revised data;

[0127] Gene mutation trajectory analysis unit tracks the decay rate of cultural characteristics and triggers gene library version updates.

[0128] Specifically, the dual-channel reinforcement learning implementation strategy is as follows:

[0129] Offline policy network: The PPO-Clip algorithm (ε=0.2) is used. 5000 sets of generated data are used in each round of training. The policy network is a 3-layer MLP (512 nodes in the hidden layer). The value function update cycle is 10 steps.

[0130] Online Optimizer: Based on real-time user feedback (such as the "pattern complexity +1" command), adjust the constraint weights within 5 seconds through Bayesian optimization (Expected Improvement acquisition function).

[0131] Gene mutation analysis: Use an LSTM network (128 hidden units) to predict the attenuation trend of cultural characteristics (for example, a version update is triggered when the average monthly usage of a pattern style drops by more than 25%), and retain differential snapshots of historical gene vectors (storage increment Δ < 10%).

[0132] Based on the same inventive concept, this application also provides an AIGC paper-cutting generation method based on multimodal constraints of cultural genes, such as Figure 2 As shown, the following steps are included:

[0133] S1. Gene feature extraction: Extracting pattern dominant features and implicit process parameters through multispectral imaging and frequency domain decomposition algorithm;

[0134] S2. Multimodal constraint injection: Converting semantic, topological, and process constraints into mathematical specifications for generating models;

[0135] S3. Collaborative generation and optimization: Dynamically update the gene library by combining the cultural adaptability evaluation of the initial generation results and user interaction correction.

[0136] This method involves a three-stage meme-driven process:

[0137] Feature extraction: The frequency domain features of the pattern (center frequency 2-10 Hz) are extracted through Morlet wavelet transform, and the DTW algorithm (dynamic time warping error less than 0.05) of motion capture data is combined to construct a process parameter-pattern morphology mapping table.

[0138] Constraint injection: The yin-yang engraving rules are encoded as a vector offset of the diffusion model (Δz = 0.3 sign(culture_feat)), imposing a directional constraint in the latent space; topological connectivity is verified through Monte Carlo sampling (1000 times / generation).

[0139] Collaborative optimization: After the first round of generation results are screened by the evaluation model (F1-score = 0.89), users can modify local features (such as sawtooth density ±15%) through the brushstroke inversion network (U-Net architecture). The modified data is then subjected to PCA dimensionality reduction and the gene library encoding matrix is updated.

[0140] like Figure 3 As shown, the gene feature extraction step S1 includes:

[0141] S11. Construct a metadata template containing 17 dimensions of cultural gene features, covering pattern composition, craft techniques, and semantic rules;

[0142] S12. Extracting spatial and temporal dual-domain features of cutting actions through temporal convolutional networks;

[0143] S13. Use cross-modal contrastive learning algorithm to fuse multi-source data and generate cultural gene encoding vectors.

[0144] That is, gene feature extraction is achieved through the following key technologies:

[0145] 17-dimensional metadata template: includes geometric features (4 dimensions: fractal dimension, symmetry index, etc.), process parameters (6 dimensions: cutting speed mean, acceleration variance, etc.), and semantic attributes (7 dimensions: auspicious meaning matching, regional style KL divergence).

[0146] Spatiotemporal feature fusion: A two-stream TCN network is used. The spatial stream processes three-dimensional point cloud data (outputting a pattern curvature distribution map), and the temporal stream analyzes the MFCC features (Mel-frequency cepstral coefficients) of motion sequences. Cross-modal alignment is achieved through the attention mechanism (scaled dot product).

[0147] Cross-modal contrastive learning: Triplet Loss (margin = 0.3) is used to optimize the feature space. The distance between positive sample pairs (same cultural semantics) is less than 0.2, and the distance between negative sample pairs is greater than 0.6.

[0148] like Figure 4 As shown, the multimodal constraint injection step S2 includes:

[0149] S21. Map the Yin-Yang engraving rules into vector offsets in the Stable Diffusion latent space;

[0150] S22. Jointly optimize the three-level constraints through a mixed integer programming model to ensure the cultural and logical consistency of the generated results;

[0151] S23. Insert a process feasibility verification layer into the potential diffusion process to block physically infeasible intermediate generation states.

[0152] The mathematical implementation of multimodal constraints is as follows:

[0153] Yin-Yang engraving mapping: Convert the engraving area constraint (more than 30%) into a hyperplane condition in the latent space (w T z+b≥0), and the generated direction is corrected by projected gradient descent (step size η=0.01).

[0154] Mixed integer programming: objective function min(λ1·L semantic +λ2·L topology +λ3·L process ), where λ1 = 0.6, λ2 = 0.3, and λ3 = 0.1; the constraints include CLIP similarity ≥ 0.8, topological breakpoints ≤ 2, and linewidth variation coefficient less than 0.2.

[0155] Process verification layer: insert differentiable physical simulation units in the 5th to 9th layers of UNet to calculate the von Mises stress of the cutting path in real time (formula: σ vm An L1 penalty (λ=0.05) is imposed on regions >15 MPa.

[0156] like Figure 5 As shown, the collaborative generation and optimization step S3 includes:

[0157] S31. Screen the initial generation results based on the cultural adaptability evaluation model. Model evaluation indicators include semantic accuracy, topological integrity, and process feasibility.

[0158] S32. Receive the user's modification instructions for the local pattern through the brush stroke correction tool, and invert the correction data to the cultural gene library;

[0159] S33. Use a dynamic weight allocation algorithm to adjust the weights of constraints to achieve closed-loop optimization of the generation system.

[0160] Specifically, the human-machine collaborative optimization mechanism is as follows:

[0161] Cultural adaptability assessment: The XGBoost model (tree depth = 6) was used. The input features included CLIP score (weight 0.5), topological connectivity index (weight 0.3), and process violation score (weight 0.2). The output comprehensive score threshold was ≥0.7.

[0162] Brushstroke inversion algorithm: The user's drawing trajectory (sampling point spacing ≤ 0.1mm) is encoded into a latent vector increment Δz, and feature space mapping is achieved through Jacobian matrix pseudo-inverse calculation (regularization coefficient 1e-4).

[0163] Dynamic weight adjustment: Based on the user correction frequency (e.g., the proportion of zigzag pattern modifications > 40%), the semantic constraint weight is adjusted by Δλ = 0.05·log(freq+1), with the constraint range λ∈[0.4,0.8].

[0164] In summary, the AIGC paper-cutting generation system and method based on multimodal constraints of cultural genes provided in this embodiment have at least the following beneficial technical effects:

[0165] 1. Multimodal Constrained Joint Optimization

[0166] By mathematically integrating semantic, topological, and process constraints (e.g., mixed integer programming models), the system addresses the dual issues of "cultural distortion" and "physical impracticality" in traditional AIGC generation. Experiments show that the paper-cut patterns generated by the system significantly outperform the unconstrained ControlNet baseline model (by approximately 35%) in both cultural semantic matching (CLIP score ≥ 0.85) and process compliance (physical simulation pass rate ≥ 92%).

[0167] 2. Dynamic cultural gene injection mechanism

[0168] The improved ControlNet architecture realizes the directional fusion of cultural gene encoding vectors through the cultural attention mechanism (multi-head cross attention, 8-head scaling factor), and dynamically adjusts the feature weights (range 0.2-0.8) during the generation process, so that the generation accuracy of pattern details (such as the continuity of hinged embroidery) is improved to the submillimeter level (error less than 0.1mm).

[0169] 3. Real-time human-machine collaborative correction capability

[0170] The interactive correction tool based on the pressure-sensitive pen (sampling rate 200Hz) is combined with the inverse diffusion inversion algorithm. User modifications to local patterns (such as adjusting the sawtooth density by ±15%) can be fed back to the system within 0.5 seconds, realizing a "creativity-generation-optimization" closed loop and improving user satisfaction (NPS score) by 40%.

[0171] In addition, it also has the following cultural heritage value:

[0172] Digital preservation of cultural genes: Through three-dimensional laser scanning (accuracy 0.01mm) and multispectral imaging (17-dimensional feature encoding), the implicit knowledge of traditional paper-cutting (such as the rule of "leaving continuous lines in relief carving") is transformed into a quantifiable and iterative digital gene library, solving the inheritance gap problem caused by the "oral transmission" of intangible cultural heritage skills.

[0173] Automated verification of cultural logic: The semantic constraint submodule maps paper-cutting cultural rules (such as "deer patterns symbolize official positions") to the text-image alignment space of the CLIP model. The cultural semantic accuracy of the generated results (expert evaluation) reaches 89%, which is 20 times more efficient than traditional manual design.

[0174] Dynamic evolution to adapt to regional styles: The dual-channel reinforcement learning framework (offline PPO + online Bayesian optimization) supports the continuous iteration of the cultural gene library. The system can adapt to different regional styles (such as the Shaanxi "Jianhua Niangzi" style and the Fujian "Liaojin" style). The style transfer accuracy (FID score below 15) is better than the existing cross-style generation model.

[0175] This patented technology demonstrates innovative value in multi-dimensional application scenarios, mainly covering the following areas:

[0176] 1. Protection and inheritance of intangible cultural heritage

[0177] Rescue of endangered skills: In view of the current situation that the number of paper-cutting inheritors decreases by 12.7% annually, the "standardization of techniques" of old artists can be achieved through the use of three-dimensional gene coding technology.

[0178] Live inheritance assistance: Provides an intelligent creation system for inheritance bases to solve the pain point of the long training cycle for young apprentices (traditionally 5-8 years).

[0179] 2. Digital transformation of the cultural and creative industries

[0180] Cultural and creative product development: supporting the intelligent reorganization and design of regional cultural elements.

[0181] Digital content production: Providing compliant cultural assets for Metaverse applications.

[0182] 3. Art education and cultural communication

[0183] Intelligent teaching system and cross-border cultural promotion.

[0184] 4. Artificial intelligence technology research and development

[0185] Development of a cultural constraint model: This project provides a transferable constraint framework for AI art generation and verifies the versatility of the technology in six types of intangible cultural heritage digitization projects, including New Year paintings and embroidery.

[0186] Innovation in the human-machine collaborative paradigm: creating a new creative model of "gene guidance-AI generation-craftsman optimization".

[0187] 5. Construction of cultural big data infrastructure

[0188] Standardization construction of gene banks and cultural security governance.

[0189] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections thereof.

[0190] As used in this specification and the appended claims, the term "if" can be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]," depending on the context.

[0191] References to "one embodiment" or "some embodiments" in this specification mean that a particular feature, structure, or characteristic described in conjunction with that embodiment is included in one or more embodiments of the present application. Thus, phrases such as "in one embodiment," "in some embodiments," "in other embodiments," and "in other embodiments" appearing in various places in this specification do not necessarily refer to the same embodiment, but rather mean "one or more but not all embodiments," unless otherwise specifically emphasized. The terms "including," "comprising," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0192] The AIGC paper-cutting generation system and method provided in the embodiments of the present application can be applied to terminal devices such as mobile phones, tablet computers, wearable devices, laptops, industrial computers, ultra-mobile personal computers (UMPCs), netbooks, and personal digital assistants (PDAs). The embodiments of the present application do not impose any restrictions on the specific types of terminal devices.

[0193] Figure 6 This is a schematic diagram of the structure of a terminal device provided by an embodiment of the present application. Figure 6 As shown, the terminal device 700 of this embodiment includes: at least one processor 702 ( Figure 6 Only one is shown), memory 701, memory 701 stores a computer program 703 that can be run on processor 702. When processor 702 executes computer program 703, it implements the steps in each of the above-mentioned paper-cut image generation method embodiments, such as Figure 3 Alternatively, when the processor 702 executes the computer program 703, the functions of the modules / units in the above-mentioned device embodiments are realized, for example Figure 1 Functionality of the modules shown.

[0194] The terminal device 700 can be a computing device such as a desktop computer, a notebook, a PDA, or a cloud server. The terminal device can include but is not limited to a memory 701 and a processor 702. Those skilled in the art will understand that Figure 6 It is merely an example of the terminal device 700 and does not constitute a limitation of the terminal device 700. The terminal device may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the terminal device may also include an input and sending device, a network access device, a bus, etc.

[0195] The processor 702 may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0196] In some embodiments, the processor 702 may be an internal storage unit of the terminal device 700, such as a hard disk or memory of the terminal device 700. The processor 702 may also be an external storage device of the terminal device 700, such as a plug-in hard disk equipped on the terminal device 700, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card, etc. Furthermore, the processor 702 may also include both the internal storage unit of the terminal device 700 and an external storage device. The processor 702 is used to store an operating system, application programs, a boot loader (BootLoader), data, and other programs, such as program code of a computer program. The processor 702 may also be used to temporarily store data that has been sent or is about to be sent.

[0197] In addition, those skilled in the art will clearly understand that for the sake of convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units in the various embodiments of the present application can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

[0198] An embodiment of the present application also provides a terminal device, which includes at least one memory, at least one processor, and a computer program stored in the at least one memory and executable on the at least one processor. When the processor executes the computer program, the terminal device implements the steps of any of the above-mentioned method embodiments.

[0199] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0200] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0201] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the scope of protection of the present application.

Claims

1. An AIGC paper-cutting generation system based on multimodal constraints of cultural genes, characterized by: include: A cultural gene library construction module is used to collect and encode the pattern characteristics, process parameters and cultural semantic rules of paper-cutting crafts; A multimodal constraint engine, which includes semantic constraint submodules, topological constraint submodules, and process constraint submodules, and is used to jointly optimize the constraints of the generation process; The AIGC generation module is based on the ControlNet neural network architecture and embeds a cultural attention mechanism to achieve dynamic injection of cultural genes; Human-computer collaborative editing module, providing pattern correction tools and feature inversion interface, supporting users to interactively optimize the generated results; and The dynamic evolution module realizes the collaborative iterative update of the cultural gene library and the generation model through a dual-channel reinforcement learning framework.

2. The AIGC paper-cutting generation system according to claim 1, characterized in that: The cultural gene library building modules include: 3D laser scanning device, used to obtain the three-dimensional deformation data and pattern topology structure of paper-cut works; Motion capture equipment, used to record the spatiotemporal motion sequence and process parameters of the cutting process; Multispectral imager, used to extract the explicit features and implicit cultural semantic features of patterns.

3. The AIGC paper-cutting generation system according to claim 1, characterized in that: The multimodal constraint engine achieves constraint optimization in the following ways: The semantic constraint submodule maps cultural semantic rules to the prompt word space of the CLIP text encoder; The topological constraint submodule generates protection masks for non-cut areas through geometric analysis and verifies the physical continuity of the ridges. The process constraint submodule converts the paper tear strength threshold into a process feasibility verification condition in the potential diffusion process.

4. The AIGC paper-cutting generation system according to claim 3, characterized in that: The process feasibility verification conditions include: The paper tear strength threshold constraint is defined based on physical experimental data; The latent vector is reversely corrected through residual connections to ensure the mathematical expressibility of the Yin-Yang engraving rules; A process parameter verification layer is implanted in the potential diffusion process to block the generation path that does not conform to the physical laws of shearing.

5. The AIGC paper-cutting generation system according to claim 1, characterized in that: The AIGC generation module includes: ControlNet architecture, embedding a multi-scale cultural attention mechanism in the UNet decoder; Dynamic weight allocation unit, which adjusts the weight ratio of cultural characteristics and generation goals according to the generation stage; The genealogy tracing unit records the cultural gene code and the source of the process parameters for each generated pattern.

6. The AIGC paper-cutting generation system according to claim 1, characterized in that: The dynamic evolution module includes: Offline training channel, which updates the encoding rules of the cultural gene library based on the generation quality assessment model; Online optimization channel, which adjusts multimodal constraint weights in real time based on user interaction data; Experience replay pool, which stores the mutation trajectory of cultural genes to optimize reinforcement learning strategies.

7. A method for generating AIGC paper-cuts based on multimodal constraints of cultural genes, characterized by: The following steps are involved: Gene feature extraction: extracting pattern dominant features and implicit process parameters through multispectral imaging and frequency domain decomposition algorithm; Multimodal constraint injection: converting semantic, topological, and process constraints into mathematical specifications for generating models; Collaborative generation and optimization: Dynamically update the gene library by combining the cultural adaptability assessment of the initial generation results and user interaction correction.

8. The AIGC paper-cutting generation method according to claim 7, characterized in that: The gene feature extraction step comprises: Construct a metadata template containing 17-dimensional cultural gene features, covering pattern composition, craft techniques and semantic rules; The spatial and temporal dual-domain features of the cutting action are extracted through a temporal convolutional network; A cross-modal contrastive learning algorithm is used to fuse multi-source data and generate cultural gene encoding vectors.

9. The AIGC paper-cutting generation method according to claim 7, characterized in that: The multimodal constraint injection step includes: Map the Yin-Yang engraving rules into vector offsets in the Stable Diffusion latent space; The three-level constraints are jointly optimized through a mixed integer programming model to ensure the cultural and logical consistency of the generated results; A process feasibility verification layer is inserted into the potential diffusion process to block physically infeasible intermediate generation states.

10. The AIGC paper-cutting generation method according to claim 7, characterized in that: The collaborative generation and optimization steps include: Initial generation results were screened based on a cultural adaptability evaluation model. Model evaluation metrics included semantic accuracy, topological integrity, and process feasibility. Receive the user's modification instructions for local patterns through the brush stroke correction tool, and invert the correction data to the cultural gene library; A dynamic weight allocation algorithm is used to adjust the weights of constraints to achieve closed-loop optimization of the generation system.

Images